glasses.nn.att package

Submodules

glasses.nn.att.CBAM module

class glasses.nn.att.CBAM.CBAM(features: int, reduction: int = 16, reduced_features: Optional[int] = None, activation: torch.nn.modules.module.Module = <class 'torch.nn.modules.activation.ReLU'>, kernel_size: int = 7)[source]

Bases: torch.nn.modules.module.Module

Implementation of Convolutional Block Attention Module proposed in CBAM: Convolutional Block Attention Module

Examples

>>> # create cbamresnet50
>>> from glasses.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import CBAM, WithAtt
>>> cbam_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=CBAM))
>>> cbam_resnet50.summary()
Parameters

  • features (int, optional) – Number of features features. Defaults to None.

  • reduction (int, optional) – Reduction ratio used to downsample the input. Defaults to 16.

  • reduced_features – If passed, use it instead of calculating the reduced features using reduction. Defaults to None.

  • kernel_size (int, optional) – kernel_size of the Conv2d to produce the 2D spatial attention map. Defaults to 7.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class glasses.nn.att.CBAM.CBAMChannelAtt(features: int, reduction: int = 16, reduced_features: Optional[int] = None, activation: torch.nn.modules.module.Module = <class 'torch.nn.modules.activation.ReLU'>)[source]

Bases: torch.nn.modules.module.Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class glasses.nn.att.CBAM.CBAMSpatialAtt(kernel_size: int = 7)[source]

Bases: torch.nn.modules.module.Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

glasses.nn.att.ECA module

class glasses.nn.att.ECA.ECA(features: int, kernel_size: int = 3, gamma: int = 2, beta: int = 1)[source]

Bases: torch.nn.modules.module.Module

Implementation of Efficient Channel Attention proposed in ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/ECA.png?raw=true

Examples

>>> # create ecaresnet50
>>> from glasses.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import ECA, WithAtt
>>> eca_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=ECA))
>>> eca_resnet50.summary()
Parameters

  • features (int, optional) – Number of features features. Defaults to None.

  • kernel_size (int, optional) – [description]. Defaults to 3.

  • gamma (int, optional) – [description]. Defaults to 2.

  • beta (int, optional) – [description]. Defaults to 1.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

glasses.nn.att.se module

class glasses.nn.att.se.ChannelSE(features: int, *args, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), **kwargs)[source]

Bases: glasses.nn.att.se.SpatialSE

Modified implementation of Squeeze and Excitation Module proposed in Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks

It squeezes channel-wise and excitates spatially.

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/ChannelSE.png?raw=true

Examples

Add ChannelSE to your own model is very simple.

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    ChannelSE(64, reduction=4)
>>>    nn.ReLU(),
>>> )

You can also direcly specify the number of features inside the module

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    ChannelSE(64, reduced_features=10)
>>>    nn.ReLU(),
>>> )

Creating the cseresnet50

>>> from glasses.nn.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import ECA, WithAtt
>>> se_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=ChannelSE))
>>> se_resnet50.summary()
Parameters

  • features (int) – Number of features

  • reduction (int, optional) – Reduction ratio used to downsample the input. Defaults to 16.

  • reduced_features (int, optional) – If passed, use it instead of calculating the reduced features using reduction. Defaults to None.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

training: bool
class glasses.nn.att.se.LegacySpatialSE(features: int, *args, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), **kwargs)[source]

Bases: glasses.nn.att.se.SpatialSE

Implementation of Squeeze and Excitation Module proposed in Squeeze-and-Excitation Networks

The idea is to apply a learned weight to rescale the channels.

It squeezes spatially and excitates channel-wise.

The authors reported a bigger performance increase where the number of features are higher.

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/se.png?raw=true

Further visualisation from Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/SpatialSE.png?raw=true

Examples

Add SpatialSE to your own model is very simple.

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    SpatialSE(64, reduction=4)
>>>    nn.ReLU(),
>>> )

You can also direcly specify the number of features inside the module

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    SpatialSE(64, reduced_features=10)
>>>    nn.ReLU(),
>>> )

The following example shows a more advance scenarion where we add Squeeze ad Excitation to a ResNetBasicBlock.

>>> class SENetBasicBlock(ResNetBasicBlock):
>>>     def __init__(self, in_features: int, out_features: int, reduction: int =16, *args, **kwargs):
>>>        super().__init__(in_features, out_features, *args, **kwargs)
>>>        # add se to the `.block`
>>>        self.block.add_module('se', SpatialSE(out_features))

Creating the original seresnet50

>>> from glasses.nn.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import ECA, WithAtt
>>> se_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=SpatialSE))
>>> se_resnet50.summary()
Parameters

  • features (int) – Number of features

  • reduction (int, optional) – Reduction ratio used to downsample the input. Defaults to 16.

  • reduced_features (int, optional) – If passed, use it instead of calculating the reduced features using reduction. Defaults to None.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

training: bool
class glasses.nn.att.se.SpatialChannelSE(*args, **kwargs)[source]

Bases: torch.nn.modules.module.Module

Implementation of Spatial and Channel Squeeze and Excitation Module proposed in Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks

This module combines booth Spatial and Channel Squeeze and Excitation

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/SpatialAndChannelSE.png?raw=true

Examples

Add SpatialChannelSE to your own model is very simple.

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    SpatialChannelSE(64, reduction=4)
>>>    nn.ReLU(),
>>> )

You can also direcly specify the number of features inside the module

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    SpatialChannelSE(64, reduced_features=10)
>>>    nn.ReLU(),
>>> )

Creating scseresnet50

>>> from glasses.nn.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import ECA, WithAtt
>>> se_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=SpatialChannelSE))
>>> se_resnet50.summary()
Parameters

  • features (int) – Number of features

  • reduction (int, optional) – Reduction ratio used to downsample the input. Defaults to 16.

  • reduced_features (int, optional) – If passed, use it instead of calculating the reduced features using reduction. Defaults to None.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class glasses.nn.att.se.SpatialSE(features: int, reduction: int = 16, reduced_features: Optional[int] = None, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True))[source]

Bases: torch.nn.modules.module.Module

Modernized Implementation of Squeeze and Excitation Module proposed in Squeeze-and-Excitation Networks with single kernel convolution instead of fully connected layers.

See LegacySpatialSE for usage.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

glasses.nn.att.utils module

class glasses.nn.att.utils.WithAtt(block: torch.nn.modules.module.Module, att: torch.nn.modules.module.Module = <class 'glasses.nn.att.se.SpatialSE'>)[source]

Bases: object

Utility class that adds an attention module after .block.

Usage
>>> WithAtt(ResNetBottleneckBlock, att=SpatialSE)
>>> WithAtt(ResNetBottleneckBlock, att=ECA)
>>> from functools import partial
>>> WithAtt(ResNetBottleneckBlock, att=partial(SpatialSE, reduction=8))

Module contents

class glasses.nn.att.CBAM(features: int, reduction: int = 16, reduced_features: Optional[int] = None, activation: torch.nn.modules.module.Module = <class 'torch.nn.modules.activation.ReLU'>, kernel_size: int = 7)[source]

Bases: torch.nn.modules.module.Module

Implementation of Convolutional Block Attention Module proposed in CBAM: Convolutional Block Attention Module

Examples

>>> # create cbamresnet50
>>> from glasses.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import CBAM, WithAtt
>>> cbam_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=CBAM))
>>> cbam_resnet50.summary()
Parameters

  • features (int, optional) – Number of features features. Defaults to None.

  • reduction (int, optional) – Reduction ratio used to downsample the input. Defaults to 16.

  • reduced_features – If passed, use it instead of calculating the reduced features using reduction. Defaults to None.

  • kernel_size (int, optional) – kernel_size of the Conv2d to produce the 2D spatial attention map. Defaults to 7.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class glasses.nn.att.ChannelSE(features: int, *args, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), **kwargs)[source]

Bases: glasses.nn.att.se.SpatialSE

Modified implementation of Squeeze and Excitation Module proposed in Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks

It squeezes channel-wise and excitates spatially.

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/ChannelSE.png?raw=true

Examples

Add ChannelSE to your own model is very simple.

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    ChannelSE(64, reduction=4)
>>>    nn.ReLU(),
>>> )

You can also direcly specify the number of features inside the module

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    ChannelSE(64, reduced_features=10)
>>>    nn.ReLU(),
>>> )

Creating the cseresnet50

>>> from glasses.nn.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import ECA, WithAtt
>>> se_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=ChannelSE))
>>> se_resnet50.summary()
Parameters

  • features (int) – Number of features

  • reduction (int, optional) – Reduction ratio used to downsample the input. Defaults to 16.

  • reduced_features (int, optional) – If passed, use it instead of calculating the reduced features using reduction. Defaults to None.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

training: bool
class glasses.nn.att.ECA(features: int, kernel_size: int = 3, gamma: int = 2, beta: int = 1)[source]

Bases: torch.nn.modules.module.Module

Implementation of Efficient Channel Attention proposed in ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/ECA.png?raw=true

Examples

>>> # create ecaresnet50
>>> from glasses.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import ECA, WithAtt
>>> eca_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=ECA))
>>> eca_resnet50.summary()
Parameters

  • features (int, optional) – Number of features features. Defaults to None.

  • kernel_size (int, optional) – [description]. Defaults to 3.

  • gamma (int, optional) – [description]. Defaults to 2.

  • beta (int, optional) – [description]. Defaults to 1.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class glasses.nn.att.LegacySpatialSE(features: int, *args, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), **kwargs)[source]

Bases: glasses.nn.att.se.SpatialSE

Implementation of Squeeze and Excitation Module proposed in Squeeze-and-Excitation Networks

The idea is to apply a learned weight to rescale the channels.

It squeezes spatially and excitates channel-wise.

The authors reported a bigger performance increase where the number of features are higher.

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/se.png?raw=true

Further visualisation from Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/SpatialSE.png?raw=true

Examples

Add SpatialSE to your own model is very simple.

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    SpatialSE(64, reduction=4)
>>>    nn.ReLU(),
>>> )

You can also direcly specify the number of features inside the module

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    SpatialSE(64, reduced_features=10)
>>>    nn.ReLU(),
>>> )

The following example shows a more advance scenarion where we add Squeeze ad Excitation to a ResNetBasicBlock.

>>> class SENetBasicBlock(ResNetBasicBlock):
>>>     def __init__(self, in_features: int, out_features: int, reduction: int =16, *args, **kwargs):
>>>        super().__init__(in_features, out_features, *args, **kwargs)
>>>        # add se to the `.block`
>>>        self.block.add_module('se', SpatialSE(out_features))

Creating the original seresnet50

>>> from glasses.nn.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import ECA, WithAtt
>>> se_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=SpatialSE))
>>> se_resnet50.summary()
Parameters

  • features (int) – Number of features

  • reduction (int, optional) – Reduction ratio used to downsample the input. Defaults to 16.

  • reduced_features (int, optional) – If passed, use it instead of calculating the reduced features using reduction. Defaults to None.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

training: bool
class glasses.nn.att.SpatialChannelSE(*args, **kwargs)[source]

Bases: torch.nn.modules.module.Module

Implementation of Spatial and Channel Squeeze and Excitation Module proposed in Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks

This module combines booth Spatial and Channel Squeeze and Excitation

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/SpatialAndChannelSE.png?raw=true

Examples

Add SpatialChannelSE to your own model is very simple.

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    SpatialChannelSE(64, reduction=4)
>>>    nn.ReLU(),
>>> )

You can also direcly specify the number of features inside the module

>>> nn.Sequential(
>>>    nn.Conv2d(32, 64, kernel_size=3),
>>>    SpatialChannelSE(64, reduced_features=10)
>>>    nn.ReLU(),
>>> )

Creating scseresnet50

>>> from glasses.nn.models.classification.resnet import ResNet, ResNetBottleneckBlock
>>> from glasses.nn.att import ECA, WithAtt
>>> se_resnet50 = ResNet.resnet50(block=WithAtt(ResNetBottleneckBlock, att=SpatialChannelSE))
>>> se_resnet50.summary()
Parameters

  • features (int) – Number of features

  • reduction (int, optional) – Reduction ratio used to downsample the input. Defaults to 16.

  • reduced_features (int, optional) – If passed, use it instead of calculating the reduced features using reduction. Defaults to None.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class glasses.nn.att.SpatialSE(features: int, reduction: int = 16, reduced_features: Optional[int] = None, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True))[source]

Bases: torch.nn.modules.module.Module

Modernized Implementation of Squeeze and Excitation Module proposed in Squeeze-and-Excitation Networks with single kernel convolution instead of fully connected layers.

See LegacySpatialSE for usage.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class glasses.nn.att.WithAtt(block: torch.nn.modules.module.Module, att: torch.nn.modules.module.Module = <class 'glasses.nn.att.se.SpatialSE'>)[source]

Bases: object

Utility class that adds an attention module after .block.

Usage
>>> WithAtt(ResNetBottleneckBlock, att=SpatialSE)
>>> WithAtt(ResNetBottleneckBlock, att=ECA)
>>> from functools import partial
>>> WithAtt(ResNetBottleneckBlock, att=partial(SpatialSE, reduction=8))