glasses.models.classification.resnet package¶

Module contents¶

class glasses.models.classification.resnet.ResNet(encoder: torch.nn.modules.module.Module = <class 'glasses.models.classification.resnet.ResNetEncoder'>, head: torch.nn.modules.module.Module = <class 'glasses.models.classification.resnet.ResNetHead'>, **kwargs)[source]¶

Bases: glasses.models.classification.base.ClassificationModule

Implementation of ResNet proposed in Deep Residual Learning for Image Recognition

ResNet.resnet18()
ResNet.resnet26()
ResNet.resnet34()
ResNet.resnet50()
ResNet.resnet101()
ResNet.resnet152()
ResNet.resnet200()

Variants (d) proposed in `Bag of Tricks for Image Classification with Convolutional Neural Networks <https://arxiv.org/pdf/1812.01187.pdf>`_

ResNet.resnet26d()
ResNet.resnet34d()
ResNet.resnet50d()
# You can construct your own one by chaning `stem` and `block`
resnet101d = ResNet.resnet101(stem=ResNetStemC, block=partial(ResNetBottleneckBlock, shortcut=ResNetShorcutD))

Examples

# change activation
ResNet.resnet18(activation = nn.SELU)
# change number of classes (default is 1000 )
ResNet.resnet18(n_classes=100)
# pass a different block
ResNet.resnet18(block=SENetBasicBlock)
# change the steam
model = ResNet.resnet18(stem=ResNetStemC)
change shortcut
model = ResNet.resnet18(block=partial(ResNetBasicBlock, shortcut=ResNetShorcutD))
# store each feature
x = torch.rand((1, 3, 224, 224))
# get features
model = ResNet.resnet18()
# first call .features, this will activate the forward hooks and tells the model you'll like to get the features
model.encoder.features
model(torch.randn((1,3,224,224)))
# get the features from the encoder
features = model.encoder.features
print([x.shape for x in features])
#[torch.Size([1, 64, 112, 112]), torch.Size([1, 64, 56, 56]), torch.Size([1, 128, 28, 28]), torch.Size([1, 256, 14, 14])]

Parameters

in_channels (int, optional) – Number of channels in the input Image (3 for RGB and 1 for Gray). Defaults to 3.
n_classes (int, optional) – Number of classes. Defaults to 1000.

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

initialize()[source]¶

classmethod resnet101(*args, block=<class 'glasses.models.classification.resnet.ResNetBottleneckBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet101 model

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/resnet/ResNet101.png?raw=true

Returns: A resnet101 model
Return type: ResNet

classmethod resnet152(*args, block=<class 'glasses.models.classification.resnet.ResNetBottleneckBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet152 model

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/resnet/ResNet152.png?raw=true

Returns: A resnet152 model
Return type: ResNet

classmethod resnet18(*args, block=<class 'glasses.models.classification.resnet.ResNetBasicBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet18 model

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/resnet/ResNet18.png?raw=true

Returns: A resnet18 model
Return type: ResNet

classmethod resnet200(*args, block=<class 'glasses.models.classification.resnet.ResNetBottleneckBlock'>, **kwargs)[source]¶

Creates a resnet200 model

Returns: A resnet200 model
Return type: ResNet

classmethod resnet26(*args, block=<class 'glasses.models.classification.resnet.ResNetBottleneckBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet26 model

Returns: A resnet26 model
Return type: ResNet

classmethod resnet26d(*args, block=<class 'glasses.models.classification.resnet.ResNetBottleneckBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet26d model

Returns: A resnet26d model
Return type: ResNet

classmethod resnet34(*args, block=<class 'glasses.models.classification.resnet.ResNetBasicBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet34 model

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/resnet/ResNet34.png?raw=true

Returns: A resnet34 model
Return type: ResNet

classmethod resnet34d(*args, block=<class 'glasses.models.classification.resnet.ResNetBasicBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet34 model

Returns: A resnet34 model
Return type: ResNet

classmethod resnet50(*args, block=<class 'glasses.models.classification.resnet.ResNetBottleneckBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet50 model

https://github.com/FrancescoSaverioZuppichini/glasses/blob/develop/docs/_static/images/resnet/ResNet50.png?raw=true

Returns: A resnet50 model
Return type: ResNet

classmethod resnet50d(*args, block=<class 'glasses.models.classification.resnet.ResNetBottleneckBlock'>, **kwargs) → glasses.models.classification.resnet.ResNet[source]¶

Creates a resnet50d model

Returns: A resnet50d model
Return type: ResNet

training: bool¶

class glasses.models.classification.resnet.ResNetBasicBlock(in_features: int, out_features: int, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), stride: int = 1, shortcut: torch.nn.modules.module.Module = <class 'glasses.models.classification.resnet.ResNetShorcut'>, **kwargs)[source]¶

Bases: torch.nn.modules.module.Module

Basic ResNet block composed by two 3x3 convs with residual connection.

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

The residual connection is showed as a black line

The output of the layer is defined as:

\(x' = F(x) + x\)

Parameters

out_features (int) – Number of input features
out_features – Number of output features
activation (nn.Module, optional) – [description]. Defaults to ReLUInPlace.
stride (int, optional) – [description]. Defaults to 1.
conv (nn.Module, optional) – [description]. Defaults to nn.Conv2d.

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.models.classification.resnet.ResNetBasicPreActBlock(in_features: int, out_features: int, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), stride: int = 1, shortcut: torch.nn.modules.module.Module = <class 'glasses.models.classification.resnet.ResNetShorcut'>, **kwargs)[source]¶

Bases: glasses.models.classification.resnet.ResNetBasicBlock

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

reduction: int = 1¶

//arxiv.org/pdf/1603.05027.pdf>`_

Parameters

out_features (int) – Number of input features
out_features – Number of ouimport inspect

Type

Pre activation ResNet basic block proposed in `Identity Mappings in Deep Residual Networks <https

class glasses.models.classification.resnet.ResNetBottleneckBlock(in_features: int, out_features: int, features: Optional[int] = None, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), reduction: int = 4, stride=1, shortcut=<class 'glasses.models.classification.resnet.ResNetShorcut'>, **kwargs)[source]¶

Bases: glasses.models.classification.resnet.ResNetBasicBlock

ResNet Bottleneck block based on the torchvision implementation.

Even if the paper says that the first conv1x1 compresses the features. We followed the original implementation that expands the out_features by a factor equal to reduction.

The stride is applied into the 3x3 conv, this improves https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch

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

The residual connection is showed as a black line

Parameters

out_features (int) – Number of input features
out_features – Number of output features
activation (nn.Module, optional) – [description]. Defaults to ReLUInPlace.
stride (int, optional) – [description]. Defaults to 1.
conv (nn.Module, optional) – [description]. Defaults to nn.Conv2d.
reduction (int, optional) – [description]. Defaults to 4.

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

training: bool¶

class glasses.models.classification.resnet.ResNetBottleneckPreActBlock(in_features: int, out_features: int, features: Optional[int] = None, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), reduction: int = 4, stride=1, shortcut=<class 'glasses.models.classification.resnet.ResNetShorcut'>, **kwargs)[source]¶

Bases: glasses.models.classification.resnet.ResNetBottleneckBlock

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

reduction: int = 4¶

//arxiv.org/pdf/1603.05027.pdf>`_

Parameters

out_features (int) – Number of input features
out_features – Number of output features
activation (nn.Module, optional) – [description]. Defaults to ReLUInPlace.
stride (int, optional) – [description]. Defaults to 1.
conv (nn.Module, optional) – [description]. Defaults to nn.Conv2d.

Type

Pre activation ResNet bottleneck block proposed in `Identity Mappings in Deep Residual Networks <https

class glasses.models.classification.resnet.ResNetEncoder(in_channels: int = 3, start_features: int = 64, widths: List[int] = [64, 128, 256, 512], depths: List[int] = [2, 2, 2, 2], activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True), block: torch.nn.modules.module.Module = <class 'glasses.models.classification.resnet.ResNetBasicBlock'>, stem: torch.nn.modules.module.Module = <class 'glasses.models.classification.resnet.ResNetStem'>, downsample_first: bool = False, **kwargs)[source]¶

Bases: glasses.models.base.Encoder

ResNet encoder composed by increasing different layers with increasing features.

Parameters

in_channels (int, optional) – [description]. Defaults to 3.
start_features (int, optional) – [description]. Defaults to 64.
widths (List[int], optional) – [description]. Defaults to [64, 128, 256, 512].
depths (List[int], optional) – [description]. Defaults to [2, 2, 2, 2].
activation (nn.Module, optional) – [description]. Defaults to ReLUInPlace.
block (nn.Module, optional) – Block used, there are several ones such as ResNetBasicBlock and ResNetBottleneckBlock . Defaults to ResNetBasicBlock.
stem (nn.Module, optional) – Stem used. Defaults to ResNetStem.

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

property features_widths¶

forward(x)[source]¶

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

property stages¶

training: bool¶

class glasses.models.classification.resnet.ResNetHead(in_features: int, n_classes: int)[source]¶

Bases: torch.nn.modules.container.Sequential

This class represents the tail of ResNet. It performs a global pooling and maps the output to the correct class by using a fully connected layer.

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

class glasses.models.classification.resnet.ResNetLayer(in_features: int, out_features: int, block: torch.nn.modules.module.Module = <class 'glasses.models.classification.resnet.ResNetBasicBlock'>, depth: int = 1, stride: int = 2, **kwargs)[source]¶

Bases: torch.nn.modules.container.Sequential

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

class glasses.models.classification.resnet.ResNetShorcut(in_features: int, out_features: int, stride: int = 2)[source]¶

Bases: torch.nn.modules.container.Sequential

Shorcut function applied by ResNet to upsample the channel when residual and output features do not match

Parameters

in_features (int) – features (channels) of the input
out_features (int) – features (channels) of the desidered output

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

class glasses.models.classification.resnet.ResNetShorcutD(in_features: int, out_features: int, stride: int = 2)[source]¶

Bases: torch.nn.modules.container.Sequential

Shorcut function proposed in Bag of Tricks for Image Classification with Convolutional Neural Networks

It applies average pool instead of stride=2 in the convolution layer

Parameters

in_features (int) – features (channels) of the input
out_features (int) – features (channels) of the desidered output

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

class glasses.models.classification.resnet.ResNetStem(in_features: int, out_features: int, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.ReLU'>, inplace=True))[source]¶

Bases: torch.nn.modules.container.Sequential

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

class glasses.models.classification.resnet.ResNetStem3x3(in_features: int, out_features: int, widths: List[int] = [32, 32], activation: torch.nn.modules.module.Module = <class 'torch.nn.modules.activation.ReLU'>)[source]¶

Bases: torch.nn.modules.container.Sequential

Modified stem proposed in Bag of Tricks for Image Classification with Convolutional Neural Networks

The observation is that the computational cost of a convolution is quadratic to the kernel width or height. A 7 × 7 convolution is 5.4 times more expensive than a 3 × 3 convolution. So this tweak replacing the 7 × 7 convolution in the input stem with three conservative 3 × 3 convolution.

Parameters

in_features (int) – [description]
out_features (int) – [description]
widths (List[int], optional) – Widths of the inner stem. Defaults to [32, 32], 32 - in, 32 - out.
activation (nn.Module, optional) – [description]. Defaults to nn.ReLU.

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