espnet2.gan_tts.parallel_wavegan.upsample.Conv2d

source

class espnet2.gan_tts.parallel_wavegan.upsample.Conv2d(*args, **kwargs)

Bases: Conv2d

Conv2d module with customized initialization.

This module extends the standard PyTorch Conv2d layer to include a custom parameter initialization method. The weights are initialized to a constant value based on the kernel size, and the bias (if present) is initialized to zero.

weight

The learnable weights of the module of shape (out_channels, in_channels, kernel_height, kernel_width).

• Type: torch.Tensor

bias

The learnable bias of the module, of shape (out_channels,) if bias is enabled, otherwise None.

• Type: torch.Tensor or None

• Parameters:

  • *args – Variable length argument list for Conv2d.
  • **kwargs – Keyword arguments for Conv2d.

reset_parameters()

Resets the parameters of the Conv2d layer.

####### Examples

>>> conv_layer = Conv2d(in_channels=3, out_channels=16, kernel_size=3)
>>> print(conv_layer.weight)  # Check initialized weights
>>> print(conv_layer.bias)     # Check initialized bias

Initialize Conv2d module.

reset_parameters()

Reset parameters.

This method initializes the weights of the Conv2d module to a constant value based on the kernel size, and sets the bias to zero if it exists. This can be useful to reset the model’s state before training or fine-tuning.

The weights are filled with the value 1.0 / np.prod(self.kernel_size), which ensures that the sum of the weights is normalized to 1 across the kernel.

NOTE

This method is typically called during the initialization of the model or before training to ensure that the weights are set to a known state.

####### Examples

>>> conv_layer = Conv2d(1, 1, kernel_size=(3, 3))
>>> conv_layer.reset_parameters()
>>> print(conv_layer.weight.data)  # Should print the initialized weights

• Raises:None –