espnet2.enh.layers.ncsnpp_utils.layers.RefineBlock

source

class espnet2.enh.layers.ncsnpp_utils.layers.RefineBlock(in_planes, features, act=ReLU(), start=False, end=False, maxpool=True)

Bases: Module

Refinement block for feature extraction and processing.

This block is designed to adaptively refine features from multiple input sources using Residual Convolutional Units (RCU), a Multi-Scale Feature (MSF) block, and a Conditional Residual Processing (CRP) block. It allows for the combination of features from different resolutions and enhances them through a series of convolutions.

adapt_convs

List of RCU blocks for adapting input features.

• Type: nn.ModuleList

output_convs

RCU block for generating output features.

• Type:RCUBlock

msf

Multi-Scale Feature block for combining features.

• Type:MSFBlock

crp

Conditional Residual Processing block for refining features.

• Type:CRPBlock

• Parameters:

  • in_planes (tuple or list) – Number of input feature planes for each block.
  • features (int) – Number of output feature planes for the final output.
  • act (callable) – Activation function to use (default: ReLU).
  • start (bool) – Flag indicating if this block is the starting block (default: False).
  • end (bool) – Flag indicating if this block is the ending block (default: False).
  • maxpool (bool) – Flag indicating if max pooling should be used (default: True).

• Raises:AssertionError – If in_planes is not a tuple or list.

####### Examples

>>> refine_block = RefineBlock((64, 128), 256)
>>> output = refine_block((input_tensor1, input_tensor2), output_shape)

NOTE

The start and end flags can be used to customize the behavior of the block based on its position in the overall network architecture.

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

forward(xs, output_shape)

RefineBlock for hierarchical feature refinement in deep networks.

This module refines input features from multiple sources through a series of convolutional layers, pooling operations, and residual connections. It is designed to adaptively fuse features from various input resolutions, enhancing the network’s ability to capture complex patterns.

n_blocks

Number of input feature blocks.

• Type: int

adapt_convs

List of Residual Convolutional Units (RCUs) for adapting input features.

• Type: nn.ModuleList

output_convs

RCU for producing output features.

• Type:RCUBlock

msf

Multi-scale feature fusion block, if applicable.

• Type:MSFBlock

crp

Contextual Refinement Pooling block.

• Type:CRPBlock

• Parameters:

  • in_planes (list or tuple) – A list or tuple of input feature dimensions.
  • features (int) – Number of output features for the final layer.
  • act (callable) – Activation function to use (default: nn.ReLU).
  • start (bool) – Whether this block is at the start of the network (default: False).
  • end (bool) – Whether this block is at the end of the network (default: False).
  • maxpool (bool) – Whether to use max pooling in CRPBlock (default: True).

####### Examples

>>> refine_block = RefineBlock(in_planes=(64, 128), features=256)
>>> input_features = (torch.randn(1, 64, 32, 32), torch.randn(1, 128, 16, 16))
>>> output = refine_block(input_features, output_shape=(16, 16))
>>> print(output.shape)
torch.Size([1, 256, 16, 16])