Skip to content

odak.learn.models

odak.learn.models

Provides necessary definitions for components used in machine learning and deep learning.

channel_gate

Bases: Module

Channel attention module with various pooling strategies. This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).

Source code in odak/learn/models/components.py 
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
class channel_gate(torch.nn.Module):
    """
    Channel attention module with various pooling strategies.
    This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).
    """
    def __init__(
                 self, 
                 gate_channels, 
                 reduction_ratio = 16, 
                 pool_types = ['avg', 'max']
                ):
        """
        Initializes the channel gate module.

        Parameters
        ----------
        gate_channels   : int
                          Number of channels of the input feature map.
        reduction_ratio : int
                          Reduction ratio for the intermediate layer.
        pool_types      : list
                          List of pooling operations to apply.
        """
        super().__init__()
        self.gate_channels = gate_channels
        hidden_channels = gate_channels // reduction_ratio
        if hidden_channels == 0:
            hidden_channels = 1
        self.mlp = torch.nn.Sequential(
                                       convolutional_block_attention.Flatten(),
                                       torch.nn.Linear(gate_channels, hidden_channels),
                                       torch.nn.ReLU(),
                                       torch.nn.Linear(hidden_channels, gate_channels)
                                      )
        self.pool_types = pool_types


    def forward(self, x):
        """
        Forward pass of the ChannelGate module.

        Applies channel-wise attention to the input tensor.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the ChannelGate module.

        Returns
        -------
        output       : torch.tensor
                       Output tensor after applying channel attention.
        """
        channel_att_sum = None
        for pool_type in self.pool_types:
            if pool_type == 'avg':
                pool = torch.nn.functional.avg_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
            elif pool_type == 'max':
                pool = torch.nn.functional.max_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
            channel_att_raw = self.mlp(pool)
            channel_att_sum = channel_att_raw if channel_att_sum is None else channel_att_sum + channel_att_raw
        scale = torch.sigmoid(channel_att_sum).unsqueeze(2).unsqueeze(3).expand_as(x)
        output = x * scale
        return output

__init__(gate_channels, reduction_ratio=16, pool_types=['avg', 'max'])

Initializes the channel gate module.

Parameters:

  • gate_channels – 
              Number of channels of the input feature map.
  • reduction_ratio (int, default: 16 ) – 
              Reduction ratio for the intermediate layer.
  • pool_types – 
              List of pooling operations to apply.
Source code in odak/learn/models/components.py 
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
def __init__(
             self, 
             gate_channels, 
             reduction_ratio = 16, 
             pool_types = ['avg', 'max']
            ):
    """
    Initializes the channel gate module.

    Parameters
    ----------
    gate_channels   : int
                      Number of channels of the input feature map.
    reduction_ratio : int
                      Reduction ratio for the intermediate layer.
    pool_types      : list
                      List of pooling operations to apply.
    """
    super().__init__()
    self.gate_channels = gate_channels
    hidden_channels = gate_channels // reduction_ratio
    if hidden_channels == 0:
        hidden_channels = 1
    self.mlp = torch.nn.Sequential(
                                   convolutional_block_attention.Flatten(),
                                   torch.nn.Linear(gate_channels, hidden_channels),
                                   torch.nn.ReLU(),
                                   torch.nn.Linear(hidden_channels, gate_channels)
                                  )
    self.pool_types = pool_types

forward(x)

Forward pass of the ChannelGate module.

Applies channel-wise attention to the input tensor.

Parameters:

  • x – 
           Input tensor to the ChannelGate module.

Returns:

  • output ( tensor ) –

    Output tensor after applying channel attention.

Source code in odak/learn/models/components.py 
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
def forward(self, x):
    """
    Forward pass of the ChannelGate module.

    Applies channel-wise attention to the input tensor.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the ChannelGate module.

    Returns
    -------
    output       : torch.tensor
                   Output tensor after applying channel attention.
    """
    channel_att_sum = None
    for pool_type in self.pool_types:
        if pool_type == 'avg':
            pool = torch.nn.functional.avg_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
        elif pool_type == 'max':
            pool = torch.nn.functional.max_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
        channel_att_raw = self.mlp(pool)
        channel_att_sum = channel_att_raw if channel_att_sum is None else channel_att_sum + channel_att_raw
    scale = torch.sigmoid(channel_att_sum).unsqueeze(2).unsqueeze(3).expand_as(x)
    output = x * scale
    return output

convolution_layer

Bases: Module

A convolution layer.

Source code in odak/learn/models/components.py 
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
class convolution_layer(torch.nn.Module):
    """
    A convolution layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 bias = False,
                 stride = 1,
                 normalization = True,
                 activation = torch.nn.ReLU()
                ):
        """
        A convolutional layer class.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        layers = [
            torch.nn.Conv2d(
                            input_channels,
                            output_channels,
                            kernel_size = kernel_size,
                            stride = stride,
                            padding = kernel_size // 2,
                            bias = bias
                           )
        ]
        if normalization:
            layers.append(torch.nn.BatchNorm2d(output_channels))
        if activation:
            layers.append(activation)
        self.model = torch.nn.Sequential(*layers)


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        result = self.model(x)
        return result

__init__(input_channels=2, output_channels=2, kernel_size=3, bias=False, stride=1, normalization=True, activation=torch.nn.ReLU())

A convolutional layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             bias = False,
             stride = 1,
             normalization = True,
             activation = torch.nn.ReLU()
            ):
    """
    A convolutional layer class.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    layers = [
        torch.nn.Conv2d(
                        input_channels,
                        output_channels,
                        kernel_size = kernel_size,
                        stride = stride,
                        padding = kernel_size // 2,
                        bias = bias
                       )
    ]
    if normalization:
        layers.append(torch.nn.BatchNorm2d(output_channels))
    if activation:
        layers.append(activation)
    self.model = torch.nn.Sequential(*layers)

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    result = self.model(x)
    return result

convolutional_block_attention

Bases: Module

Convolutional Block Attention Module (CBAM) class. This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).

Source code in odak/learn/models/components.py 
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
class convolutional_block_attention(torch.nn.Module):
    """
    Convolutional Block Attention Module (CBAM) class. 
    This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).
    """
    def __init__(
                 self, 
                 gate_channels, 
                 reduction_ratio = 16, 
                 pool_types = ['avg', 'max'], 
                 no_spatial = False
                ):
        """
        Initializes the convolutional block attention module.

        Parameters
        ----------
        gate_channels   : int
                          Number of channels of the input feature map.
        reduction_ratio : int
                          Reduction ratio for the channel attention.
        pool_types      : list
                          List of pooling operations to apply for channel attention.
        no_spatial      : bool
                          If True, spatial attention is not applied.
        """
        super(convolutional_block_attention, self).__init__()
        self.channel_gate = channel_gate(gate_channels, reduction_ratio, pool_types)
        self.no_spatial = no_spatial
        if not no_spatial:
            self.spatial_gate = spatial_gate()


    class Flatten(torch.nn.Module):
        """
        Flattens the input tensor to a 2D matrix.
        """
        def forward(self, x):
            return x.view(x.size(0), -1)


    def forward(self, x):
        """
        Forward pass of the convolutional block attention module.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the CBAM module.

        Returns
        -------
        x_out        : torch.tensor
                       Output tensor after applying channel and spatial attention.
        """
        x_out = self.channel_gate(x)
        if not self.no_spatial:
            x_out = self.spatial_gate(x_out)
        return x_out

Flatten

Bases: Module

Flattens the input tensor to a 2D matrix.

Source code in odak/learn/models/components.py 
757
758
759
760
761
762
class Flatten(torch.nn.Module):
    """
    Flattens the input tensor to a 2D matrix.
    """
    def forward(self, x):
        return x.view(x.size(0), -1)

__init__(gate_channels, reduction_ratio=16, pool_types=['avg', 'max'], no_spatial=False)

Initializes the convolutional block attention module.

Parameters:

  • gate_channels – 
              Number of channels of the input feature map.
  • reduction_ratio (int, default: 16 ) – 
              Reduction ratio for the channel attention.
  • pool_types – 
              List of pooling operations to apply for channel attention.
  • no_spatial – 
              If True, spatial attention is not applied.
Source code in odak/learn/models/components.py 
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
def __init__(
             self, 
             gate_channels, 
             reduction_ratio = 16, 
             pool_types = ['avg', 'max'], 
             no_spatial = False
            ):
    """
    Initializes the convolutional block attention module.

    Parameters
    ----------
    gate_channels   : int
                      Number of channels of the input feature map.
    reduction_ratio : int
                      Reduction ratio for the channel attention.
    pool_types      : list
                      List of pooling operations to apply for channel attention.
    no_spatial      : bool
                      If True, spatial attention is not applied.
    """
    super(convolutional_block_attention, self).__init__()
    self.channel_gate = channel_gate(gate_channels, reduction_ratio, pool_types)
    self.no_spatial = no_spatial
    if not no_spatial:
        self.spatial_gate = spatial_gate()

forward(x)

Forward pass of the convolutional block attention module.

Parameters:

  • x – 
           Input tensor to the CBAM module.

Returns:

  • x_out ( tensor ) –

    Output tensor after applying channel and spatial attention.

Source code in odak/learn/models/components.py 
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
def forward(self, x):
    """
    Forward pass of the convolutional block attention module.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the CBAM module.

    Returns
    -------
    x_out        : torch.tensor
                   Output tensor after applying channel and spatial attention.
    """
    x_out = self.channel_gate(x)
    if not self.no_spatial:
        x_out = self.spatial_gate(x_out)
    return x_out

double_convolution

Bases: Module

A double convolution layer.

Source code in odak/learn/models/components.py 
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
class double_convolution(torch.nn.Module):
    """
    A double convolution layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 mid_channels = None,
                 output_channels = 2,
                 kernel_size = 3, 
                 bias = False,
                 normalization = True,
                 activation = torch.nn.ReLU()
                ):
        """
        Double convolution model.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels    : int
                          Number of channels in the hidden layer between two convolutions.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        if isinstance(mid_channels, type(None)):
            mid_channels = output_channels
        self.activation = activation
        self.model = torch.nn.Sequential(
                                         convolution_layer(
                                                           input_channels = input_channels,
                                                           output_channels = mid_channels,
                                                           kernel_size = kernel_size,
                                                           bias = bias,
                                                           normalization = normalization,
                                                           activation = self.activation
                                                          ),
                                         convolution_layer(
                                                           input_channels = mid_channels,
                                                           output_channels = output_channels,
                                                           kernel_size = kernel_size,
                                                           bias = bias,
                                                           normalization = normalization,
                                                           activation = self.activation
                                                          )
                                        )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        result = self.model(x)
        return result

__init__(input_channels=2, mid_channels=None, output_channels=2, kernel_size=3, bias=False, normalization=True, activation=torch.nn.ReLU())

Double convolution model.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of channels in the hidden layer between two convolutions.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
def __init__(
             self,
             input_channels = 2,
             mid_channels = None,
             output_channels = 2,
             kernel_size = 3, 
             bias = False,
             normalization = True,
             activation = torch.nn.ReLU()
            ):
    """
    Double convolution model.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels    : int
                      Number of channels in the hidden layer between two convolutions.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    if isinstance(mid_channels, type(None)):
        mid_channels = output_channels
    self.activation = activation
    self.model = torch.nn.Sequential(
                                     convolution_layer(
                                                       input_channels = input_channels,
                                                       output_channels = mid_channels,
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       normalization = normalization,
                                                       activation = self.activation
                                                      ),
                                     convolution_layer(
                                                       input_channels = mid_channels,
                                                       output_channels = output_channels,
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       normalization = normalization,
                                                       activation = self.activation
                                                      )
                                    )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    result = self.model(x)
    return result

downsample_layer

Bases: Module

A downscaling component followed by a double convolution.

Source code in odak/learn/models/components.py 
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
class downsample_layer(torch.nn.Module):
    """
    A downscaling component followed by a double convolution.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.maxpool_conv = torch.nn.Sequential(
                                                torch.nn.MaxPool2d(2),
                                                double_convolution(
                                                                   input_channels = input_channels,
                                                                   mid_channels = output_channels,
                                                                   output_channels = output_channels,
                                                                   kernel_size = kernel_size,
                                                                   bias = bias,
                                                                   activation = activation
                                                                  )
                                               )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x              : torch.tensor
                         First input data.



        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        result = self.maxpool_conv(x)
        return result

__init__(input_channels, output_channels, kernel_size=3, bias=False, activation=torch.nn.ReLU())

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.maxpool_conv = torch.nn.Sequential(
                                            torch.nn.MaxPool2d(2),
                                            double_convolution(
                                                               input_channels = input_channels,
                                                               mid_channels = output_channels,
                                                               output_channels = output_channels,
                                                               kernel_size = kernel_size,
                                                               bias = bias,
                                                               activation = activation
                                                              )
                                           )

forward(x)

Forward model.

Parameters:

  • x – 
             First input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x              : torch.tensor
                     First input data.



    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    result = self.maxpool_conv(x)
    return result

global_feature_module

Bases: Module

A global feature layer that processes global features from input channels and applies them to another input tensor via learned transformations.

Source code in odak/learn/models/components.py 
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
class global_feature_module(torch.nn.Module):
    """
    A global feature layer that processes global features from input channels and
    applies them to another input tensor via learned transformations.
    """
    def __init__(
                 self,
                 input_channels,
                 mid_channels,
                 output_channels,
                 kernel_size,
                 bias = False,
                 normalization = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A global feature layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels  : int
                          Number of mid channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.transformations_1 = global_transformations(input_channels, output_channels)
        self.global_features_1 = double_convolution(
                                                    input_channels = input_channels,
                                                    mid_channels = mid_channels,
                                                    output_channels = output_channels,
                                                    kernel_size = kernel_size,
                                                    bias = bias,
                                                    normalization = normalization,
                                                    activation = activation
                                                   )
        self.global_features_2 = double_convolution(
                                                    input_channels = input_channels,
                                                    mid_channels = mid_channels,
                                                    output_channels = output_channels,
                                                    kernel_size = kernel_size,
                                                    bias = bias,
                                                    normalization = normalization,
                                                    activation = activation
                                                   )
        self.transformations_2 = global_transformations(input_channels, output_channels)


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.

        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        global_tensor_1 = self.transformations_1(x1, x2)
        y1 = self.global_features_1(global_tensor_1)
        y2 = self.global_features_2(y1)
        global_tensor_2 = self.transformations_2(y1, y2)
        return global_tensor_2

__init__(input_channels, mid_channels, output_channels, kernel_size, bias=False, normalization=False, activation=torch.nn.ReLU())

A global feature layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of mid channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
def __init__(
             self,
             input_channels,
             mid_channels,
             output_channels,
             kernel_size,
             bias = False,
             normalization = False,
             activation = torch.nn.ReLU()
            ):
    """
    A global feature layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels  : int
                      Number of mid channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.transformations_1 = global_transformations(input_channels, output_channels)
    self.global_features_1 = double_convolution(
                                                input_channels = input_channels,
                                                mid_channels = mid_channels,
                                                output_channels = output_channels,
                                                kernel_size = kernel_size,
                                                bias = bias,
                                                normalization = normalization,
                                                activation = activation
                                               )
    self.global_features_2 = double_convolution(
                                                input_channels = input_channels,
                                                mid_channels = mid_channels,
                                                output_channels = output_channels,
                                                kernel_size = kernel_size,
                                                bias = bias,
                                                normalization = normalization,
                                                activation = activation
                                               )
    self.transformations_2 = global_transformations(input_channels, output_channels)

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.

    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    global_tensor_1 = self.transformations_1(x1, x2)
    y1 = self.global_features_1(global_tensor_1)
    y2 = self.global_features_2(y1)
    global_tensor_2 = self.transformations_2(y1, y2)
    return global_tensor_2

global_transformations

Bases: Module

A global feature layer that processes global features from input channels and applies learned transformations to another input tensor.

This implementation is adapted from RSGUnet: https://github.com/MTLab/rsgunet_image_enhance.

Reference: J. Huang, P. Zhu, M. Geng et al. "Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices."

Source code in odak/learn/models/components.py 
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
class global_transformations(torch.nn.Module):
    """
    A global feature layer that processes global features from input channels and
    applies learned transformations to another input tensor.

    This implementation is adapted from RSGUnet:
    https://github.com/MTLab/rsgunet_image_enhance.

    Reference:
    J. Huang, P. Zhu, M. Geng et al. "Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices."
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels
                ):
        """
        A global feature layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        """
        super().__init__()
        self.global_feature_1 = torch.nn.Sequential(
            torch.nn.Linear(input_channels, output_channels),
            torch.nn.LeakyReLU(0.2, inplace = True),
        )
        self.global_feature_2 = torch.nn.Sequential(
            torch.nn.Linear(output_channels, output_channels),
            torch.nn.LeakyReLU(0.2, inplace = True)
        )


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.

        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        y = torch.mean(x2, dim = (2, 3))
        y1 = self.global_feature_1(y)
        y2 = self.global_feature_2(y1)
        y1 = y1.unsqueeze(2).unsqueeze(3)
        y2 = y2.unsqueeze(2).unsqueeze(3)
        result = x1 * y1 + y2
        return result

__init__(input_channels, output_channels)

A global feature layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
Source code in odak/learn/models/components.py 
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
def __init__(
             self,
             input_channels,
             output_channels
            ):
    """
    A global feature layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    """
    super().__init__()
    self.global_feature_1 = torch.nn.Sequential(
        torch.nn.Linear(input_channels, output_channels),
        torch.nn.LeakyReLU(0.2, inplace = True),
    )
    self.global_feature_2 = torch.nn.Sequential(
        torch.nn.Linear(output_channels, output_channels),
        torch.nn.LeakyReLU(0.2, inplace = True)
    )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.

    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    y = torch.mean(x2, dim = (2, 3))
    y1 = self.global_feature_1(y)
    y2 = self.global_feature_2(y1)
    y1 = y1.unsqueeze(2).unsqueeze(3)
    y2 = y2.unsqueeze(2).unsqueeze(3)
    result = x1 * y1 + y2
    return result

multi_layer_perceptron

Bases: Module

A multi-layer perceptron model.

Source code in odak/learn/models/models.py 
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
class multi_layer_perceptron(torch.nn.Module):
    """
    A multi-layer perceptron model.
    """

    def __init__(self,
                 dimensions,
                 activation = torch.nn.ReLU(),
                 bias = False,
                 model_type = 'conventional',
                 siren_multiplier = 1.,
                 input_multiplier = None
                ):
        """
        Parameters
        ----------
        dimensions        : list
                            List of integers representing the dimensions of each layer (e.g., [2, 10, 1], where the first layer has two channels and last one has one channel.).
        activation        : torch.nn
                            Nonlinear activation function.
                            Default is `torch.nn.ReLU()`.
        bias              : bool
                            If set to True, linear layers will include biases.
        siren_multiplier  : float
                            When using `SIREN` model type, this parameter functions as a hyperparameter.
                            The original SIREN work uses 30.
                            You can bypass this parameter by providing input that are not normalized and larger then one.
        input_multiplier  : float
                            Initial value of the input multiplier before the very first layer.
        model_type        : str
                            Model type: `conventional`, `swish`, `SIREN`, `FILM SIREN`, `Gaussian`.
                            `conventional` refers to a standard multi layer perceptron.
                            For `SIREN,` see: Sitzmann, Vincent, et al. "Implicit neural representations with periodic activation functions." Advances in neural information processing systems 33 (2020): 7462-7473.
                            For `Swish,` see: Ramachandran, Prajit, Barret Zoph, and Quoc V. Le. "Searching for activation functions." arXiv preprint arXiv:1710.05941 (2017). 
                            For `FILM SIREN,` see: Chan, Eric R., et al. "pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
                            For `Gaussian,` see: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.
        """
        super(multi_layer_perceptron, self).__init__()
        self.activation = activation
        self.bias = bias
        self.model_type = model_type
        self.layers = torch.nn.ModuleList()
        self.siren_multiplier = siren_multiplier
        self.dimensions = dimensions
        for i in range(len(self.dimensions) - 1):
            self.layers.append(torch.nn.Linear(self.dimensions[i], self.dimensions[i + 1], bias = self.bias))
        if not isinstance(input_multiplier, type(None)):
            self.input_multiplier = torch.nn.ParameterList()
            self.input_multiplier.append(torch.nn.Parameter(torch.ones(1, self.dimensions[0]) * input_multiplier))
        if self.model_type == 'FILM SIREN':
            self.alpha = torch.nn.ParameterList()
            for j in self.dimensions[1:-1]:
                self.alpha.append(torch.nn.Parameter(torch.randn(2, 1, j)))
        if self.model_type == 'Gaussian':
            self.alpha = torch.nn.ParameterList()
            for j in self.dimensions[1:-1]:
                self.alpha.append(torch.nn.Parameter(torch.randn(1, 1, j)))


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        if hasattr(self, 'input_multiplier'):
            result = x * self.input_multiplier[0]
        else:
            result = x
        for layer_id, layer in enumerate(self.layers[:-1]):
            result = layer(result)
            if self.model_type == 'conventional':
                result = self.activation(result)
            elif self.model_type == 'swish':
                result = swish(result)
            elif self.model_type == 'SIREN':
                result = torch.sin(result * self.siren_multiplier)
            elif self.model_type == 'FILM SIREN':
                result = torch.sin(self.alpha[layer_id][0] * result + self.alpha[layer_id][1])
            elif self.model_type == 'Gaussian': 
                result = gaussian(result, self.alpha[layer_id][0])
        result = self.layers[-1](result)
        return result

__init__(dimensions, activation=torch.nn.ReLU(), bias=False, model_type='conventional', siren_multiplier=1.0, input_multiplier=None)

Parameters:

  • dimensions – 
                List of integers representing the dimensions of each layer (e.g., [2, 10, 1], where the first layer has two channels and last one has one channel.).
  • activation – 
                Nonlinear activation function.
            Default is `torch.nn.ReLU()`.
  • bias – 
                If set to True, linear layers will include biases.
  • siren_multiplier – 
                When using `SIREN` model type, this parameter functions as a hyperparameter.
            The original SIREN work uses 30.
            You can bypass this parameter by providing input that are not normalized and larger then one.
  • input_multiplier – 
                Initial value of the input multiplier before the very first layer.
  • model_type – 
                Model type: `conventional`, `swish`, `SIREN`, `FILM SIREN`, `Gaussian`.
            `conventional` refers to a standard multi layer perceptron.
            For `SIREN,` see: Sitzmann, Vincent, et al. "Implicit neural representations with periodic activation functions." Advances in neural information processing systems 33 (2020): 7462-7473.
            For `Swish,` see: Ramachandran, Prajit, Barret Zoph, and Quoc V. Le. "Searching for activation functions." arXiv preprint arXiv:1710.05941 (2017). 
            For `FILM SIREN,` see: Chan, Eric R., et al. "pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
            For `Gaussian,` see: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.
Source code in odak/learn/models/models.py 
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
def __init__(self,
             dimensions,
             activation = torch.nn.ReLU(),
             bias = False,
             model_type = 'conventional',
             siren_multiplier = 1.,
             input_multiplier = None
            ):
    """
    Parameters
    ----------
    dimensions        : list
                        List of integers representing the dimensions of each layer (e.g., [2, 10, 1], where the first layer has two channels and last one has one channel.).
    activation        : torch.nn
                        Nonlinear activation function.
                        Default is `torch.nn.ReLU()`.
    bias              : bool
                        If set to True, linear layers will include biases.
    siren_multiplier  : float
                        When using `SIREN` model type, this parameter functions as a hyperparameter.
                        The original SIREN work uses 30.
                        You can bypass this parameter by providing input that are not normalized and larger then one.
    input_multiplier  : float
                        Initial value of the input multiplier before the very first layer.
    model_type        : str
                        Model type: `conventional`, `swish`, `SIREN`, `FILM SIREN`, `Gaussian`.
                        `conventional` refers to a standard multi layer perceptron.
                        For `SIREN,` see: Sitzmann, Vincent, et al. "Implicit neural representations with periodic activation functions." Advances in neural information processing systems 33 (2020): 7462-7473.
                        For `Swish,` see: Ramachandran, Prajit, Barret Zoph, and Quoc V. Le. "Searching for activation functions." arXiv preprint arXiv:1710.05941 (2017). 
                        For `FILM SIREN,` see: Chan, Eric R., et al. "pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
                        For `Gaussian,` see: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.
    """
    super(multi_layer_perceptron, self).__init__()
    self.activation = activation
    self.bias = bias
    self.model_type = model_type
    self.layers = torch.nn.ModuleList()
    self.siren_multiplier = siren_multiplier
    self.dimensions = dimensions
    for i in range(len(self.dimensions) - 1):
        self.layers.append(torch.nn.Linear(self.dimensions[i], self.dimensions[i + 1], bias = self.bias))
    if not isinstance(input_multiplier, type(None)):
        self.input_multiplier = torch.nn.ParameterList()
        self.input_multiplier.append(torch.nn.Parameter(torch.ones(1, self.dimensions[0]) * input_multiplier))
    if self.model_type == 'FILM SIREN':
        self.alpha = torch.nn.ParameterList()
        for j in self.dimensions[1:-1]:
            self.alpha.append(torch.nn.Parameter(torch.randn(2, 1, j)))
    if self.model_type == 'Gaussian':
        self.alpha = torch.nn.ParameterList()
        for j in self.dimensions[1:-1]:
            self.alpha.append(torch.nn.Parameter(torch.randn(1, 1, j)))

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/models.py 
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    if hasattr(self, 'input_multiplier'):
        result = x * self.input_multiplier[0]
    else:
        result = x
    for layer_id, layer in enumerate(self.layers[:-1]):
        result = layer(result)
        if self.model_type == 'conventional':
            result = self.activation(result)
        elif self.model_type == 'swish':
            result = swish(result)
        elif self.model_type == 'SIREN':
            result = torch.sin(result * self.siren_multiplier)
        elif self.model_type == 'FILM SIREN':
            result = torch.sin(self.alpha[layer_id][0] * result + self.alpha[layer_id][1])
        elif self.model_type == 'Gaussian': 
            result = gaussian(result, self.alpha[layer_id][0])
    result = self.layers[-1](result)
    return result

non_local_layer

Bases: Module

Self-Attention Layer [zi = Wzyi + xi] (non-local block : ref https://arxiv.org/abs/1711.07971)

Source code in odak/learn/models/components.py 
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
class non_local_layer(torch.nn.Module):
    """
    Self-Attention Layer [zi = Wzyi + xi] (non-local block : ref https://arxiv.org/abs/1711.07971)
    """
    def __init__(
                 self,
                 input_channels = 1024,
                 bottleneck_channels = 512,
                 kernel_size = 1,
                 bias = False,
                ):
        """

        Parameters
        ----------
        input_channels      : int
                              Number of input channels.
        bottleneck_channels : int
                              Number of middle channels.
        kernel_size         : int
                              Kernel size.
        bias                : bool 
                              Set to True to let convolutional layers have bias term.
        """
        super(non_local_layer, self).__init__()
        self.input_channels = input_channels
        self.bottleneck_channels = bottleneck_channels
        self.g = torch.nn.Conv2d(
                                 self.input_channels, 
                                 self.bottleneck_channels,
                                 kernel_size = kernel_size,
                                 padding = kernel_size // 2,
                                 bias = bias
                                )
        self.W_z = torch.nn.Sequential(
                                       torch.nn.Conv2d(
                                                       self.bottleneck_channels,
                                                       self.input_channels, 
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       padding = kernel_size // 2
                                                      ),
                                       torch.nn.BatchNorm2d(self.input_channels)
                                      )
        torch.nn.init.constant_(self.W_z[1].weight, 0)   
        torch.nn.init.constant_(self.W_z[1].bias, 0)


    def forward(self, x):
        """
        Forward model [zi = Wzyi + xi]

        Parameters
        ----------
        x               : torch.tensor
                          First input data.                       


        Returns
        ----------
        z               : torch.tensor
                          Estimated output.
        """
        batch_size, channels, height, width = x.size()
        theta = x.view(batch_size, channels, -1).permute(0, 2, 1)
        phi = x.view(batch_size, channels, -1).permute(0, 2, 1)
        g = self.g(x).view(batch_size, self.bottleneck_channels, -1).permute(0, 2, 1)
        attn = torch.bmm(theta, phi.transpose(1, 2)) / (height * width)
        attn = torch.nn.functional.softmax(attn, dim=-1)
        y = torch.bmm(attn, g).permute(0, 2, 1).contiguous().view(batch_size, self.bottleneck_channels, height, width)
        W_y = self.W_z(y)
        z = W_y + x
        return z

__init__(input_channels=1024, bottleneck_channels=512, kernel_size=1, bias=False)

Parameters:

  • input_channels – 
                  Number of input channels.
  • bottleneck_channels (int, default: 512 ) – 
                  Number of middle channels.
  • kernel_size – 
                  Kernel size.
  • bias – 
                  Set to True to let convolutional layers have bias term.
Source code in odak/learn/models/components.py 
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
def __init__(
             self,
             input_channels = 1024,
             bottleneck_channels = 512,
             kernel_size = 1,
             bias = False,
            ):
    """

    Parameters
    ----------
    input_channels      : int
                          Number of input channels.
    bottleneck_channels : int
                          Number of middle channels.
    kernel_size         : int
                          Kernel size.
    bias                : bool 
                          Set to True to let convolutional layers have bias term.
    """
    super(non_local_layer, self).__init__()
    self.input_channels = input_channels
    self.bottleneck_channels = bottleneck_channels
    self.g = torch.nn.Conv2d(
                             self.input_channels, 
                             self.bottleneck_channels,
                             kernel_size = kernel_size,
                             padding = kernel_size // 2,
                             bias = bias
                            )
    self.W_z = torch.nn.Sequential(
                                   torch.nn.Conv2d(
                                                   self.bottleneck_channels,
                                                   self.input_channels, 
                                                   kernel_size = kernel_size,
                                                   bias = bias,
                                                   padding = kernel_size // 2
                                                  ),
                                   torch.nn.BatchNorm2d(self.input_channels)
                                  )
    torch.nn.init.constant_(self.W_z[1].weight, 0)   
    torch.nn.init.constant_(self.W_z[1].bias, 0)

forward(x)

Forward model [zi = Wzyi + xi]

Parameters:

  • x – 
              First input data.

Returns:

  • z ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
def forward(self, x):
    """
    Forward model [zi = Wzyi + xi]

    Parameters
    ----------
    x               : torch.tensor
                      First input data.                       


    Returns
    ----------
    z               : torch.tensor
                      Estimated output.
    """
    batch_size, channels, height, width = x.size()
    theta = x.view(batch_size, channels, -1).permute(0, 2, 1)
    phi = x.view(batch_size, channels, -1).permute(0, 2, 1)
    g = self.g(x).view(batch_size, self.bottleneck_channels, -1).permute(0, 2, 1)
    attn = torch.bmm(theta, phi.transpose(1, 2)) / (height * width)
    attn = torch.nn.functional.softmax(attn, dim=-1)
    y = torch.bmm(attn, g).permute(0, 2, 1).contiguous().view(batch_size, self.bottleneck_channels, height, width)
    W_y = self.W_z(y)
    z = W_y + x
    return z

normalization

Bases: Module

A normalization layer.

Source code in odak/learn/models/components.py 
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
class normalization(torch.nn.Module):
    """
    A normalization layer.
    """
    def __init__(
                 self,
                 dim = 1,
                ):
        """
        Normalization layer.


        Parameters
        ----------
        dim             : int
                          Dimension (axis) to normalize.
        """
        super().__init__()
        self.k = torch.nn.Parameter(torch.ones(1, dim, 1, 1))


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        eps = 1e-5 if x.dtype == torch.float32 else 1e-3
        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
        mean = torch.mean(x, dim = 1, keepdim = True)
        result =  (x - mean) * (var + eps).rsqrt() * self.k
        return result

__init__(dim=1)

Normalization layer.

Parameters:

  • dim – 
              Dimension (axis) to normalize.
Source code in odak/learn/models/components.py 
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
def __init__(
             self,
             dim = 1,
            ):
    """
    Normalization layer.


    Parameters
    ----------
    dim             : int
                      Dimension (axis) to normalize.
    """
    super().__init__()
    self.k = torch.nn.Parameter(torch.ones(1, dim, 1, 1))

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    eps = 1e-5 if x.dtype == torch.float32 else 1e-3
    var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
    mean = torch.mean(x, dim = 1, keepdim = True)
    result =  (x - mean) * (var + eps).rsqrt() * self.k
    return result

positional_encoder

Bases: Module

A positional encoder module.

Source code in odak/learn/models/components.py 
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
class positional_encoder(torch.nn.Module):
    """
    A positional encoder module.
    """

    def __init__(self, L):
        """
        A positional encoder module.

        Parameters
        ----------
        L                   : int
                              Positional encoding level.
        """
        super(positional_encoder, self).__init__()
        self.L = L


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x               : torch.tensor
                          Input data.

        Returns
        ----------
        result          : torch.tensor
                          Result of the forward operation
        """
        B, C = x.shape
        x = x.view(B, C, 1)
        results = [x]
        for i in range(1, self.L + 1):
            freq = (2 ** i) * torch.pi
            cos_x = torch.cos(freq * x)
            sin_x = torch.sin(freq * x)
            results.append(cos_x)
            results.append(sin_x)
        results = torch.cat(results, dim=2)
        results = results.permute(0, 2, 1)
        results = results.reshape(B, -1)
        return results

__init__(L)

A positional encoder module.

Parameters:

  • L – 
                  Positional encoding level.
Source code in odak/learn/models/components.py 
790
791
792
793
794
795
796
797
798
799
800
def __init__(self, L):
    """
    A positional encoder module.

    Parameters
    ----------
    L                   : int
                          Positional encoding level.
    """
    super(positional_encoder, self).__init__()
    self.L = L

forward(x)

Forward model.

Parameters:

  • x – 
              Input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x               : torch.tensor
                      Input data.

    Returns
    ----------
    result          : torch.tensor
                      Result of the forward operation
    """
    B, C = x.shape
    x = x.view(B, C, 1)
    results = [x]
    for i in range(1, self.L + 1):
        freq = (2 ** i) * torch.pi
        cos_x = torch.cos(freq * x)
        sin_x = torch.sin(freq * x)
        results.append(cos_x)
        results.append(sin_x)
    results = torch.cat(results, dim=2)
    results = results.permute(0, 2, 1)
    results = results.reshape(B, -1)
    return results

residual_attention_layer

Bases: Module

A residual block with an attention layer.

Source code in odak/learn/models/components.py 
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
class residual_attention_layer(torch.nn.Module):
    """
    A residual block with an attention layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 1,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        An attention layer class.


        Parameters
        ----------
        input_channels  : int or optioal
                          Number of input channels.
        output_channels : int or optional
                          Number of middle channels.
        kernel_size     : int or optional
                          Kernel size.
        bias            : bool or optional
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn or optional
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.activation = activation
        self.convolution0 = torch.nn.Sequential(
                                                torch.nn.Conv2d(
                                                                input_channels,
                                                                output_channels,
                                                                kernel_size = kernel_size,
                                                                padding = kernel_size // 2,
                                                                bias = bias
                                                               ),
                                                torch.nn.BatchNorm2d(output_channels)
                                               )
        self.convolution1 = torch.nn.Sequential(
                                                torch.nn.Conv2d(
                                                                input_channels,
                                                                output_channels,
                                                                kernel_size = kernel_size,
                                                                padding = kernel_size // 2,
                                                                bias = bias
                                                               ),
                                                torch.nn.BatchNorm2d(output_channels)
                                               )
        self.final_layer = torch.nn.Sequential(
                                               self.activation,
                                               torch.nn.Conv2d(
                                                               output_channels,
                                                               output_channels,
                                                               kernel_size = kernel_size,
                                                               padding = kernel_size // 2,
                                                               bias = bias
                                                              )
                                              )


    def forward(self, x0, x1):
        """
        Forward model.

        Parameters
        ----------
        x0             : torch.tensor
                         First input data.

        x1             : torch.tensor
                         Seconnd input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        y0 = self.convolution0(x0)
        y1 = self.convolution1(x1)
        y2 = torch.add(y0, y1)
        result = self.final_layer(y2) * x0
        return result

__init__(input_channels=2, output_channels=2, kernel_size=1, bias=False, activation=torch.nn.ReLU())

An attention layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int or optional, default: 2 ) – 
              Number of middle channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 1,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    An attention layer class.


    Parameters
    ----------
    input_channels  : int or optioal
                      Number of input channels.
    output_channels : int or optional
                      Number of middle channels.
    kernel_size     : int or optional
                      Kernel size.
    bias            : bool or optional
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn or optional
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.activation = activation
    self.convolution0 = torch.nn.Sequential(
                                            torch.nn.Conv2d(
                                                            input_channels,
                                                            output_channels,
                                                            kernel_size = kernel_size,
                                                            padding = kernel_size // 2,
                                                            bias = bias
                                                           ),
                                            torch.nn.BatchNorm2d(output_channels)
                                           )
    self.convolution1 = torch.nn.Sequential(
                                            torch.nn.Conv2d(
                                                            input_channels,
                                                            output_channels,
                                                            kernel_size = kernel_size,
                                                            padding = kernel_size // 2,
                                                            bias = bias
                                                           ),
                                            torch.nn.BatchNorm2d(output_channels)
                                           )
    self.final_layer = torch.nn.Sequential(
                                           self.activation,
                                           torch.nn.Conv2d(
                                                           output_channels,
                                                           output_channels,
                                                           kernel_size = kernel_size,
                                                           padding = kernel_size // 2,
                                                           bias = bias
                                                          )
                                          )

forward(x0, x1)

Forward model.

Parameters:

  • x0 – 
             First input data.
  • x1 – 
             Seconnd input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
def forward(self, x0, x1):
    """
    Forward model.

    Parameters
    ----------
    x0             : torch.tensor
                     First input data.

    x1             : torch.tensor
                     Seconnd input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    y0 = self.convolution0(x0)
    y1 = self.convolution1(x1)
    y2 = torch.add(y0, y1)
    result = self.final_layer(y2) * x0
    return result

residual_layer

Bases: Module

A residual layer.

Source code in odak/learn/models/components.py 
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
class residual_layer(torch.nn.Module):
    """
    A residual layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 mid_channels = 16,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A convolutional layer class.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels    : int
                          Number of middle channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.activation = activation
        self.convolution = double_convolution(
                                              input_channels,
                                              mid_channels = mid_channels,
                                              output_channels = input_channels,
                                              kernel_size = kernel_size,
                                              bias = bias,
                                              activation = activation
                                             )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        x0 = self.convolution(x)
        return x + x0

__init__(input_channels=2, mid_channels=16, kernel_size=3, bias=False, activation=torch.nn.ReLU())

A convolutional layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of middle channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
def __init__(
             self,
             input_channels = 2,
             mid_channels = 16,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    A convolutional layer class.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels    : int
                      Number of middle channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.activation = activation
    self.convolution = double_convolution(
                                          input_channels,
                                          mid_channels = mid_channels,
                                          output_channels = input_channels,
                                          kernel_size = kernel_size,
                                          bias = bias,
                                          activation = activation
                                         )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    x0 = self.convolution(x)
    return x + x0

spatial_gate

Bases: Module

Spatial attention module that applies a convolution layer after channel pooling. This class is heavily inspired by https://github.com/Jongchan/attention-module/blob/master/MODELS/cbam.py.

Source code in odak/learn/models/components.py 
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
class spatial_gate(torch.nn.Module):
    """
    Spatial attention module that applies a convolution layer after channel pooling.
    This class is heavily inspired by https://github.com/Jongchan/attention-module/blob/master/MODELS/cbam.py.
    """
    def __init__(self):
        """
        Initializes the spatial gate module.
        """
        super().__init__()
        kernel_size = 7
        self.spatial = convolution_layer(2, 1, kernel_size, bias = False, activation = torch.nn.Identity())


    def channel_pool(self, x):
        """
        Applies max and average pooling on the channels.

        Parameters
        ----------
        x             : torch.tensor
                        Input tensor.

        Returns
        -------
        output        : torch.tensor
                        Output tensor.
        """
        max_pool = torch.max(x, 1)[0].unsqueeze(1)
        avg_pool = torch.mean(x, 1).unsqueeze(1)
        output = torch.cat((max_pool, avg_pool), dim=1)
        return output


    def forward(self, x):
        """
        Forward pass of the SpatialGate module.

        Applies spatial attention to the input tensor.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the SpatialGate module.

        Returns
        -------
        scaled_x     : torch.tensor
                       Output tensor after applying spatial attention.
        """
        x_compress = self.channel_pool(x)
        x_out = self.spatial(x_compress)
        scale = torch.sigmoid(x_out)
        scaled_x = x * scale
        return scaled_x

__init__()

Initializes the spatial gate module.

Source code in odak/learn/models/components.py 
672
673
674
675
676
677
678
def __init__(self):
    """
    Initializes the spatial gate module.
    """
    super().__init__()
    kernel_size = 7
    self.spatial = convolution_layer(2, 1, kernel_size, bias = False, activation = torch.nn.Identity())

channel_pool(x)

Applies max and average pooling on the channels.

Parameters:

  • x – 
            Input tensor.

Returns:

  • output ( tensor ) –

    Output tensor.

Source code in odak/learn/models/components.py 
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
def channel_pool(self, x):
    """
    Applies max and average pooling on the channels.

    Parameters
    ----------
    x             : torch.tensor
                    Input tensor.

    Returns
    -------
    output        : torch.tensor
                    Output tensor.
    """
    max_pool = torch.max(x, 1)[0].unsqueeze(1)
    avg_pool = torch.mean(x, 1).unsqueeze(1)
    output = torch.cat((max_pool, avg_pool), dim=1)
    return output

forward(x)

Forward pass of the SpatialGate module.

Applies spatial attention to the input tensor.

Parameters:

  • x – 
           Input tensor to the SpatialGate module.

Returns:

  • scaled_x ( tensor ) –

    Output tensor after applying spatial attention.

Source code in odak/learn/models/components.py 
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
def forward(self, x):
    """
    Forward pass of the SpatialGate module.

    Applies spatial attention to the input tensor.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the SpatialGate module.

    Returns
    -------
    scaled_x     : torch.tensor
                   Output tensor after applying spatial attention.
    """
    x_compress = self.channel_pool(x)
    x_out = self.spatial(x_compress)
    scale = torch.sigmoid(x_out)
    scaled_x = x * scale
    return scaled_x

spatially_adaptive_convolution

Bases: Module

A spatially adaptive convolution layer.

References

C. Zheng et al. "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions." C. Xu et al. "Squeezesegv3: Spatially-adaptive Convolution for Efficient Point-Cloud Segmentation." C. Zheng et al. "Windowing Decomposition Convolutional Neural Network for Image Enhancement."

Source code in odak/learn/models/components.py 
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
class spatially_adaptive_convolution(torch.nn.Module):
    """
    A spatially adaptive convolution layer.

    References
    ----------

    C. Zheng et al. "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions."
    C. Xu et al. "Squeezesegv3: Spatially-adaptive Convolution for Efficient Point-Cloud Segmentation."
    C. Zheng et al. "Windowing Decomposition Convolutional Neural Network for Image Enhancement."
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 stride = 1,
                 padding = 1,
                 bias = False,
                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                ):
        """
        Initializes a spatially adaptive convolution layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Size of the convolution kernel.
        stride          : int
                          Stride of the convolution.
        padding         : int
                          Padding added to both sides of the input.
        bias            : bool
                          If True, includes a bias term in the convolution.
        activation      : torch.nn.Module
                          Activation function to apply. If None, no activation is applied.
        """
        super(spatially_adaptive_convolution, self).__init__()
        self.kernel_size = kernel_size
        self.input_channels = input_channels
        self.output_channels = output_channels
        self.stride = stride
        self.padding = padding
        self.standard_convolution = torch.nn.Conv2d(
                                                    in_channels = input_channels,
                                                    out_channels = self.output_channels,
                                                    kernel_size = kernel_size,
                                                    stride = stride,
                                                    padding = padding,
                                                    bias = bias
                                                   )
        self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
        self.activation = activation


    def forward(self, x, sv_kernel_feature):
        """
        Forward pass for the spatially adaptive convolution layer.

        Parameters
        ----------
        x                  : torch.tensor
                            Input data tensor.
                            Dimension: (1, C, H, W)
        sv_kernel_feature   : torch.tensor
                            Spatially varying kernel features.
                            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

        Returns
        -------
        sa_output          : torch.tensor
                            Estimated output tensor.
                            Dimension: (1, output_channels, H_out, W_out)
        """
        # Pad input and sv_kernel_feature if necessary
        if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
                -2) * self.stride != x.size(-2):
            diffY = sv_kernel_feature.size(-2) % self.stride
            diffX = sv_kernel_feature.size(-1) % self.stride
            sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                            diffY // 2, diffY - diffY // 2))
            diffY = x.size(-2) % self.stride
            diffX = x.size(-1) % self.stride
            x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                            diffY // 2, diffY - diffY // 2))

        # Unfold the input tensor for matrix multiplication
        input_feature = torch.nn.functional.unfold(
                                                   x,
                                                   kernel_size = (self.kernel_size, self.kernel_size),
                                                   stride = self.stride,
                                                   padding = self.padding
                                                  )

        # Resize sv_kernel_feature to match the input feature
        sv_kernel = sv_kernel_feature.reshape(
                                              1,
                                              self.input_channels * self.kernel_size * self.kernel_size,
                                              (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                             )

        # Resize weight to match the input channels and kernel size
        si_kernel = self.weight.reshape(
                                        self.weight_output_channels,
                                        self.input_channels * self.kernel_size * self.kernel_size
                                       )

        # Apply spatially varying kernels
        sv_feature = input_feature * sv_kernel

        # Perform matrix multiplication
        sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                                1, self.weight_output_channels,
                                                                (x.size(-2) // self.stride),
                                                                (x.size(-1) // self.stride)
                                                               )
        return sa_output

__init__(input_channels=2, output_channels=2, kernel_size=3, stride=1, padding=1, bias=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

Initializes a spatially adaptive convolution layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Size of the convolution kernel.
  • stride – 
              Stride of the convolution.
  • padding – 
              Padding added to both sides of the input.
  • bias – 
              If True, includes a bias term in the convolution.
  • activation – 
              Activation function to apply. If None, no activation is applied.
Source code in odak/learn/models/components.py 
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             stride = 1,
             padding = 1,
             bias = False,
             activation = torch.nn.LeakyReLU(0.2, inplace = True)
            ):
    """
    Initializes a spatially adaptive convolution layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Size of the convolution kernel.
    stride          : int
                      Stride of the convolution.
    padding         : int
                      Padding added to both sides of the input.
    bias            : bool
                      If True, includes a bias term in the convolution.
    activation      : torch.nn.Module
                      Activation function to apply. If None, no activation is applied.
    """
    super(spatially_adaptive_convolution, self).__init__()
    self.kernel_size = kernel_size
    self.input_channels = input_channels
    self.output_channels = output_channels
    self.stride = stride
    self.padding = padding
    self.standard_convolution = torch.nn.Conv2d(
                                                in_channels = input_channels,
                                                out_channels = self.output_channels,
                                                kernel_size = kernel_size,
                                                stride = stride,
                                                padding = padding,
                                                bias = bias
                                               )
    self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
    self.activation = activation

forward(x, sv_kernel_feature)

Forward pass for the spatially adaptive convolution layer.

Parameters:

  • x – 
                Input data tensor.
            Dimension: (1, C, H, W)
  • sv_kernel_feature – 
                Spatially varying kernel features.
            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

Returns:

  • sa_output ( tensor ) –

    Estimated output tensor. Dimension: (1, output_channels, H_out, W_out)

Source code in odak/learn/models/components.py 
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
def forward(self, x, sv_kernel_feature):
    """
    Forward pass for the spatially adaptive convolution layer.

    Parameters
    ----------
    x                  : torch.tensor
                        Input data tensor.
                        Dimension: (1, C, H, W)
    sv_kernel_feature   : torch.tensor
                        Spatially varying kernel features.
                        Dimension: (1, C_i * kernel_size * kernel_size, H, W)

    Returns
    -------
    sa_output          : torch.tensor
                        Estimated output tensor.
                        Dimension: (1, output_channels, H_out, W_out)
    """
    # Pad input and sv_kernel_feature if necessary
    if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
            -2) * self.stride != x.size(-2):
        diffY = sv_kernel_feature.size(-2) % self.stride
        diffX = sv_kernel_feature.size(-1) % self.stride
        sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                        diffY // 2, diffY - diffY // 2))
        diffY = x.size(-2) % self.stride
        diffX = x.size(-1) % self.stride
        x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                        diffY // 2, diffY - diffY // 2))

    # Unfold the input tensor for matrix multiplication
    input_feature = torch.nn.functional.unfold(
                                               x,
                                               kernel_size = (self.kernel_size, self.kernel_size),
                                               stride = self.stride,
                                               padding = self.padding
                                              )

    # Resize sv_kernel_feature to match the input feature
    sv_kernel = sv_kernel_feature.reshape(
                                          1,
                                          self.input_channels * self.kernel_size * self.kernel_size,
                                          (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                         )

    # Resize weight to match the input channels and kernel size
    si_kernel = self.weight.reshape(
                                    self.weight_output_channels,
                                    self.input_channels * self.kernel_size * self.kernel_size
                                   )

    # Apply spatially varying kernels
    sv_feature = input_feature * sv_kernel

    # Perform matrix multiplication
    sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                            1, self.weight_output_channels,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )
    return sa_output

spatially_adaptive_module

Bases: Module

A spatially adaptive module that combines learned spatially adaptive convolutions.

References

Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.

Source code in odak/learn/models/components.py 
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
class spatially_adaptive_module(torch.nn.Module):
    """
    A spatially adaptive module that combines learned spatially adaptive convolutions.

    References
    ----------

    Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 stride = 1,
                 padding = 1,
                 bias = False,
                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                ):
        """
        Initializes a spatially adaptive module.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Size of the convolution kernel.
        stride          : int
                          Stride of the convolution.
        padding         : int
                          Padding added to both sides of the input.
        bias            : bool
                          If True, includes a bias term in the convolution.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super(spatially_adaptive_module, self).__init__()
        self.kernel_size = kernel_size
        self.input_channels = input_channels
        self.output_channels = output_channels
        self.stride = stride
        self.padding = padding
        self.weight_output_channels = self.output_channels - 1
        self.standard_convolution = torch.nn.Conv2d(
                                                    in_channels = input_channels,
                                                    out_channels = self.weight_output_channels,
                                                    kernel_size = kernel_size,
                                                    stride = stride,
                                                    padding = padding,
                                                    bias = bias
                                                   )
        self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
        self.activation = activation


    def forward(self, x, sv_kernel_feature):
        """
        Forward pass for the spatially adaptive module.

        Parameters
        ----------
        x                  : torch.tensor
                            Input data tensor.
                            Dimension: (1, C, H, W)
        sv_kernel_feature   : torch.tensor
                            Spatially varying kernel features.
                            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

        Returns
        -------
        output             : torch.tensor
                            Combined output tensor from standard and spatially adaptive convolutions.
                            Dimension: (1, output_channels, H_out, W_out)
        """
        # Pad input and sv_kernel_feature if necessary
        if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
                -2) * self.stride != x.size(-2):
            diffY = sv_kernel_feature.size(-2) % self.stride
            diffX = sv_kernel_feature.size(-1) % self.stride
            sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                            diffY // 2, diffY - diffY // 2))
            diffY = x.size(-2) % self.stride
            diffX = x.size(-1) % self.stride
            x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                            diffY // 2, diffY - diffY // 2))

        # Unfold the input tensor for matrix multiplication
        input_feature = torch.nn.functional.unfold(
                                                   x,
                                                   kernel_size = (self.kernel_size, self.kernel_size),
                                                   stride = self.stride,
                                                   padding = self.padding
                                                  )

        # Resize sv_kernel_feature to match the input feature
        sv_kernel = sv_kernel_feature.reshape(
                                              1,
                                              self.input_channels * self.kernel_size * self.kernel_size,
                                              (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                             )

        # Apply sv_kernel to the input_feature
        sv_feature = input_feature * sv_kernel

        # Original spatially varying convolution output
        sv_output = torch.sum(sv_feature, dim = 1).reshape(
                                                           1,
                                                            1,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )

        # Reshape weight for spatially adaptive convolution
        si_kernel = self.weight.reshape(
                                        self.weight_output_channels,
                                        self.input_channels * self.kernel_size * self.kernel_size
                                       )

        # Apply si_kernel on sv convolution output
        sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                                1, self.weight_output_channels,
                                                                (x.size(-2) // self.stride),
                                                                (x.size(-1) // self.stride)
                                                               )

        # Combine the outputs and apply activation function
        output = self.activation(torch.cat((sv_output, sa_output), dim = 1))
        return output

__init__(input_channels=2, output_channels=2, kernel_size=3, stride=1, padding=1, bias=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

Initializes a spatially adaptive module.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Size of the convolution kernel.
  • stride – 
              Stride of the convolution.
  • padding – 
              Padding added to both sides of the input.
  • bias – 
              If True, includes a bias term in the convolution.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             stride = 1,
             padding = 1,
             bias = False,
             activation = torch.nn.LeakyReLU(0.2, inplace = True)
            ):
    """
    Initializes a spatially adaptive module.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Size of the convolution kernel.
    stride          : int
                      Stride of the convolution.
    padding         : int
                      Padding added to both sides of the input.
    bias            : bool
                      If True, includes a bias term in the convolution.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super(spatially_adaptive_module, self).__init__()
    self.kernel_size = kernel_size
    self.input_channels = input_channels
    self.output_channels = output_channels
    self.stride = stride
    self.padding = padding
    self.weight_output_channels = self.output_channels - 1
    self.standard_convolution = torch.nn.Conv2d(
                                                in_channels = input_channels,
                                                out_channels = self.weight_output_channels,
                                                kernel_size = kernel_size,
                                                stride = stride,
                                                padding = padding,
                                                bias = bias
                                               )
    self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
    self.activation = activation

forward(x, sv_kernel_feature)

Forward pass for the spatially adaptive module.

Parameters:

  • x – 
                Input data tensor.
            Dimension: (1, C, H, W)
  • sv_kernel_feature – 
                Spatially varying kernel features.
            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

Returns:

  • output ( tensor ) –

    Combined output tensor from standard and spatially adaptive convolutions. Dimension: (1, output_channels, H_out, W_out)

Source code in odak/learn/models/components.py 
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
def forward(self, x, sv_kernel_feature):
    """
    Forward pass for the spatially adaptive module.

    Parameters
    ----------
    x                  : torch.tensor
                        Input data tensor.
                        Dimension: (1, C, H, W)
    sv_kernel_feature   : torch.tensor
                        Spatially varying kernel features.
                        Dimension: (1, C_i * kernel_size * kernel_size, H, W)

    Returns
    -------
    output             : torch.tensor
                        Combined output tensor from standard and spatially adaptive convolutions.
                        Dimension: (1, output_channels, H_out, W_out)
    """
    # Pad input and sv_kernel_feature if necessary
    if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
            -2) * self.stride != x.size(-2):
        diffY = sv_kernel_feature.size(-2) % self.stride
        diffX = sv_kernel_feature.size(-1) % self.stride
        sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                        diffY // 2, diffY - diffY // 2))
        diffY = x.size(-2) % self.stride
        diffX = x.size(-1) % self.stride
        x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                        diffY // 2, diffY - diffY // 2))

    # Unfold the input tensor for matrix multiplication
    input_feature = torch.nn.functional.unfold(
                                               x,
                                               kernel_size = (self.kernel_size, self.kernel_size),
                                               stride = self.stride,
                                               padding = self.padding
                                              )

    # Resize sv_kernel_feature to match the input feature
    sv_kernel = sv_kernel_feature.reshape(
                                          1,
                                          self.input_channels * self.kernel_size * self.kernel_size,
                                          (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                         )

    # Apply sv_kernel to the input_feature
    sv_feature = input_feature * sv_kernel

    # Original spatially varying convolution output
    sv_output = torch.sum(sv_feature, dim = 1).reshape(
                                                       1,
                                                        1,
                                                        (x.size(-2) // self.stride),
                                                        (x.size(-1) // self.stride)
                                                       )

    # Reshape weight for spatially adaptive convolution
    si_kernel = self.weight.reshape(
                                    self.weight_output_channels,
                                    self.input_channels * self.kernel_size * self.kernel_size
                                   )

    # Apply si_kernel on sv convolution output
    sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                            1, self.weight_output_channels,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )

    # Combine the outputs and apply activation function
    output = self.activation(torch.cat((sv_output, sa_output), dim = 1))
    return output

spatially_adaptive_unet

Bases: Module

Spatially varying U-Net model based on spatially adaptive convolution.

References

Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.

Source code in odak/learn/models/models.py 
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
class spatially_adaptive_unet(torch.nn.Module):
    """
    Spatially varying U-Net model based on spatially adaptive convolution.

    References
    ----------

    Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.
    """
    def __init__(
                 self,
                 depth=3,
                 dimensions=8,
                 input_channels=6,
                 out_channels=6,
                 kernel_size=3,
                 bias=True,
                 normalization=False,
                 activation=torch.nn.LeakyReLU(0.2, inplace=True)
                ):
        """
        U-Net model.

        Parameters
        ----------
        depth          : int
                         Number of upsampling and downsampling layers.
        dimensions     : int
                         Number of dimensions.
        input_channels : int
                         Number of input channels.
        out_channels   : int
                         Number of output channels.
        bias           : bool
                         Set to True to let convolutional layers learn a bias term.
        normalization  : bool
                         If True, adds a Batch Normalization layer after the convolutional layer.
        activation     : torch.nn
                         Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
        """
        super().__init__()
        self.depth = depth
        self.out_channels = out_channels
        self.inc = convolution_layer(
                                     input_channels=input_channels,
                                     output_channels=dimensions,
                                     kernel_size=kernel_size,
                                     bias=bias,
                                     normalization=normalization,
                                     activation=activation
                                    )

        self.encoder = torch.nn.ModuleList()
        for i in range(self.depth + 1):  # Downsampling layers
            down_in_channels = dimensions * (2 ** i)
            down_out_channels = 2 * down_in_channels
            pooling_layer = torch.nn.AvgPool2d(2)
            double_convolution_layer = double_convolution(
                                                          input_channels=down_in_channels,
                                                          mid_channels=down_in_channels,
                                                          output_channels=down_in_channels,
                                                          kernel_size=kernel_size,
                                                          bias=bias,
                                                          normalization=normalization,
                                                          activation=activation
                                                         )
            sam = spatially_adaptive_module(
                                            input_channels=down_in_channels,
                                            output_channels=down_out_channels,
                                            kernel_size=kernel_size,
                                            bias=bias,
                                            activation=activation
                                           )
            self.encoder.append(torch.nn.ModuleList([pooling_layer, double_convolution_layer, sam]))
        self.global_feature_module = torch.nn.ModuleList()
        double_convolution_layer = double_convolution(
                                                      input_channels=dimensions * (2 ** (depth + 1)),
                                                      mid_channels=dimensions * (2 ** (depth + 1)),
                                                      output_channels=dimensions * (2 ** (depth + 1)),
                                                      kernel_size=kernel_size,
                                                      bias=bias,
                                                      normalization=normalization,
                                                      activation=activation
                                                     )
        global_feature_layer = global_feature_module(
                                                     input_channels=dimensions * (2 ** (depth + 1)),
                                                     mid_channels=dimensions * (2 ** (depth + 1)),
                                                     output_channels=dimensions * (2 ** (depth + 1)),
                                                     kernel_size=kernel_size,
                                                     bias=bias,
                                                     activation=torch.nn.LeakyReLU(0.2, inplace=True)
                                                    )
        self.global_feature_module.append(torch.nn.ModuleList([double_convolution_layer, global_feature_layer]))
        self.decoder = torch.nn.ModuleList()
        for i in range(depth, -1, -1):
            up_in_channels = dimensions * (2 ** (i + 1))
            up_mid_channels = up_in_channels // 2
            if i == 0:
                up_out_channels = self.out_channels
                upsample_layer = upsample_convtranspose2d_layer(
                                                                input_channels=up_in_channels,
                                                                output_channels=up_mid_channels,
                                                                kernel_size=2,
                                                                stride=2,
                                                                bias=bias,
                                                               )
                conv_layer = torch.nn.Sequential(
                    convolution_layer(
                                      input_channels=up_mid_channels,
                                      output_channels=up_mid_channels,
                                      kernel_size=kernel_size,
                                      bias=bias,
                                      normalization=normalization,
                                      activation=activation,
                                     ),
                    convolution_layer(
                                      input_channels=up_mid_channels,
                                      output_channels=up_out_channels,
                                      kernel_size=1,
                                      bias=bias,
                                      normalization=normalization,
                                      activation=None,
                                     )
                )
                self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))
            else:
                up_out_channels = up_in_channels // 2
                upsample_layer = upsample_convtranspose2d_layer(
                                                                input_channels=up_in_channels,
                                                                output_channels=up_mid_channels,
                                                                kernel_size=2,
                                                                stride=2,
                                                                bias=bias,
                                                               )
                conv_layer = double_convolution(
                                                input_channels=up_mid_channels,
                                                mid_channels=up_mid_channels,
                                                output_channels=up_out_channels,
                                                kernel_size=kernel_size,
                                                bias=bias,
                                                normalization=normalization,
                                                activation=activation,
                                               )
                self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))


    def forward(self, sv_kernel, field):
        """
        Forward model.

        Parameters
        ----------
        sv_kernel : list of torch.tensor
                    Learned spatially varying kernels.
                    Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
                    where C_i, H_i, and W_i represent the channel, height, and width
                    of each feature at a certain scale.

        field     : torch.tensor
                    Input field data.
                    Dimension: (1, 6, H, W)

        Returns
        -------
        target_field : torch.tensor
                       Estimated output.
                       Dimension: (1, 6, H, W)
        """
        x = self.inc(field)
        downsampling_outputs = [x]
        for i, down_layer in enumerate(self.encoder):
            x_down = down_layer[0](downsampling_outputs[-1])
            downsampling_outputs.append(x_down)
            sam_output = down_layer[2](x_down + down_layer[1](x_down), sv_kernel[self.depth - i])
            downsampling_outputs.append(sam_output)
        global_feature = self.global_feature_module[0][0](downsampling_outputs[-1])
        global_feature = self.global_feature_module[0][1](downsampling_outputs[-1], global_feature)
        downsampling_outputs.append(global_feature)
        x_up = downsampling_outputs[-1]
        for i, up_layer in enumerate(self.decoder):
            x_up = up_layer[0](x_up, downsampling_outputs[2 * (self.depth - i)])
            x_up = up_layer[1](x_up)
        result = x_up
        return result

__init__(depth=3, dimensions=8, input_channels=6, out_channels=6, kernel_size=3, bias=True, normalization=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

U-Net model.

Parameters:

  • depth – 
             Number of upsampling and downsampling layers.
  • dimensions – 
             Number of dimensions.
  • input_channels (int, default: 6 ) – 
             Number of input channels.
  • out_channels – 
             Number of output channels.
  • bias – 
             Set to True to let convolutional layers learn a bias term.
  • normalization – 
             If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
             Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
Source code in odak/learn/models/models.py 
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
def __init__(
             self,
             depth=3,
             dimensions=8,
             input_channels=6,
             out_channels=6,
             kernel_size=3,
             bias=True,
             normalization=False,
             activation=torch.nn.LeakyReLU(0.2, inplace=True)
            ):
    """
    U-Net model.

    Parameters
    ----------
    depth          : int
                     Number of upsampling and downsampling layers.
    dimensions     : int
                     Number of dimensions.
    input_channels : int
                     Number of input channels.
    out_channels   : int
                     Number of output channels.
    bias           : bool
                     Set to True to let convolutional layers learn a bias term.
    normalization  : bool
                     If True, adds a Batch Normalization layer after the convolutional layer.
    activation     : torch.nn
                     Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
    """
    super().__init__()
    self.depth = depth
    self.out_channels = out_channels
    self.inc = convolution_layer(
                                 input_channels=input_channels,
                                 output_channels=dimensions,
                                 kernel_size=kernel_size,
                                 bias=bias,
                                 normalization=normalization,
                                 activation=activation
                                )

    self.encoder = torch.nn.ModuleList()
    for i in range(self.depth + 1):  # Downsampling layers
        down_in_channels = dimensions * (2 ** i)
        down_out_channels = 2 * down_in_channels
        pooling_layer = torch.nn.AvgPool2d(2)
        double_convolution_layer = double_convolution(
                                                      input_channels=down_in_channels,
                                                      mid_channels=down_in_channels,
                                                      output_channels=down_in_channels,
                                                      kernel_size=kernel_size,
                                                      bias=bias,
                                                      normalization=normalization,
                                                      activation=activation
                                                     )
        sam = spatially_adaptive_module(
                                        input_channels=down_in_channels,
                                        output_channels=down_out_channels,
                                        kernel_size=kernel_size,
                                        bias=bias,
                                        activation=activation
                                       )
        self.encoder.append(torch.nn.ModuleList([pooling_layer, double_convolution_layer, sam]))
    self.global_feature_module = torch.nn.ModuleList()
    double_convolution_layer = double_convolution(
                                                  input_channels=dimensions * (2 ** (depth + 1)),
                                                  mid_channels=dimensions * (2 ** (depth + 1)),
                                                  output_channels=dimensions * (2 ** (depth + 1)),
                                                  kernel_size=kernel_size,
                                                  bias=bias,
                                                  normalization=normalization,
                                                  activation=activation
                                                 )
    global_feature_layer = global_feature_module(
                                                 input_channels=dimensions * (2 ** (depth + 1)),
                                                 mid_channels=dimensions * (2 ** (depth + 1)),
                                                 output_channels=dimensions * (2 ** (depth + 1)),
                                                 kernel_size=kernel_size,
                                                 bias=bias,
                                                 activation=torch.nn.LeakyReLU(0.2, inplace=True)
                                                )
    self.global_feature_module.append(torch.nn.ModuleList([double_convolution_layer, global_feature_layer]))
    self.decoder = torch.nn.ModuleList()
    for i in range(depth, -1, -1):
        up_in_channels = dimensions * (2 ** (i + 1))
        up_mid_channels = up_in_channels // 2
        if i == 0:
            up_out_channels = self.out_channels
            upsample_layer = upsample_convtranspose2d_layer(
                                                            input_channels=up_in_channels,
                                                            output_channels=up_mid_channels,
                                                            kernel_size=2,
                                                            stride=2,
                                                            bias=bias,
                                                           )
            conv_layer = torch.nn.Sequential(
                convolution_layer(
                                  input_channels=up_mid_channels,
                                  output_channels=up_mid_channels,
                                  kernel_size=kernel_size,
                                  bias=bias,
                                  normalization=normalization,
                                  activation=activation,
                                 ),
                convolution_layer(
                                  input_channels=up_mid_channels,
                                  output_channels=up_out_channels,
                                  kernel_size=1,
                                  bias=bias,
                                  normalization=normalization,
                                  activation=None,
                                 )
            )
            self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))
        else:
            up_out_channels = up_in_channels // 2
            upsample_layer = upsample_convtranspose2d_layer(
                                                            input_channels=up_in_channels,
                                                            output_channels=up_mid_channels,
                                                            kernel_size=2,
                                                            stride=2,
                                                            bias=bias,
                                                           )
            conv_layer = double_convolution(
                                            input_channels=up_mid_channels,
                                            mid_channels=up_mid_channels,
                                            output_channels=up_out_channels,
                                            kernel_size=kernel_size,
                                            bias=bias,
                                            normalization=normalization,
                                            activation=activation,
                                           )
            self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))

forward(sv_kernel, field)

Forward model.

Parameters:

  • sv_kernel (list of torch.tensor) – 
        Learned spatially varying kernels.
    Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
    where C_i, H_i, and W_i represent the channel, height, and width
    of each feature at a certain scale.
  • field – 
        Input field data.
    Dimension: (1, 6, H, W)

Returns:

  • target_field ( tensor ) –

    Estimated output. Dimension: (1, 6, H, W)

Source code in odak/learn/models/models.py 
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
def forward(self, sv_kernel, field):
    """
    Forward model.

    Parameters
    ----------
    sv_kernel : list of torch.tensor
                Learned spatially varying kernels.
                Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
                where C_i, H_i, and W_i represent the channel, height, and width
                of each feature at a certain scale.

    field     : torch.tensor
                Input field data.
                Dimension: (1, 6, H, W)

    Returns
    -------
    target_field : torch.tensor
                   Estimated output.
                   Dimension: (1, 6, H, W)
    """
    x = self.inc(field)
    downsampling_outputs = [x]
    for i, down_layer in enumerate(self.encoder):
        x_down = down_layer[0](downsampling_outputs[-1])
        downsampling_outputs.append(x_down)
        sam_output = down_layer[2](x_down + down_layer[1](x_down), sv_kernel[self.depth - i])
        downsampling_outputs.append(sam_output)
    global_feature = self.global_feature_module[0][0](downsampling_outputs[-1])
    global_feature = self.global_feature_module[0][1](downsampling_outputs[-1], global_feature)
    downsampling_outputs.append(global_feature)
    x_up = downsampling_outputs[-1]
    for i, up_layer in enumerate(self.decoder):
        x_up = up_layer[0](x_up, downsampling_outputs[2 * (self.depth - i)])
        x_up = up_layer[1](x_up)
    result = x_up
    return result

spatially_varying_kernel_generation_model

Bases: Module

Spatially_varying_kernel_generation_model revised from RSGUnet: https://github.com/MTLab/rsgunet_image_enhance.

Refer to: J. Huang, P. Zhu, M. Geng et al. Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices.

Source code in odak/learn/models/models.py 
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
class spatially_varying_kernel_generation_model(torch.nn.Module):
    """
    Spatially_varying_kernel_generation_model revised from RSGUnet:
    https://github.com/MTLab/rsgunet_image_enhance.

    Refer to:
    J. Huang, P. Zhu, M. Geng et al. Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices.
    """

    def __init__(
                 self,
                 depth = 3,
                 dimensions = 8,
                 input_channels = 7,
                 kernel_size = 3,
                 bias = True,
                 normalization = False,
                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                ):
        """
        U-Net model.

        Parameters
        ----------
        depth          : int
                         Number of upsampling and downsampling layers.
        dimensions     : int
                         Number of dimensions.
        input_channels : int
                         Number of input channels.
        bias           : bool
                         Set to True to let convolutional layers learn a bias term.
        normalization  : bool
                         If True, adds a Batch Normalization layer after the convolutional layer.
        activation     : torch.nn
                         Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
        """
        super().__init__()
        self.depth = depth
        self.inc = convolution_layer(
                                     input_channels = input_channels,
                                     output_channels = dimensions,
                                     kernel_size = kernel_size,
                                     bias = bias,
                                     normalization = normalization,
                                     activation = activation
                                    )
        self.encoder = torch.nn.ModuleList()
        for i in range(depth + 1):  # downsampling layers
            if i == 0:
                in_channels = dimensions * (2 ** i)
                out_channels = dimensions * (2 ** i)
            elif i == depth:
                in_channels = dimensions * (2 ** (i - 1))
                out_channels = dimensions * (2 ** (i - 1))
            else:
                in_channels = dimensions * (2 ** (i - 1))
                out_channels = 2 * in_channels
            pooling_layer = torch.nn.AvgPool2d(2)
            double_convolution_layer = double_convolution(
                                                          input_channels = in_channels,
                                                          mid_channels = in_channels,
                                                          output_channels = out_channels,
                                                          kernel_size = kernel_size,
                                                          bias = bias,
                                                          normalization = normalization,
                                                          activation = activation
                                                         )
            self.encoder.append(pooling_layer)
            self.encoder.append(double_convolution_layer)
        self.spatially_varying_feature = torch.nn.ModuleList()  # for kernel generation
        for i in range(depth, -1, -1):
            if i == 1:
                svf_in_channels = dimensions + 2 ** (self.depth + i) + 1
            else:
                svf_in_channels = 2 ** (self.depth + i) + 1
            svf_out_channels = (2 ** (self.depth + i)) * (kernel_size * kernel_size)
            svf_mid_channels = dimensions * (2 ** (self.depth - 1))
            spatially_varying_kernel_generation = torch.nn.ModuleList()
            for j in range(i, -1, -1):
                pooling_layer = torch.nn.AvgPool2d(2 ** (j + 1))
                spatially_varying_kernel_generation.append(pooling_layer)
            kernel_generation_block = torch.nn.Sequential(
                torch.nn.Conv2d(
                                in_channels = svf_in_channels,
                                out_channels = svf_mid_channels,
                                kernel_size = kernel_size,
                                padding = kernel_size // 2,
                                bias = bias
                               ),
                activation,
                torch.nn.Conv2d(
                                in_channels = svf_mid_channels,
                                out_channels = svf_mid_channels,
                                kernel_size = kernel_size,
                                padding = kernel_size // 2,
                                bias = bias
                               ),
                activation,
                torch.nn.Conv2d(
                                in_channels = svf_mid_channels,
                                out_channels = svf_out_channels,
                                kernel_size = kernel_size,
                                padding = kernel_size // 2,
                                bias = bias
                               ),
            )
            spatially_varying_kernel_generation.append(kernel_generation_block)
            self.spatially_varying_feature.append(spatially_varying_kernel_generation)
        self.decoder = torch.nn.ModuleList()
        global_feature_layer = global_feature_module(  # global feature layer
                                                     input_channels = dimensions * (2 ** (depth - 1)),
                                                     mid_channels = dimensions * (2 ** (depth - 1)),
                                                     output_channels = dimensions * (2 ** (depth - 1)),
                                                     kernel_size = kernel_size,
                                                     bias = bias,
                                                     activation = torch.nn.LeakyReLU(0.2, inplace = True)
                                                    )
        self.decoder.append(global_feature_layer)
        for i in range(depth, 0, -1):
            if i == 2:
                up_in_channels = (dimensions // 2) * (2 ** i)
                up_out_channels = up_in_channels
                up_mid_channels = up_in_channels
            elif i == 1:
                up_in_channels = dimensions * 2
                up_out_channels = dimensions
                up_mid_channels = up_out_channels
            else:
                up_in_channels = (dimensions // 2) * (2 ** i)
                up_out_channels = up_in_channels // 2
                up_mid_channels = up_in_channels
            upsample_layer = upsample_convtranspose2d_layer(
                                                            input_channels = up_in_channels,
                                                            output_channels = up_mid_channels,
                                                            kernel_size = 2,
                                                            stride = 2,
                                                            bias = bias,
                                                           )
            conv_layer = double_convolution(
                                            input_channels = up_mid_channels,
                                            output_channels = up_out_channels,
                                            kernel_size = kernel_size,
                                            bias = bias,
                                            normalization = normalization,
                                            activation = activation,
                                           )
            self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))


    def forward(self, focal_surface, field):
        """
        Forward model.

        Parameters
        ----------
        focal_surface : torch.tensor
                        Input focal surface data.
                        Dimension: (1, 1, H, W)

        field         : torch.tensor
                        Input field data.
                        Dimension: (1, 6, H, W)

        Returns
        -------
        sv_kernel : list of torch.tensor
                    Learned spatially varying kernels.
                    Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
                    where C_i, H_i, and W_i represent the channel, height, and width
                    of each feature at a certain scale.
        """
        x = self.inc(torch.cat((focal_surface, field), dim = 1))
        downsampling_outputs = [focal_surface]
        downsampling_outputs.append(x)
        for i, down_layer in enumerate(self.encoder):
            x_down = down_layer(downsampling_outputs[-1])
            downsampling_outputs.append(x_down)
        sv_kernels = []
        for i, (up_layer, svf_layer) in enumerate(zip(self.decoder, self.spatially_varying_feature)):
            if i == 0:
                global_feature = up_layer(downsampling_outputs[-2], downsampling_outputs[-1])
                downsampling_outputs[-1] = global_feature
                sv_feature = [global_feature, downsampling_outputs[0]]
                for j in range(self.depth - i + 1):
                    sv_feature[1] = svf_layer[self.depth - i](sv_feature[1])
                    if j > 0:
                        sv_feature.append(svf_layer[j](downsampling_outputs[2 * j]))
                sv_feature = [sv_feature[0], sv_feature[1], sv_feature[4], sv_feature[2],
                              sv_feature[3]]
                sv_kernel = svf_layer[-1](torch.cat(sv_feature, dim = 1))
                sv_kernels.append(sv_kernel)
            else:
                x_up = up_layer[0](downsampling_outputs[-1],
                                   downsampling_outputs[2 * (self.depth + 1 - i) + 1])
                x_up = up_layer[1](x_up)
                downsampling_outputs[-1] = x_up
                sv_feature = [x_up, downsampling_outputs[0]]
                for j in range(self.depth - i + 1):
                    sv_feature[1] = svf_layer[self.depth - i](sv_feature[1])
                    if j > 0:
                        sv_feature.append(svf_layer[j](downsampling_outputs[2 * j]))
                if i == 1:
                    sv_feature = [sv_feature[0], sv_feature[1], sv_feature[3], sv_feature[2]]
                sv_kernel = svf_layer[-1](torch.cat(sv_feature, dim = 1))
                sv_kernels.append(sv_kernel)
        return sv_kernels

__init__(depth=3, dimensions=8, input_channels=7, kernel_size=3, bias=True, normalization=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

U-Net model.

Parameters:

  • depth – 
             Number of upsampling and downsampling layers.
  • dimensions – 
             Number of dimensions.
  • input_channels (int, default: 7 ) – 
             Number of input channels.
  • bias – 
             Set to True to let convolutional layers learn a bias term.
  • normalization – 
             If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
             Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
Source code in odak/learn/models/models.py 
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
def __init__(
             self,
             depth = 3,
             dimensions = 8,
             input_channels = 7,
             kernel_size = 3,
             bias = True,
             normalization = False,
             activation = torch.nn.LeakyReLU(0.2, inplace = True)
            ):
    """
    U-Net model.

    Parameters
    ----------
    depth          : int
                     Number of upsampling and downsampling layers.
    dimensions     : int
                     Number of dimensions.
    input_channels : int
                     Number of input channels.
    bias           : bool
                     Set to True to let convolutional layers learn a bias term.
    normalization  : bool
                     If True, adds a Batch Normalization layer after the convolutional layer.
    activation     : torch.nn
                     Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
    """
    super().__init__()
    self.depth = depth
    self.inc = convolution_layer(
                                 input_channels = input_channels,
                                 output_channels = dimensions,
                                 kernel_size = kernel_size,
                                 bias = bias,
                                 normalization = normalization,
                                 activation = activation
                                )
    self.encoder = torch.nn.ModuleList()
    for i in range(depth + 1):  # downsampling layers
        if i == 0:
            in_channels = dimensions * (2 ** i)
            out_channels = dimensions * (2 ** i)
        elif i == depth:
            in_channels = dimensions * (2 ** (i - 1))
            out_channels = dimensions * (2 ** (i - 1))
        else:
            in_channels = dimensions * (2 ** (i - 1))
            out_channels = 2 * in_channels
        pooling_layer = torch.nn.AvgPool2d(2)
        double_convolution_layer = double_convolution(
                                                      input_channels = in_channels,
                                                      mid_channels = in_channels,
                                                      output_channels = out_channels,
                                                      kernel_size = kernel_size,
                                                      bias = bias,
                                                      normalization = normalization,
                                                      activation = activation
                                                     )
        self.encoder.append(pooling_layer)
        self.encoder.append(double_convolution_layer)
    self.spatially_varying_feature = torch.nn.ModuleList()  # for kernel generation
    for i in range(depth, -1, -1):
        if i == 1:
            svf_in_channels = dimensions + 2 ** (self.depth + i) + 1
        else:
            svf_in_channels = 2 ** (self.depth + i) + 1
        svf_out_channels = (2 ** (self.depth + i)) * (kernel_size * kernel_size)
        svf_mid_channels = dimensions * (2 ** (self.depth - 1))
        spatially_varying_kernel_generation = torch.nn.ModuleList()
        for j in range(i, -1, -1):
            pooling_layer = torch.nn.AvgPool2d(2 ** (j + 1))
            spatially_varying_kernel_generation.append(pooling_layer)
        kernel_generation_block = torch.nn.Sequential(
            torch.nn.Conv2d(
                            in_channels = svf_in_channels,
                            out_channels = svf_mid_channels,
                            kernel_size = kernel_size,
                            padding = kernel_size // 2,
                            bias = bias
                           ),
            activation,
            torch.nn.Conv2d(
                            in_channels = svf_mid_channels,
                            out_channels = svf_mid_channels,
                            kernel_size = kernel_size,
                            padding = kernel_size // 2,
                            bias = bias
                           ),
            activation,
            torch.nn.Conv2d(
                            in_channels = svf_mid_channels,
                            out_channels = svf_out_channels,
                            kernel_size = kernel_size,
                            padding = kernel_size // 2,
                            bias = bias
                           ),
        )
        spatially_varying_kernel_generation.append(kernel_generation_block)
        self.spatially_varying_feature.append(spatially_varying_kernel_generation)
    self.decoder = torch.nn.ModuleList()
    global_feature_layer = global_feature_module(  # global feature layer
                                                 input_channels = dimensions * (2 ** (depth - 1)),
                                                 mid_channels = dimensions * (2 ** (depth - 1)),
                                                 output_channels = dimensions * (2 ** (depth - 1)),
                                                 kernel_size = kernel_size,
                                                 bias = bias,
                                                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                                                )
    self.decoder.append(global_feature_layer)
    for i in range(depth, 0, -1):
        if i == 2:
            up_in_channels = (dimensions // 2) * (2 ** i)
            up_out_channels = up_in_channels
            up_mid_channels = up_in_channels
        elif i == 1:
            up_in_channels = dimensions * 2
            up_out_channels = dimensions
            up_mid_channels = up_out_channels
        else:
            up_in_channels = (dimensions // 2) * (2 ** i)
            up_out_channels = up_in_channels // 2
            up_mid_channels = up_in_channels
        upsample_layer = upsample_convtranspose2d_layer(
                                                        input_channels = up_in_channels,
                                                        output_channels = up_mid_channels,
                                                        kernel_size = 2,
                                                        stride = 2,
                                                        bias = bias,
                                                       )
        conv_layer = double_convolution(
                                        input_channels = up_mid_channels,
                                        output_channels = up_out_channels,
                                        kernel_size = kernel_size,
                                        bias = bias,
                                        normalization = normalization,
                                        activation = activation,
                                       )
        self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))

forward(focal_surface, field)

Forward model.

Parameters:

  • focal_surface (tensor) – 
            Input focal surface data.
        Dimension: (1, 1, H, W)
  • field – 
            Input field data.
        Dimension: (1, 6, H, W)

Returns:

  • sv_kernel ( list of torch.tensor ) –

    Learned spatially varying kernels. Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i), where C_i, H_i, and W_i represent the channel, height, and width of each feature at a certain scale.

Source code in odak/learn/models/models.py 
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
def forward(self, focal_surface, field):
    """
    Forward model.

    Parameters
    ----------
    focal_surface : torch.tensor
                    Input focal surface data.
                    Dimension: (1, 1, H, W)

    field         : torch.tensor
                    Input field data.
                    Dimension: (1, 6, H, W)

    Returns
    -------
    sv_kernel : list of torch.tensor
                Learned spatially varying kernels.
                Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
                where C_i, H_i, and W_i represent the channel, height, and width
                of each feature at a certain scale.
    """
    x = self.inc(torch.cat((focal_surface, field), dim = 1))
    downsampling_outputs = [focal_surface]
    downsampling_outputs.append(x)
    for i, down_layer in enumerate(self.encoder):
        x_down = down_layer(downsampling_outputs[-1])
        downsampling_outputs.append(x_down)
    sv_kernels = []
    for i, (up_layer, svf_layer) in enumerate(zip(self.decoder, self.spatially_varying_feature)):
        if i == 0:
            global_feature = up_layer(downsampling_outputs[-2], downsampling_outputs[-1])
            downsampling_outputs[-1] = global_feature
            sv_feature = [global_feature, downsampling_outputs[0]]
            for j in range(self.depth - i + 1):
                sv_feature[1] = svf_layer[self.depth - i](sv_feature[1])
                if j > 0:
                    sv_feature.append(svf_layer[j](downsampling_outputs[2 * j]))
            sv_feature = [sv_feature[0], sv_feature[1], sv_feature[4], sv_feature[2],
                          sv_feature[3]]
            sv_kernel = svf_layer[-1](torch.cat(sv_feature, dim = 1))
            sv_kernels.append(sv_kernel)
        else:
            x_up = up_layer[0](downsampling_outputs[-1],
                               downsampling_outputs[2 * (self.depth + 1 - i) + 1])
            x_up = up_layer[1](x_up)
            downsampling_outputs[-1] = x_up
            sv_feature = [x_up, downsampling_outputs[0]]
            for j in range(self.depth - i + 1):
                sv_feature[1] = svf_layer[self.depth - i](sv_feature[1])
                if j > 0:
                    sv_feature.append(svf_layer[j](downsampling_outputs[2 * j]))
            if i == 1:
                sv_feature = [sv_feature[0], sv_feature[1], sv_feature[3], sv_feature[2]]
            sv_kernel = svf_layer[-1](torch.cat(sv_feature, dim = 1))
            sv_kernels.append(sv_kernel)
    return sv_kernels

unet

Bases: Module

A U-Net model, heavily inspired from https://github.com/milesial/Pytorch-UNet/tree/master/unet and more can be read from Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Publishing, 2015.

Source code in odak/learn/models/models.py 
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
class unet(torch.nn.Module):
    """
    A U-Net model, heavily inspired from `https://github.com/milesial/Pytorch-UNet/tree/master/unet` and more can be read from Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Publishing, 2015.
    """

    def __init__(
                 self, 
                 depth = 4,
                 dimensions = 64, 
                 input_channels = 2, 
                 output_channels = 1, 
                 bilinear = False,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU(inplace = True),
                ):
        """
        U-Net model.

        Parameters
        ----------
        depth             : int
                            Number of upsampling and downsampling
        dimensions        : int
                            Number of dimensions.
        input_channels    : int
                            Number of input channels.
        output_channels   : int
                            Number of output channels.
        bilinear          : bool
                            Uses bilinear upsampling in upsampling layers when set True.
        bias              : bool
                            Set True to let convolutional layers learn a bias term.
        activation        : torch.nn
                            Non-linear activation layer to be used (e.g., torch.nn.ReLU(), torch.nn.Sigmoid().
        """
        super(unet, self).__init__()
        self.inc = double_convolution(
                                      input_channels = input_channels,
                                      mid_channels = dimensions,
                                      output_channels = dimensions,
                                      kernel_size = kernel_size,
                                      bias = bias,
                                      activation = activation
                                     )      

        self.downsampling_layers = torch.nn.ModuleList()
        self.upsampling_layers = torch.nn.ModuleList()
        for i in range(depth): # downsampling layers
            in_channels = dimensions * (2 ** i)
            out_channels = dimensions * (2 ** (i + 1))
            down_layer = downsample_layer(in_channels,
                                            out_channels,
                                            kernel_size=kernel_size,
                                            bias=bias,
                                            activation=activation
                                            )
            self.downsampling_layers.append(down_layer)      

        for i in range(depth - 1, -1, -1):  # upsampling layers
            up_in_channels = dimensions * (2 ** (i + 1))  
            up_out_channels = dimensions * (2 ** i) 
            up_layer = upsample_layer(up_in_channels, up_out_channels, kernel_size=kernel_size, bias=bias, activation=activation, bilinear=bilinear)
            self.upsampling_layers.append(up_layer)
        self.outc = torch.nn.Conv2d(
                                    dimensions, 
                                    output_channels,
                                    kernel_size = kernel_size,
                                    padding = kernel_size // 2,
                                    bias = bias
                                   )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        downsampling_outputs = [self.inc(x)]
        for down_layer in self.downsampling_layers:
            x_down = down_layer(downsampling_outputs[-1])
            downsampling_outputs.append(x_down)
        x_up = downsampling_outputs[-1]
        for i, up_layer in enumerate((self.upsampling_layers)):
            x_up = up_layer(x_up, downsampling_outputs[-(i + 2)])       
        result = self.outc(x_up)
        return result

__init__(depth=4, dimensions=64, input_channels=2, output_channels=1, bilinear=False, kernel_size=3, bias=False, activation=torch.nn.ReLU(inplace=True))

U-Net model.

Parameters:

  • depth – 
                Number of upsampling and downsampling
  • dimensions – 
                Number of dimensions.
  • input_channels – 
                Number of input channels.
  • output_channels – 
                Number of output channels.
  • bilinear – 
                Uses bilinear upsampling in upsampling layers when set True.
  • bias – 
                Set True to let convolutional layers learn a bias term.
  • activation – 
                Non-linear activation layer to be used (e.g., torch.nn.ReLU(), torch.nn.Sigmoid().
Source code in odak/learn/models/models.py 
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
def __init__(
             self, 
             depth = 4,
             dimensions = 64, 
             input_channels = 2, 
             output_channels = 1, 
             bilinear = False,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU(inplace = True),
            ):
    """
    U-Net model.

    Parameters
    ----------
    depth             : int
                        Number of upsampling and downsampling
    dimensions        : int
                        Number of dimensions.
    input_channels    : int
                        Number of input channels.
    output_channels   : int
                        Number of output channels.
    bilinear          : bool
                        Uses bilinear upsampling in upsampling layers when set True.
    bias              : bool
                        Set True to let convolutional layers learn a bias term.
    activation        : torch.nn
                        Non-linear activation layer to be used (e.g., torch.nn.ReLU(), torch.nn.Sigmoid().
    """
    super(unet, self).__init__()
    self.inc = double_convolution(
                                  input_channels = input_channels,
                                  mid_channels = dimensions,
                                  output_channels = dimensions,
                                  kernel_size = kernel_size,
                                  bias = bias,
                                  activation = activation
                                 )      

    self.downsampling_layers = torch.nn.ModuleList()
    self.upsampling_layers = torch.nn.ModuleList()
    for i in range(depth): # downsampling layers
        in_channels = dimensions * (2 ** i)
        out_channels = dimensions * (2 ** (i + 1))
        down_layer = downsample_layer(in_channels,
                                        out_channels,
                                        kernel_size=kernel_size,
                                        bias=bias,
                                        activation=activation
                                        )
        self.downsampling_layers.append(down_layer)      

    for i in range(depth - 1, -1, -1):  # upsampling layers
        up_in_channels = dimensions * (2 ** (i + 1))  
        up_out_channels = dimensions * (2 ** i) 
        up_layer = upsample_layer(up_in_channels, up_out_channels, kernel_size=kernel_size, bias=bias, activation=activation, bilinear=bilinear)
        self.upsampling_layers.append(up_layer)
    self.outc = torch.nn.Conv2d(
                                dimensions, 
                                output_channels,
                                kernel_size = kernel_size,
                                padding = kernel_size // 2,
                                bias = bias
                               )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/models.py 
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    downsampling_outputs = [self.inc(x)]
    for down_layer in self.downsampling_layers:
        x_down = down_layer(downsampling_outputs[-1])
        downsampling_outputs.append(x_down)
    x_up = downsampling_outputs[-1]
    for i, up_layer in enumerate((self.upsampling_layers)):
        x_up = up_layer(x_up, downsampling_outputs[-(i + 2)])       
    result = self.outc(x_up)
    return result

upsample_convtranspose2d_layer

Bases: Module

An upsampling convtranspose2d layer.

Source code in odak/learn/models/components.py 
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
class upsample_convtranspose2d_layer(torch.nn.Module):
    """
    An upsampling convtranspose2d layer.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 2,
                 stride = 2,
                 bias = False,
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        """
        super().__init__()
        self.up = torch.nn.ConvTranspose2d(
                                           in_channels = input_channels,
                                           out_channels = output_channels,
                                           bias = bias,
                                           kernel_size = kernel_size,
                                           stride = stride
                                          )

    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.


        Returns
        ----------
        result        : torch.tensor
                        Result of the forward operation
        """
        x1 = self.up(x1)
        diffY = x2.size()[2] - x1.size()[2]
        diffX = x2.size()[3] - x1.size()[3]
        x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                          diffY // 2, diffY - diffY // 2])
        result = x1 + x2
        return result

__init__(input_channels, output_channels, kernel_size=2, stride=2, bias=False)

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
Source code in odak/learn/models/components.py 
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 2,
             stride = 2,
             bias = False,
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    """
    super().__init__()
    self.up = torch.nn.ConvTranspose2d(
                                       in_channels = input_channels,
                                       out_channels = output_channels,
                                       bias = bias,
                                       kernel_size = kernel_size,
                                       stride = stride
                                      )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.


    Returns
    ----------
    result        : torch.tensor
                    Result of the forward operation
    """
    x1 = self.up(x1)
    diffY = x2.size()[2] - x1.size()[2]
    diffX = x2.size()[3] - x1.size()[3]
    x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                      diffY // 2, diffY - diffY // 2])
    result = x1 + x2
    return result

upsample_layer

Bases: Module

An upsampling convolutional layer.

Source code in odak/learn/models/components.py 
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
class upsample_layer(torch.nn.Module):
    """
    An upsampling convolutional layer.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU(),
                 bilinear = True
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        bilinear        : bool
                          If set to True, bilinear sampling is used.
        """
        super(upsample_layer, self).__init__()
        if bilinear:
            self.up = torch.nn.Upsample(scale_factor = 2, mode = 'bilinear', align_corners = True)
            self.conv = double_convolution(
                                           input_channels = input_channels + output_channels,
                                           mid_channels = input_channels // 2,
                                           output_channels = output_channels,
                                           kernel_size = kernel_size,
                                           bias = bias,
                                           activation = activation
                                          )
        else:
            self.up = torch.nn.ConvTranspose2d(input_channels , input_channels // 2, kernel_size = 2, stride = 2)
            self.conv = double_convolution(
                                           input_channels = input_channels,
                                           mid_channels = output_channels,
                                           output_channels = output_channels,
                                           kernel_size = kernel_size,
                                           bias = bias,
                                           activation = activation
                                          )


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.


        Returns
        ----------
        result        : torch.tensor
                        Result of the forward operation
        """ 
        x1 = self.up(x1)
        diffY = x2.size()[2] - x1.size()[2]
        diffX = x2.size()[3] - x1.size()[3]
        x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                          diffY // 2, diffY - diffY // 2])
        x = torch.cat([x2, x1], dim = 1)
        result = self.conv(x)
        return result

__init__(input_channels, output_channels, kernel_size=3, bias=False, activation=torch.nn.ReLU(), bilinear=True)

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
  • bilinear – 
              If set to True, bilinear sampling is used.
Source code in odak/learn/models/components.py 
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU(),
             bilinear = True
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    bilinear        : bool
                      If set to True, bilinear sampling is used.
    """
    super(upsample_layer, self).__init__()
    if bilinear:
        self.up = torch.nn.Upsample(scale_factor = 2, mode = 'bilinear', align_corners = True)
        self.conv = double_convolution(
                                       input_channels = input_channels + output_channels,
                                       mid_channels = input_channels // 2,
                                       output_channels = output_channels,
                                       kernel_size = kernel_size,
                                       bias = bias,
                                       activation = activation
                                      )
    else:
        self.up = torch.nn.ConvTranspose2d(input_channels , input_channels // 2, kernel_size = 2, stride = 2)
        self.conv = double_convolution(
                                       input_channels = input_channels,
                                       mid_channels = output_channels,
                                       output_channels = output_channels,
                                       kernel_size = kernel_size,
                                       bias = bias,
                                       activation = activation
                                      )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.


    Returns
    ----------
    result        : torch.tensor
                    Result of the forward operation
    """ 
    x1 = self.up(x1)
    diffY = x2.size()[2] - x1.size()[2]
    diffX = x2.size()[3] - x1.size()[3]
    x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                      diffY // 2, diffY - diffY // 2])
    x = torch.cat([x2, x1], dim = 1)
    result = self.conv(x)
    return result

gaussian(x, multiplier=1.0)

A Gaussian non-linear activation. For more details: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.

Parameters:

  • x – 
           Input data.
  • multiplier – 
           Multiplier.

Returns:

  • result ( float or tensor ) –

    Ouput data.

Source code in odak/learn/models/components.py 
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
def gaussian(x, multiplier = 1.):
    """
    A Gaussian non-linear activation.
    For more details: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.

    Parameters
    ----------
    x            : float or torch.tensor
                   Input data.
    multiplier   : float or torch.tensor
                   Multiplier.

    Returns
    -------
    result       : float or torch.tensor
                   Ouput data.
    """
    result = torch.exp(- (multiplier * x) ** 2)
    return result

swish(x)

A swish non-linear activation. For more details: https://en.wikipedia.org/wiki/Swish_function

Parameters:

  • x – 
             Input.

Returns:

  • out ( float or tensor ) –

    Output.

Source code in odak/learn/models/components.py 
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def swish(x):
    """
    A swish non-linear activation.
    For more details: https://en.wikipedia.org/wiki/Swish_function

    Parameters
    -----------
    x              : float or torch.tensor
                     Input.

    Returns
    -------
    out            : float or torch.tensor
                     Output.
    """
    out = x * torch.sigmoid(x)
    return out

channel_gate

Bases: Module

Channel attention module with various pooling strategies. This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).

Source code in odak/learn/models/components.py 
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
class channel_gate(torch.nn.Module):
    """
    Channel attention module with various pooling strategies.
    This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).
    """
    def __init__(
                 self, 
                 gate_channels, 
                 reduction_ratio = 16, 
                 pool_types = ['avg', 'max']
                ):
        """
        Initializes the channel gate module.

        Parameters
        ----------
        gate_channels   : int
                          Number of channels of the input feature map.
        reduction_ratio : int
                          Reduction ratio for the intermediate layer.
        pool_types      : list
                          List of pooling operations to apply.
        """
        super().__init__()
        self.gate_channels = gate_channels
        hidden_channels = gate_channels // reduction_ratio
        if hidden_channels == 0:
            hidden_channels = 1
        self.mlp = torch.nn.Sequential(
                                       convolutional_block_attention.Flatten(),
                                       torch.nn.Linear(gate_channels, hidden_channels),
                                       torch.nn.ReLU(),
                                       torch.nn.Linear(hidden_channels, gate_channels)
                                      )
        self.pool_types = pool_types


    def forward(self, x):
        """
        Forward pass of the ChannelGate module.

        Applies channel-wise attention to the input tensor.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the ChannelGate module.

        Returns
        -------
        output       : torch.tensor
                       Output tensor after applying channel attention.
        """
        channel_att_sum = None
        for pool_type in self.pool_types:
            if pool_type == 'avg':
                pool = torch.nn.functional.avg_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
            elif pool_type == 'max':
                pool = torch.nn.functional.max_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
            channel_att_raw = self.mlp(pool)
            channel_att_sum = channel_att_raw if channel_att_sum is None else channel_att_sum + channel_att_raw
        scale = torch.sigmoid(channel_att_sum).unsqueeze(2).unsqueeze(3).expand_as(x)
        output = x * scale
        return output

__init__(gate_channels, reduction_ratio=16, pool_types=['avg', 'max'])

Initializes the channel gate module.

Parameters:

  • gate_channels – 
              Number of channels of the input feature map.
  • reduction_ratio (int, default: 16 ) – 
              Reduction ratio for the intermediate layer.
  • pool_types – 
              List of pooling operations to apply.
Source code in odak/learn/models/components.py 
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
def __init__(
             self, 
             gate_channels, 
             reduction_ratio = 16, 
             pool_types = ['avg', 'max']
            ):
    """
    Initializes the channel gate module.

    Parameters
    ----------
    gate_channels   : int
                      Number of channels of the input feature map.
    reduction_ratio : int
                      Reduction ratio for the intermediate layer.
    pool_types      : list
                      List of pooling operations to apply.
    """
    super().__init__()
    self.gate_channels = gate_channels
    hidden_channels = gate_channels // reduction_ratio
    if hidden_channels == 0:
        hidden_channels = 1
    self.mlp = torch.nn.Sequential(
                                   convolutional_block_attention.Flatten(),
                                   torch.nn.Linear(gate_channels, hidden_channels),
                                   torch.nn.ReLU(),
                                   torch.nn.Linear(hidden_channels, gate_channels)
                                  )
    self.pool_types = pool_types

forward(x)

Forward pass of the ChannelGate module.

Applies channel-wise attention to the input tensor.

Parameters:

  • x – 
           Input tensor to the ChannelGate module.

Returns:

  • output ( tensor ) –

    Output tensor after applying channel attention.

Source code in odak/learn/models/components.py 
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
def forward(self, x):
    """
    Forward pass of the ChannelGate module.

    Applies channel-wise attention to the input tensor.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the ChannelGate module.

    Returns
    -------
    output       : torch.tensor
                   Output tensor after applying channel attention.
    """
    channel_att_sum = None
    for pool_type in self.pool_types:
        if pool_type == 'avg':
            pool = torch.nn.functional.avg_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
        elif pool_type == 'max':
            pool = torch.nn.functional.max_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
        channel_att_raw = self.mlp(pool)
        channel_att_sum = channel_att_raw if channel_att_sum is None else channel_att_sum + channel_att_raw
    scale = torch.sigmoid(channel_att_sum).unsqueeze(2).unsqueeze(3).expand_as(x)
    output = x * scale
    return output

convolution_layer

Bases: Module

A convolution layer.

Source code in odak/learn/models/components.py 
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
class convolution_layer(torch.nn.Module):
    """
    A convolution layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 bias = False,
                 stride = 1,
                 normalization = True,
                 activation = torch.nn.ReLU()
                ):
        """
        A convolutional layer class.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        layers = [
            torch.nn.Conv2d(
                            input_channels,
                            output_channels,
                            kernel_size = kernel_size,
                            stride = stride,
                            padding = kernel_size // 2,
                            bias = bias
                           )
        ]
        if normalization:
            layers.append(torch.nn.BatchNorm2d(output_channels))
        if activation:
            layers.append(activation)
        self.model = torch.nn.Sequential(*layers)


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        result = self.model(x)
        return result

__init__(input_channels=2, output_channels=2, kernel_size=3, bias=False, stride=1, normalization=True, activation=torch.nn.ReLU())

A convolutional layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             bias = False,
             stride = 1,
             normalization = True,
             activation = torch.nn.ReLU()
            ):
    """
    A convolutional layer class.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    layers = [
        torch.nn.Conv2d(
                        input_channels,
                        output_channels,
                        kernel_size = kernel_size,
                        stride = stride,
                        padding = kernel_size // 2,
                        bias = bias
                       )
    ]
    if normalization:
        layers.append(torch.nn.BatchNorm2d(output_channels))
    if activation:
        layers.append(activation)
    self.model = torch.nn.Sequential(*layers)

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    result = self.model(x)
    return result

convolutional_block_attention

Bases: Module

Convolutional Block Attention Module (CBAM) class. This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).

Source code in odak/learn/models/components.py 
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
class convolutional_block_attention(torch.nn.Module):
    """
    Convolutional Block Attention Module (CBAM) class. 
    This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).
    """
    def __init__(
                 self, 
                 gate_channels, 
                 reduction_ratio = 16, 
                 pool_types = ['avg', 'max'], 
                 no_spatial = False
                ):
        """
        Initializes the convolutional block attention module.

        Parameters
        ----------
        gate_channels   : int
                          Number of channels of the input feature map.
        reduction_ratio : int
                          Reduction ratio for the channel attention.
        pool_types      : list
                          List of pooling operations to apply for channel attention.
        no_spatial      : bool
                          If True, spatial attention is not applied.
        """
        super(convolutional_block_attention, self).__init__()
        self.channel_gate = channel_gate(gate_channels, reduction_ratio, pool_types)
        self.no_spatial = no_spatial
        if not no_spatial:
            self.spatial_gate = spatial_gate()


    class Flatten(torch.nn.Module):
        """
        Flattens the input tensor to a 2D matrix.
        """
        def forward(self, x):
            return x.view(x.size(0), -1)


    def forward(self, x):
        """
        Forward pass of the convolutional block attention module.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the CBAM module.

        Returns
        -------
        x_out        : torch.tensor
                       Output tensor after applying channel and spatial attention.
        """
        x_out = self.channel_gate(x)
        if not self.no_spatial:
            x_out = self.spatial_gate(x_out)
        return x_out

Flatten

Bases: Module

Flattens the input tensor to a 2D matrix.

Source code in odak/learn/models/components.py 
757
758
759
760
761
762
class Flatten(torch.nn.Module):
    """
    Flattens the input tensor to a 2D matrix.
    """
    def forward(self, x):
        return x.view(x.size(0), -1)

__init__(gate_channels, reduction_ratio=16, pool_types=['avg', 'max'], no_spatial=False)

Initializes the convolutional block attention module.

Parameters:

  • gate_channels – 
              Number of channels of the input feature map.
  • reduction_ratio (int, default: 16 ) – 
              Reduction ratio for the channel attention.
  • pool_types – 
              List of pooling operations to apply for channel attention.
  • no_spatial – 
              If True, spatial attention is not applied.
Source code in odak/learn/models/components.py 
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
def __init__(
             self, 
             gate_channels, 
             reduction_ratio = 16, 
             pool_types = ['avg', 'max'], 
             no_spatial = False
            ):
    """
    Initializes the convolutional block attention module.

    Parameters
    ----------
    gate_channels   : int
                      Number of channels of the input feature map.
    reduction_ratio : int
                      Reduction ratio for the channel attention.
    pool_types      : list
                      List of pooling operations to apply for channel attention.
    no_spatial      : bool
                      If True, spatial attention is not applied.
    """
    super(convolutional_block_attention, self).__init__()
    self.channel_gate = channel_gate(gate_channels, reduction_ratio, pool_types)
    self.no_spatial = no_spatial
    if not no_spatial:
        self.spatial_gate = spatial_gate()

forward(x)

Forward pass of the convolutional block attention module.

Parameters:

  • x – 
           Input tensor to the CBAM module.

Returns:

  • x_out ( tensor ) –

    Output tensor after applying channel and spatial attention.

Source code in odak/learn/models/components.py 
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
def forward(self, x):
    """
    Forward pass of the convolutional block attention module.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the CBAM module.

    Returns
    -------
    x_out        : torch.tensor
                   Output tensor after applying channel and spatial attention.
    """
    x_out = self.channel_gate(x)
    if not self.no_spatial:
        x_out = self.spatial_gate(x_out)
    return x_out

double_convolution

Bases: Module

A double convolution layer.

Source code in odak/learn/models/components.py 
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
class double_convolution(torch.nn.Module):
    """
    A double convolution layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 mid_channels = None,
                 output_channels = 2,
                 kernel_size = 3, 
                 bias = False,
                 normalization = True,
                 activation = torch.nn.ReLU()
                ):
        """
        Double convolution model.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels    : int
                          Number of channels in the hidden layer between two convolutions.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        if isinstance(mid_channels, type(None)):
            mid_channels = output_channels
        self.activation = activation
        self.model = torch.nn.Sequential(
                                         convolution_layer(
                                                           input_channels = input_channels,
                                                           output_channels = mid_channels,
                                                           kernel_size = kernel_size,
                                                           bias = bias,
                                                           normalization = normalization,
                                                           activation = self.activation
                                                          ),
                                         convolution_layer(
                                                           input_channels = mid_channels,
                                                           output_channels = output_channels,
                                                           kernel_size = kernel_size,
                                                           bias = bias,
                                                           normalization = normalization,
                                                           activation = self.activation
                                                          )
                                        )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        result = self.model(x)
        return result

__init__(input_channels=2, mid_channels=None, output_channels=2, kernel_size=3, bias=False, normalization=True, activation=torch.nn.ReLU())

Double convolution model.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of channels in the hidden layer between two convolutions.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
def __init__(
             self,
             input_channels = 2,
             mid_channels = None,
             output_channels = 2,
             kernel_size = 3, 
             bias = False,
             normalization = True,
             activation = torch.nn.ReLU()
            ):
    """
    Double convolution model.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels    : int
                      Number of channels in the hidden layer between two convolutions.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    if isinstance(mid_channels, type(None)):
        mid_channels = output_channels
    self.activation = activation
    self.model = torch.nn.Sequential(
                                     convolution_layer(
                                                       input_channels = input_channels,
                                                       output_channels = mid_channels,
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       normalization = normalization,
                                                       activation = self.activation
                                                      ),
                                     convolution_layer(
                                                       input_channels = mid_channels,
                                                       output_channels = output_channels,
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       normalization = normalization,
                                                       activation = self.activation
                                                      )
                                    )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    result = self.model(x)
    return result

downsample_layer

Bases: Module

A downscaling component followed by a double convolution.

Source code in odak/learn/models/components.py 
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
class downsample_layer(torch.nn.Module):
    """
    A downscaling component followed by a double convolution.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.maxpool_conv = torch.nn.Sequential(
                                                torch.nn.MaxPool2d(2),
                                                double_convolution(
                                                                   input_channels = input_channels,
                                                                   mid_channels = output_channels,
                                                                   output_channels = output_channels,
                                                                   kernel_size = kernel_size,
                                                                   bias = bias,
                                                                   activation = activation
                                                                  )
                                               )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x              : torch.tensor
                         First input data.



        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        result = self.maxpool_conv(x)
        return result

__init__(input_channels, output_channels, kernel_size=3, bias=False, activation=torch.nn.ReLU())

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.maxpool_conv = torch.nn.Sequential(
                                            torch.nn.MaxPool2d(2),
                                            double_convolution(
                                                               input_channels = input_channels,
                                                               mid_channels = output_channels,
                                                               output_channels = output_channels,
                                                               kernel_size = kernel_size,
                                                               bias = bias,
                                                               activation = activation
                                                              )
                                           )

forward(x)

Forward model.

Parameters:

  • x – 
             First input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x              : torch.tensor
                     First input data.



    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    result = self.maxpool_conv(x)
    return result

global_feature_module

Bases: Module

A global feature layer that processes global features from input channels and applies them to another input tensor via learned transformations.

Source code in odak/learn/models/components.py 
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
class global_feature_module(torch.nn.Module):
    """
    A global feature layer that processes global features from input channels and
    applies them to another input tensor via learned transformations.
    """
    def __init__(
                 self,
                 input_channels,
                 mid_channels,
                 output_channels,
                 kernel_size,
                 bias = False,
                 normalization = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A global feature layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels  : int
                          Number of mid channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.transformations_1 = global_transformations(input_channels, output_channels)
        self.global_features_1 = double_convolution(
                                                    input_channels = input_channels,
                                                    mid_channels = mid_channels,
                                                    output_channels = output_channels,
                                                    kernel_size = kernel_size,
                                                    bias = bias,
                                                    normalization = normalization,
                                                    activation = activation
                                                   )
        self.global_features_2 = double_convolution(
                                                    input_channels = input_channels,
                                                    mid_channels = mid_channels,
                                                    output_channels = output_channels,
                                                    kernel_size = kernel_size,
                                                    bias = bias,
                                                    normalization = normalization,
                                                    activation = activation
                                                   )
        self.transformations_2 = global_transformations(input_channels, output_channels)


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.

        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        global_tensor_1 = self.transformations_1(x1, x2)
        y1 = self.global_features_1(global_tensor_1)
        y2 = self.global_features_2(y1)
        global_tensor_2 = self.transformations_2(y1, y2)
        return global_tensor_2

__init__(input_channels, mid_channels, output_channels, kernel_size, bias=False, normalization=False, activation=torch.nn.ReLU())

A global feature layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of mid channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
def __init__(
             self,
             input_channels,
             mid_channels,
             output_channels,
             kernel_size,
             bias = False,
             normalization = False,
             activation = torch.nn.ReLU()
            ):
    """
    A global feature layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels  : int
                      Number of mid channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.transformations_1 = global_transformations(input_channels, output_channels)
    self.global_features_1 = double_convolution(
                                                input_channels = input_channels,
                                                mid_channels = mid_channels,
                                                output_channels = output_channels,
                                                kernel_size = kernel_size,
                                                bias = bias,
                                                normalization = normalization,
                                                activation = activation
                                               )
    self.global_features_2 = double_convolution(
                                                input_channels = input_channels,
                                                mid_channels = mid_channels,
                                                output_channels = output_channels,
                                                kernel_size = kernel_size,
                                                bias = bias,
                                                normalization = normalization,
                                                activation = activation
                                               )
    self.transformations_2 = global_transformations(input_channels, output_channels)

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.

    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    global_tensor_1 = self.transformations_1(x1, x2)
    y1 = self.global_features_1(global_tensor_1)
    y2 = self.global_features_2(y1)
    global_tensor_2 = self.transformations_2(y1, y2)
    return global_tensor_2

global_transformations

Bases: Module

A global feature layer that processes global features from input channels and applies learned transformations to another input tensor.

This implementation is adapted from RSGUnet: https://github.com/MTLab/rsgunet_image_enhance.

Reference: J. Huang, P. Zhu, M. Geng et al. "Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices."

Source code in odak/learn/models/components.py 
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
class global_transformations(torch.nn.Module):
    """
    A global feature layer that processes global features from input channels and
    applies learned transformations to another input tensor.

    This implementation is adapted from RSGUnet:
    https://github.com/MTLab/rsgunet_image_enhance.

    Reference:
    J. Huang, P. Zhu, M. Geng et al. "Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices."
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels
                ):
        """
        A global feature layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        """
        super().__init__()
        self.global_feature_1 = torch.nn.Sequential(
            torch.nn.Linear(input_channels, output_channels),
            torch.nn.LeakyReLU(0.2, inplace = True),
        )
        self.global_feature_2 = torch.nn.Sequential(
            torch.nn.Linear(output_channels, output_channels),
            torch.nn.LeakyReLU(0.2, inplace = True)
        )


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.

        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        y = torch.mean(x2, dim = (2, 3))
        y1 = self.global_feature_1(y)
        y2 = self.global_feature_2(y1)
        y1 = y1.unsqueeze(2).unsqueeze(3)
        y2 = y2.unsqueeze(2).unsqueeze(3)
        result = x1 * y1 + y2
        return result

__init__(input_channels, output_channels)

A global feature layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
Source code in odak/learn/models/components.py 
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
def __init__(
             self,
             input_channels,
             output_channels
            ):
    """
    A global feature layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    """
    super().__init__()
    self.global_feature_1 = torch.nn.Sequential(
        torch.nn.Linear(input_channels, output_channels),
        torch.nn.LeakyReLU(0.2, inplace = True),
    )
    self.global_feature_2 = torch.nn.Sequential(
        torch.nn.Linear(output_channels, output_channels),
        torch.nn.LeakyReLU(0.2, inplace = True)
    )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.

    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    y = torch.mean(x2, dim = (2, 3))
    y1 = self.global_feature_1(y)
    y2 = self.global_feature_2(y1)
    y1 = y1.unsqueeze(2).unsqueeze(3)
    y2 = y2.unsqueeze(2).unsqueeze(3)
    result = x1 * y1 + y2
    return result

non_local_layer

Bases: Module

Self-Attention Layer [zi = Wzyi + xi] (non-local block : ref https://arxiv.org/abs/1711.07971)

Source code in odak/learn/models/components.py 
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
class non_local_layer(torch.nn.Module):
    """
    Self-Attention Layer [zi = Wzyi + xi] (non-local block : ref https://arxiv.org/abs/1711.07971)
    """
    def __init__(
                 self,
                 input_channels = 1024,
                 bottleneck_channels = 512,
                 kernel_size = 1,
                 bias = False,
                ):
        """

        Parameters
        ----------
        input_channels      : int
                              Number of input channels.
        bottleneck_channels : int
                              Number of middle channels.
        kernel_size         : int
                              Kernel size.
        bias                : bool 
                              Set to True to let convolutional layers have bias term.
        """
        super(non_local_layer, self).__init__()
        self.input_channels = input_channels
        self.bottleneck_channels = bottleneck_channels
        self.g = torch.nn.Conv2d(
                                 self.input_channels, 
                                 self.bottleneck_channels,
                                 kernel_size = kernel_size,
                                 padding = kernel_size // 2,
                                 bias = bias
                                )
        self.W_z = torch.nn.Sequential(
                                       torch.nn.Conv2d(
                                                       self.bottleneck_channels,
                                                       self.input_channels, 
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       padding = kernel_size // 2
                                                      ),
                                       torch.nn.BatchNorm2d(self.input_channels)
                                      )
        torch.nn.init.constant_(self.W_z[1].weight, 0)   
        torch.nn.init.constant_(self.W_z[1].bias, 0)


    def forward(self, x):
        """
        Forward model [zi = Wzyi + xi]

        Parameters
        ----------
        x               : torch.tensor
                          First input data.                       


        Returns
        ----------
        z               : torch.tensor
                          Estimated output.
        """
        batch_size, channels, height, width = x.size()
        theta = x.view(batch_size, channels, -1).permute(0, 2, 1)
        phi = x.view(batch_size, channels, -1).permute(0, 2, 1)
        g = self.g(x).view(batch_size, self.bottleneck_channels, -1).permute(0, 2, 1)
        attn = torch.bmm(theta, phi.transpose(1, 2)) / (height * width)
        attn = torch.nn.functional.softmax(attn, dim=-1)
        y = torch.bmm(attn, g).permute(0, 2, 1).contiguous().view(batch_size, self.bottleneck_channels, height, width)
        W_y = self.W_z(y)
        z = W_y + x
        return z

__init__(input_channels=1024, bottleneck_channels=512, kernel_size=1, bias=False)

Parameters:

  • input_channels – 
                  Number of input channels.
  • bottleneck_channels (int, default: 512 ) – 
                  Number of middle channels.
  • kernel_size – 
                  Kernel size.
  • bias – 
                  Set to True to let convolutional layers have bias term.
Source code in odak/learn/models/components.py 
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
def __init__(
             self,
             input_channels = 1024,
             bottleneck_channels = 512,
             kernel_size = 1,
             bias = False,
            ):
    """

    Parameters
    ----------
    input_channels      : int
                          Number of input channels.
    bottleneck_channels : int
                          Number of middle channels.
    kernel_size         : int
                          Kernel size.
    bias                : bool 
                          Set to True to let convolutional layers have bias term.
    """
    super(non_local_layer, self).__init__()
    self.input_channels = input_channels
    self.bottleneck_channels = bottleneck_channels
    self.g = torch.nn.Conv2d(
                             self.input_channels, 
                             self.bottleneck_channels,
                             kernel_size = kernel_size,
                             padding = kernel_size // 2,
                             bias = bias
                            )
    self.W_z = torch.nn.Sequential(
                                   torch.nn.Conv2d(
                                                   self.bottleneck_channels,
                                                   self.input_channels, 
                                                   kernel_size = kernel_size,
                                                   bias = bias,
                                                   padding = kernel_size // 2
                                                  ),
                                   torch.nn.BatchNorm2d(self.input_channels)
                                  )
    torch.nn.init.constant_(self.W_z[1].weight, 0)   
    torch.nn.init.constant_(self.W_z[1].bias, 0)

forward(x)

Forward model [zi = Wzyi + xi]

Parameters:

  • x – 
              First input data.

Returns:

  • z ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
def forward(self, x):
    """
    Forward model [zi = Wzyi + xi]

    Parameters
    ----------
    x               : torch.tensor
                      First input data.                       


    Returns
    ----------
    z               : torch.tensor
                      Estimated output.
    """
    batch_size, channels, height, width = x.size()
    theta = x.view(batch_size, channels, -1).permute(0, 2, 1)
    phi = x.view(batch_size, channels, -1).permute(0, 2, 1)
    g = self.g(x).view(batch_size, self.bottleneck_channels, -1).permute(0, 2, 1)
    attn = torch.bmm(theta, phi.transpose(1, 2)) / (height * width)
    attn = torch.nn.functional.softmax(attn, dim=-1)
    y = torch.bmm(attn, g).permute(0, 2, 1).contiguous().view(batch_size, self.bottleneck_channels, height, width)
    W_y = self.W_z(y)
    z = W_y + x
    return z

normalization

Bases: Module

A normalization layer.

Source code in odak/learn/models/components.py 
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
class normalization(torch.nn.Module):
    """
    A normalization layer.
    """
    def __init__(
                 self,
                 dim = 1,
                ):
        """
        Normalization layer.


        Parameters
        ----------
        dim             : int
                          Dimension (axis) to normalize.
        """
        super().__init__()
        self.k = torch.nn.Parameter(torch.ones(1, dim, 1, 1))


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        eps = 1e-5 if x.dtype == torch.float32 else 1e-3
        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
        mean = torch.mean(x, dim = 1, keepdim = True)
        result =  (x - mean) * (var + eps).rsqrt() * self.k
        return result

__init__(dim=1)

Normalization layer.

Parameters:

  • dim – 
              Dimension (axis) to normalize.
Source code in odak/learn/models/components.py 
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
def __init__(
             self,
             dim = 1,
            ):
    """
    Normalization layer.


    Parameters
    ----------
    dim             : int
                      Dimension (axis) to normalize.
    """
    super().__init__()
    self.k = torch.nn.Parameter(torch.ones(1, dim, 1, 1))

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    eps = 1e-5 if x.dtype == torch.float32 else 1e-3
    var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
    mean = torch.mean(x, dim = 1, keepdim = True)
    result =  (x - mean) * (var + eps).rsqrt() * self.k
    return result

positional_encoder

Bases: Module

A positional encoder module.

Source code in odak/learn/models/components.py 
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
class positional_encoder(torch.nn.Module):
    """
    A positional encoder module.
    """

    def __init__(self, L):
        """
        A positional encoder module.

        Parameters
        ----------
        L                   : int
                              Positional encoding level.
        """
        super(positional_encoder, self).__init__()
        self.L = L


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x               : torch.tensor
                          Input data.

        Returns
        ----------
        result          : torch.tensor
                          Result of the forward operation
        """
        B, C = x.shape
        x = x.view(B, C, 1)
        results = [x]
        for i in range(1, self.L + 1):
            freq = (2 ** i) * torch.pi
            cos_x = torch.cos(freq * x)
            sin_x = torch.sin(freq * x)
            results.append(cos_x)
            results.append(sin_x)
        results = torch.cat(results, dim=2)
        results = results.permute(0, 2, 1)
        results = results.reshape(B, -1)
        return results

__init__(L)

A positional encoder module.

Parameters:

  • L – 
                  Positional encoding level.
Source code in odak/learn/models/components.py 
790
791
792
793
794
795
796
797
798
799
800
def __init__(self, L):
    """
    A positional encoder module.

    Parameters
    ----------
    L                   : int
                          Positional encoding level.
    """
    super(positional_encoder, self).__init__()
    self.L = L

forward(x)

Forward model.

Parameters:

  • x – 
              Input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x               : torch.tensor
                      Input data.

    Returns
    ----------
    result          : torch.tensor
                      Result of the forward operation
    """
    B, C = x.shape
    x = x.view(B, C, 1)
    results = [x]
    for i in range(1, self.L + 1):
        freq = (2 ** i) * torch.pi
        cos_x = torch.cos(freq * x)
        sin_x = torch.sin(freq * x)
        results.append(cos_x)
        results.append(sin_x)
    results = torch.cat(results, dim=2)
    results = results.permute(0, 2, 1)
    results = results.reshape(B, -1)
    return results

residual_attention_layer

Bases: Module

A residual block with an attention layer.

Source code in odak/learn/models/components.py 
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
class residual_attention_layer(torch.nn.Module):
    """
    A residual block with an attention layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 1,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        An attention layer class.


        Parameters
        ----------
        input_channels  : int or optioal
                          Number of input channels.
        output_channels : int or optional
                          Number of middle channels.
        kernel_size     : int or optional
                          Kernel size.
        bias            : bool or optional
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn or optional
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.activation = activation
        self.convolution0 = torch.nn.Sequential(
                                                torch.nn.Conv2d(
                                                                input_channels,
                                                                output_channels,
                                                                kernel_size = kernel_size,
                                                                padding = kernel_size // 2,
                                                                bias = bias
                                                               ),
                                                torch.nn.BatchNorm2d(output_channels)
                                               )
        self.convolution1 = torch.nn.Sequential(
                                                torch.nn.Conv2d(
                                                                input_channels,
                                                                output_channels,
                                                                kernel_size = kernel_size,
                                                                padding = kernel_size // 2,
                                                                bias = bias
                                                               ),
                                                torch.nn.BatchNorm2d(output_channels)
                                               )
        self.final_layer = torch.nn.Sequential(
                                               self.activation,
                                               torch.nn.Conv2d(
                                                               output_channels,
                                                               output_channels,
                                                               kernel_size = kernel_size,
                                                               padding = kernel_size // 2,
                                                               bias = bias
                                                              )
                                              )


    def forward(self, x0, x1):
        """
        Forward model.

        Parameters
        ----------
        x0             : torch.tensor
                         First input data.

        x1             : torch.tensor
                         Seconnd input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        y0 = self.convolution0(x0)
        y1 = self.convolution1(x1)
        y2 = torch.add(y0, y1)
        result = self.final_layer(y2) * x0
        return result

__init__(input_channels=2, output_channels=2, kernel_size=1, bias=False, activation=torch.nn.ReLU())

An attention layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int or optional, default: 2 ) – 
              Number of middle channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 1,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    An attention layer class.


    Parameters
    ----------
    input_channels  : int or optioal
                      Number of input channels.
    output_channels : int or optional
                      Number of middle channels.
    kernel_size     : int or optional
                      Kernel size.
    bias            : bool or optional
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn or optional
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.activation = activation
    self.convolution0 = torch.nn.Sequential(
                                            torch.nn.Conv2d(
                                                            input_channels,
                                                            output_channels,
                                                            kernel_size = kernel_size,
                                                            padding = kernel_size // 2,
                                                            bias = bias
                                                           ),
                                            torch.nn.BatchNorm2d(output_channels)
                                           )
    self.convolution1 = torch.nn.Sequential(
                                            torch.nn.Conv2d(
                                                            input_channels,
                                                            output_channels,
                                                            kernel_size = kernel_size,
                                                            padding = kernel_size // 2,
                                                            bias = bias
                                                           ),
                                            torch.nn.BatchNorm2d(output_channels)
                                           )
    self.final_layer = torch.nn.Sequential(
                                           self.activation,
                                           torch.nn.Conv2d(
                                                           output_channels,
                                                           output_channels,
                                                           kernel_size = kernel_size,
                                                           padding = kernel_size // 2,
                                                           bias = bias
                                                          )
                                          )

forward(x0, x1)

Forward model.

Parameters:

  • x0 – 
             First input data.
  • x1 – 
             Seconnd input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
def forward(self, x0, x1):
    """
    Forward model.

    Parameters
    ----------
    x0             : torch.tensor
                     First input data.

    x1             : torch.tensor
                     Seconnd input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    y0 = self.convolution0(x0)
    y1 = self.convolution1(x1)
    y2 = torch.add(y0, y1)
    result = self.final_layer(y2) * x0
    return result

residual_layer

Bases: Module

A residual layer.

Source code in odak/learn/models/components.py 
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
class residual_layer(torch.nn.Module):
    """
    A residual layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 mid_channels = 16,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A convolutional layer class.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels    : int
                          Number of middle channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.activation = activation
        self.convolution = double_convolution(
                                              input_channels,
                                              mid_channels = mid_channels,
                                              output_channels = input_channels,
                                              kernel_size = kernel_size,
                                              bias = bias,
                                              activation = activation
                                             )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        x0 = self.convolution(x)
        return x + x0

__init__(input_channels=2, mid_channels=16, kernel_size=3, bias=False, activation=torch.nn.ReLU())

A convolutional layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of middle channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
def __init__(
             self,
             input_channels = 2,
             mid_channels = 16,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    A convolutional layer class.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels    : int
                      Number of middle channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.activation = activation
    self.convolution = double_convolution(
                                          input_channels,
                                          mid_channels = mid_channels,
                                          output_channels = input_channels,
                                          kernel_size = kernel_size,
                                          bias = bias,
                                          activation = activation
                                         )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    x0 = self.convolution(x)
    return x + x0

spatial_gate

Bases: Module

Spatial attention module that applies a convolution layer after channel pooling. This class is heavily inspired by https://github.com/Jongchan/attention-module/blob/master/MODELS/cbam.py.

Source code in odak/learn/models/components.py 
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
class spatial_gate(torch.nn.Module):
    """
    Spatial attention module that applies a convolution layer after channel pooling.
    This class is heavily inspired by https://github.com/Jongchan/attention-module/blob/master/MODELS/cbam.py.
    """
    def __init__(self):
        """
        Initializes the spatial gate module.
        """
        super().__init__()
        kernel_size = 7
        self.spatial = convolution_layer(2, 1, kernel_size, bias = False, activation = torch.nn.Identity())


    def channel_pool(self, x):
        """
        Applies max and average pooling on the channels.

        Parameters
        ----------
        x             : torch.tensor
                        Input tensor.

        Returns
        -------
        output        : torch.tensor
                        Output tensor.
        """
        max_pool = torch.max(x, 1)[0].unsqueeze(1)
        avg_pool = torch.mean(x, 1).unsqueeze(1)
        output = torch.cat((max_pool, avg_pool), dim=1)
        return output


    def forward(self, x):
        """
        Forward pass of the SpatialGate module.

        Applies spatial attention to the input tensor.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the SpatialGate module.

        Returns
        -------
        scaled_x     : torch.tensor
                       Output tensor after applying spatial attention.
        """
        x_compress = self.channel_pool(x)
        x_out = self.spatial(x_compress)
        scale = torch.sigmoid(x_out)
        scaled_x = x * scale
        return scaled_x

__init__()

Initializes the spatial gate module.

Source code in odak/learn/models/components.py 
672
673
674
675
676
677
678
def __init__(self):
    """
    Initializes the spatial gate module.
    """
    super().__init__()
    kernel_size = 7
    self.spatial = convolution_layer(2, 1, kernel_size, bias = False, activation = torch.nn.Identity())

channel_pool(x)

Applies max and average pooling on the channels.

Parameters:

  • x – 
            Input tensor.

Returns:

  • output ( tensor ) –

    Output tensor.

Source code in odak/learn/models/components.py 
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
def channel_pool(self, x):
    """
    Applies max and average pooling on the channels.

    Parameters
    ----------
    x             : torch.tensor
                    Input tensor.

    Returns
    -------
    output        : torch.tensor
                    Output tensor.
    """
    max_pool = torch.max(x, 1)[0].unsqueeze(1)
    avg_pool = torch.mean(x, 1).unsqueeze(1)
    output = torch.cat((max_pool, avg_pool), dim=1)
    return output

forward(x)

Forward pass of the SpatialGate module.

Applies spatial attention to the input tensor.

Parameters:

  • x – 
           Input tensor to the SpatialGate module.

Returns:

  • scaled_x ( tensor ) –

    Output tensor after applying spatial attention.

Source code in odak/learn/models/components.py 
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
def forward(self, x):
    """
    Forward pass of the SpatialGate module.

    Applies spatial attention to the input tensor.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the SpatialGate module.

    Returns
    -------
    scaled_x     : torch.tensor
                   Output tensor after applying spatial attention.
    """
    x_compress = self.channel_pool(x)
    x_out = self.spatial(x_compress)
    scale = torch.sigmoid(x_out)
    scaled_x = x * scale
    return scaled_x

spatially_adaptive_convolution

Bases: Module

A spatially adaptive convolution layer.

References

C. Zheng et al. "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions." C. Xu et al. "Squeezesegv3: Spatially-adaptive Convolution for Efficient Point-Cloud Segmentation." C. Zheng et al. "Windowing Decomposition Convolutional Neural Network for Image Enhancement."

Source code in odak/learn/models/components.py 
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
class spatially_adaptive_convolution(torch.nn.Module):
    """
    A spatially adaptive convolution layer.

    References
    ----------

    C. Zheng et al. "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions."
    C. Xu et al. "Squeezesegv3: Spatially-adaptive Convolution for Efficient Point-Cloud Segmentation."
    C. Zheng et al. "Windowing Decomposition Convolutional Neural Network for Image Enhancement."
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 stride = 1,
                 padding = 1,
                 bias = False,
                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                ):
        """
        Initializes a spatially adaptive convolution layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Size of the convolution kernel.
        stride          : int
                          Stride of the convolution.
        padding         : int
                          Padding added to both sides of the input.
        bias            : bool
                          If True, includes a bias term in the convolution.
        activation      : torch.nn.Module
                          Activation function to apply. If None, no activation is applied.
        """
        super(spatially_adaptive_convolution, self).__init__()
        self.kernel_size = kernel_size
        self.input_channels = input_channels
        self.output_channels = output_channels
        self.stride = stride
        self.padding = padding
        self.standard_convolution = torch.nn.Conv2d(
                                                    in_channels = input_channels,
                                                    out_channels = self.output_channels,
                                                    kernel_size = kernel_size,
                                                    stride = stride,
                                                    padding = padding,
                                                    bias = bias
                                                   )
        self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
        self.activation = activation


    def forward(self, x, sv_kernel_feature):
        """
        Forward pass for the spatially adaptive convolution layer.

        Parameters
        ----------
        x                  : torch.tensor
                            Input data tensor.
                            Dimension: (1, C, H, W)
        sv_kernel_feature   : torch.tensor
                            Spatially varying kernel features.
                            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

        Returns
        -------
        sa_output          : torch.tensor
                            Estimated output tensor.
                            Dimension: (1, output_channels, H_out, W_out)
        """
        # Pad input and sv_kernel_feature if necessary
        if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
                -2) * self.stride != x.size(-2):
            diffY = sv_kernel_feature.size(-2) % self.stride
            diffX = sv_kernel_feature.size(-1) % self.stride
            sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                            diffY // 2, diffY - diffY // 2))
            diffY = x.size(-2) % self.stride
            diffX = x.size(-1) % self.stride
            x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                            diffY // 2, diffY - diffY // 2))

        # Unfold the input tensor for matrix multiplication
        input_feature = torch.nn.functional.unfold(
                                                   x,
                                                   kernel_size = (self.kernel_size, self.kernel_size),
                                                   stride = self.stride,
                                                   padding = self.padding
                                                  )

        # Resize sv_kernel_feature to match the input feature
        sv_kernel = sv_kernel_feature.reshape(
                                              1,
                                              self.input_channels * self.kernel_size * self.kernel_size,
                                              (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                             )

        # Resize weight to match the input channels and kernel size
        si_kernel = self.weight.reshape(
                                        self.weight_output_channels,
                                        self.input_channels * self.kernel_size * self.kernel_size
                                       )

        # Apply spatially varying kernels
        sv_feature = input_feature * sv_kernel

        # Perform matrix multiplication
        sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                                1, self.weight_output_channels,
                                                                (x.size(-2) // self.stride),
                                                                (x.size(-1) // self.stride)
                                                               )
        return sa_output

__init__(input_channels=2, output_channels=2, kernel_size=3, stride=1, padding=1, bias=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

Initializes a spatially adaptive convolution layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Size of the convolution kernel.
  • stride – 
              Stride of the convolution.
  • padding – 
              Padding added to both sides of the input.
  • bias – 
              If True, includes a bias term in the convolution.
  • activation – 
              Activation function to apply. If None, no activation is applied.
Source code in odak/learn/models/components.py 
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             stride = 1,
             padding = 1,
             bias = False,
             activation = torch.nn.LeakyReLU(0.2, inplace = True)
            ):
    """
    Initializes a spatially adaptive convolution layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Size of the convolution kernel.
    stride          : int
                      Stride of the convolution.
    padding         : int
                      Padding added to both sides of the input.
    bias            : bool
                      If True, includes a bias term in the convolution.
    activation      : torch.nn.Module
                      Activation function to apply. If None, no activation is applied.
    """
    super(spatially_adaptive_convolution, self).__init__()
    self.kernel_size = kernel_size
    self.input_channels = input_channels
    self.output_channels = output_channels
    self.stride = stride
    self.padding = padding
    self.standard_convolution = torch.nn.Conv2d(
                                                in_channels = input_channels,
                                                out_channels = self.output_channels,
                                                kernel_size = kernel_size,
                                                stride = stride,
                                                padding = padding,
                                                bias = bias
                                               )
    self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
    self.activation = activation

forward(x, sv_kernel_feature)

Forward pass for the spatially adaptive convolution layer.

Parameters:

  • x – 
                Input data tensor.
            Dimension: (1, C, H, W)
  • sv_kernel_feature – 
                Spatially varying kernel features.
            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

Returns:

  • sa_output ( tensor ) –

    Estimated output tensor. Dimension: (1, output_channels, H_out, W_out)

Source code in odak/learn/models/components.py 
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
def forward(self, x, sv_kernel_feature):
    """
    Forward pass for the spatially adaptive convolution layer.

    Parameters
    ----------
    x                  : torch.tensor
                        Input data tensor.
                        Dimension: (1, C, H, W)
    sv_kernel_feature   : torch.tensor
                        Spatially varying kernel features.
                        Dimension: (1, C_i * kernel_size * kernel_size, H, W)

    Returns
    -------
    sa_output          : torch.tensor
                        Estimated output tensor.
                        Dimension: (1, output_channels, H_out, W_out)
    """
    # Pad input and sv_kernel_feature if necessary
    if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
            -2) * self.stride != x.size(-2):
        diffY = sv_kernel_feature.size(-2) % self.stride
        diffX = sv_kernel_feature.size(-1) % self.stride
        sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                        diffY // 2, diffY - diffY // 2))
        diffY = x.size(-2) % self.stride
        diffX = x.size(-1) % self.stride
        x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                        diffY // 2, diffY - diffY // 2))

    # Unfold the input tensor for matrix multiplication
    input_feature = torch.nn.functional.unfold(
                                               x,
                                               kernel_size = (self.kernel_size, self.kernel_size),
                                               stride = self.stride,
                                               padding = self.padding
                                              )

    # Resize sv_kernel_feature to match the input feature
    sv_kernel = sv_kernel_feature.reshape(
                                          1,
                                          self.input_channels * self.kernel_size * self.kernel_size,
                                          (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                         )

    # Resize weight to match the input channels and kernel size
    si_kernel = self.weight.reshape(
                                    self.weight_output_channels,
                                    self.input_channels * self.kernel_size * self.kernel_size
                                   )

    # Apply spatially varying kernels
    sv_feature = input_feature * sv_kernel

    # Perform matrix multiplication
    sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                            1, self.weight_output_channels,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )
    return sa_output

spatially_adaptive_module

Bases: Module

A spatially adaptive module that combines learned spatially adaptive convolutions.

References

Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.

Source code in odak/learn/models/components.py 
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
class spatially_adaptive_module(torch.nn.Module):
    """
    A spatially adaptive module that combines learned spatially adaptive convolutions.

    References
    ----------

    Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 stride = 1,
                 padding = 1,
                 bias = False,
                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                ):
        """
        Initializes a spatially adaptive module.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Size of the convolution kernel.
        stride          : int
                          Stride of the convolution.
        padding         : int
                          Padding added to both sides of the input.
        bias            : bool
                          If True, includes a bias term in the convolution.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super(spatially_adaptive_module, self).__init__()
        self.kernel_size = kernel_size
        self.input_channels = input_channels
        self.output_channels = output_channels
        self.stride = stride
        self.padding = padding
        self.weight_output_channels = self.output_channels - 1
        self.standard_convolution = torch.nn.Conv2d(
                                                    in_channels = input_channels,
                                                    out_channels = self.weight_output_channels,
                                                    kernel_size = kernel_size,
                                                    stride = stride,
                                                    padding = padding,
                                                    bias = bias
                                                   )
        self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
        self.activation = activation


    def forward(self, x, sv_kernel_feature):
        """
        Forward pass for the spatially adaptive module.

        Parameters
        ----------
        x                  : torch.tensor
                            Input data tensor.
                            Dimension: (1, C, H, W)
        sv_kernel_feature   : torch.tensor
                            Spatially varying kernel features.
                            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

        Returns
        -------
        output             : torch.tensor
                            Combined output tensor from standard and spatially adaptive convolutions.
                            Dimension: (1, output_channels, H_out, W_out)
        """
        # Pad input and sv_kernel_feature if necessary
        if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
                -2) * self.stride != x.size(-2):
            diffY = sv_kernel_feature.size(-2) % self.stride
            diffX = sv_kernel_feature.size(-1) % self.stride
            sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                            diffY // 2, diffY - diffY // 2))
            diffY = x.size(-2) % self.stride
            diffX = x.size(-1) % self.stride
            x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                            diffY // 2, diffY - diffY // 2))

        # Unfold the input tensor for matrix multiplication
        input_feature = torch.nn.functional.unfold(
                                                   x,
                                                   kernel_size = (self.kernel_size, self.kernel_size),
                                                   stride = self.stride,
                                                   padding = self.padding
                                                  )

        # Resize sv_kernel_feature to match the input feature
        sv_kernel = sv_kernel_feature.reshape(
                                              1,
                                              self.input_channels * self.kernel_size * self.kernel_size,
                                              (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                             )

        # Apply sv_kernel to the input_feature
        sv_feature = input_feature * sv_kernel

        # Original spatially varying convolution output
        sv_output = torch.sum(sv_feature, dim = 1).reshape(
                                                           1,
                                                            1,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )

        # Reshape weight for spatially adaptive convolution
        si_kernel = self.weight.reshape(
                                        self.weight_output_channels,
                                        self.input_channels * self.kernel_size * self.kernel_size
                                       )

        # Apply si_kernel on sv convolution output
        sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                                1, self.weight_output_channels,
                                                                (x.size(-2) // self.stride),
                                                                (x.size(-1) // self.stride)
                                                               )

        # Combine the outputs and apply activation function
        output = self.activation(torch.cat((sv_output, sa_output), dim = 1))
        return output

__init__(input_channels=2, output_channels=2, kernel_size=3, stride=1, padding=1, bias=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

Initializes a spatially adaptive module.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Size of the convolution kernel.
  • stride – 
              Stride of the convolution.
  • padding – 
              Padding added to both sides of the input.
  • bias – 
              If True, includes a bias term in the convolution.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             stride = 1,
             padding = 1,
             bias = False,
             activation = torch.nn.LeakyReLU(0.2, inplace = True)
            ):
    """
    Initializes a spatially adaptive module.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Size of the convolution kernel.
    stride          : int
                      Stride of the convolution.
    padding         : int
                      Padding added to both sides of the input.
    bias            : bool
                      If True, includes a bias term in the convolution.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super(spatially_adaptive_module, self).__init__()
    self.kernel_size = kernel_size
    self.input_channels = input_channels
    self.output_channels = output_channels
    self.stride = stride
    self.padding = padding
    self.weight_output_channels = self.output_channels - 1
    self.standard_convolution = torch.nn.Conv2d(
                                                in_channels = input_channels,
                                                out_channels = self.weight_output_channels,
                                                kernel_size = kernel_size,
                                                stride = stride,
                                                padding = padding,
                                                bias = bias
                                               )
    self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
    self.activation = activation

forward(x, sv_kernel_feature)

Forward pass for the spatially adaptive module.

Parameters:

  • x – 
                Input data tensor.
            Dimension: (1, C, H, W)
  • sv_kernel_feature – 
                Spatially varying kernel features.
            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

Returns:

  • output ( tensor ) –

    Combined output tensor from standard and spatially adaptive convolutions. Dimension: (1, output_channels, H_out, W_out)

Source code in odak/learn/models/components.py 
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
def forward(self, x, sv_kernel_feature):
    """
    Forward pass for the spatially adaptive module.

    Parameters
    ----------
    x                  : torch.tensor
                        Input data tensor.
                        Dimension: (1, C, H, W)
    sv_kernel_feature   : torch.tensor
                        Spatially varying kernel features.
                        Dimension: (1, C_i * kernel_size * kernel_size, H, W)

    Returns
    -------
    output             : torch.tensor
                        Combined output tensor from standard and spatially adaptive convolutions.
                        Dimension: (1, output_channels, H_out, W_out)
    """
    # Pad input and sv_kernel_feature if necessary
    if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
            -2) * self.stride != x.size(-2):
        diffY = sv_kernel_feature.size(-2) % self.stride
        diffX = sv_kernel_feature.size(-1) % self.stride
        sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                        diffY // 2, diffY - diffY // 2))
        diffY = x.size(-2) % self.stride
        diffX = x.size(-1) % self.stride
        x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                        diffY // 2, diffY - diffY // 2))

    # Unfold the input tensor for matrix multiplication
    input_feature = torch.nn.functional.unfold(
                                               x,
                                               kernel_size = (self.kernel_size, self.kernel_size),
                                               stride = self.stride,
                                               padding = self.padding
                                              )

    # Resize sv_kernel_feature to match the input feature
    sv_kernel = sv_kernel_feature.reshape(
                                          1,
                                          self.input_channels * self.kernel_size * self.kernel_size,
                                          (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                         )

    # Apply sv_kernel to the input_feature
    sv_feature = input_feature * sv_kernel

    # Original spatially varying convolution output
    sv_output = torch.sum(sv_feature, dim = 1).reshape(
                                                       1,
                                                        1,
                                                        (x.size(-2) // self.stride),
                                                        (x.size(-1) // self.stride)
                                                       )

    # Reshape weight for spatially adaptive convolution
    si_kernel = self.weight.reshape(
                                    self.weight_output_channels,
                                    self.input_channels * self.kernel_size * self.kernel_size
                                   )

    # Apply si_kernel on sv convolution output
    sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                            1, self.weight_output_channels,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )

    # Combine the outputs and apply activation function
    output = self.activation(torch.cat((sv_output, sa_output), dim = 1))
    return output

upsample_convtranspose2d_layer

Bases: Module

An upsampling convtranspose2d layer.

Source code in odak/learn/models/components.py 
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
class upsample_convtranspose2d_layer(torch.nn.Module):
    """
    An upsampling convtranspose2d layer.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 2,
                 stride = 2,
                 bias = False,
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        """
        super().__init__()
        self.up = torch.nn.ConvTranspose2d(
                                           in_channels = input_channels,
                                           out_channels = output_channels,
                                           bias = bias,
                                           kernel_size = kernel_size,
                                           stride = stride
                                          )

    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.


        Returns
        ----------
        result        : torch.tensor
                        Result of the forward operation
        """
        x1 = self.up(x1)
        diffY = x2.size()[2] - x1.size()[2]
        diffX = x2.size()[3] - x1.size()[3]
        x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                          diffY // 2, diffY - diffY // 2])
        result = x1 + x2
        return result

__init__(input_channels, output_channels, kernel_size=2, stride=2, bias=False)

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
Source code in odak/learn/models/components.py 
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 2,
             stride = 2,
             bias = False,
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    """
    super().__init__()
    self.up = torch.nn.ConvTranspose2d(
                                       in_channels = input_channels,
                                       out_channels = output_channels,
                                       bias = bias,
                                       kernel_size = kernel_size,
                                       stride = stride
                                      )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.


    Returns
    ----------
    result        : torch.tensor
                    Result of the forward operation
    """
    x1 = self.up(x1)
    diffY = x2.size()[2] - x1.size()[2]
    diffX = x2.size()[3] - x1.size()[3]
    x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                      diffY // 2, diffY - diffY // 2])
    result = x1 + x2
    return result

upsample_layer

Bases: Module

An upsampling convolutional layer.

Source code in odak/learn/models/components.py 
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
class upsample_layer(torch.nn.Module):
    """
    An upsampling convolutional layer.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU(),
                 bilinear = True
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        bilinear        : bool
                          If set to True, bilinear sampling is used.
        """
        super(upsample_layer, self).__init__()
        if bilinear:
            self.up = torch.nn.Upsample(scale_factor = 2, mode = 'bilinear', align_corners = True)
            self.conv = double_convolution(
                                           input_channels = input_channels + output_channels,
                                           mid_channels = input_channels // 2,
                                           output_channels = output_channels,
                                           kernel_size = kernel_size,
                                           bias = bias,
                                           activation = activation
                                          )
        else:
            self.up = torch.nn.ConvTranspose2d(input_channels , input_channels // 2, kernel_size = 2, stride = 2)
            self.conv = double_convolution(
                                           input_channels = input_channels,
                                           mid_channels = output_channels,
                                           output_channels = output_channels,
                                           kernel_size = kernel_size,
                                           bias = bias,
                                           activation = activation
                                          )


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.


        Returns
        ----------
        result        : torch.tensor
                        Result of the forward operation
        """ 
        x1 = self.up(x1)
        diffY = x2.size()[2] - x1.size()[2]
        diffX = x2.size()[3] - x1.size()[3]
        x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                          diffY // 2, diffY - diffY // 2])
        x = torch.cat([x2, x1], dim = 1)
        result = self.conv(x)
        return result

__init__(input_channels, output_channels, kernel_size=3, bias=False, activation=torch.nn.ReLU(), bilinear=True)

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
  • bilinear – 
              If set to True, bilinear sampling is used.
Source code in odak/learn/models/components.py 
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU(),
             bilinear = True
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    bilinear        : bool
                      If set to True, bilinear sampling is used.
    """
    super(upsample_layer, self).__init__()
    if bilinear:
        self.up = torch.nn.Upsample(scale_factor = 2, mode = 'bilinear', align_corners = True)
        self.conv = double_convolution(
                                       input_channels = input_channels + output_channels,
                                       mid_channels = input_channels // 2,
                                       output_channels = output_channels,
                                       kernel_size = kernel_size,
                                       bias = bias,
                                       activation = activation
                                      )
    else:
        self.up = torch.nn.ConvTranspose2d(input_channels , input_channels // 2, kernel_size = 2, stride = 2)
        self.conv = double_convolution(
                                       input_channels = input_channels,
                                       mid_channels = output_channels,
                                       output_channels = output_channels,
                                       kernel_size = kernel_size,
                                       bias = bias,
                                       activation = activation
                                      )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.


    Returns
    ----------
    result        : torch.tensor
                    Result of the forward operation
    """ 
    x1 = self.up(x1)
    diffY = x2.size()[2] - x1.size()[2]
    diffX = x2.size()[3] - x1.size()[3]
    x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                      diffY // 2, diffY - diffY // 2])
    x = torch.cat([x2, x1], dim = 1)
    result = self.conv(x)
    return result

gaussian(x, multiplier=1.0)

A Gaussian non-linear activation. For more details: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.

Parameters:

  • x – 
           Input data.
  • multiplier – 
           Multiplier.

Returns:

  • result ( float or tensor ) –

    Ouput data.

Source code in odak/learn/models/components.py 
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
def gaussian(x, multiplier = 1.):
    """
    A Gaussian non-linear activation.
    For more details: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.

    Parameters
    ----------
    x            : float or torch.tensor
                   Input data.
    multiplier   : float or torch.tensor
                   Multiplier.

    Returns
    -------
    result       : float or torch.tensor
                   Ouput data.
    """
    result = torch.exp(- (multiplier * x) ** 2)
    return result

swish(x)

A swish non-linear activation. For more details: https://en.wikipedia.org/wiki/Swish_function

Parameters:

  • x – 
             Input.

Returns:

  • out ( float or tensor ) –

    Output.

Source code in odak/learn/models/components.py 
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def swish(x):
    """
    A swish non-linear activation.
    For more details: https://en.wikipedia.org/wiki/Swish_function

    Parameters
    -----------
    x              : float or torch.tensor
                     Input.

    Returns
    -------
    out            : float or torch.tensor
                     Output.
    """
    out = x * torch.sigmoid(x)
    return out

channel_gate

Bases: Module

Channel attention module with various pooling strategies. This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).

Source code in odak/learn/models/components.py 
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
class channel_gate(torch.nn.Module):
    """
    Channel attention module with various pooling strategies.
    This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).
    """
    def __init__(
                 self, 
                 gate_channels, 
                 reduction_ratio = 16, 
                 pool_types = ['avg', 'max']
                ):
        """
        Initializes the channel gate module.

        Parameters
        ----------
        gate_channels   : int
                          Number of channels of the input feature map.
        reduction_ratio : int
                          Reduction ratio for the intermediate layer.
        pool_types      : list
                          List of pooling operations to apply.
        """
        super().__init__()
        self.gate_channels = gate_channels
        hidden_channels = gate_channels // reduction_ratio
        if hidden_channels == 0:
            hidden_channels = 1
        self.mlp = torch.nn.Sequential(
                                       convolutional_block_attention.Flatten(),
                                       torch.nn.Linear(gate_channels, hidden_channels),
                                       torch.nn.ReLU(),
                                       torch.nn.Linear(hidden_channels, gate_channels)
                                      )
        self.pool_types = pool_types


    def forward(self, x):
        """
        Forward pass of the ChannelGate module.

        Applies channel-wise attention to the input tensor.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the ChannelGate module.

        Returns
        -------
        output       : torch.tensor
                       Output tensor after applying channel attention.
        """
        channel_att_sum = None
        for pool_type in self.pool_types:
            if pool_type == 'avg':
                pool = torch.nn.functional.avg_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
            elif pool_type == 'max':
                pool = torch.nn.functional.max_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
            channel_att_raw = self.mlp(pool)
            channel_att_sum = channel_att_raw if channel_att_sum is None else channel_att_sum + channel_att_raw
        scale = torch.sigmoid(channel_att_sum).unsqueeze(2).unsqueeze(3).expand_as(x)
        output = x * scale
        return output

__init__(gate_channels, reduction_ratio=16, pool_types=['avg', 'max'])

Initializes the channel gate module.

Parameters:

  • gate_channels – 
              Number of channels of the input feature map.
  • reduction_ratio (int, default: 16 ) – 
              Reduction ratio for the intermediate layer.
  • pool_types – 
              List of pooling operations to apply.
Source code in odak/learn/models/components.py 
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
def __init__(
             self, 
             gate_channels, 
             reduction_ratio = 16, 
             pool_types = ['avg', 'max']
            ):
    """
    Initializes the channel gate module.

    Parameters
    ----------
    gate_channels   : int
                      Number of channels of the input feature map.
    reduction_ratio : int
                      Reduction ratio for the intermediate layer.
    pool_types      : list
                      List of pooling operations to apply.
    """
    super().__init__()
    self.gate_channels = gate_channels
    hidden_channels = gate_channels // reduction_ratio
    if hidden_channels == 0:
        hidden_channels = 1
    self.mlp = torch.nn.Sequential(
                                   convolutional_block_attention.Flatten(),
                                   torch.nn.Linear(gate_channels, hidden_channels),
                                   torch.nn.ReLU(),
                                   torch.nn.Linear(hidden_channels, gate_channels)
                                  )
    self.pool_types = pool_types

forward(x)

Forward pass of the ChannelGate module.

Applies channel-wise attention to the input tensor.

Parameters:

  • x – 
           Input tensor to the ChannelGate module.

Returns:

  • output ( tensor ) –

    Output tensor after applying channel attention.

Source code in odak/learn/models/components.py 
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
def forward(self, x):
    """
    Forward pass of the ChannelGate module.

    Applies channel-wise attention to the input tensor.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the ChannelGate module.

    Returns
    -------
    output       : torch.tensor
                   Output tensor after applying channel attention.
    """
    channel_att_sum = None
    for pool_type in self.pool_types:
        if pool_type == 'avg':
            pool = torch.nn.functional.avg_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
        elif pool_type == 'max':
            pool = torch.nn.functional.max_pool2d(x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
        channel_att_raw = self.mlp(pool)
        channel_att_sum = channel_att_raw if channel_att_sum is None else channel_att_sum + channel_att_raw
    scale = torch.sigmoid(channel_att_sum).unsqueeze(2).unsqueeze(3).expand_as(x)
    output = x * scale
    return output

convolution_layer

Bases: Module

A convolution layer.

Source code in odak/learn/models/components.py 
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
class convolution_layer(torch.nn.Module):
    """
    A convolution layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 bias = False,
                 stride = 1,
                 normalization = True,
                 activation = torch.nn.ReLU()
                ):
        """
        A convolutional layer class.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        layers = [
            torch.nn.Conv2d(
                            input_channels,
                            output_channels,
                            kernel_size = kernel_size,
                            stride = stride,
                            padding = kernel_size // 2,
                            bias = bias
                           )
        ]
        if normalization:
            layers.append(torch.nn.BatchNorm2d(output_channels))
        if activation:
            layers.append(activation)
        self.model = torch.nn.Sequential(*layers)


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        result = self.model(x)
        return result

__init__(input_channels=2, output_channels=2, kernel_size=3, bias=False, stride=1, normalization=True, activation=torch.nn.ReLU())

A convolutional layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             bias = False,
             stride = 1,
             normalization = True,
             activation = torch.nn.ReLU()
            ):
    """
    A convolutional layer class.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    layers = [
        torch.nn.Conv2d(
                        input_channels,
                        output_channels,
                        kernel_size = kernel_size,
                        stride = stride,
                        padding = kernel_size // 2,
                        bias = bias
                       )
    ]
    if normalization:
        layers.append(torch.nn.BatchNorm2d(output_channels))
    if activation:
        layers.append(activation)
    self.model = torch.nn.Sequential(*layers)

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    result = self.model(x)
    return result

convolutional_block_attention

Bases: Module

Convolutional Block Attention Module (CBAM) class. This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).

Source code in odak/learn/models/components.py 
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
class convolutional_block_attention(torch.nn.Module):
    """
    Convolutional Block Attention Module (CBAM) class. 
    This class is heavily inspired https://github.com/Jongchan/attention-module/commit/e4ee180f1335c09db14d39a65d97c8ca3d1f7b16 (MIT License).
    """
    def __init__(
                 self, 
                 gate_channels, 
                 reduction_ratio = 16, 
                 pool_types = ['avg', 'max'], 
                 no_spatial = False
                ):
        """
        Initializes the convolutional block attention module.

        Parameters
        ----------
        gate_channels   : int
                          Number of channels of the input feature map.
        reduction_ratio : int
                          Reduction ratio for the channel attention.
        pool_types      : list
                          List of pooling operations to apply for channel attention.
        no_spatial      : bool
                          If True, spatial attention is not applied.
        """
        super(convolutional_block_attention, self).__init__()
        self.channel_gate = channel_gate(gate_channels, reduction_ratio, pool_types)
        self.no_spatial = no_spatial
        if not no_spatial:
            self.spatial_gate = spatial_gate()


    class Flatten(torch.nn.Module):
        """
        Flattens the input tensor to a 2D matrix.
        """
        def forward(self, x):
            return x.view(x.size(0), -1)


    def forward(self, x):
        """
        Forward pass of the convolutional block attention module.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the CBAM module.

        Returns
        -------
        x_out        : torch.tensor
                       Output tensor after applying channel and spatial attention.
        """
        x_out = self.channel_gate(x)
        if not self.no_spatial:
            x_out = self.spatial_gate(x_out)
        return x_out

Flatten

Bases: Module

Flattens the input tensor to a 2D matrix.

Source code in odak/learn/models/components.py 
757
758
759
760
761
762
class Flatten(torch.nn.Module):
    """
    Flattens the input tensor to a 2D matrix.
    """
    def forward(self, x):
        return x.view(x.size(0), -1)

__init__(gate_channels, reduction_ratio=16, pool_types=['avg', 'max'], no_spatial=False)

Initializes the convolutional block attention module.

Parameters:

  • gate_channels – 
              Number of channels of the input feature map.
  • reduction_ratio (int, default: 16 ) – 
              Reduction ratio for the channel attention.
  • pool_types – 
              List of pooling operations to apply for channel attention.
  • no_spatial – 
              If True, spatial attention is not applied.
Source code in odak/learn/models/components.py 
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
def __init__(
             self, 
             gate_channels, 
             reduction_ratio = 16, 
             pool_types = ['avg', 'max'], 
             no_spatial = False
            ):
    """
    Initializes the convolutional block attention module.

    Parameters
    ----------
    gate_channels   : int
                      Number of channels of the input feature map.
    reduction_ratio : int
                      Reduction ratio for the channel attention.
    pool_types      : list
                      List of pooling operations to apply for channel attention.
    no_spatial      : bool
                      If True, spatial attention is not applied.
    """
    super(convolutional_block_attention, self).__init__()
    self.channel_gate = channel_gate(gate_channels, reduction_ratio, pool_types)
    self.no_spatial = no_spatial
    if not no_spatial:
        self.spatial_gate = spatial_gate()

forward(x)

Forward pass of the convolutional block attention module.

Parameters:

  • x – 
           Input tensor to the CBAM module.

Returns:

  • x_out ( tensor ) –

    Output tensor after applying channel and spatial attention.

Source code in odak/learn/models/components.py 
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
def forward(self, x):
    """
    Forward pass of the convolutional block attention module.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the CBAM module.

    Returns
    -------
    x_out        : torch.tensor
                   Output tensor after applying channel and spatial attention.
    """
    x_out = self.channel_gate(x)
    if not self.no_spatial:
        x_out = self.spatial_gate(x_out)
    return x_out

double_convolution

Bases: Module

A double convolution layer.

Source code in odak/learn/models/components.py 
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
class double_convolution(torch.nn.Module):
    """
    A double convolution layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 mid_channels = None,
                 output_channels = 2,
                 kernel_size = 3, 
                 bias = False,
                 normalization = True,
                 activation = torch.nn.ReLU()
                ):
        """
        Double convolution model.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels    : int
                          Number of channels in the hidden layer between two convolutions.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        if isinstance(mid_channels, type(None)):
            mid_channels = output_channels
        self.activation = activation
        self.model = torch.nn.Sequential(
                                         convolution_layer(
                                                           input_channels = input_channels,
                                                           output_channels = mid_channels,
                                                           kernel_size = kernel_size,
                                                           bias = bias,
                                                           normalization = normalization,
                                                           activation = self.activation
                                                          ),
                                         convolution_layer(
                                                           input_channels = mid_channels,
                                                           output_channels = output_channels,
                                                           kernel_size = kernel_size,
                                                           bias = bias,
                                                           normalization = normalization,
                                                           activation = self.activation
                                                          )
                                        )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        result = self.model(x)
        return result

__init__(input_channels=2, mid_channels=None, output_channels=2, kernel_size=3, bias=False, normalization=True, activation=torch.nn.ReLU())

Double convolution model.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of channels in the hidden layer between two convolutions.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
def __init__(
             self,
             input_channels = 2,
             mid_channels = None,
             output_channels = 2,
             kernel_size = 3, 
             bias = False,
             normalization = True,
             activation = torch.nn.ReLU()
            ):
    """
    Double convolution model.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels    : int
                      Number of channels in the hidden layer between two convolutions.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    if isinstance(mid_channels, type(None)):
        mid_channels = output_channels
    self.activation = activation
    self.model = torch.nn.Sequential(
                                     convolution_layer(
                                                       input_channels = input_channels,
                                                       output_channels = mid_channels,
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       normalization = normalization,
                                                       activation = self.activation
                                                      ),
                                     convolution_layer(
                                                       input_channels = mid_channels,
                                                       output_channels = output_channels,
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       normalization = normalization,
                                                       activation = self.activation
                                                      )
                                    )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    result = self.model(x)
    return result

downsample_layer

Bases: Module

A downscaling component followed by a double convolution.

Source code in odak/learn/models/components.py 
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
class downsample_layer(torch.nn.Module):
    """
    A downscaling component followed by a double convolution.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.maxpool_conv = torch.nn.Sequential(
                                                torch.nn.MaxPool2d(2),
                                                double_convolution(
                                                                   input_channels = input_channels,
                                                                   mid_channels = output_channels,
                                                                   output_channels = output_channels,
                                                                   kernel_size = kernel_size,
                                                                   bias = bias,
                                                                   activation = activation
                                                                  )
                                               )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x              : torch.tensor
                         First input data.



        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        result = self.maxpool_conv(x)
        return result

__init__(input_channels, output_channels, kernel_size=3, bias=False, activation=torch.nn.ReLU())

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.maxpool_conv = torch.nn.Sequential(
                                            torch.nn.MaxPool2d(2),
                                            double_convolution(
                                                               input_channels = input_channels,
                                                               mid_channels = output_channels,
                                                               output_channels = output_channels,
                                                               kernel_size = kernel_size,
                                                               bias = bias,
                                                               activation = activation
                                                              )
                                           )

forward(x)

Forward model.

Parameters:

  • x – 
             First input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x              : torch.tensor
                     First input data.



    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    result = self.maxpool_conv(x)
    return result

global_feature_module

Bases: Module

A global feature layer that processes global features from input channels and applies them to another input tensor via learned transformations.

Source code in odak/learn/models/components.py 
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
class global_feature_module(torch.nn.Module):
    """
    A global feature layer that processes global features from input channels and
    applies them to another input tensor via learned transformations.
    """
    def __init__(
                 self,
                 input_channels,
                 mid_channels,
                 output_channels,
                 kernel_size,
                 bias = False,
                 normalization = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A global feature layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels  : int
                          Number of mid channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        normalization   : bool
                          If True, adds a Batch Normalization layer after the convolutional layer.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.transformations_1 = global_transformations(input_channels, output_channels)
        self.global_features_1 = double_convolution(
                                                    input_channels = input_channels,
                                                    mid_channels = mid_channels,
                                                    output_channels = output_channels,
                                                    kernel_size = kernel_size,
                                                    bias = bias,
                                                    normalization = normalization,
                                                    activation = activation
                                                   )
        self.global_features_2 = double_convolution(
                                                    input_channels = input_channels,
                                                    mid_channels = mid_channels,
                                                    output_channels = output_channels,
                                                    kernel_size = kernel_size,
                                                    bias = bias,
                                                    normalization = normalization,
                                                    activation = activation
                                                   )
        self.transformations_2 = global_transformations(input_channels, output_channels)


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.

        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        global_tensor_1 = self.transformations_1(x1, x2)
        y1 = self.global_features_1(global_tensor_1)
        y2 = self.global_features_2(y1)
        global_tensor_2 = self.transformations_2(y1, y2)
        return global_tensor_2

__init__(input_channels, mid_channels, output_channels, kernel_size, bias=False, normalization=False, activation=torch.nn.ReLU())

A global feature layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of mid channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • normalization – 
              If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
def __init__(
             self,
             input_channels,
             mid_channels,
             output_channels,
             kernel_size,
             bias = False,
             normalization = False,
             activation = torch.nn.ReLU()
            ):
    """
    A global feature layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels  : int
                      Number of mid channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    normalization   : bool
                      If True, adds a Batch Normalization layer after the convolutional layer.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.transformations_1 = global_transformations(input_channels, output_channels)
    self.global_features_1 = double_convolution(
                                                input_channels = input_channels,
                                                mid_channels = mid_channels,
                                                output_channels = output_channels,
                                                kernel_size = kernel_size,
                                                bias = bias,
                                                normalization = normalization,
                                                activation = activation
                                               )
    self.global_features_2 = double_convolution(
                                                input_channels = input_channels,
                                                mid_channels = mid_channels,
                                                output_channels = output_channels,
                                                kernel_size = kernel_size,
                                                bias = bias,
                                                normalization = normalization,
                                                activation = activation
                                               )
    self.transformations_2 = global_transformations(input_channels, output_channels)

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.

    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    global_tensor_1 = self.transformations_1(x1, x2)
    y1 = self.global_features_1(global_tensor_1)
    y2 = self.global_features_2(y1)
    global_tensor_2 = self.transformations_2(y1, y2)
    return global_tensor_2

global_transformations

Bases: Module

A global feature layer that processes global features from input channels and applies learned transformations to another input tensor.

This implementation is adapted from RSGUnet: https://github.com/MTLab/rsgunet_image_enhance.

Reference: J. Huang, P. Zhu, M. Geng et al. "Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices."

Source code in odak/learn/models/components.py 
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
class global_transformations(torch.nn.Module):
    """
    A global feature layer that processes global features from input channels and
    applies learned transformations to another input tensor.

    This implementation is adapted from RSGUnet:
    https://github.com/MTLab/rsgunet_image_enhance.

    Reference:
    J. Huang, P. Zhu, M. Geng et al. "Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices."
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels
                ):
        """
        A global feature layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        """
        super().__init__()
        self.global_feature_1 = torch.nn.Sequential(
            torch.nn.Linear(input_channels, output_channels),
            torch.nn.LeakyReLU(0.2, inplace = True),
        )
        self.global_feature_2 = torch.nn.Sequential(
            torch.nn.Linear(output_channels, output_channels),
            torch.nn.LeakyReLU(0.2, inplace = True)
        )


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.

        Returns
        ----------
        result        : torch.tensor
                        Estimated output.
        """
        y = torch.mean(x2, dim = (2, 3))
        y1 = self.global_feature_1(y)
        y2 = self.global_feature_2(y1)
        y1 = y1.unsqueeze(2).unsqueeze(3)
        y2 = y2.unsqueeze(2).unsqueeze(3)
        result = x1 * y1 + y2
        return result

__init__(input_channels, output_channels)

A global feature layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
Source code in odak/learn/models/components.py 
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
def __init__(
             self,
             input_channels,
             output_channels
            ):
    """
    A global feature layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    """
    super().__init__()
    self.global_feature_1 = torch.nn.Sequential(
        torch.nn.Linear(input_channels, output_channels),
        torch.nn.LeakyReLU(0.2, inplace = True),
    )
    self.global_feature_2 = torch.nn.Sequential(
        torch.nn.Linear(output_channels, output_channels),
        torch.nn.LeakyReLU(0.2, inplace = True)
    )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.

    Returns
    ----------
    result        : torch.tensor
                    Estimated output.
    """
    y = torch.mean(x2, dim = (2, 3))
    y1 = self.global_feature_1(y)
    y2 = self.global_feature_2(y1)
    y1 = y1.unsqueeze(2).unsqueeze(3)
    y2 = y2.unsqueeze(2).unsqueeze(3)
    result = x1 * y1 + y2
    return result

multi_layer_perceptron

Bases: Module

A multi-layer perceptron model.

Source code in odak/learn/models/models.py 
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
class multi_layer_perceptron(torch.nn.Module):
    """
    A multi-layer perceptron model.
    """

    def __init__(self,
                 dimensions,
                 activation = torch.nn.ReLU(),
                 bias = False,
                 model_type = 'conventional',
                 siren_multiplier = 1.,
                 input_multiplier = None
                ):
        """
        Parameters
        ----------
        dimensions        : list
                            List of integers representing the dimensions of each layer (e.g., [2, 10, 1], where the first layer has two channels and last one has one channel.).
        activation        : torch.nn
                            Nonlinear activation function.
                            Default is `torch.nn.ReLU()`.
        bias              : bool
                            If set to True, linear layers will include biases.
        siren_multiplier  : float
                            When using `SIREN` model type, this parameter functions as a hyperparameter.
                            The original SIREN work uses 30.
                            You can bypass this parameter by providing input that are not normalized and larger then one.
        input_multiplier  : float
                            Initial value of the input multiplier before the very first layer.
        model_type        : str
                            Model type: `conventional`, `swish`, `SIREN`, `FILM SIREN`, `Gaussian`.
                            `conventional` refers to a standard multi layer perceptron.
                            For `SIREN,` see: Sitzmann, Vincent, et al. "Implicit neural representations with periodic activation functions." Advances in neural information processing systems 33 (2020): 7462-7473.
                            For `Swish,` see: Ramachandran, Prajit, Barret Zoph, and Quoc V. Le. "Searching for activation functions." arXiv preprint arXiv:1710.05941 (2017). 
                            For `FILM SIREN,` see: Chan, Eric R., et al. "pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
                            For `Gaussian,` see: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.
        """
        super(multi_layer_perceptron, self).__init__()
        self.activation = activation
        self.bias = bias
        self.model_type = model_type
        self.layers = torch.nn.ModuleList()
        self.siren_multiplier = siren_multiplier
        self.dimensions = dimensions
        for i in range(len(self.dimensions) - 1):
            self.layers.append(torch.nn.Linear(self.dimensions[i], self.dimensions[i + 1], bias = self.bias))
        if not isinstance(input_multiplier, type(None)):
            self.input_multiplier = torch.nn.ParameterList()
            self.input_multiplier.append(torch.nn.Parameter(torch.ones(1, self.dimensions[0]) * input_multiplier))
        if self.model_type == 'FILM SIREN':
            self.alpha = torch.nn.ParameterList()
            for j in self.dimensions[1:-1]:
                self.alpha.append(torch.nn.Parameter(torch.randn(2, 1, j)))
        if self.model_type == 'Gaussian':
            self.alpha = torch.nn.ParameterList()
            for j in self.dimensions[1:-1]:
                self.alpha.append(torch.nn.Parameter(torch.randn(1, 1, j)))


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        if hasattr(self, 'input_multiplier'):
            result = x * self.input_multiplier[0]
        else:
            result = x
        for layer_id, layer in enumerate(self.layers[:-1]):
            result = layer(result)
            if self.model_type == 'conventional':
                result = self.activation(result)
            elif self.model_type == 'swish':
                result = swish(result)
            elif self.model_type == 'SIREN':
                result = torch.sin(result * self.siren_multiplier)
            elif self.model_type == 'FILM SIREN':
                result = torch.sin(self.alpha[layer_id][0] * result + self.alpha[layer_id][1])
            elif self.model_type == 'Gaussian': 
                result = gaussian(result, self.alpha[layer_id][0])
        result = self.layers[-1](result)
        return result

__init__(dimensions, activation=torch.nn.ReLU(), bias=False, model_type='conventional', siren_multiplier=1.0, input_multiplier=None)

Parameters:

  • dimensions – 
                List of integers representing the dimensions of each layer (e.g., [2, 10, 1], where the first layer has two channels and last one has one channel.).
  • activation – 
                Nonlinear activation function.
            Default is `torch.nn.ReLU()`.
  • bias – 
                If set to True, linear layers will include biases.
  • siren_multiplier – 
                When using `SIREN` model type, this parameter functions as a hyperparameter.
            The original SIREN work uses 30.
            You can bypass this parameter by providing input that are not normalized and larger then one.
  • input_multiplier – 
                Initial value of the input multiplier before the very first layer.
  • model_type – 
                Model type: `conventional`, `swish`, `SIREN`, `FILM SIREN`, `Gaussian`.
            `conventional` refers to a standard multi layer perceptron.
            For `SIREN,` see: Sitzmann, Vincent, et al. "Implicit neural representations with periodic activation functions." Advances in neural information processing systems 33 (2020): 7462-7473.
            For `Swish,` see: Ramachandran, Prajit, Barret Zoph, and Quoc V. Le. "Searching for activation functions." arXiv preprint arXiv:1710.05941 (2017). 
            For `FILM SIREN,` see: Chan, Eric R., et al. "pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
            For `Gaussian,` see: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.
Source code in odak/learn/models/models.py 
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
def __init__(self,
             dimensions,
             activation = torch.nn.ReLU(),
             bias = False,
             model_type = 'conventional',
             siren_multiplier = 1.,
             input_multiplier = None
            ):
    """
    Parameters
    ----------
    dimensions        : list
                        List of integers representing the dimensions of each layer (e.g., [2, 10, 1], where the first layer has two channels and last one has one channel.).
    activation        : torch.nn
                        Nonlinear activation function.
                        Default is `torch.nn.ReLU()`.
    bias              : bool
                        If set to True, linear layers will include biases.
    siren_multiplier  : float
                        When using `SIREN` model type, this parameter functions as a hyperparameter.
                        The original SIREN work uses 30.
                        You can bypass this parameter by providing input that are not normalized and larger then one.
    input_multiplier  : float
                        Initial value of the input multiplier before the very first layer.
    model_type        : str
                        Model type: `conventional`, `swish`, `SIREN`, `FILM SIREN`, `Gaussian`.
                        `conventional` refers to a standard multi layer perceptron.
                        For `SIREN,` see: Sitzmann, Vincent, et al. "Implicit neural representations with periodic activation functions." Advances in neural information processing systems 33 (2020): 7462-7473.
                        For `Swish,` see: Ramachandran, Prajit, Barret Zoph, and Quoc V. Le. "Searching for activation functions." arXiv preprint arXiv:1710.05941 (2017). 
                        For `FILM SIREN,` see: Chan, Eric R., et al. "pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
                        For `Gaussian,` see: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.
    """
    super(multi_layer_perceptron, self).__init__()
    self.activation = activation
    self.bias = bias
    self.model_type = model_type
    self.layers = torch.nn.ModuleList()
    self.siren_multiplier = siren_multiplier
    self.dimensions = dimensions
    for i in range(len(self.dimensions) - 1):
        self.layers.append(torch.nn.Linear(self.dimensions[i], self.dimensions[i + 1], bias = self.bias))
    if not isinstance(input_multiplier, type(None)):
        self.input_multiplier = torch.nn.ParameterList()
        self.input_multiplier.append(torch.nn.Parameter(torch.ones(1, self.dimensions[0]) * input_multiplier))
    if self.model_type == 'FILM SIREN':
        self.alpha = torch.nn.ParameterList()
        for j in self.dimensions[1:-1]:
            self.alpha.append(torch.nn.Parameter(torch.randn(2, 1, j)))
    if self.model_type == 'Gaussian':
        self.alpha = torch.nn.ParameterList()
        for j in self.dimensions[1:-1]:
            self.alpha.append(torch.nn.Parameter(torch.randn(1, 1, j)))

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/models.py 
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    if hasattr(self, 'input_multiplier'):
        result = x * self.input_multiplier[0]
    else:
        result = x
    for layer_id, layer in enumerate(self.layers[:-1]):
        result = layer(result)
        if self.model_type == 'conventional':
            result = self.activation(result)
        elif self.model_type == 'swish':
            result = swish(result)
        elif self.model_type == 'SIREN':
            result = torch.sin(result * self.siren_multiplier)
        elif self.model_type == 'FILM SIREN':
            result = torch.sin(self.alpha[layer_id][0] * result + self.alpha[layer_id][1])
        elif self.model_type == 'Gaussian': 
            result = gaussian(result, self.alpha[layer_id][0])
    result = self.layers[-1](result)
    return result

non_local_layer

Bases: Module

Self-Attention Layer [zi = Wzyi + xi] (non-local block : ref https://arxiv.org/abs/1711.07971)

Source code in odak/learn/models/components.py 
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
class non_local_layer(torch.nn.Module):
    """
    Self-Attention Layer [zi = Wzyi + xi] (non-local block : ref https://arxiv.org/abs/1711.07971)
    """
    def __init__(
                 self,
                 input_channels = 1024,
                 bottleneck_channels = 512,
                 kernel_size = 1,
                 bias = False,
                ):
        """

        Parameters
        ----------
        input_channels      : int
                              Number of input channels.
        bottleneck_channels : int
                              Number of middle channels.
        kernel_size         : int
                              Kernel size.
        bias                : bool 
                              Set to True to let convolutional layers have bias term.
        """
        super(non_local_layer, self).__init__()
        self.input_channels = input_channels
        self.bottleneck_channels = bottleneck_channels
        self.g = torch.nn.Conv2d(
                                 self.input_channels, 
                                 self.bottleneck_channels,
                                 kernel_size = kernel_size,
                                 padding = kernel_size // 2,
                                 bias = bias
                                )
        self.W_z = torch.nn.Sequential(
                                       torch.nn.Conv2d(
                                                       self.bottleneck_channels,
                                                       self.input_channels, 
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       padding = kernel_size // 2
                                                      ),
                                       torch.nn.BatchNorm2d(self.input_channels)
                                      )
        torch.nn.init.constant_(self.W_z[1].weight, 0)   
        torch.nn.init.constant_(self.W_z[1].bias, 0)


    def forward(self, x):
        """
        Forward model [zi = Wzyi + xi]

        Parameters
        ----------
        x               : torch.tensor
                          First input data.                       


        Returns
        ----------
        z               : torch.tensor
                          Estimated output.
        """
        batch_size, channels, height, width = x.size()
        theta = x.view(batch_size, channels, -1).permute(0, 2, 1)
        phi = x.view(batch_size, channels, -1).permute(0, 2, 1)
        g = self.g(x).view(batch_size, self.bottleneck_channels, -1).permute(0, 2, 1)
        attn = torch.bmm(theta, phi.transpose(1, 2)) / (height * width)
        attn = torch.nn.functional.softmax(attn, dim=-1)
        y = torch.bmm(attn, g).permute(0, 2, 1).contiguous().view(batch_size, self.bottleneck_channels, height, width)
        W_y = self.W_z(y)
        z = W_y + x
        return z

__init__(input_channels=1024, bottleneck_channels=512, kernel_size=1, bias=False)

Parameters:

  • input_channels – 
                  Number of input channels.
  • bottleneck_channels (int, default: 512 ) – 
                  Number of middle channels.
  • kernel_size – 
                  Kernel size.
  • bias – 
                  Set to True to let convolutional layers have bias term.
Source code in odak/learn/models/components.py 
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
def __init__(
             self,
             input_channels = 1024,
             bottleneck_channels = 512,
             kernel_size = 1,
             bias = False,
            ):
    """

    Parameters
    ----------
    input_channels      : int
                          Number of input channels.
    bottleneck_channels : int
                          Number of middle channels.
    kernel_size         : int
                          Kernel size.
    bias                : bool 
                          Set to True to let convolutional layers have bias term.
    """
    super(non_local_layer, self).__init__()
    self.input_channels = input_channels
    self.bottleneck_channels = bottleneck_channels
    self.g = torch.nn.Conv2d(
                             self.input_channels, 
                             self.bottleneck_channels,
                             kernel_size = kernel_size,
                             padding = kernel_size // 2,
                             bias = bias
                            )
    self.W_z = torch.nn.Sequential(
                                   torch.nn.Conv2d(
                                                   self.bottleneck_channels,
                                                   self.input_channels, 
                                                   kernel_size = kernel_size,
                                                   bias = bias,
                                                   padding = kernel_size // 2
                                                  ),
                                   torch.nn.BatchNorm2d(self.input_channels)
                                  )
    torch.nn.init.constant_(self.W_z[1].weight, 0)   
    torch.nn.init.constant_(self.W_z[1].bias, 0)

forward(x)

Forward model [zi = Wzyi + xi]

Parameters:

  • x – 
              First input data.

Returns:

  • z ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
def forward(self, x):
    """
    Forward model [zi = Wzyi + xi]

    Parameters
    ----------
    x               : torch.tensor
                      First input data.                       


    Returns
    ----------
    z               : torch.tensor
                      Estimated output.
    """
    batch_size, channels, height, width = x.size()
    theta = x.view(batch_size, channels, -1).permute(0, 2, 1)
    phi = x.view(batch_size, channels, -1).permute(0, 2, 1)
    g = self.g(x).view(batch_size, self.bottleneck_channels, -1).permute(0, 2, 1)
    attn = torch.bmm(theta, phi.transpose(1, 2)) / (height * width)
    attn = torch.nn.functional.softmax(attn, dim=-1)
    y = torch.bmm(attn, g).permute(0, 2, 1).contiguous().view(batch_size, self.bottleneck_channels, height, width)
    W_y = self.W_z(y)
    z = W_y + x
    return z

normalization

Bases: Module

A normalization layer.

Source code in odak/learn/models/components.py 
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
class normalization(torch.nn.Module):
    """
    A normalization layer.
    """
    def __init__(
                 self,
                 dim = 1,
                ):
        """
        Normalization layer.


        Parameters
        ----------
        dim             : int
                          Dimension (axis) to normalize.
        """
        super().__init__()
        self.k = torch.nn.Parameter(torch.ones(1, dim, 1, 1))


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        eps = 1e-5 if x.dtype == torch.float32 else 1e-3
        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
        mean = torch.mean(x, dim = 1, keepdim = True)
        result =  (x - mean) * (var + eps).rsqrt() * self.k
        return result

__init__(dim=1)

Normalization layer.

Parameters:

  • dim – 
              Dimension (axis) to normalize.
Source code in odak/learn/models/components.py 
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
def __init__(
             self,
             dim = 1,
            ):
    """
    Normalization layer.


    Parameters
    ----------
    dim             : int
                      Dimension (axis) to normalize.
    """
    super().__init__()
    self.k = torch.nn.Parameter(torch.ones(1, dim, 1, 1))

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    eps = 1e-5 if x.dtype == torch.float32 else 1e-3
    var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
    mean = torch.mean(x, dim = 1, keepdim = True)
    result =  (x - mean) * (var + eps).rsqrt() * self.k
    return result

positional_encoder

Bases: Module

A positional encoder module.

Source code in odak/learn/models/components.py 
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
class positional_encoder(torch.nn.Module):
    """
    A positional encoder module.
    """

    def __init__(self, L):
        """
        A positional encoder module.

        Parameters
        ----------
        L                   : int
                              Positional encoding level.
        """
        super(positional_encoder, self).__init__()
        self.L = L


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x               : torch.tensor
                          Input data.

        Returns
        ----------
        result          : torch.tensor
                          Result of the forward operation
        """
        B, C = x.shape
        x = x.view(B, C, 1)
        results = [x]
        for i in range(1, self.L + 1):
            freq = (2 ** i) * torch.pi
            cos_x = torch.cos(freq * x)
            sin_x = torch.sin(freq * x)
            results.append(cos_x)
            results.append(sin_x)
        results = torch.cat(results, dim=2)
        results = results.permute(0, 2, 1)
        results = results.reshape(B, -1)
        return results

__init__(L)

A positional encoder module.

Parameters:

  • L – 
                  Positional encoding level.
Source code in odak/learn/models/components.py 
790
791
792
793
794
795
796
797
798
799
800
def __init__(self, L):
    """
    A positional encoder module.

    Parameters
    ----------
    L                   : int
                          Positional encoding level.
    """
    super(positional_encoder, self).__init__()
    self.L = L

forward(x)

Forward model.

Parameters:

  • x – 
              Input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x               : torch.tensor
                      Input data.

    Returns
    ----------
    result          : torch.tensor
                      Result of the forward operation
    """
    B, C = x.shape
    x = x.view(B, C, 1)
    results = [x]
    for i in range(1, self.L + 1):
        freq = (2 ** i) * torch.pi
        cos_x = torch.cos(freq * x)
        sin_x = torch.sin(freq * x)
        results.append(cos_x)
        results.append(sin_x)
    results = torch.cat(results, dim=2)
    results = results.permute(0, 2, 1)
    results = results.reshape(B, -1)
    return results

residual_attention_layer

Bases: Module

A residual block with an attention layer.

Source code in odak/learn/models/components.py 
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
class residual_attention_layer(torch.nn.Module):
    """
    A residual block with an attention layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 1,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        An attention layer class.


        Parameters
        ----------
        input_channels  : int or optioal
                          Number of input channels.
        output_channels : int or optional
                          Number of middle channels.
        kernel_size     : int or optional
                          Kernel size.
        bias            : bool or optional
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn or optional
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.activation = activation
        self.convolution0 = torch.nn.Sequential(
                                                torch.nn.Conv2d(
                                                                input_channels,
                                                                output_channels,
                                                                kernel_size = kernel_size,
                                                                padding = kernel_size // 2,
                                                                bias = bias
                                                               ),
                                                torch.nn.BatchNorm2d(output_channels)
                                               )
        self.convolution1 = torch.nn.Sequential(
                                                torch.nn.Conv2d(
                                                                input_channels,
                                                                output_channels,
                                                                kernel_size = kernel_size,
                                                                padding = kernel_size // 2,
                                                                bias = bias
                                                               ),
                                                torch.nn.BatchNorm2d(output_channels)
                                               )
        self.final_layer = torch.nn.Sequential(
                                               self.activation,
                                               torch.nn.Conv2d(
                                                               output_channels,
                                                               output_channels,
                                                               kernel_size = kernel_size,
                                                               padding = kernel_size // 2,
                                                               bias = bias
                                                              )
                                              )


    def forward(self, x0, x1):
        """
        Forward model.

        Parameters
        ----------
        x0             : torch.tensor
                         First input data.

        x1             : torch.tensor
                         Seconnd input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        y0 = self.convolution0(x0)
        y1 = self.convolution1(x1)
        y2 = torch.add(y0, y1)
        result = self.final_layer(y2) * x0
        return result

__init__(input_channels=2, output_channels=2, kernel_size=1, bias=False, activation=torch.nn.ReLU())

An attention layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int or optional, default: 2 ) – 
              Number of middle channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 1,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    An attention layer class.


    Parameters
    ----------
    input_channels  : int or optioal
                      Number of input channels.
    output_channels : int or optional
                      Number of middle channels.
    kernel_size     : int or optional
                      Kernel size.
    bias            : bool or optional
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn or optional
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.activation = activation
    self.convolution0 = torch.nn.Sequential(
                                            torch.nn.Conv2d(
                                                            input_channels,
                                                            output_channels,
                                                            kernel_size = kernel_size,
                                                            padding = kernel_size // 2,
                                                            bias = bias
                                                           ),
                                            torch.nn.BatchNorm2d(output_channels)
                                           )
    self.convolution1 = torch.nn.Sequential(
                                            torch.nn.Conv2d(
                                                            input_channels,
                                                            output_channels,
                                                            kernel_size = kernel_size,
                                                            padding = kernel_size // 2,
                                                            bias = bias
                                                           ),
                                            torch.nn.BatchNorm2d(output_channels)
                                           )
    self.final_layer = torch.nn.Sequential(
                                           self.activation,
                                           torch.nn.Conv2d(
                                                           output_channels,
                                                           output_channels,
                                                           kernel_size = kernel_size,
                                                           padding = kernel_size // 2,
                                                           bias = bias
                                                          )
                                          )

forward(x0, x1)

Forward model.

Parameters:

  • x0 – 
             First input data.
  • x1 – 
             Seconnd input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
def forward(self, x0, x1):
    """
    Forward model.

    Parameters
    ----------
    x0             : torch.tensor
                     First input data.

    x1             : torch.tensor
                     Seconnd input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    y0 = self.convolution0(x0)
    y1 = self.convolution1(x1)
    y2 = torch.add(y0, y1)
    result = self.final_layer(y2) * x0
    return result

residual_layer

Bases: Module

A residual layer.

Source code in odak/learn/models/components.py 
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
class residual_layer(torch.nn.Module):
    """
    A residual layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 mid_channels = 16,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU()
                ):
        """
        A convolutional layer class.


        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        mid_channels    : int
                          Number of middle channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.activation = activation
        self.convolution = double_convolution(
                                              input_channels,
                                              mid_channels = mid_channels,
                                              output_channels = input_channels,
                                              kernel_size = kernel_size,
                                              bias = bias,
                                              activation = activation
                                             )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        x0 = self.convolution(x)
        return x + x0

__init__(input_channels=2, mid_channels=16, kernel_size=3, bias=False, activation=torch.nn.ReLU())

A convolutional layer class.

Parameters:

  • input_channels – 
              Number of input channels.
  • mid_channels – 
              Number of middle channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
def __init__(
             self,
             input_channels = 2,
             mid_channels = 16,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU()
            ):
    """
    A convolutional layer class.


    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    mid_channels    : int
                      Number of middle channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.activation = activation
    self.convolution = double_convolution(
                                          input_channels,
                                          mid_channels = mid_channels,
                                          output_channels = input_channels,
                                          kernel_size = kernel_size,
                                          bias = bias,
                                          activation = activation
                                         )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    x0 = self.convolution(x)
    return x + x0

spatial_gate

Bases: Module

Spatial attention module that applies a convolution layer after channel pooling. This class is heavily inspired by https://github.com/Jongchan/attention-module/blob/master/MODELS/cbam.py.

Source code in odak/learn/models/components.py 
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
class spatial_gate(torch.nn.Module):
    """
    Spatial attention module that applies a convolution layer after channel pooling.
    This class is heavily inspired by https://github.com/Jongchan/attention-module/blob/master/MODELS/cbam.py.
    """
    def __init__(self):
        """
        Initializes the spatial gate module.
        """
        super().__init__()
        kernel_size = 7
        self.spatial = convolution_layer(2, 1, kernel_size, bias = False, activation = torch.nn.Identity())


    def channel_pool(self, x):
        """
        Applies max and average pooling on the channels.

        Parameters
        ----------
        x             : torch.tensor
                        Input tensor.

        Returns
        -------
        output        : torch.tensor
                        Output tensor.
        """
        max_pool = torch.max(x, 1)[0].unsqueeze(1)
        avg_pool = torch.mean(x, 1).unsqueeze(1)
        output = torch.cat((max_pool, avg_pool), dim=1)
        return output


    def forward(self, x):
        """
        Forward pass of the SpatialGate module.

        Applies spatial attention to the input tensor.

        Parameters
        ----------
        x            : torch.tensor
                       Input tensor to the SpatialGate module.

        Returns
        -------
        scaled_x     : torch.tensor
                       Output tensor after applying spatial attention.
        """
        x_compress = self.channel_pool(x)
        x_out = self.spatial(x_compress)
        scale = torch.sigmoid(x_out)
        scaled_x = x * scale
        return scaled_x

__init__()

Initializes the spatial gate module.

Source code in odak/learn/models/components.py 
672
673
674
675
676
677
678
def __init__(self):
    """
    Initializes the spatial gate module.
    """
    super().__init__()
    kernel_size = 7
    self.spatial = convolution_layer(2, 1, kernel_size, bias = False, activation = torch.nn.Identity())

channel_pool(x)

Applies max and average pooling on the channels.

Parameters:

  • x – 
            Input tensor.

Returns:

  • output ( tensor ) –

    Output tensor.

Source code in odak/learn/models/components.py 
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
def channel_pool(self, x):
    """
    Applies max and average pooling on the channels.

    Parameters
    ----------
    x             : torch.tensor
                    Input tensor.

    Returns
    -------
    output        : torch.tensor
                    Output tensor.
    """
    max_pool = torch.max(x, 1)[0].unsqueeze(1)
    avg_pool = torch.mean(x, 1).unsqueeze(1)
    output = torch.cat((max_pool, avg_pool), dim=1)
    return output

forward(x)

Forward pass of the SpatialGate module.

Applies spatial attention to the input tensor.

Parameters:

  • x – 
           Input tensor to the SpatialGate module.

Returns:

  • scaled_x ( tensor ) –

    Output tensor after applying spatial attention.

Source code in odak/learn/models/components.py 
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
def forward(self, x):
    """
    Forward pass of the SpatialGate module.

    Applies spatial attention to the input tensor.

    Parameters
    ----------
    x            : torch.tensor
                   Input tensor to the SpatialGate module.

    Returns
    -------
    scaled_x     : torch.tensor
                   Output tensor after applying spatial attention.
    """
    x_compress = self.channel_pool(x)
    x_out = self.spatial(x_compress)
    scale = torch.sigmoid(x_out)
    scaled_x = x * scale
    return scaled_x

spatially_adaptive_convolution

Bases: Module

A spatially adaptive convolution layer.

References

C. Zheng et al. "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions." C. Xu et al. "Squeezesegv3: Spatially-adaptive Convolution for Efficient Point-Cloud Segmentation." C. Zheng et al. "Windowing Decomposition Convolutional Neural Network for Image Enhancement."

Source code in odak/learn/models/components.py 
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
class spatially_adaptive_convolution(torch.nn.Module):
    """
    A spatially adaptive convolution layer.

    References
    ----------

    C. Zheng et al. "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions."
    C. Xu et al. "Squeezesegv3: Spatially-adaptive Convolution for Efficient Point-Cloud Segmentation."
    C. Zheng et al. "Windowing Decomposition Convolutional Neural Network for Image Enhancement."
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 stride = 1,
                 padding = 1,
                 bias = False,
                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                ):
        """
        Initializes a spatially adaptive convolution layer.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Size of the convolution kernel.
        stride          : int
                          Stride of the convolution.
        padding         : int
                          Padding added to both sides of the input.
        bias            : bool
                          If True, includes a bias term in the convolution.
        activation      : torch.nn.Module
                          Activation function to apply. If None, no activation is applied.
        """
        super(spatially_adaptive_convolution, self).__init__()
        self.kernel_size = kernel_size
        self.input_channels = input_channels
        self.output_channels = output_channels
        self.stride = stride
        self.padding = padding
        self.standard_convolution = torch.nn.Conv2d(
                                                    in_channels = input_channels,
                                                    out_channels = self.output_channels,
                                                    kernel_size = kernel_size,
                                                    stride = stride,
                                                    padding = padding,
                                                    bias = bias
                                                   )
        self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
        self.activation = activation


    def forward(self, x, sv_kernel_feature):
        """
        Forward pass for the spatially adaptive convolution layer.

        Parameters
        ----------
        x                  : torch.tensor
                            Input data tensor.
                            Dimension: (1, C, H, W)
        sv_kernel_feature   : torch.tensor
                            Spatially varying kernel features.
                            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

        Returns
        -------
        sa_output          : torch.tensor
                            Estimated output tensor.
                            Dimension: (1, output_channels, H_out, W_out)
        """
        # Pad input and sv_kernel_feature if necessary
        if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
                -2) * self.stride != x.size(-2):
            diffY = sv_kernel_feature.size(-2) % self.stride
            diffX = sv_kernel_feature.size(-1) % self.stride
            sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                            diffY // 2, diffY - diffY // 2))
            diffY = x.size(-2) % self.stride
            diffX = x.size(-1) % self.stride
            x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                            diffY // 2, diffY - diffY // 2))

        # Unfold the input tensor for matrix multiplication
        input_feature = torch.nn.functional.unfold(
                                                   x,
                                                   kernel_size = (self.kernel_size, self.kernel_size),
                                                   stride = self.stride,
                                                   padding = self.padding
                                                  )

        # Resize sv_kernel_feature to match the input feature
        sv_kernel = sv_kernel_feature.reshape(
                                              1,
                                              self.input_channels * self.kernel_size * self.kernel_size,
                                              (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                             )

        # Resize weight to match the input channels and kernel size
        si_kernel = self.weight.reshape(
                                        self.weight_output_channels,
                                        self.input_channels * self.kernel_size * self.kernel_size
                                       )

        # Apply spatially varying kernels
        sv_feature = input_feature * sv_kernel

        # Perform matrix multiplication
        sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                                1, self.weight_output_channels,
                                                                (x.size(-2) // self.stride),
                                                                (x.size(-1) // self.stride)
                                                               )
        return sa_output

__init__(input_channels=2, output_channels=2, kernel_size=3, stride=1, padding=1, bias=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

Initializes a spatially adaptive convolution layer.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Size of the convolution kernel.
  • stride – 
              Stride of the convolution.
  • padding – 
              Padding added to both sides of the input.
  • bias – 
              If True, includes a bias term in the convolution.
  • activation – 
              Activation function to apply. If None, no activation is applied.
Source code in odak/learn/models/components.py 
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             stride = 1,
             padding = 1,
             bias = False,
             activation = torch.nn.LeakyReLU(0.2, inplace = True)
            ):
    """
    Initializes a spatially adaptive convolution layer.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Size of the convolution kernel.
    stride          : int
                      Stride of the convolution.
    padding         : int
                      Padding added to both sides of the input.
    bias            : bool
                      If True, includes a bias term in the convolution.
    activation      : torch.nn.Module
                      Activation function to apply. If None, no activation is applied.
    """
    super(spatially_adaptive_convolution, self).__init__()
    self.kernel_size = kernel_size
    self.input_channels = input_channels
    self.output_channels = output_channels
    self.stride = stride
    self.padding = padding
    self.standard_convolution = torch.nn.Conv2d(
                                                in_channels = input_channels,
                                                out_channels = self.output_channels,
                                                kernel_size = kernel_size,
                                                stride = stride,
                                                padding = padding,
                                                bias = bias
                                               )
    self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
    self.activation = activation

forward(x, sv_kernel_feature)

Forward pass for the spatially adaptive convolution layer.

Parameters:

  • x – 
                Input data tensor.
            Dimension: (1, C, H, W)
  • sv_kernel_feature – 
                Spatially varying kernel features.
            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

Returns:

  • sa_output ( tensor ) –

    Estimated output tensor. Dimension: (1, output_channels, H_out, W_out)

Source code in odak/learn/models/components.py 
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
def forward(self, x, sv_kernel_feature):
    """
    Forward pass for the spatially adaptive convolution layer.

    Parameters
    ----------
    x                  : torch.tensor
                        Input data tensor.
                        Dimension: (1, C, H, W)
    sv_kernel_feature   : torch.tensor
                        Spatially varying kernel features.
                        Dimension: (1, C_i * kernel_size * kernel_size, H, W)

    Returns
    -------
    sa_output          : torch.tensor
                        Estimated output tensor.
                        Dimension: (1, output_channels, H_out, W_out)
    """
    # Pad input and sv_kernel_feature if necessary
    if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
            -2) * self.stride != x.size(-2):
        diffY = sv_kernel_feature.size(-2) % self.stride
        diffX = sv_kernel_feature.size(-1) % self.stride
        sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                        diffY // 2, diffY - diffY // 2))
        diffY = x.size(-2) % self.stride
        diffX = x.size(-1) % self.stride
        x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                        diffY // 2, diffY - diffY // 2))

    # Unfold the input tensor for matrix multiplication
    input_feature = torch.nn.functional.unfold(
                                               x,
                                               kernel_size = (self.kernel_size, self.kernel_size),
                                               stride = self.stride,
                                               padding = self.padding
                                              )

    # Resize sv_kernel_feature to match the input feature
    sv_kernel = sv_kernel_feature.reshape(
                                          1,
                                          self.input_channels * self.kernel_size * self.kernel_size,
                                          (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                         )

    # Resize weight to match the input channels and kernel size
    si_kernel = self.weight.reshape(
                                    self.weight_output_channels,
                                    self.input_channels * self.kernel_size * self.kernel_size
                                   )

    # Apply spatially varying kernels
    sv_feature = input_feature * sv_kernel

    # Perform matrix multiplication
    sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                            1, self.weight_output_channels,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )
    return sa_output

spatially_adaptive_module

Bases: Module

A spatially adaptive module that combines learned spatially adaptive convolutions.

References

Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.

Source code in odak/learn/models/components.py 
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
class spatially_adaptive_module(torch.nn.Module):
    """
    A spatially adaptive module that combines learned spatially adaptive convolutions.

    References
    ----------

    Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 stride = 1,
                 padding = 1,
                 bias = False,
                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                ):
        """
        Initializes a spatially adaptive module.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Size of the convolution kernel.
        stride          : int
                          Stride of the convolution.
        padding         : int
                          Padding added to both sides of the input.
        bias            : bool
                          If True, includes a bias term in the convolution.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super(spatially_adaptive_module, self).__init__()
        self.kernel_size = kernel_size
        self.input_channels = input_channels
        self.output_channels = output_channels
        self.stride = stride
        self.padding = padding
        self.weight_output_channels = self.output_channels - 1
        self.standard_convolution = torch.nn.Conv2d(
                                                    in_channels = input_channels,
                                                    out_channels = self.weight_output_channels,
                                                    kernel_size = kernel_size,
                                                    stride = stride,
                                                    padding = padding,
                                                    bias = bias
                                                   )
        self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
        self.activation = activation


    def forward(self, x, sv_kernel_feature):
        """
        Forward pass for the spatially adaptive module.

        Parameters
        ----------
        x                  : torch.tensor
                            Input data tensor.
                            Dimension: (1, C, H, W)
        sv_kernel_feature   : torch.tensor
                            Spatially varying kernel features.
                            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

        Returns
        -------
        output             : torch.tensor
                            Combined output tensor from standard and spatially adaptive convolutions.
                            Dimension: (1, output_channels, H_out, W_out)
        """
        # Pad input and sv_kernel_feature if necessary
        if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
                -2) * self.stride != x.size(-2):
            diffY = sv_kernel_feature.size(-2) % self.stride
            diffX = sv_kernel_feature.size(-1) % self.stride
            sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                            diffY // 2, diffY - diffY // 2))
            diffY = x.size(-2) % self.stride
            diffX = x.size(-1) % self.stride
            x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                            diffY // 2, diffY - diffY // 2))

        # Unfold the input tensor for matrix multiplication
        input_feature = torch.nn.functional.unfold(
                                                   x,
                                                   kernel_size = (self.kernel_size, self.kernel_size),
                                                   stride = self.stride,
                                                   padding = self.padding
                                                  )

        # Resize sv_kernel_feature to match the input feature
        sv_kernel = sv_kernel_feature.reshape(
                                              1,
                                              self.input_channels * self.kernel_size * self.kernel_size,
                                              (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                             )

        # Apply sv_kernel to the input_feature
        sv_feature = input_feature * sv_kernel

        # Original spatially varying convolution output
        sv_output = torch.sum(sv_feature, dim = 1).reshape(
                                                           1,
                                                            1,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )

        # Reshape weight for spatially adaptive convolution
        si_kernel = self.weight.reshape(
                                        self.weight_output_channels,
                                        self.input_channels * self.kernel_size * self.kernel_size
                                       )

        # Apply si_kernel on sv convolution output
        sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                                1, self.weight_output_channels,
                                                                (x.size(-2) // self.stride),
                                                                (x.size(-1) // self.stride)
                                                               )

        # Combine the outputs and apply activation function
        output = self.activation(torch.cat((sv_output, sa_output), dim = 1))
        return output

__init__(input_channels=2, output_channels=2, kernel_size=3, stride=1, padding=1, bias=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

Initializes a spatially adaptive module.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int, default: 2 ) – 
              Number of output channels.
  • kernel_size – 
              Size of the convolution kernel.
  • stride – 
              Stride of the convolution.
  • padding – 
              Padding added to both sides of the input.
  • bias – 
              If True, includes a bias term in the convolution.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             stride = 1,
             padding = 1,
             bias = False,
             activation = torch.nn.LeakyReLU(0.2, inplace = True)
            ):
    """
    Initializes a spatially adaptive module.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Size of the convolution kernel.
    stride          : int
                      Stride of the convolution.
    padding         : int
                      Padding added to both sides of the input.
    bias            : bool
                      If True, includes a bias term in the convolution.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super(spatially_adaptive_module, self).__init__()
    self.kernel_size = kernel_size
    self.input_channels = input_channels
    self.output_channels = output_channels
    self.stride = stride
    self.padding = padding
    self.weight_output_channels = self.output_channels - 1
    self.standard_convolution = torch.nn.Conv2d(
                                                in_channels = input_channels,
                                                out_channels = self.weight_output_channels,
                                                kernel_size = kernel_size,
                                                stride = stride,
                                                padding = padding,
                                                bias = bias
                                               )
    self.weight = torch.nn.Parameter(data = self.standard_convolution.weight, requires_grad = True)
    self.activation = activation

forward(x, sv_kernel_feature)

Forward pass for the spatially adaptive module.

Parameters:

  • x – 
                Input data tensor.
            Dimension: (1, C, H, W)
  • sv_kernel_feature – 
                Spatially varying kernel features.
            Dimension: (1, C_i * kernel_size * kernel_size, H, W)

Returns:

  • output ( tensor ) –

    Combined output tensor from standard and spatially adaptive convolutions. Dimension: (1, output_channels, H_out, W_out)

Source code in odak/learn/models/components.py 
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
def forward(self, x, sv_kernel_feature):
    """
    Forward pass for the spatially adaptive module.

    Parameters
    ----------
    x                  : torch.tensor
                        Input data tensor.
                        Dimension: (1, C, H, W)
    sv_kernel_feature   : torch.tensor
                        Spatially varying kernel features.
                        Dimension: (1, C_i * kernel_size * kernel_size, H, W)

    Returns
    -------
    output             : torch.tensor
                        Combined output tensor from standard and spatially adaptive convolutions.
                        Dimension: (1, output_channels, H_out, W_out)
    """
    # Pad input and sv_kernel_feature if necessary
    if sv_kernel_feature.size(-1) * self.stride != x.size(-1) or sv_kernel_feature.size(
            -2) * self.stride != x.size(-2):
        diffY = sv_kernel_feature.size(-2) % self.stride
        diffX = sv_kernel_feature.size(-1) % self.stride
        sv_kernel_feature = torch.nn.functional.pad(sv_kernel_feature, (diffX // 2, diffX - diffX // 2,
                                                                        diffY // 2, diffY - diffY // 2))
        diffY = x.size(-2) % self.stride
        diffX = x.size(-1) % self.stride
        x = torch.nn.functional.pad(x, (diffX // 2, diffX - diffX // 2,
                                        diffY // 2, diffY - diffY // 2))

    # Unfold the input tensor for matrix multiplication
    input_feature = torch.nn.functional.unfold(
                                               x,
                                               kernel_size = (self.kernel_size, self.kernel_size),
                                               stride = self.stride,
                                               padding = self.padding
                                              )

    # Resize sv_kernel_feature to match the input feature
    sv_kernel = sv_kernel_feature.reshape(
                                          1,
                                          self.input_channels * self.kernel_size * self.kernel_size,
                                          (x.size(-2) // self.stride) * (x.size(-1) // self.stride)
                                         )

    # Apply sv_kernel to the input_feature
    sv_feature = input_feature * sv_kernel

    # Original spatially varying convolution output
    sv_output = torch.sum(sv_feature, dim = 1).reshape(
                                                       1,
                                                        1,
                                                        (x.size(-2) // self.stride),
                                                        (x.size(-1) // self.stride)
                                                       )

    # Reshape weight for spatially adaptive convolution
    si_kernel = self.weight.reshape(
                                    self.weight_output_channels,
                                    self.input_channels * self.kernel_size * self.kernel_size
                                   )

    # Apply si_kernel on sv convolution output
    sa_output = torch.matmul(si_kernel, sv_feature).reshape(
                                                            1, self.weight_output_channels,
                                                            (x.size(-2) // self.stride),
                                                            (x.size(-1) // self.stride)
                                                           )

    # Combine the outputs and apply activation function
    output = self.activation(torch.cat((sv_output, sa_output), dim = 1))
    return output

spatially_adaptive_unet

Bases: Module

Spatially varying U-Net model based on spatially adaptive convolution.

References

Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.

Source code in odak/learn/models/models.py 
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
class spatially_adaptive_unet(torch.nn.Module):
    """
    Spatially varying U-Net model based on spatially adaptive convolution.

    References
    ----------

    Chuanjun Zheng, Yicheng Zhan, Liang Shi, Ozan Cakmakci, and Kaan Akşit, "Focal Surface Holographic Light Transport using Learned Spatially Adaptive Convolutions," SIGGRAPH Asia 2024 Technical Communications (SA Technical Communications '24), December, 2024.
    """
    def __init__(
                 self,
                 depth=3,
                 dimensions=8,
                 input_channels=6,
                 out_channels=6,
                 kernel_size=3,
                 bias=True,
                 normalization=False,
                 activation=torch.nn.LeakyReLU(0.2, inplace=True)
                ):
        """
        U-Net model.

        Parameters
        ----------
        depth          : int
                         Number of upsampling and downsampling layers.
        dimensions     : int
                         Number of dimensions.
        input_channels : int
                         Number of input channels.
        out_channels   : int
                         Number of output channels.
        bias           : bool
                         Set to True to let convolutional layers learn a bias term.
        normalization  : bool
                         If True, adds a Batch Normalization layer after the convolutional layer.
        activation     : torch.nn
                         Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
        """
        super().__init__()
        self.depth = depth
        self.out_channels = out_channels
        self.inc = convolution_layer(
                                     input_channels=input_channels,
                                     output_channels=dimensions,
                                     kernel_size=kernel_size,
                                     bias=bias,
                                     normalization=normalization,
                                     activation=activation
                                    )

        self.encoder = torch.nn.ModuleList()
        for i in range(self.depth + 1):  # Downsampling layers
            down_in_channels = dimensions * (2 ** i)
            down_out_channels = 2 * down_in_channels
            pooling_layer = torch.nn.AvgPool2d(2)
            double_convolution_layer = double_convolution(
                                                          input_channels=down_in_channels,
                                                          mid_channels=down_in_channels,
                                                          output_channels=down_in_channels,
                                                          kernel_size=kernel_size,
                                                          bias=bias,
                                                          normalization=normalization,
                                                          activation=activation
                                                         )
            sam = spatially_adaptive_module(
                                            input_channels=down_in_channels,
                                            output_channels=down_out_channels,
                                            kernel_size=kernel_size,
                                            bias=bias,
                                            activation=activation
                                           )
            self.encoder.append(torch.nn.ModuleList([pooling_layer, double_convolution_layer, sam]))
        self.global_feature_module = torch.nn.ModuleList()
        double_convolution_layer = double_convolution(
                                                      input_channels=dimensions * (2 ** (depth + 1)),
                                                      mid_channels=dimensions * (2 ** (depth + 1)),
                                                      output_channels=dimensions * (2 ** (depth + 1)),
                                                      kernel_size=kernel_size,
                                                      bias=bias,
                                                      normalization=normalization,
                                                      activation=activation
                                                     )
        global_feature_layer = global_feature_module(
                                                     input_channels=dimensions * (2 ** (depth + 1)),
                                                     mid_channels=dimensions * (2 ** (depth + 1)),
                                                     output_channels=dimensions * (2 ** (depth + 1)),
                                                     kernel_size=kernel_size,
                                                     bias=bias,
                                                     activation=torch.nn.LeakyReLU(0.2, inplace=True)
                                                    )
        self.global_feature_module.append(torch.nn.ModuleList([double_convolution_layer, global_feature_layer]))
        self.decoder = torch.nn.ModuleList()
        for i in range(depth, -1, -1):
            up_in_channels = dimensions * (2 ** (i + 1))
            up_mid_channels = up_in_channels // 2
            if i == 0:
                up_out_channels = self.out_channels
                upsample_layer = upsample_convtranspose2d_layer(
                                                                input_channels=up_in_channels,
                                                                output_channels=up_mid_channels,
                                                                kernel_size=2,
                                                                stride=2,
                                                                bias=bias,
                                                               )
                conv_layer = torch.nn.Sequential(
                    convolution_layer(
                                      input_channels=up_mid_channels,
                                      output_channels=up_mid_channels,
                                      kernel_size=kernel_size,
                                      bias=bias,
                                      normalization=normalization,
                                      activation=activation,
                                     ),
                    convolution_layer(
                                      input_channels=up_mid_channels,
                                      output_channels=up_out_channels,
                                      kernel_size=1,
                                      bias=bias,
                                      normalization=normalization,
                                      activation=None,
                                     )
                )
                self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))
            else:
                up_out_channels = up_in_channels // 2
                upsample_layer = upsample_convtranspose2d_layer(
                                                                input_channels=up_in_channels,
                                                                output_channels=up_mid_channels,
                                                                kernel_size=2,
                                                                stride=2,
                                                                bias=bias,
                                                               )
                conv_layer = double_convolution(
                                                input_channels=up_mid_channels,
                                                mid_channels=up_mid_channels,
                                                output_channels=up_out_channels,
                                                kernel_size=kernel_size,
                                                bias=bias,
                                                normalization=normalization,
                                                activation=activation,
                                               )
                self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))


    def forward(self, sv_kernel, field):
        """
        Forward model.

        Parameters
        ----------
        sv_kernel : list of torch.tensor
                    Learned spatially varying kernels.
                    Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
                    where C_i, H_i, and W_i represent the channel, height, and width
                    of each feature at a certain scale.

        field     : torch.tensor
                    Input field data.
                    Dimension: (1, 6, H, W)

        Returns
        -------
        target_field : torch.tensor
                       Estimated output.
                       Dimension: (1, 6, H, W)
        """
        x = self.inc(field)
        downsampling_outputs = [x]
        for i, down_layer in enumerate(self.encoder):
            x_down = down_layer[0](downsampling_outputs[-1])
            downsampling_outputs.append(x_down)
            sam_output = down_layer[2](x_down + down_layer[1](x_down), sv_kernel[self.depth - i])
            downsampling_outputs.append(sam_output)
        global_feature = self.global_feature_module[0][0](downsampling_outputs[-1])
        global_feature = self.global_feature_module[0][1](downsampling_outputs[-1], global_feature)
        downsampling_outputs.append(global_feature)
        x_up = downsampling_outputs[-1]
        for i, up_layer in enumerate(self.decoder):
            x_up = up_layer[0](x_up, downsampling_outputs[2 * (self.depth - i)])
            x_up = up_layer[1](x_up)
        result = x_up
        return result

__init__(depth=3, dimensions=8, input_channels=6, out_channels=6, kernel_size=3, bias=True, normalization=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

U-Net model.

Parameters:

  • depth – 
             Number of upsampling and downsampling layers.
  • dimensions – 
             Number of dimensions.
  • input_channels (int, default: 6 ) – 
             Number of input channels.
  • out_channels – 
             Number of output channels.
  • bias – 
             Set to True to let convolutional layers learn a bias term.
  • normalization – 
             If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
             Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
Source code in odak/learn/models/models.py 
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
def __init__(
             self,
             depth=3,
             dimensions=8,
             input_channels=6,
             out_channels=6,
             kernel_size=3,
             bias=True,
             normalization=False,
             activation=torch.nn.LeakyReLU(0.2, inplace=True)
            ):
    """
    U-Net model.

    Parameters
    ----------
    depth          : int
                     Number of upsampling and downsampling layers.
    dimensions     : int
                     Number of dimensions.
    input_channels : int
                     Number of input channels.
    out_channels   : int
                     Number of output channels.
    bias           : bool
                     Set to True to let convolutional layers learn a bias term.
    normalization  : bool
                     If True, adds a Batch Normalization layer after the convolutional layer.
    activation     : torch.nn
                     Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
    """
    super().__init__()
    self.depth = depth
    self.out_channels = out_channels
    self.inc = convolution_layer(
                                 input_channels=input_channels,
                                 output_channels=dimensions,
                                 kernel_size=kernel_size,
                                 bias=bias,
                                 normalization=normalization,
                                 activation=activation
                                )

    self.encoder = torch.nn.ModuleList()
    for i in range(self.depth + 1):  # Downsampling layers
        down_in_channels = dimensions * (2 ** i)
        down_out_channels = 2 * down_in_channels
        pooling_layer = torch.nn.AvgPool2d(2)
        double_convolution_layer = double_convolution(
                                                      input_channels=down_in_channels,
                                                      mid_channels=down_in_channels,
                                                      output_channels=down_in_channels,
                                                      kernel_size=kernel_size,
                                                      bias=bias,
                                                      normalization=normalization,
                                                      activation=activation
                                                     )
        sam = spatially_adaptive_module(
                                        input_channels=down_in_channels,
                                        output_channels=down_out_channels,
                                        kernel_size=kernel_size,
                                        bias=bias,
                                        activation=activation
                                       )
        self.encoder.append(torch.nn.ModuleList([pooling_layer, double_convolution_layer, sam]))
    self.global_feature_module = torch.nn.ModuleList()
    double_convolution_layer = double_convolution(
                                                  input_channels=dimensions * (2 ** (depth + 1)),
                                                  mid_channels=dimensions * (2 ** (depth + 1)),
                                                  output_channels=dimensions * (2 ** (depth + 1)),
                                                  kernel_size=kernel_size,
                                                  bias=bias,
                                                  normalization=normalization,
                                                  activation=activation
                                                 )
    global_feature_layer = global_feature_module(
                                                 input_channels=dimensions * (2 ** (depth + 1)),
                                                 mid_channels=dimensions * (2 ** (depth + 1)),
                                                 output_channels=dimensions * (2 ** (depth + 1)),
                                                 kernel_size=kernel_size,
                                                 bias=bias,
                                                 activation=torch.nn.LeakyReLU(0.2, inplace=True)
                                                )
    self.global_feature_module.append(torch.nn.ModuleList([double_convolution_layer, global_feature_layer]))
    self.decoder = torch.nn.ModuleList()
    for i in range(depth, -1, -1):
        up_in_channels = dimensions * (2 ** (i + 1))
        up_mid_channels = up_in_channels // 2
        if i == 0:
            up_out_channels = self.out_channels
            upsample_layer = upsample_convtranspose2d_layer(
                                                            input_channels=up_in_channels,
                                                            output_channels=up_mid_channels,
                                                            kernel_size=2,
                                                            stride=2,
                                                            bias=bias,
                                                           )
            conv_layer = torch.nn.Sequential(
                convolution_layer(
                                  input_channels=up_mid_channels,
                                  output_channels=up_mid_channels,
                                  kernel_size=kernel_size,
                                  bias=bias,
                                  normalization=normalization,
                                  activation=activation,
                                 ),
                convolution_layer(
                                  input_channels=up_mid_channels,
                                  output_channels=up_out_channels,
                                  kernel_size=1,
                                  bias=bias,
                                  normalization=normalization,
                                  activation=None,
                                 )
            )
            self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))
        else:
            up_out_channels = up_in_channels // 2
            upsample_layer = upsample_convtranspose2d_layer(
                                                            input_channels=up_in_channels,
                                                            output_channels=up_mid_channels,
                                                            kernel_size=2,
                                                            stride=2,
                                                            bias=bias,
                                                           )
            conv_layer = double_convolution(
                                            input_channels=up_mid_channels,
                                            mid_channels=up_mid_channels,
                                            output_channels=up_out_channels,
                                            kernel_size=kernel_size,
                                            bias=bias,
                                            normalization=normalization,
                                            activation=activation,
                                           )
            self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))

forward(sv_kernel, field)

Forward model.

Parameters:

  • sv_kernel (list of torch.tensor) – 
        Learned spatially varying kernels.
    Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
    where C_i, H_i, and W_i represent the channel, height, and width
    of each feature at a certain scale.
  • field – 
        Input field data.
    Dimension: (1, 6, H, W)

Returns:

  • target_field ( tensor ) –

    Estimated output. Dimension: (1, 6, H, W)

Source code in odak/learn/models/models.py 
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
def forward(self, sv_kernel, field):
    """
    Forward model.

    Parameters
    ----------
    sv_kernel : list of torch.tensor
                Learned spatially varying kernels.
                Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
                where C_i, H_i, and W_i represent the channel, height, and width
                of each feature at a certain scale.

    field     : torch.tensor
                Input field data.
                Dimension: (1, 6, H, W)

    Returns
    -------
    target_field : torch.tensor
                   Estimated output.
                   Dimension: (1, 6, H, W)
    """
    x = self.inc(field)
    downsampling_outputs = [x]
    for i, down_layer in enumerate(self.encoder):
        x_down = down_layer[0](downsampling_outputs[-1])
        downsampling_outputs.append(x_down)
        sam_output = down_layer[2](x_down + down_layer[1](x_down), sv_kernel[self.depth - i])
        downsampling_outputs.append(sam_output)
    global_feature = self.global_feature_module[0][0](downsampling_outputs[-1])
    global_feature = self.global_feature_module[0][1](downsampling_outputs[-1], global_feature)
    downsampling_outputs.append(global_feature)
    x_up = downsampling_outputs[-1]
    for i, up_layer in enumerate(self.decoder):
        x_up = up_layer[0](x_up, downsampling_outputs[2 * (self.depth - i)])
        x_up = up_layer[1](x_up)
    result = x_up
    return result

spatially_varying_kernel_generation_model

Bases: Module

Spatially_varying_kernel_generation_model revised from RSGUnet: https://github.com/MTLab/rsgunet_image_enhance.

Refer to: J. Huang, P. Zhu, M. Geng et al. Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices.

Source code in odak/learn/models/models.py 
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
class spatially_varying_kernel_generation_model(torch.nn.Module):
    """
    Spatially_varying_kernel_generation_model revised from RSGUnet:
    https://github.com/MTLab/rsgunet_image_enhance.

    Refer to:
    J. Huang, P. Zhu, M. Geng et al. Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices.
    """

    def __init__(
                 self,
                 depth = 3,
                 dimensions = 8,
                 input_channels = 7,
                 kernel_size = 3,
                 bias = True,
                 normalization = False,
                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                ):
        """
        U-Net model.

        Parameters
        ----------
        depth          : int
                         Number of upsampling and downsampling layers.
        dimensions     : int
                         Number of dimensions.
        input_channels : int
                         Number of input channels.
        bias           : bool
                         Set to True to let convolutional layers learn a bias term.
        normalization  : bool
                         If True, adds a Batch Normalization layer after the convolutional layer.
        activation     : torch.nn
                         Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
        """
        super().__init__()
        self.depth = depth
        self.inc = convolution_layer(
                                     input_channels = input_channels,
                                     output_channels = dimensions,
                                     kernel_size = kernel_size,
                                     bias = bias,
                                     normalization = normalization,
                                     activation = activation
                                    )
        self.encoder = torch.nn.ModuleList()
        for i in range(depth + 1):  # downsampling layers
            if i == 0:
                in_channels = dimensions * (2 ** i)
                out_channels = dimensions * (2 ** i)
            elif i == depth:
                in_channels = dimensions * (2 ** (i - 1))
                out_channels = dimensions * (2 ** (i - 1))
            else:
                in_channels = dimensions * (2 ** (i - 1))
                out_channels = 2 * in_channels
            pooling_layer = torch.nn.AvgPool2d(2)
            double_convolution_layer = double_convolution(
                                                          input_channels = in_channels,
                                                          mid_channels = in_channels,
                                                          output_channels = out_channels,
                                                          kernel_size = kernel_size,
                                                          bias = bias,
                                                          normalization = normalization,
                                                          activation = activation
                                                         )
            self.encoder.append(pooling_layer)
            self.encoder.append(double_convolution_layer)
        self.spatially_varying_feature = torch.nn.ModuleList()  # for kernel generation
        for i in range(depth, -1, -1):
            if i == 1:
                svf_in_channels = dimensions + 2 ** (self.depth + i) + 1
            else:
                svf_in_channels = 2 ** (self.depth + i) + 1
            svf_out_channels = (2 ** (self.depth + i)) * (kernel_size * kernel_size)
            svf_mid_channels = dimensions * (2 ** (self.depth - 1))
            spatially_varying_kernel_generation = torch.nn.ModuleList()
            for j in range(i, -1, -1):
                pooling_layer = torch.nn.AvgPool2d(2 ** (j + 1))
                spatially_varying_kernel_generation.append(pooling_layer)
            kernel_generation_block = torch.nn.Sequential(
                torch.nn.Conv2d(
                                in_channels = svf_in_channels,
                                out_channels = svf_mid_channels,
                                kernel_size = kernel_size,
                                padding = kernel_size // 2,
                                bias = bias
                               ),
                activation,
                torch.nn.Conv2d(
                                in_channels = svf_mid_channels,
                                out_channels = svf_mid_channels,
                                kernel_size = kernel_size,
                                padding = kernel_size // 2,
                                bias = bias
                               ),
                activation,
                torch.nn.Conv2d(
                                in_channels = svf_mid_channels,
                                out_channels = svf_out_channels,
                                kernel_size = kernel_size,
                                padding = kernel_size // 2,
                                bias = bias
                               ),
            )
            spatially_varying_kernel_generation.append(kernel_generation_block)
            self.spatially_varying_feature.append(spatially_varying_kernel_generation)
        self.decoder = torch.nn.ModuleList()
        global_feature_layer = global_feature_module(  # global feature layer
                                                     input_channels = dimensions * (2 ** (depth - 1)),
                                                     mid_channels = dimensions * (2 ** (depth - 1)),
                                                     output_channels = dimensions * (2 ** (depth - 1)),
                                                     kernel_size = kernel_size,
                                                     bias = bias,
                                                     activation = torch.nn.LeakyReLU(0.2, inplace = True)
                                                    )
        self.decoder.append(global_feature_layer)
        for i in range(depth, 0, -1):
            if i == 2:
                up_in_channels = (dimensions // 2) * (2 ** i)
                up_out_channels = up_in_channels
                up_mid_channels = up_in_channels
            elif i == 1:
                up_in_channels = dimensions * 2
                up_out_channels = dimensions
                up_mid_channels = up_out_channels
            else:
                up_in_channels = (dimensions // 2) * (2 ** i)
                up_out_channels = up_in_channels // 2
                up_mid_channels = up_in_channels
            upsample_layer = upsample_convtranspose2d_layer(
                                                            input_channels = up_in_channels,
                                                            output_channels = up_mid_channels,
                                                            kernel_size = 2,
                                                            stride = 2,
                                                            bias = bias,
                                                           )
            conv_layer = double_convolution(
                                            input_channels = up_mid_channels,
                                            output_channels = up_out_channels,
                                            kernel_size = kernel_size,
                                            bias = bias,
                                            normalization = normalization,
                                            activation = activation,
                                           )
            self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))


    def forward(self, focal_surface, field):
        """
        Forward model.

        Parameters
        ----------
        focal_surface : torch.tensor
                        Input focal surface data.
                        Dimension: (1, 1, H, W)

        field         : torch.tensor
                        Input field data.
                        Dimension: (1, 6, H, W)

        Returns
        -------
        sv_kernel : list of torch.tensor
                    Learned spatially varying kernels.
                    Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
                    where C_i, H_i, and W_i represent the channel, height, and width
                    of each feature at a certain scale.
        """
        x = self.inc(torch.cat((focal_surface, field), dim = 1))
        downsampling_outputs = [focal_surface]
        downsampling_outputs.append(x)
        for i, down_layer in enumerate(self.encoder):
            x_down = down_layer(downsampling_outputs[-1])
            downsampling_outputs.append(x_down)
        sv_kernels = []
        for i, (up_layer, svf_layer) in enumerate(zip(self.decoder, self.spatially_varying_feature)):
            if i == 0:
                global_feature = up_layer(downsampling_outputs[-2], downsampling_outputs[-1])
                downsampling_outputs[-1] = global_feature
                sv_feature = [global_feature, downsampling_outputs[0]]
                for j in range(self.depth - i + 1):
                    sv_feature[1] = svf_layer[self.depth - i](sv_feature[1])
                    if j > 0:
                        sv_feature.append(svf_layer[j](downsampling_outputs[2 * j]))
                sv_feature = [sv_feature[0], sv_feature[1], sv_feature[4], sv_feature[2],
                              sv_feature[3]]
                sv_kernel = svf_layer[-1](torch.cat(sv_feature, dim = 1))
                sv_kernels.append(sv_kernel)
            else:
                x_up = up_layer[0](downsampling_outputs[-1],
                                   downsampling_outputs[2 * (self.depth + 1 - i) + 1])
                x_up = up_layer[1](x_up)
                downsampling_outputs[-1] = x_up
                sv_feature = [x_up, downsampling_outputs[0]]
                for j in range(self.depth - i + 1):
                    sv_feature[1] = svf_layer[self.depth - i](sv_feature[1])
                    if j > 0:
                        sv_feature.append(svf_layer[j](downsampling_outputs[2 * j]))
                if i == 1:
                    sv_feature = [sv_feature[0], sv_feature[1], sv_feature[3], sv_feature[2]]
                sv_kernel = svf_layer[-1](torch.cat(sv_feature, dim = 1))
                sv_kernels.append(sv_kernel)
        return sv_kernels

__init__(depth=3, dimensions=8, input_channels=7, kernel_size=3, bias=True, normalization=False, activation=torch.nn.LeakyReLU(0.2, inplace=True))

U-Net model.

Parameters:

  • depth – 
             Number of upsampling and downsampling layers.
  • dimensions – 
             Number of dimensions.
  • input_channels (int, default: 7 ) – 
             Number of input channels.
  • bias – 
             Set to True to let convolutional layers learn a bias term.
  • normalization – 
             If True, adds a Batch Normalization layer after the convolutional layer.
  • activation – 
             Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
Source code in odak/learn/models/models.py 
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
def __init__(
             self,
             depth = 3,
             dimensions = 8,
             input_channels = 7,
             kernel_size = 3,
             bias = True,
             normalization = False,
             activation = torch.nn.LeakyReLU(0.2, inplace = True)
            ):
    """
    U-Net model.

    Parameters
    ----------
    depth          : int
                     Number of upsampling and downsampling layers.
    dimensions     : int
                     Number of dimensions.
    input_channels : int
                     Number of input channels.
    bias           : bool
                     Set to True to let convolutional layers learn a bias term.
    normalization  : bool
                     If True, adds a Batch Normalization layer after the convolutional layer.
    activation     : torch.nn
                     Non-linear activation layer (e.g., torch.nn.ReLU(), torch.nn.Sigmoid()).
    """
    super().__init__()
    self.depth = depth
    self.inc = convolution_layer(
                                 input_channels = input_channels,
                                 output_channels = dimensions,
                                 kernel_size = kernel_size,
                                 bias = bias,
                                 normalization = normalization,
                                 activation = activation
                                )
    self.encoder = torch.nn.ModuleList()
    for i in range(depth + 1):  # downsampling layers
        if i == 0:
            in_channels = dimensions * (2 ** i)
            out_channels = dimensions * (2 ** i)
        elif i == depth:
            in_channels = dimensions * (2 ** (i - 1))
            out_channels = dimensions * (2 ** (i - 1))
        else:
            in_channels = dimensions * (2 ** (i - 1))
            out_channels = 2 * in_channels
        pooling_layer = torch.nn.AvgPool2d(2)
        double_convolution_layer = double_convolution(
                                                      input_channels = in_channels,
                                                      mid_channels = in_channels,
                                                      output_channels = out_channels,
                                                      kernel_size = kernel_size,
                                                      bias = bias,
                                                      normalization = normalization,
                                                      activation = activation
                                                     )
        self.encoder.append(pooling_layer)
        self.encoder.append(double_convolution_layer)
    self.spatially_varying_feature = torch.nn.ModuleList()  # for kernel generation
    for i in range(depth, -1, -1):
        if i == 1:
            svf_in_channels = dimensions + 2 ** (self.depth + i) + 1
        else:
            svf_in_channels = 2 ** (self.depth + i) + 1
        svf_out_channels = (2 ** (self.depth + i)) * (kernel_size * kernel_size)
        svf_mid_channels = dimensions * (2 ** (self.depth - 1))
        spatially_varying_kernel_generation = torch.nn.ModuleList()
        for j in range(i, -1, -1):
            pooling_layer = torch.nn.AvgPool2d(2 ** (j + 1))
            spatially_varying_kernel_generation.append(pooling_layer)
        kernel_generation_block = torch.nn.Sequential(
            torch.nn.Conv2d(
                            in_channels = svf_in_channels,
                            out_channels = svf_mid_channels,
                            kernel_size = kernel_size,
                            padding = kernel_size // 2,
                            bias = bias
                           ),
            activation,
            torch.nn.Conv2d(
                            in_channels = svf_mid_channels,
                            out_channels = svf_mid_channels,
                            kernel_size = kernel_size,
                            padding = kernel_size // 2,
                            bias = bias
                           ),
            activation,
            torch.nn.Conv2d(
                            in_channels = svf_mid_channels,
                            out_channels = svf_out_channels,
                            kernel_size = kernel_size,
                            padding = kernel_size // 2,
                            bias = bias
                           ),
        )
        spatially_varying_kernel_generation.append(kernel_generation_block)
        self.spatially_varying_feature.append(spatially_varying_kernel_generation)
    self.decoder = torch.nn.ModuleList()
    global_feature_layer = global_feature_module(  # global feature layer
                                                 input_channels = dimensions * (2 ** (depth - 1)),
                                                 mid_channels = dimensions * (2 ** (depth - 1)),
                                                 output_channels = dimensions * (2 ** (depth - 1)),
                                                 kernel_size = kernel_size,
                                                 bias = bias,
                                                 activation = torch.nn.LeakyReLU(0.2, inplace = True)
                                                )
    self.decoder.append(global_feature_layer)
    for i in range(depth, 0, -1):
        if i == 2:
            up_in_channels = (dimensions // 2) * (2 ** i)
            up_out_channels = up_in_channels
            up_mid_channels = up_in_channels
        elif i == 1:
            up_in_channels = dimensions * 2
            up_out_channels = dimensions
            up_mid_channels = up_out_channels
        else:
            up_in_channels = (dimensions // 2) * (2 ** i)
            up_out_channels = up_in_channels // 2
            up_mid_channels = up_in_channels
        upsample_layer = upsample_convtranspose2d_layer(
                                                        input_channels = up_in_channels,
                                                        output_channels = up_mid_channels,
                                                        kernel_size = 2,
                                                        stride = 2,
                                                        bias = bias,
                                                       )
        conv_layer = double_convolution(
                                        input_channels = up_mid_channels,
                                        output_channels = up_out_channels,
                                        kernel_size = kernel_size,
                                        bias = bias,
                                        normalization = normalization,
                                        activation = activation,
                                       )
        self.decoder.append(torch.nn.ModuleList([upsample_layer, conv_layer]))

forward(focal_surface, field)

Forward model.

Parameters:

  • focal_surface (tensor) – 
            Input focal surface data.
        Dimension: (1, 1, H, W)
  • field – 
            Input field data.
        Dimension: (1, 6, H, W)

Returns:

  • sv_kernel ( list of torch.tensor ) –

    Learned spatially varying kernels. Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i), where C_i, H_i, and W_i represent the channel, height, and width of each feature at a certain scale.

Source code in odak/learn/models/models.py 
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
def forward(self, focal_surface, field):
    """
    Forward model.

    Parameters
    ----------
    focal_surface : torch.tensor
                    Input focal surface data.
                    Dimension: (1, 1, H, W)

    field         : torch.tensor
                    Input field data.
                    Dimension: (1, 6, H, W)

    Returns
    -------
    sv_kernel : list of torch.tensor
                Learned spatially varying kernels.
                Dimension of each element in the list: (1, C_i * kernel_size * kernel_size, H_i, W_i),
                where C_i, H_i, and W_i represent the channel, height, and width
                of each feature at a certain scale.
    """
    x = self.inc(torch.cat((focal_surface, field), dim = 1))
    downsampling_outputs = [focal_surface]
    downsampling_outputs.append(x)
    for i, down_layer in enumerate(self.encoder):
        x_down = down_layer(downsampling_outputs[-1])
        downsampling_outputs.append(x_down)
    sv_kernels = []
    for i, (up_layer, svf_layer) in enumerate(zip(self.decoder, self.spatially_varying_feature)):
        if i == 0:
            global_feature = up_layer(downsampling_outputs[-2], downsampling_outputs[-1])
            downsampling_outputs[-1] = global_feature
            sv_feature = [global_feature, downsampling_outputs[0]]
            for j in range(self.depth - i + 1):
                sv_feature[1] = svf_layer[self.depth - i](sv_feature[1])
                if j > 0:
                    sv_feature.append(svf_layer[j](downsampling_outputs[2 * j]))
            sv_feature = [sv_feature[0], sv_feature[1], sv_feature[4], sv_feature[2],
                          sv_feature[3]]
            sv_kernel = svf_layer[-1](torch.cat(sv_feature, dim = 1))
            sv_kernels.append(sv_kernel)
        else:
            x_up = up_layer[0](downsampling_outputs[-1],
                               downsampling_outputs[2 * (self.depth + 1 - i) + 1])
            x_up = up_layer[1](x_up)
            downsampling_outputs[-1] = x_up
            sv_feature = [x_up, downsampling_outputs[0]]
            for j in range(self.depth - i + 1):
                sv_feature[1] = svf_layer[self.depth - i](sv_feature[1])
                if j > 0:
                    sv_feature.append(svf_layer[j](downsampling_outputs[2 * j]))
            if i == 1:
                sv_feature = [sv_feature[0], sv_feature[1], sv_feature[3], sv_feature[2]]
            sv_kernel = svf_layer[-1](torch.cat(sv_feature, dim = 1))
            sv_kernels.append(sv_kernel)
    return sv_kernels

unet

Bases: Module

A U-Net model, heavily inspired from https://github.com/milesial/Pytorch-UNet/tree/master/unet and more can be read from Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Publishing, 2015.

Source code in odak/learn/models/models.py 
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
class unet(torch.nn.Module):
    """
    A U-Net model, heavily inspired from `https://github.com/milesial/Pytorch-UNet/tree/master/unet` and more can be read from Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Publishing, 2015.
    """

    def __init__(
                 self, 
                 depth = 4,
                 dimensions = 64, 
                 input_channels = 2, 
                 output_channels = 1, 
                 bilinear = False,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU(inplace = True),
                ):
        """
        U-Net model.

        Parameters
        ----------
        depth             : int
                            Number of upsampling and downsampling
        dimensions        : int
                            Number of dimensions.
        input_channels    : int
                            Number of input channels.
        output_channels   : int
                            Number of output channels.
        bilinear          : bool
                            Uses bilinear upsampling in upsampling layers when set True.
        bias              : bool
                            Set True to let convolutional layers learn a bias term.
        activation        : torch.nn
                            Non-linear activation layer to be used (e.g., torch.nn.ReLU(), torch.nn.Sigmoid().
        """
        super(unet, self).__init__()
        self.inc = double_convolution(
                                      input_channels = input_channels,
                                      mid_channels = dimensions,
                                      output_channels = dimensions,
                                      kernel_size = kernel_size,
                                      bias = bias,
                                      activation = activation
                                     )      

        self.downsampling_layers = torch.nn.ModuleList()
        self.upsampling_layers = torch.nn.ModuleList()
        for i in range(depth): # downsampling layers
            in_channels = dimensions * (2 ** i)
            out_channels = dimensions * (2 ** (i + 1))
            down_layer = downsample_layer(in_channels,
                                            out_channels,
                                            kernel_size=kernel_size,
                                            bias=bias,
                                            activation=activation
                                            )
            self.downsampling_layers.append(down_layer)      

        for i in range(depth - 1, -1, -1):  # upsampling layers
            up_in_channels = dimensions * (2 ** (i + 1))  
            up_out_channels = dimensions * (2 ** i) 
            up_layer = upsample_layer(up_in_channels, up_out_channels, kernel_size=kernel_size, bias=bias, activation=activation, bilinear=bilinear)
            self.upsampling_layers.append(up_layer)
        self.outc = torch.nn.Conv2d(
                                    dimensions, 
                                    output_channels,
                                    kernel_size = kernel_size,
                                    padding = kernel_size // 2,
                                    bias = bias
                                   )


    def forward(self, x):
        """
        Forward model.

        Parameters
        ----------
        x             : torch.tensor
                        Input data.


        Returns
        ----------
        result        : torch.tensor
                        Estimated output.      
        """
        downsampling_outputs = [self.inc(x)]
        for down_layer in self.downsampling_layers:
            x_down = down_layer(downsampling_outputs[-1])
            downsampling_outputs.append(x_down)
        x_up = downsampling_outputs[-1]
        for i, up_layer in enumerate((self.upsampling_layers)):
            x_up = up_layer(x_up, downsampling_outputs[-(i + 2)])       
        result = self.outc(x_up)
        return result

__init__(depth=4, dimensions=64, input_channels=2, output_channels=1, bilinear=False, kernel_size=3, bias=False, activation=torch.nn.ReLU(inplace=True))

U-Net model.

Parameters:

  • depth – 
                Number of upsampling and downsampling
  • dimensions – 
                Number of dimensions.
  • input_channels – 
                Number of input channels.
  • output_channels – 
                Number of output channels.
  • bilinear – 
                Uses bilinear upsampling in upsampling layers when set True.
  • bias – 
                Set True to let convolutional layers learn a bias term.
  • activation – 
                Non-linear activation layer to be used (e.g., torch.nn.ReLU(), torch.nn.Sigmoid().
Source code in odak/learn/models/models.py 
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
def __init__(
             self, 
             depth = 4,
             dimensions = 64, 
             input_channels = 2, 
             output_channels = 1, 
             bilinear = False,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU(inplace = True),
            ):
    """
    U-Net model.

    Parameters
    ----------
    depth             : int
                        Number of upsampling and downsampling
    dimensions        : int
                        Number of dimensions.
    input_channels    : int
                        Number of input channels.
    output_channels   : int
                        Number of output channels.
    bilinear          : bool
                        Uses bilinear upsampling in upsampling layers when set True.
    bias              : bool
                        Set True to let convolutional layers learn a bias term.
    activation        : torch.nn
                        Non-linear activation layer to be used (e.g., torch.nn.ReLU(), torch.nn.Sigmoid().
    """
    super(unet, self).__init__()
    self.inc = double_convolution(
                                  input_channels = input_channels,
                                  mid_channels = dimensions,
                                  output_channels = dimensions,
                                  kernel_size = kernel_size,
                                  bias = bias,
                                  activation = activation
                                 )      

    self.downsampling_layers = torch.nn.ModuleList()
    self.upsampling_layers = torch.nn.ModuleList()
    for i in range(depth): # downsampling layers
        in_channels = dimensions * (2 ** i)
        out_channels = dimensions * (2 ** (i + 1))
        down_layer = downsample_layer(in_channels,
                                        out_channels,
                                        kernel_size=kernel_size,
                                        bias=bias,
                                        activation=activation
                                        )
        self.downsampling_layers.append(down_layer)      

    for i in range(depth - 1, -1, -1):  # upsampling layers
        up_in_channels = dimensions * (2 ** (i + 1))  
        up_out_channels = dimensions * (2 ** i) 
        up_layer = upsample_layer(up_in_channels, up_out_channels, kernel_size=kernel_size, bias=bias, activation=activation, bilinear=bilinear)
        self.upsampling_layers.append(up_layer)
    self.outc = torch.nn.Conv2d(
                                dimensions, 
                                output_channels,
                                kernel_size = kernel_size,
                                padding = kernel_size // 2,
                                bias = bias
                               )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/models.py 
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
def forward(self, x):
    """
    Forward model.

    Parameters
    ----------
    x             : torch.tensor
                    Input data.


    Returns
    ----------
    result        : torch.tensor
                    Estimated output.      
    """
    downsampling_outputs = [self.inc(x)]
    for down_layer in self.downsampling_layers:
        x_down = down_layer(downsampling_outputs[-1])
        downsampling_outputs.append(x_down)
    x_up = downsampling_outputs[-1]
    for i, up_layer in enumerate((self.upsampling_layers)):
        x_up = up_layer(x_up, downsampling_outputs[-(i + 2)])       
    result = self.outc(x_up)
    return result

upsample_convtranspose2d_layer

Bases: Module

An upsampling convtranspose2d layer.

Source code in odak/learn/models/components.py 
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
class upsample_convtranspose2d_layer(torch.nn.Module):
    """
    An upsampling convtranspose2d layer.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 2,
                 stride = 2,
                 bias = False,
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool
                          Set to True to let convolutional layers have bias term.
        """
        super().__init__()
        self.up = torch.nn.ConvTranspose2d(
                                           in_channels = input_channels,
                                           out_channels = output_channels,
                                           bias = bias,
                                           kernel_size = kernel_size,
                                           stride = stride
                                          )

    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.


        Returns
        ----------
        result        : torch.tensor
                        Result of the forward operation
        """
        x1 = self.up(x1)
        diffY = x2.size()[2] - x1.size()[2]
        diffX = x2.size()[3] - x1.size()[3]
        x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                          diffY // 2, diffY - diffY // 2])
        result = x1 + x2
        return result

__init__(input_channels, output_channels, kernel_size=2, stride=2, bias=False)

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
Source code in odak/learn/models/components.py 
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 2,
             stride = 2,
             bias = False,
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool
                      Set to True to let convolutional layers have bias term.
    """
    super().__init__()
    self.up = torch.nn.ConvTranspose2d(
                                       in_channels = input_channels,
                                       out_channels = output_channels,
                                       bias = bias,
                                       kernel_size = kernel_size,
                                       stride = stride
                                      )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.


    Returns
    ----------
    result        : torch.tensor
                    Result of the forward operation
    """
    x1 = self.up(x1)
    diffY = x2.size()[2] - x1.size()[2]
    diffX = x2.size()[3] - x1.size()[3]
    x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                      diffY // 2, diffY - diffY // 2])
    result = x1 + x2
    return result

upsample_layer

Bases: Module

An upsampling convolutional layer.

Source code in odak/learn/models/components.py 
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
class upsample_layer(torch.nn.Module):
    """
    An upsampling convolutional layer.
    """
    def __init__(
                 self,
                 input_channels,
                 output_channels,
                 kernel_size = 3,
                 bias = False,
                 activation = torch.nn.ReLU(),
                 bilinear = True
                ):
        """
        A downscaling component with a double convolution.

        Parameters
        ----------
        input_channels  : int
                          Number of input channels.
        output_channels : int
                          Number of output channels.
        kernel_size     : int
                          Kernel size.
        bias            : bool 
                          Set to True to let convolutional layers have bias term.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        bilinear        : bool
                          If set to True, bilinear sampling is used.
        """
        super(upsample_layer, self).__init__()
        if bilinear:
            self.up = torch.nn.Upsample(scale_factor = 2, mode = 'bilinear', align_corners = True)
            self.conv = double_convolution(
                                           input_channels = input_channels + output_channels,
                                           mid_channels = input_channels // 2,
                                           output_channels = output_channels,
                                           kernel_size = kernel_size,
                                           bias = bias,
                                           activation = activation
                                          )
        else:
            self.up = torch.nn.ConvTranspose2d(input_channels , input_channels // 2, kernel_size = 2, stride = 2)
            self.conv = double_convolution(
                                           input_channels = input_channels,
                                           mid_channels = output_channels,
                                           output_channels = output_channels,
                                           kernel_size = kernel_size,
                                           bias = bias,
                                           activation = activation
                                          )


    def forward(self, x1, x2):
        """
        Forward model.

        Parameters
        ----------
        x1             : torch.tensor
                         First input data.
        x2             : torch.tensor
                         Second input data.


        Returns
        ----------
        result        : torch.tensor
                        Result of the forward operation
        """ 
        x1 = self.up(x1)
        diffY = x2.size()[2] - x1.size()[2]
        diffX = x2.size()[3] - x1.size()[3]
        x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                          diffY // 2, diffY - diffY // 2])
        x = torch.cat([x2, x1], dim = 1)
        result = self.conv(x)
        return result

__init__(input_channels, output_channels, kernel_size=3, bias=False, activation=torch.nn.ReLU(), bilinear=True)

A downscaling component with a double convolution.

Parameters:

  • input_channels – 
              Number of input channels.
  • output_channels (int) – 
              Number of output channels.
  • kernel_size – 
              Kernel size.
  • bias – 
              Set to True to let convolutional layers have bias term.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
  • bilinear – 
              If set to True, bilinear sampling is used.
Source code in odak/learn/models/components.py 
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
def __init__(
             self,
             input_channels,
             output_channels,
             kernel_size = 3,
             bias = False,
             activation = torch.nn.ReLU(),
             bilinear = True
            ):
    """
    A downscaling component with a double convolution.

    Parameters
    ----------
    input_channels  : int
                      Number of input channels.
    output_channels : int
                      Number of output channels.
    kernel_size     : int
                      Kernel size.
    bias            : bool 
                      Set to True to let convolutional layers have bias term.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    bilinear        : bool
                      If set to True, bilinear sampling is used.
    """
    super(upsample_layer, self).__init__()
    if bilinear:
        self.up = torch.nn.Upsample(scale_factor = 2, mode = 'bilinear', align_corners = True)
        self.conv = double_convolution(
                                       input_channels = input_channels + output_channels,
                                       mid_channels = input_channels // 2,
                                       output_channels = output_channels,
                                       kernel_size = kernel_size,
                                       bias = bias,
                                       activation = activation
                                      )
    else:
        self.up = torch.nn.ConvTranspose2d(input_channels , input_channels // 2, kernel_size = 2, stride = 2)
        self.conv = double_convolution(
                                       input_channels = input_channels,
                                       mid_channels = output_channels,
                                       output_channels = output_channels,
                                       kernel_size = kernel_size,
                                       bias = bias,
                                       activation = activation
                                      )

forward(x1, x2)

Forward model.

Parameters:

  • x1 – 
             First input data.
  • x2 – 
             Second input data.

Returns:

  • result ( tensor ) –

    Result of the forward operation

Source code in odak/learn/models/components.py 
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
def forward(self, x1, x2):
    """
    Forward model.

    Parameters
    ----------
    x1             : torch.tensor
                     First input data.
    x2             : torch.tensor
                     Second input data.


    Returns
    ----------
    result        : torch.tensor
                    Result of the forward operation
    """ 
    x1 = self.up(x1)
    diffY = x2.size()[2] - x1.size()[2]
    diffX = x2.size()[3] - x1.size()[3]
    x1 = torch.nn.functional.pad(x1, [diffX // 2, diffX - diffX // 2,
                                      diffY // 2, diffY - diffY // 2])
    x = torch.cat([x2, x1], dim = 1)
    result = self.conv(x)
    return result

gaussian(x, multiplier=1.0)

A Gaussian non-linear activation. For more details: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.

Parameters:

  • x – 
           Input data.
  • multiplier – 
           Multiplier.

Returns:

  • result ( float or tensor ) –

    Ouput data.

Source code in odak/learn/models/components.py 
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
def gaussian(x, multiplier = 1.):
    """
    A Gaussian non-linear activation.
    For more details: Ramasinghe, Sameera, and Simon Lucey. "Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps." In European Conference on Computer Vision, pp. 142-158. Cham: Springer Nature Switzerland, 2022.

    Parameters
    ----------
    x            : float or torch.tensor
                   Input data.
    multiplier   : float or torch.tensor
                   Multiplier.

    Returns
    -------
    result       : float or torch.tensor
                   Ouput data.
    """
    result = torch.exp(- (multiplier * x) ** 2)
    return result

swish(x)

A swish non-linear activation. For more details: https://en.wikipedia.org/wiki/Swish_function

Parameters:

  • x – 
             Input.

Returns:

  • out ( float or tensor ) –

    Output.

Source code in odak/learn/models/components.py 
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def swish(x):
    """
    A swish non-linear activation.
    For more details: https://en.wikipedia.org/wiki/Swish_function

    Parameters
    -----------
    x              : float or torch.tensor
                     Input.

    Returns
    -------
    out            : float or torch.tensor
                     Output.
    """
    out = x * torch.sigmoid(x)
    return out