espnet2.tts.prodiff.denoiser.ResidualBlock

source

class espnet2.tts.prodiff.denoiser.ResidualBlock(adim: int, channels: int, dilation: int)

Bases: Module

Residual Block for Diffusion Denoiser.

This class implements a residual block used in the diffusion denoiser architecture. It consists of a convolutional layer, linear projection, and several activation functions, facilitating the transformation of input tensors through skip connections and conditioning inputs.

conv

1D convolutional layer for processing input channels.

• Type: nn.Conv1d

diff_proj

Linear projection for diffusion step.

• Type: nn.Linear

cond_proj

Convolutional layer for conditioning inputs.

• Type: nn.Conv1d

out_proj

Convolutional layer for output processing.

• Type: nn.Conv1d

• Parameters:

  • adim (int) – Size of dimensions.
  • channels (int) – Number of channels.
  • dilation (int) – Size of dilations.

• Returns: Tuple of output tensor and skip connection.

• Return type: Union[torch.Tensor, torch.Tensor]

####### Examples

>>> block = ResidualBlock(adim=256, channels=64, dilation=2)
>>> x = torch.randn(1, 64, 100)  # Input tensor
>>> condition = torch.randn(1, 256, 100)  # Conditioning tensor
>>> step = torch.randn(1, 256)  # Diffusion step
>>> output, skip = block(x, condition, step)
>>> print(output.shape, skip.shape)
(torch.Size([1, 64, 100]), torch.Size([1, 128, 100]))

Initialization.

• Parameters:
  • adim (int) – Size of dimensions.
  • channels (int) – Number of channels.
  • dilation (int) – Size of dilations.

forward(x: Tensor, condition: Tensor, step: Tensor) → Tensor

Calculate forward propagation.

This method performs the forward pass of the Residual Block, incorporating the input tensor, conditioning tensor, and the diffusion step tensor. The output consists of a residual and a skip connection.

• Parameters:
  • x (torch.Tensor) – Input tensor.
  • condition (torch.Tensor) – Conditioning tensor.
  • step (torch.Tensor) – Number of diffusion step.
• Returns: A tuple containing: : - Output tensor (residual)
  • Skip connection tensor.
• Return type: Union[torch.Tensor, torch.Tensor]

####### Examples

>>> block = ResidualBlock(adim=128, channels=256, dilation=2)
>>> x = torch.randn(32, 256, 100)  # Batch of 32, 256 channels, 100 time steps
>>> condition = torch.randn(32, 128, 100)  # Conditioning tensor
>>> step = torch.randn(32, 1)  # Diffusion step
>>> output, skip = block(x, condition, step)
>>> output.shape
torch.Size([32, 256, 100])
>>> skip.shape
torch.Size([32, 256, 100])