espnet2.gan_codec.shared.discriminator.msstft_discriminator.MultiDiscriminator

source

class espnet2.gan_codec.shared.discriminator.msstft_discriminator.MultiDiscriminator

Bases: ABC, Module

Base implementation for discriminators composed of sub-discriminators acting at different scales.

This class serves as a base for implementing multi-scale discriminators that utilize several sub-discriminators to analyze input signals at different frequency scales. It defines the interface that derived classes must implement.

None

• Parameters:None
• Returns: None
• Yields: None
• Raises:
  • NotImplementedError – If the forward method is not implemented in a derived class.
  • NotImplementedError – If the num_discriminators property is not implemented in a derived class.

######### Examples

To create a custom multi-discriminator, derive from this class and implement the forward method and the num_discriminators property:

``

`

python class CustomMultiDiscriminator(MultiDiscriminator):

def forward(self, x: torch.Tensor): : # Custom forward implementation pass

@property def num_discriminators(self) -> int:

return 2 # Example number of discriminators

``

`

####### NOTE This is an abstract class and cannot be instantiated directly.

abstract forward(x: Tensor)

Forward pass for the MultiScaleSTFTDiscriminator.

This method processes the input tensor x through multiple sub-discriminators to produce outputs that capture various frequency and time scales. Each sub-discriminator computes its own feature maps and final logits.

• Parameters:x (torch.Tensor) – Input tensor of shape (B, C, T), where:
  • B: Batch size
  • C: Number of input channels
  • T: Length of the input sequence
• Returns: A list containing feature maps and logits from each sub-discriminator. Each entry in the list corresponds to a sub-discriminator and contains:

  • Feature maps from each layer of the sub-discriminator
  • Final logits from the last layer of the sub-discriminator

• Return type: List[List[torch.Tensor]]

######### Examples

>>> discriminator = MultiScaleSTFTDiscriminator(filters=64)
>>> input_tensor = torch.randn(8, 1, 16000)  # Batch of 8, 1 channel, 16000 samples
>>> output = discriminator(input_tensor)
>>> print(len(output))  # Should print the number of sub-discriminators
>>> print(len(output[0]))  # Should print the number of feature maps + 1 for logits

####### NOTE The input tensor should be preprocessed to match the expected shape before passing it to this method. Each sub-discriminator processes the input independently and the outputs are collected into a list.

• Raises:
  • RuntimeError – If the input tensor does not have the correct number of
  • dimensions or if it contains invalid values. –

abstract property num_discriminators : int

Base implementation for discriminators composed of sub-discriminators acting at different scales.

This class serves as a template for creating multi-scale discriminators, which are composed of several sub-discriminators. Each sub-discriminator operates at a different scale, allowing for a more comprehensive analysis of the input data.

None

• Parameters:None
• Returns: None
• Yields: None
• Raises:
  • NotImplementedError – If the forward method is not implemented.
  • NotImplementedError – If the num_discriminators property is not
  • implemented. –

######### Examples

To create a derived class from MultiDiscriminator, implement the forward method and the num_discriminators property.

class MyDiscriminator(MultiDiscriminator): : def forward(self, x: torch.Tensor): : # Custom forward implementation pass <br/> @property def num_discriminators(self) -> int: <br/>

return 3 # Example number of discriminators

####### NOTE This class cannot be instantiated directly, as it is intended to be a base class for specific multi-discriminator implementations.