espnet2.gan_tts.vits.flow.FlipFlow

source

class espnet2.gan_tts.vits.flow.FlipFlow(*args, **kwargs)

Bases: Module

FlipFlow is a flip flow module that applies a simple flipping operation on

the input tensor during forward propagation. This module is part of the VITS architecture used for generative adversarial networks in text-to-speech tasks.

None

• Parameters:
  • x (torch.Tensor) – Input tensor of shape (B, channels, T).
  • inverse (bool) – Whether to perform the inverse operation. Default is False.
• Returns:
  • If inverse is False: : - Tensor: Flipped tensor of shape (B, channels, T).
    • Tensor: Log-determinant tensor for negative log-likelihood (NLL) of shape (B,).
  • If inverse is True: : - Tensor: The original input tensor after flipping.
• Return type: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]

####### Examples

>>> flip_flow = FlipFlow()
>>> x = torch.randn(4, 3, 5)  # Example input tensor
>>> flipped, logdet = flip_flow(x)  # Forward operation
>>> original = flip_flow(x, inverse=True)  # Inverse operation

NOTE

This module assumes that the input tensor has at least 3 dimensions corresponding to the batch size, number of channels, and sequence length.

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

forward(x: Tensor, *args, inverse: bool = False, **kwargs) → Tensor | Tuple[Tensor, Tensor]

Calculate forward propagation.

This method performs the forward pass of the FlipFlow module, which applies a flipping transformation to the input tensor. It can also perform the inverse operation based on the inverse flag.

• Parameters:
  • x (torch.Tensor) – Input tensor of shape (B, channels, T).
  • *args – Additional positional arguments (unused).
  • inverse (bool , optional) – Whether to inverse the flow. Defaults to False.
  • **kwargs – Additional keyword arguments (unused).
• Returns:
  • If inverse is False, returns a tuple containing: : - Flipped tensor of shape (B, channels, T).
    • Log-determinant tensor for negative log-likelihood (NLL) of shape (B,).
  • If inverse is True, returns the tensor after applying the inverse transformation.
• Return type: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]

####### Examples

>>> flip_flow = FlipFlow()
>>> input_tensor = torch.randn(5, 3, 10)  # Example input
>>> output_tensor, logdet = flip_flow(input_tensor)
>>> inverted_tensor = flip_flow(input_tensor, inverse=True)