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 
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
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
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 
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
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 
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
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 
 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
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
class convolution_layer(torch.nn.Module):
    """
    A convolution layer.
    """
    def __init__(
                 self,
                 input_channels = 2,
                 output_channels = 2,
                 kernel_size = 3,
                 bias = False,
                 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.
        activation      : torch.nn
                          Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
        """
        super().__init__()
        self.activation = activation
        self.model = 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.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, output_channels=2, kernel_size=3, bias=False, 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.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
 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
102
103
104
105
106
def __init__(
             self,
             input_channels = 2,
             output_channels = 2,
             kernel_size = 3,
             bias = False,
             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.
    activation      : torch.nn
                      Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
    """
    super().__init__()
    self.activation = activation
    self.model = 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.activation
                                    )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
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 
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
722
723
724
725
726
727
728
729
730
731
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 
706
707
708
709
710
711
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 
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
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 
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
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 
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
196
197
198
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,
                 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.
        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,
                                                           activation = self.activation
                                                          ),
                                         convolution_layer(
                                                           input_channels = mid_channels,
                                                           output_channels = output_channels,
                                                           kernel_size = kernel_size,
                                                           bias = bias,
                                                           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, 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.
  • activation – 
              Nonlinear activation layer to be used. If None, uses torch.nn.ReLU().
Source code in odak/learn/models/components.py 
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
def __init__(
             self,
             input_channels = 2,
             mid_channels = None,
             output_channels = 2,
             kernel_size = 3, 
             bias = False,
             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.
    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,
                                                       activation = self.activation
                                                      ),
                                     convolution_layer(
                                                       input_channels = mid_channels,
                                                       output_channels = output_channels,
                                                       kernel_size = kernel_size,
                                                       bias = bias,
                                                       activation = self.activation
                                                      )
                                    )

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/components.py 
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
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 
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
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 
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
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 
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
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

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 
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 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 
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
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 
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, 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 
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
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 
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
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 
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
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 
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
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) * math.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 
739
740
741
742
743
744
745
746
747
748
749
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 
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
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) * math.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 
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
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 
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
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 
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
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 
 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
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 
 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
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 
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
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 
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
665
666
667
668
669
670
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 
621
622
623
624
625
626
627
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 
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
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 
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
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

upsample_layer

Bases: Module

An upsampling convolutional layer.

Source code in odak/learn/models/components.py 
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
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 
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
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 
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
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

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
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`, `SIREN`, `FILM SIREN`.
                            `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 `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.
        """
        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)))


    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(self.siren_multiplier * result)
            elif self.model_type == 'SIREN':
                result = torch.sin(result)
            elif self.model_type == 'FILM SIREN':
                result = torch.sin(self.alpha[layer_id][0] * result + self.alpha[layer_id][1])
        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`, `SIREN`, `FILM SIREN`.
            `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 `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.
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
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`, `SIREN`, `FILM SIREN`.
                        `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 `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.
    """
    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)))

forward(x)

Forward model.

Parameters:

  • x – 
            Input data.

Returns:

  • result ( tensor ) –

    Estimated output.

Source code in odak/learn/models/models.py 
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
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(self.siren_multiplier * result)
        elif self.model_type == 'SIREN':
            result = torch.sin(result)
        elif self.model_type == 'FILM SIREN':
            result = torch.sin(self.alpha[layer_id][0] * result + self.alpha[layer_id][1])
    result = self.layers[-1](result)
    return result

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 
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 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
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
                                     )
        factor = 2 if bilinear else 1        
        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 
 94
 95
 96
 97
 98
 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
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
                                 )
    factor = 2 if bilinear else 1        
    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 
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
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