neural_de.external.derain package

Submodules

neural_de.external.derain.blocks module

2D convolution class

class neural_de.external.derain.blocks.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, activation_func=LeakyReLU(negative_slope=0.1, inplace=True), norm_layer=<class 'torch.nn.modules.batchnorm.BatchNorm2d'>, use_bias=False, padding_type='reflect')

Bases: Module

2D convolution class Args: in_channels : int - Number of input channels out_channels : int - Number of output channels kernel_size : int - Size of kernel stride : int - Stride of convolution activation_func : func - Activation function after convolution norm_layer : functools.partial - Normalization layer use_bias : bool - If set, then use bias padding_type : str - The name of padding layer: reflect | replicate | zero

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.

training: bool

class neural_de.external.derain.blocks.DecoderBlock(in_channels, skip_channels, out_channels, activation_func=LeakyReLU(negative_slope=0.1, inplace=True), norm_layer=<class 'torch.nn.modules.batchnorm.BatchNorm2d'>, use_bias=False, padding_type='reflect', upsample_mode='transpose')

Bases: Module

Decoder block with skip connections Args: in_channels : int - number of input channels skip_channels : int - number of skip connection channels out_channels : int - number of output channels activation_func : func - activation function after convolution norm_layer : functools.partial - normalization layer use_bias : bool - if set, then use bias padding_type : str - the name of padding layer: reflect | replicate | zero upsample_mode : str - the mode for interpolation: transpose | bilinear | nearest

forward(x, skip=None)

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 neural_de.external.derain.blocks.DeformableConv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)

Bases: Module

2D deformable convolution class :type in_channels: :param in_channels: int - number of input channels :type out_channels: :param out_channels: int - number of output channels :type kernel_size: :param kernel_size: int - size of kernel :type stride: :param stride: int - stride of convolution :type padding: :param padding: int - padding :param use_bias: bool - if set, then use bias

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.

training: bool

class neural_de.external.derain.blocks.DeformableResnetBlock(dim, padding_type, norm_layer, use_dropout, use_bias, activation_func)

Bases: Module

Define a Resnet block with deformable convolutions

build_conv_block(dim, padding_type, norm_layer, use_dropout, use_bias, activation_func)

Construct a convolutional block. :type dim: :param dim: :type dim: int :type padding_type: :param padding_type: reflect | replicate | zero :type padding_type: str :param norm_layer – normalization layer: :type use_dropout: :param use_dropout: :type use_dropout: bool :type use_bias: :param use_bias: :type use_bias: bool :type activation_func: :param activation_func: :type activation_func: func

Returns a conv block (with a conv layer, a normalization layer, and a non-linearity layer)

forward(x)

Forward function (with skip connections)

training: bool

class neural_de.external.derain.blocks.ResNetModified(input_nc, output_nc, ngf=64, norm_layer=<class 'torch.nn.modules.batchnorm.BatchNorm2d'>, activation_func=LeakyReLU(negative_slope=0.1, inplace=True), use_dropout=False, n_blocks=6, padding_type='reflect', upsample_mode='bilinear')

Bases: Module

Resnet-based generator that consists of deformable Resnet blocks.

forward(input)

Standard forward

training: bool

class neural_de.external.derain.blocks.UpConv2d(in_channels, out_channels, kernel_size=3, activation_func=LeakyReLU(negative_slope=0.1, inplace=True), norm_layer=<class 'torch.nn.modules.batchnorm.BatchNorm2d'>, use_bias=False, padding_type='reflect', interpolate_mode='bilinear')

Bases: Module

Up-convolution (upsample + convolution) block class Args: in_channels : int - number of input channels out_channels : int - number of output channels kernel_size : int - size of kernel (k x k) activation_func : func - activation function after convolution norm_layer : functools.partial - normalization layer use_bias : bool - if set, then use bias padding_type : str - the name of padding layer: reflect | replicate | zero interpolate_mode : str - the mode for interpolation: bilinear | nearest

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.

training: bool

Module contents

Code adapted from the Gt-Rain implementation: https://github.com/UCLA-VMG/GT-RAIN .