espnet2.enh.layers.ncsnpp_utils.normalization.ConditionalInstanceNorm2dPlus

source

class espnet2.enh.layers.ncsnpp_utils.normalization.ConditionalInstanceNorm2dPlus(num_features, num_classes, bias=True)

Bases: Module

Conditional Instance Normalization layer with additional scaling and bias.

This class implements a conditional instance normalization layer that normalizes input features using learned parameters based on the provided class labels. The layer includes the capability to apply an additional scaling factor and a bias term for more flexible output.

num_features

The number of input features (channels).

• Type: int

bias

Indicates whether to include a bias term in the layer.

• Type: bool

instance_norm

Instance normalization module.

• Type: nn.InstanceNorm2d

embed

Embedding layer for learning scaling and bias.

• Type: nn.Embedding

• Parameters:

  • num_features (int) – Number of input features (channels).
  • num_classes (int) – Number of classes for the conditional embedding.
  • bias (bool) – If True, includes a bias term (default: True).

• Returns: Normalized output tensor.

• Return type: Tensor

####### Examples

>>> layer = ConditionalInstanceNorm2dPlus(num_features=64, num_classes=10)
>>> x = torch.randn(8, 64, 32, 32)  # Batch of 8 images
>>> y = torch.randint(0, 10, (8,))  # Random class labels for the batch
>>> output = layer(x, y)

NOTE

The input tensor x is expected to have the shape (batch_size, num_features, height, width). The class labels y should have the shape (batch_size,).

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

forward(x, y)

Applies the Conditional Instance Normalization operation.

This method performs Conditional Instance Normalization on the input tensor x based on the class label y. The operation normalizes the input tensor using the instance normalization technique and applies learnable scaling and shifting parameters based on the class embedding.

• Parameters:
  • x (torch.Tensor) – The input tensor of shape (N, C, H, W) where:
    • N is the batch size,
    • C is the number of channels,
    • H is the height,
    • W is the width.
  • y (torch.Tensor) – The class label tensor of shape (N,) where each element corresponds to the class index for the respective input in x.
• Returns: The output tensor after applying Conditional Instance Normalization, with the same shape as the input tensor x.
• Return type: torch.Tensor

####### Examples

>>> norm_layer = ConditionalInstanceNorm2dPlus(num_features=64, num_classes=10)
>>> input_tensor = torch.randn(8, 64, 32, 32)  # Example input
>>> class_labels = torch.randint(0, 10, (8,))  # Random class labels
>>> output_tensor = norm_layer(input_tensor, class_labels)
>>> print(output_tensor.shape)  # Should be [8, 64, 32, 32]

NOTE

The input tensor x must have 4 dimensions, and the class label y must be a 1-dimensional tensor with class indices ranging from 0 to num_classes - 1.

• Raises:
  • IndexError – If the class index in y is out of bounds for the embedding
  • layer. –