espnet2.enh.layers.ncsnpp_utils.layers.CondCRPBlock

source

class espnet2.enh.layers.ncsnpp_utils.layers.CondCRPBlock(features, n_stages, num_classes, normalizer, act=ReLU())

Bases: Module

Conditional Residual Processing Block.

This class implements a conditional residual processing block that applies a series of convolutional layers, along with conditional normalization and pooling operations. It is designed to facilitate enhanced feature extraction and integration for tasks requiring class-specific adaptations.

convs

List of convolutional layers.

• Type: nn.ModuleList

norms

List of normalization layers.

• Type: nn.ModuleList

normalizer

Function used for normalization.

• Type: callable

n_stages

Number of stages in the block.

• Type: int

pool

Average pooling layer.

• Type: nn.AvgPool2d

act

Activation function applied in the block.

• Type: callable

• Parameters:

  • features (int) – Number of input and output features for the convolutional layers.
  • n_stages (int) – Number of stages (layers) to be applied in the block.
  • num_classes (int) – Number of classes for conditional normalization.
  • normalizer (callable) – Normalization function to be applied to the input.
  • act (callable , optional) – Activation function (default is nn.ReLU()).

####### Examples

>>> features = 64
>>> n_stages = 3
>>> num_classes = 10
>>> normalizer = ConditionalInstanceNorm2dPlus
>>> block = CondCRPBlock(features, n_stages, num_classes, normalizer)
>>> x = torch.randn(1, 64, 32, 32)  # Example input
>>> y = torch.randint(0, num_classes, (1,))  # Example class input
>>> output = block(x, y)

Forward: : The forward method takes an input tensor x and a conditional input y, applies the activation function, processes the input through the normalization and convolution layers, and returns the output tensor.

• Raises:
  • ValueError – If the input tensor dimensions do not match the expected
  • dimensions. –

NOTE

Ensure that the input tensor x and the conditional tensor y are appropriately shaped to match the model’s expectations.

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

forward(x, y)

Forward pass through the Conditional CRP Block.

This method applies a series of conditional normalization and convolution operations to the input tensor x, conditioned on the input tensor y. The operations are repeated over a specified number of stages, with intermediate pooling applied at each stage.

• Parameters:
  • x (torch.Tensor) – The input tensor of shape (batch_size, num_features, height, width).
  • y (torch.Tensor) – The conditional input tensor of shape (batch_size, num_classes).
• Returns: The output tensor of the same shape as x, which incorporates the results of the convolutional operations.
• Return type: torch.Tensor

####### Examples

>>> cond_crp_block = CondCRPBlock(features=64, n_stages=3,
...                                 num_classes=10,
...                                 normalizer=ConditionalInstanceNorm2dPlus)
>>> x = torch.randn(8, 64, 32, 32)  # Batch of 8, 64 features
>>> y = torch.randint(0, 10, (8,))   # Random class labels
>>> output = cond_crp_block(x, y)
>>> output.shape
torch.Size([8, 64, 32, 32])