espnet2.diar.attractor.abs_attractor.AbsAttractor

source

class espnet2.diar.attractor.abs_attractor.AbsAttractor(*args, **kwargs)

Bases: Module, ABC

Abstract base class for implementing attractor mechanisms in a neural network.

This class serves as a blueprint for creating specific types of attractors used in speech processing tasks, particularly in the context of speaker diarization. It inherits from torch.nn.Module and requires subclasses to implement the forward method.

None

• Parameters:
  • enc_input (torch.Tensor) – The encoded input tensor, typically containing features from a speech signal.
  • ilens (torch.Tensor) – A tensor containing the lengths of the input sequences.
  • dec_input (torch.Tensor) – The input tensor for the decoder, which may include previous decoder outputs or additional context.
• Returns: A tuple containing: : - output (torch.Tensor): The output tensor from the attractor.
  • attention_weights (torch.Tensor): The attention weights <br/> corresponding to the attractor’s output.
• Return type: Tuple[torch.Tensor, torch.Tensor]
• Raises:
  • NotImplementedError – If the forward method is not implemented in a
  • subclass. –

####### Examples

To create a specific attractor, you would subclass AbsAttractor and implement the forward method as follows:

``

`

python class MyAttractor(AbsAttractor):

def forward(self, enc_input, ilens, dec_input): : # Implement the forward logic here return output, attention_weights

``

`

NOTE

This class is intended to be subclassed. Direct instantiation of AbsAttractor will result in an error due to the unimplemented forward method.

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

abstract forward(enc_input: Tensor, ilens: Tensor, dec_input: Tensor) → Tuple[Tensor, Tensor]

Computes the forward pass for the AbsAttractor model.

This method is an abstract method that must be implemented by any subclass of the AbsAttractor class. It processes the input tensors to produce the output tensors necessary for the model’s operation. The specifics of the computation will depend on the implementation in the derived class.

• Parameters:
  • enc_input (torch.Tensor) – The encoded input tensor, typically from an encoder network.
  • ilens (torch.Tensor) – A tensor representing the lengths of the input sequences. This is used to handle variable-length sequences.
  • dec_input (torch.Tensor) – The input tensor for the decoder, usually containing the previous output tokens in the decoding process.
• Returns: A tuple containing two tensors: : - The first tensor is the output of the model after processing the <br/> input.
  • The second tensor may represent additional information, such as hidden states or attention weights.
• Return type: Tuple[torch.Tensor, torch.Tensor]
• Raises:NotImplementedError – This method must be implemented by subclasses of AbsAttractor.

####### Examples

class MyAttractor(AbsAttractor): : def forward(self, enc_input, ilens, dec_input): : # Implement specific forward pass logic here return output, additional_info

model = MyAttractor() output, info = model(enc_input, ilens, dec_input)

NOTE

This method is intended to be overridden in derived classes to provide specific functionality for different types of attractors.