espnet2.gan_tts.vits.flow.LogFlow

source

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

Bases: Module

LogFlow module for calculating forward propagation in flow-based models.

This class implements a log flow transformation as part of the Variational Inference Text-to-Speech (VITS) model. It computes the log of the input tensor while applying a mask and calculates the log-determinant necessary for negative log-likelihood (NLL) computation.

• Parameters:
  • x (Tensor) – Input tensor of shape (B, channels, T).
  • x_mask (Tensor) – Mask tensor of shape (B, 1, T).
  • inverse (bool) – Whether to compute the inverse of the flow.
  • eps (float) – A small value to prevent log(0). Default is 1e-5.
• Returns:
  • If inverse is False: : - Output tensor of shape (B, channels, T).
    • Log-determinant tensor for NLL of shape (B,).
  • If inverse is True: : - Output tensor of shape (B, channels, T).
• Return type: Union[Tensor, Tuple[Tensor, Tensor]]

Examples

>>> log_flow = LogFlow()
>>> x = torch.tensor([[[0.1, 0.2, 0.3]]])  # Shape: (1, 1, 3)
>>> x_mask = torch.tensor([[[1, 1, 1]]])  # Shape: (1, 1, 3)
>>> y, logdet = log_flow(x, x_mask)  # Forward propagation
>>> x_inv = log_flow(y, x_mask, inverse=True)  # Inverse propagation

NOTE

The log transformation applied here is element-wise, and the log-determinant is computed by summing the negative log of the transformed input tensor over the time and channel dimensions.

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

forward(x: Tensor, x_mask: Tensor, inverse: bool = False, eps: float = 1e-05, **kwargs) → Tensor | Tuple[Tensor, Tensor]

Calculate forward propagation.

• Parameters:
  • x (Tensor) – Input tensor (B, channels, T).
  • x_mask (Tensor) – Mask tensor (B, 1, T).
  • inverse (bool) – Whether to inverse the flow.
  • eps (float) – Epsilon for log.
• Returns: Output tensor (B, channels, T). Tensor: Log-determinant tensor for NLL (B,) if not inverse.
• Return type: Tensor