espnet2.gan_tts.parallel_wavegan.upsample.UpsampleNetwork

source

class espnet2.gan_tts.parallel_wavegan.upsample.UpsampleNetwork(upsample_scales: List[int], nonlinear_activation: str | None = None, nonlinear_activation_params: Dict[str, Any] = {}, interpolate_mode: str = 'nearest', freq_axis_kernel_size: int = 1)

Bases: Module

Upsampling network module.

This module performs upsampling on input tensors through a series of interpolation and convolution layers. It allows for customizable non-linear activation functions and scaling factors.

• Parameters:
  • upsample_scales (List *[*int ]) – List of upsampling scales.
  • nonlinear_activation (Optional *[*str ]) – Activation function name.
  • nonlinear_activation_params (Dict *[*str , Any ]) – Arguments for the specified activation function.
  • interpolate_mode (str) – Interpolation mode for upsampling.
  • freq_axis_kernel_size (int) – Kernel size in the direction of frequency axis.

####### Examples

>>> upsample_network = UpsampleNetwork(
...     upsample_scales=[2, 2],
...     nonlinear_activation='ReLU',
...     nonlinear_activation_params={},
...     interpolate_mode='nearest',
...     freq_axis_kernel_size=3
... )
>>> input_tensor = torch.randn(1, 80, 50)  # (B, C, T_feats)
>>> output_tensor = upsample_network(input_tensor)
>>> output_tensor.shape
torch.Size([1, 80, 200])  # (B, C, T_wav)

Initialize UpsampleNetwork module.

• Parameters:
  • upsample_scales (List *[*int ]) – List of upsampling scales.
  • nonlinear_activation (Optional *[*str ]) – Activation function name.
  • nonlinear_activation_params (Dict *[*str , Any ]) – Arguments for the specified activation function.
  • interpolate_mode (str) – Interpolation mode.
  • freq_axis_kernel_size (int) – Kernel size in the direction of frequency axis.

forward(c: Tensor) → Tensor

Calculate forward propagation.

This method processes the input tensor through a series of upsampling layers and returns the upsampled tensor.

• Parameters:c (Tensor) – Input tensor of shape (B, C, T_feats), where:
  • B: Batch size
  • C: Number of channels
  • T_feats: Number of feature dimensions
• Returns: Upsampled tensor of shape (B, C, T_wav), where: : - T_wav = T_feats * prod(upsample_scales), representing the total length after upsampling.
• Return type: Tensor

####### Examples

>>> model = UpsampleNetwork(upsample_scales=[2, 2])
>>> input_tensor = torch.randn(1, 10, 5)  # Example input
>>> output_tensor = model(input_tensor)
>>> output_tensor.shape
torch.Size([1, 10, 20])  # Example output shape after upsampling