espnet2.gan_tts.parallel_wavegan.upsample.Stretch2d

source

class espnet2.gan_tts.parallel_wavegan.upsample.Stretch2d(x_scale: int, y_scale: int, mode: str = 'nearest')

Bases: Module

Stretch2d module.

This module performs 2D stretching (upsampling) on input tensors, typically used in audio processing applications such as spectrogram manipulation.

The code is modified from https://github.com/kan-bayashi/ParallelWaveGAN.

x_scale

X scaling factor (Time axis in spectrogram).

• Type: int

y_scale

Y scaling factor (Frequency axis in spectrogram).

• Type: int

mode

Interpolation mode for upsampling.

• Type: str

• Parameters:

  • x_scale (int) – X scaling factor (Time axis in spectrogram).
  • y_scale (int) – Y scaling factor (Frequency axis in spectrogram).
  • mode (str) – Interpolation mode.

• Returns: None

####### Examples

>>> import torch
>>> stretch = Stretch2d(x_scale=2, y_scale=3, mode='nearest')
>>> input_tensor = torch.randn(1, 1, 4, 4)  # (B, C, F, T)
>>> output_tensor = stretch(input_tensor)
>>> output_tensor.shape
torch.Size([1, 1, 12, 8])  # (B, C, F * y_scale, T * x_scale)

Initialize Stretch2d module.

• Parameters:
  • x_scale (int) – X scaling factor (Time axis in spectrogram).
  • y_scale (int) – Y scaling factor (Frequency axis in spectrogram).
  • mode (str) – Interpolation mode.

forward(x: Tensor) → Tensor

Calculate forward propagation.

This method performs the forward pass of the Stretch2d module by applying upsampling to the input tensor using the specified scaling factors and interpolation mode.

• Parameters:x (Tensor) – Input tensor of shape (B, C, F, T), where: B - Batch size C - Number of channels F - Frequency bins T - Time steps
• Returns: Interpolated tensor of shape (B, C, F * y_scale, T * x_scale), : where y_scale and x_scale are the scaling factors defined during initialization.
• Return type: Tensor

####### Examples

>>> stretch = Stretch2d(x_scale=2, y_scale=3, mode='linear')
>>> input_tensor = torch.randn(1, 1, 4, 4)  # Example input
>>> output_tensor = stretch(input_tensor)
>>> output_tensor.shape
torch.Size([1, 1, 12, 8])  # Output shape after upsampling