Encoder

Convolution Encoder

class ezflow.encoder.conv_encoder.BasicConvEncoder(in_channels=3, config=[64, 128, 256, 512], norm=None)[source]

A Basic Convolution Encoder with a fixed size kernel = 3, padding=1 and dilation = 1. Every alternate layer has stride = 1 followed by stride = 2.

Parameters

  • in_channels (int) – Number of input channels

  • config (list of int) – Configuration for the layers in the encoder

  • norm (str) – Type of normalization to use. Can be None, ‘batch’, ‘group’, ‘instance’

forward(x)[source]

Performs forward pass.

Parameters

x (torch.Tensor) – Input tensor

Returns

List of all the output convolutions from each encoder layer

Return type

List[torch.Tensor],

class ezflow.encoder.conv_encoder.FlowNetConvEncoder(in_channels=3, config=[64, 128, 256, 512], norm=None)[source]

Convolutional encoder based on the FlowNet architecture Used in FlowNet: Learning Optical Flow with Convolutional Networks (https://arxiv.org/abs/1504.06852)

Parameters

  • in_channels (int) – Number of input channels

  • config (list of int) – Configuration for the layers in the encoder

  • norm (str) – Type of normalization to use. Can be None, ‘batch’, ‘group’, ‘instance’

GANet

class ezflow.encoder.ganet.GANetBackbone(in_channels=3, out_channels=32)[source]

Feature extractor backbone used in GA-Net: Guided Aggregation Net for End-to-end Stereo Matching (https://arxiv.org/abs/1904.06587)

Parameters

  • in_channels (int) – Number of input channels

  • out_channels (int) – Number of output channels

forward(x)[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.

PSPNet

class ezflow.encoder.pspnet.PSPNetBackbone(is_proj=True, groups=1, in_channels=3, norm=True)[source]

PSPNet feature extractor backbone (https://arxiv.org/abs/1612.01105) Used in Volumetric Correspondence Networks for Optical Flow (https://papers.nips.cc/paper/2019/hash/bbf94b34eb32268ada57a3be5062fe7d-Abstract.html)

Parameters

  • is_proj (bool) – Whether to use projection pooling or not

  • groups (int) – Number of groups in the convolutional

  • in_channels (int) – Number of input channels

  • norm (bool) – Whether to use batch normalization

forward(x)[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.

class ezflow.encoder.pspnet.PyramidPooling(in_channels, levels=4, norm=True)[source]

Pyramid pooling module for the PSPNet feature extractor

Parameters

  • in_channels (int) – Number of input channels

  • levels (int) – Number of levels in the pyramid

  • norm (bool) – Whether to use batch normalization

forward(x)[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.

class ezflow.encoder.pspnet.ResidualBlock(in_channels, out_channels, stride=1, downsample=None, dilation=1, norm=True)[source]
forward(x)[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.

Pyramid

class ezflow.encoder.pyramid.PyramidEncoder(in_channels=3, config=[16, 32, 64, 96, 128, 196])[source]

Pyramid encoder which returns a hierarchy of features Used in PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume (https://arxiv.org/abs/1709.02371)

Parameters

  • in_channels (int) – Number of input channels

  • config (list of int) – Configuration of the pyramid encoder’s layers

forward(img)[source]

Performs forward pass.

Parameters

img (torch.Tensor) – Input tensor

Returns

List of all the output convolutions from each encoder layer

Return type

List[torch.Tensor],

Residual

class ezflow.encoder.residual.BasicEncoder(in_channels=3, norm='batch', p_dropout=0.0, layer_config=(64, 96, 128), num_residual_layers=(2, 2, 2), intermediate_features=False)[source]

ResNet-style encoder with basic residual blocks

Parameters

  • in_channels (int) – Number of input channels

  • out_channels (int) – Number of output channels

  • norm (str) – Normalization layer to use. One of “batch”, “instance”, “group”, or None

  • p_dropout (float) – Dropout probability

  • layer_config (list of int or tuple of int) – Number of output features per layer

  • num_residual_layers (list of int or tuple of int) – Number of residual blocks features per layer

  • intermediate_features (bool, default False) – Whether to return intermediate features to get a feature hierarchy

forward(x)[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.

class ezflow.encoder.residual.BottleneckEncoder(in_channels=3, norm='batch', p_dropout=0.0, layer_config=(32, 64, 96), num_residual_layers=(2, 2, 2), intermediate_features=False)[source]

ResNet-style encoder with bottleneck residual blocks

Parameters

  • in_channels (int) – Number of input channels

  • out_channels (int) – Number of output channels

  • norm (str) – Normalization layer to use. One of “batch”, “instance”, “group”, or None

  • p_dropout (float) – Dropout probability

  • layer_config (list of int or tuple of int) – Configuration of encoder’s layers

  • intermediate_features (bool, default False) – Whether to return intermediate features to get a feature hierarchy

forward(x)[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.

Builder

ezflow.encoder.build.build_encoder(cfg_grp=None, name=None, instantiate=True, **kwargs)[source]

Build an encoder from a registered encoder name

Parameters

  • cfg (CfgNode) – Config to pass to the encoder

  • name (str) – Name of the registered encoder

  • instantiate (bool) – Whether to instantiate the encoder

  • kwargs (dict) – Additional keyword arguments to pass to the encoder

Returns

The encoder object

Return type

torch.nn.Module