inferno.extensions.layers package

Submodules

inferno.extensions.layers.activations module

class inferno.extensions.layers.activations.SELU

Bases: torch.nn.modules.module.Module

forward(input)

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.

static selu(x)

inferno.extensions.layers.building_blocks module

class inferno.extensions.layers.building_blocks.ResBlockBase(in_channels, out_channels, dim, size=2, force_skip_op=False, activated=True)

Bases: torch.nn.modules.module.Module

activated_skip_op_factory(in_channels, out_channels)

activation_op_factory(index)

forward(input)

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.

nonactivated_conv_op_factory(in_channels, out_channels, index)

class inferno.extensions.layers.building_blocks.ResBlock(in_channels, out_channels, dim, size=2, activated=True, activation='ReLU', batchnorm=True, force_skip_op=False, conv_kwargs=None)

Bases: inferno.extensions.layers.building_blocks.ResBlockBase

activated_skip_op_factory(in_channels, out_channels)

activation_op_factory(index)

batchnorm_op_factory(in_channels)

nonactivated_conv_op_factory(in_channels, out_channels, index)

inferno.extensions.layers.convolutional module

class inferno.extensions.layers.convolutional.ConvActivation(in_channels, out_channels, kernel_size, dim, activation, stride=1, dilation=1, groups=None, depthwise=False, bias=True, deconv=False, initialization=None)

Bases: torch.nn.modules.module.Module

Convolutional layer with ‘SAME’ padding followed by an activation.

forward(input)

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.

get_padding(kernel_size, dilation)

class inferno.extensions.layers.convolutional.ConvELU2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D Convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.ConvELU3D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D Convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.ConvSigmoid2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D Convolutional layer with ‘SAME’ padding, Sigmoid and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.ConvSigmoid3D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D Convolutional layer with ‘SAME’ padding, Sigmoid and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.DeconvELU2D(in_channels, out_channels, kernel_size=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D deconvolutional layer with ELU and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.DeconvELU3D(in_channels, out_channels, kernel_size=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D deconvolutional layer with ELU and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.StridedConvELU2D(in_channels, out_channels, kernel_size, stride=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D strided convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.StridedConvELU3D(in_channels, out_channels, kernel_size, stride=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D strided convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.DilatedConvELU2D(in_channels, out_channels, kernel_size, dilation=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D dilated convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.DilatedConvELU3D(in_channels, out_channels, kernel_size, dilation=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D dilated convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.convolutional.Conv2D(in_channels, out_channels, kernel_size, dilation=1, stride=1, activation=None)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D convolutional layer with same padding and orthogonal weight initialization. By default, this layer does not apply an activation function.

class inferno.extensions.layers.convolutional.Conv3D(in_channels, out_channels, kernel_size, dilation=1, stride=1, activation=None)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D convolutional layer with same padding and orthogonal weight initialization. By default, this layer does not apply an activation function.

class inferno.extensions.layers.convolutional.BNReLUConv2D(in_channels, out_channels, kernel_size, stride=1)

Bases: inferno.extensions.layers.convolutional._BNReLUSomeConv, inferno.extensions.layers.convolutional.ConvActivation

2D BN-ReLU-Conv layer with ‘SAME’ padding and He weight initialization.

class inferno.extensions.layers.convolutional.BNReLUConv3D(in_channels, out_channels, kernel_size, stride=1)

Bases: inferno.extensions.layers.convolutional._BNReLUSomeConv, inferno.extensions.layers.convolutional.ConvActivation

3D BN-ReLU-Conv layer with ‘SAME’ padding and He weight initialization.

class inferno.extensions.layers.convolutional.BNReLUDepthwiseConv2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional._BNReLUSomeConv, inferno.extensions.layers.convolutional.ConvActivation

2D BN-ReLU-Conv layer with ‘SAME’ padding, He weight initialization and depthwise convolution. Note that depthwise convolutions require in_channels == out_channels.

class inferno.extensions.layers.convolutional.ConvSELU2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D Convolutional layer with SELU activation and the appropriate weight initialization.

class inferno.extensions.layers.convolutional.ConvSELU3D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D Convolutional layer with SELU activation and the appropriate weight initialization.

class inferno.extensions.layers.convolutional.ConvReLU2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D Convolutional layer with ‘SAME’ padding, ReLU and Kaiming normal weight initialization.

class inferno.extensions.layers.convolutional.ConvReLU3D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D Convolutional layer with ‘SAME’ padding, ReLU and Kaiming normal weight initialization.

inferno.extensions.layers.device module

class inferno.extensions.layers.device.DeviceTransfer(target_device, device_ordinal=None, async=False)

Bases: torch.nn.modules.module.Module

Layer to transfer variables to a specified device.

forward(*inputs)

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.

class inferno.extensions.layers.device.OnDevice(module, target_device, device_ordinal=None, async=False)

Bases: torch.nn.modules.module.Module

Moves a module to a device. The advantage of using this over torch.nn.Module.cuda is that the inputs are transferred to the same device as the module, enabling easy model parallelism.

forward(*inputs)

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.

transfer_module(module)

inferno.extensions.layers.identity module

inferno.extensions.layers.prefab module

class inferno.extensions.layers.prefab.PreActSimpleResidualBlock(in_channels, num_hidden_channels, upsample=False, downsample=False)

Bases: inferno.extensions.layers.prefab.ResidualBlock

class inferno.extensions.layers.prefab.ResidualBlock(layers, resample=None)

Bases: torch.nn.modules.module.Module

forward(input)

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.

inferno.extensions.layers.res_unet module

class inferno.extensions.layers.res_unet.ResBlockUNet(in_channels, dim, out_channels, unet_kwargs=None, res_block_kwargs=None, activated=True, side_out_parts=None)

Bases: inferno.extensions.layers.unet_base.UNetBase

TODO.

ACCC

activated

TYPE – Description

dim

TYPE – Description

res_block_kwargs

TYPE – Description

side_out_parts

TYPE – Description

unet_kwargs

TYPE – Description

conv_op_factory(in_channels, out_channels, part, index)

inferno.extensions.layers.reshape module

class inferno.extensions.layers.reshape.View(as_shape)

Bases: torch.nn.modules.module.Module

forward(input)

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.

validate_as_shape(as_shape)

class inferno.extensions.layers.reshape.AsMatrix

Bases: inferno.extensions.layers.reshape.View

class inferno.extensions.layers.reshape.Flatten

Bases: inferno.extensions.layers.reshape.View

class inferno.extensions.layers.reshape.As3D(channel_as_z=False, num_channels_or_num_z_slices=1)

Bases: torch.nn.modules.module.Module

forward(input)

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.

class inferno.extensions.layers.reshape.As2D(z_as_channel=True)

Bases: torch.nn.modules.module.Module

forward(input)

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.

class inferno.extensions.layers.reshape.Concatenate(dim=1)

Bases: torch.nn.modules.module.Module

Concatenate input tensors along a specified dimension.

forward(*inputs)

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.

class inferno.extensions.layers.reshape.Cat(dim=1)

Bases: inferno.extensions.layers.reshape.Concatenate

An alias for Concatenate. Hey, everyone knows who Cat is.

class inferno.extensions.layers.reshape.ResizeAndConcatenate(target_size, pool_mode='average', dim=1)

Bases: torch.nn.modules.module.Module

Resize input tensors spatially (to a specified target size) before concatenating them along the a given dim (channel, i.e. 1 by default). The down-sampling mode can be specified (‘average’ or ‘max’), but the up-sampling is always ‘nearest’.

POOL_MODE_MAPPING = {'average': 'avg', 'avg': 'avg', 'max': 'max', 'mean': 'avg'}

forward(*inputs)

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.

class inferno.extensions.layers.reshape.PoolCat(target_size, pool_mode='average', dim=1)

Bases: inferno.extensions.layers.reshape.ResizeAndConcatenate

Alias for ResizeAndConcatenate, just to annoy snarky web developers.

class inferno.extensions.layers.reshape.GlobalMeanPooling

Bases: inferno.extensions.layers.reshape.ResizeAndConcatenate

Global mean pooling layer.

class inferno.extensions.layers.reshape.GlobalMaxPooling

Bases: inferno.extensions.layers.reshape.ResizeAndConcatenate

Global max pooling layer.

class inferno.extensions.layers.reshape.Sum

Bases: torch.nn.modules.module.Module

Sum all inputs.

forward(*inputs)

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.

class inferno.extensions.layers.reshape.SplitChannels(channel_index)

Bases: torch.nn.modules.module.Module

Split input at a given index along the channel axis.

forward(input)

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.

class inferno.extensions.layers.reshape.Squeeze

Bases: torch.nn.modules.module.Module

forward(x)

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.

class inferno.extensions.layers.reshape.RemoveSingletonDimension(dim=1)

Bases: torch.nn.modules.module.Module

forward(x)

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.

inferno.extensions.layers.sampling module

class inferno.extensions.layers.sampling.AnisotropicUpsample(scale_factor)

Bases: torch.nn.modules.module.Module

forward(input)

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.

class inferno.extensions.layers.sampling.AnisotropicPool(downscale_factor)

Bases: torch.nn.modules.pooling.MaxPool3d

inferno.extensions.layers.unet_base module

class inferno.extensions.layers.unet_base.UNetBase(in_channels, out_channels, dim, depth=3, gain=2, residual=False, upsample_mode=None, p_dropout=None)

Bases: torch.nn.modules.module.Module

Base class for implementing UNets.

The depth and dimension of the UNet is flexible. The deriving classes must implement conv_op_factory and can implement upsample_op_factory and downsample_op_factory.

in_channels

int – Description

out_channels

int – Description

dim

int – Spatial dimension of data (must be 2 or 3)

depth

int – How many down-sampling / up-sampling steps shall be performed

gain

int – Multiplicative increase of channels while going down in the UNet. The same factor is used to decrease the number of channels while going up in the UNet.

residual

bool – If residual is true, the output of the down-streams are added to the up-stream results. Otherwise the results are concatenated.

conv_op_factory(in_channels, out_channels, part, index)

downsample_op_factory(index)

forward(input)

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.

post_conv_op_regularizer_factory(in_channels, out_channels, part, index)

pre_conv_op_regularizer_factory(in_channels, out_channels, part, index)

upsample_op_factory(index)

Module contents

class inferno.extensions.layers.SELU

Bases: torch.nn.modules.module.Module

forward(input)

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.

static selu(x)

class inferno.extensions.layers.ConvActivation(in_channels, out_channels, kernel_size, dim, activation, stride=1, dilation=1, groups=None, depthwise=False, bias=True, deconv=False, initialization=None)

Bases: torch.nn.modules.module.Module

Convolutional layer with ‘SAME’ padding followed by an activation.

forward(input)

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.

get_padding(kernel_size, dilation)

class inferno.extensions.layers.ConvELU2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D Convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.ConvELU3D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D Convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.ConvSigmoid2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D Convolutional layer with ‘SAME’ padding, Sigmoid and orthogonal weight initialization.

class inferno.extensions.layers.ConvSigmoid3D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D Convolutional layer with ‘SAME’ padding, Sigmoid and orthogonal weight initialization.

class inferno.extensions.layers.DeconvELU2D(in_channels, out_channels, kernel_size=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D deconvolutional layer with ELU and orthogonal weight initialization.

class inferno.extensions.layers.DeconvELU3D(in_channels, out_channels, kernel_size=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D deconvolutional layer with ELU and orthogonal weight initialization.

class inferno.extensions.layers.StridedConvELU2D(in_channels, out_channels, kernel_size, stride=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D strided convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.StridedConvELU3D(in_channels, out_channels, kernel_size, stride=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D strided convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.DilatedConvELU2D(in_channels, out_channels, kernel_size, dilation=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D dilated convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.DilatedConvELU3D(in_channels, out_channels, kernel_size, dilation=2)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D dilated convolutional layer with ‘SAME’ padding, ELU and orthogonal weight initialization.

class inferno.extensions.layers.Conv2D(in_channels, out_channels, kernel_size, dilation=1, stride=1, activation=None)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D convolutional layer with same padding and orthogonal weight initialization. By default, this layer does not apply an activation function.

class inferno.extensions.layers.Conv3D(in_channels, out_channels, kernel_size, dilation=1, stride=1, activation=None)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D convolutional layer with same padding and orthogonal weight initialization. By default, this layer does not apply an activation function.

class inferno.extensions.layers.BNReLUConv2D(in_channels, out_channels, kernel_size, stride=1)

Bases: inferno.extensions.layers.convolutional._BNReLUSomeConv, inferno.extensions.layers.convolutional.ConvActivation

2D BN-ReLU-Conv layer with ‘SAME’ padding and He weight initialization.

class inferno.extensions.layers.BNReLUConv3D(in_channels, out_channels, kernel_size, stride=1)

Bases: inferno.extensions.layers.convolutional._BNReLUSomeConv, inferno.extensions.layers.convolutional.ConvActivation

3D BN-ReLU-Conv layer with ‘SAME’ padding and He weight initialization.

class inferno.extensions.layers.BNReLUDepthwiseConv2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional._BNReLUSomeConv, inferno.extensions.layers.convolutional.ConvActivation

2D BN-ReLU-Conv layer with ‘SAME’ padding, He weight initialization and depthwise convolution. Note that depthwise convolutions require in_channels == out_channels.

class inferno.extensions.layers.ConvSELU2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D Convolutional layer with SELU activation and the appropriate weight initialization.

class inferno.extensions.layers.ConvSELU3D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D Convolutional layer with SELU activation and the appropriate weight initialization.

class inferno.extensions.layers.ConvReLU2D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

2D Convolutional layer with ‘SAME’ padding, ReLU and Kaiming normal weight initialization.

class inferno.extensions.layers.ConvReLU3D(in_channels, out_channels, kernel_size)

Bases: inferno.extensions.layers.convolutional.ConvActivation

3D Convolutional layer with ‘SAME’ padding, ReLU and Kaiming normal weight initialization.

class inferno.extensions.layers.DeviceTransfer(target_device, device_ordinal=None, async=False)

Bases: torch.nn.modules.module.Module

Layer to transfer variables to a specified device.

forward(*inputs)

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.

class inferno.extensions.layers.OnDevice(module, target_device, device_ordinal=None, async=False)

Bases: torch.nn.modules.module.Module

Moves a module to a device. The advantage of using this over torch.nn.Module.cuda is that the inputs are transferred to the same device as the module, enabling easy model parallelism.

forward(*inputs)

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.

transfer_module(module)

class inferno.extensions.layers.View(as_shape)

Bases: torch.nn.modules.module.Module

forward(input)

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.

validate_as_shape(as_shape)

class inferno.extensions.layers.AsMatrix

Bases: inferno.extensions.layers.reshape.View

class inferno.extensions.layers.Flatten

Bases: inferno.extensions.layers.reshape.View

class inferno.extensions.layers.As3D(channel_as_z=False, num_channels_or_num_z_slices=1)

Bases: torch.nn.modules.module.Module

forward(input)

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.

class inferno.extensions.layers.As2D(z_as_channel=True)

Bases: torch.nn.modules.module.Module

forward(input)

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.

class inferno.extensions.layers.Concatenate(dim=1)

Bases: torch.nn.modules.module.Module

Concatenate input tensors along a specified dimension.

forward(*inputs)

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.

class inferno.extensions.layers.Cat(dim=1)

Bases: inferno.extensions.layers.reshape.Concatenate

An alias for Concatenate. Hey, everyone knows who Cat is.

class inferno.extensions.layers.ResizeAndConcatenate(target_size, pool_mode='average', dim=1)

Bases: torch.nn.modules.module.Module

Resize input tensors spatially (to a specified target size) before concatenating them along the a given dim (channel, i.e. 1 by default). The down-sampling mode can be specified (‘average’ or ‘max’), but the up-sampling is always ‘nearest’.

POOL_MODE_MAPPING = {'average': 'avg', 'avg': 'avg', 'max': 'max', 'mean': 'avg'}

forward(*inputs)

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.

class inferno.extensions.layers.PoolCat(target_size, pool_mode='average', dim=1)

Bases: inferno.extensions.layers.reshape.ResizeAndConcatenate

Alias for ResizeAndConcatenate, just to annoy snarky web developers.

class inferno.extensions.layers.GlobalMeanPooling

Bases: inferno.extensions.layers.reshape.ResizeAndConcatenate

Global mean pooling layer.

class inferno.extensions.layers.GlobalMaxPooling

Bases: inferno.extensions.layers.reshape.ResizeAndConcatenate

Global max pooling layer.

class inferno.extensions.layers.Sum

Bases: torch.nn.modules.module.Module

Sum all inputs.

forward(*inputs)

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.

class inferno.extensions.layers.SplitChannels(channel_index)

Bases: torch.nn.modules.module.Module

Split input at a given index along the channel axis.

forward(input)

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.

class inferno.extensions.layers.Squeeze

Bases: torch.nn.modules.module.Module

forward(x)

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.

class inferno.extensions.layers.RemoveSingletonDimension(dim=1)

Bases: torch.nn.modules.module.Module

forward(x)

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.

class inferno.extensions.layers.UNetBase(in_channels, out_channels, dim, depth=3, gain=2, residual=False, upsample_mode=None, p_dropout=None)

Bases: torch.nn.modules.module.Module

Base class for implementing UNets.

The depth and dimension of the UNet is flexible. The deriving classes must implement conv_op_factory and can implement upsample_op_factory and downsample_op_factory.

in_channels

int – Description

out_channels

int – Description

dim

int – Spatial dimension of data (must be 2 or 3)

depth

int – How many down-sampling / up-sampling steps shall be performed

gain

int – Multiplicative increase of channels while going down in the UNet. The same factor is used to decrease the number of channels while going up in the UNet.

residual

bool – If residual is true, the output of the down-streams are added to the up-stream results. Otherwise the results are concatenated.

conv_op_factory(in_channels, out_channels, part, index)

downsample_op_factory(index)

forward(input)

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.

post_conv_op_regularizer_factory(in_channels, out_channels, part, index)

pre_conv_op_regularizer_factory(in_channels, out_channels, part, index)

upsample_op_factory(index)

class inferno.extensions.layers.ResBlockUNet(in_channels, dim, out_channels, unet_kwargs=None, res_block_kwargs=None, activated=True, side_out_parts=None)

Bases: inferno.extensions.layers.unet_base.UNetBase

TODO.

ACCC

activated

TYPE – Description

dim

TYPE – Description

res_block_kwargs

TYPE – Description

side_out_parts

TYPE – Description

unet_kwargs

TYPE – Description

conv_op_factory(in_channels, out_channels, part, index)

class inferno.extensions.layers.ResBlockBase(in_channels, out_channels, dim, size=2, force_skip_op=False, activated=True)

Bases: torch.nn.modules.module.Module

activated_skip_op_factory(in_channels, out_channels)

activation_op_factory(index)

forward(input)

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.

nonactivated_conv_op_factory(in_channels, out_channels, index)

class inferno.extensions.layers.ResBlock(in_channels, out_channels, dim, size=2, activated=True, activation='ReLU', batchnorm=True, force_skip_op=False, conv_kwargs=None)

Bases: inferno.extensions.layers.building_blocks.ResBlockBase

activated_skip_op_factory(in_channels, out_channels)

activation_op_factory(index)

batchnorm_op_factory(in_channels)

nonactivated_conv_op_factory(in_channels, out_channels, index)