espnet2.diar.label_processor.LabelProcessor

source

class espnet2.diar.label_processor.LabelProcessor(win_length: int = 512, hop_length: int = 128, center: bool = True)

Bases: Module

LabelProcessor is a PyTorch module that aggregates labels for speaker diarization tasks.

This class utilizes the LabelAggregate layer to perform label aggregation over specified window and hop lengths. It is designed to process input label tensors and their corresponding lengths, outputting aggregated labels along with their lengths.

label_aggregator

An instance of the LabelAggregate

• Type:LabelAggregate

class responsible for aggregating labels based on specified parameters.

• Parameters:
  • win_length (int) – The length of the window for aggregation. Defaults to 512.
  • hop_length (int) – The hop length for the sliding window. Defaults to 128.
  • center (bool) – Whether to center the window. Defaults to True.
• Returns: A tuple containing: : - output (torch.Tensor): The aggregated label tensor of shape <br/> (Batch, Frames, Label_dim).
  • olens (torch.Tensor): The lengths of the output sequences of shape (Batch).
• Return type: Tuple[torch.Tensor, torch.Tensor]

####### Examples

>>> label_processor = LabelProcessor(win_length=256, hop_length=64)
>>> input_tensor = torch.randn(10, 1000, 20)  # Batch of 10, 1000 samples, 20 labels
>>> ilens_tensor = torch.tensor([1000] * 10)  # Lengths for each batch
>>> output, olens = label_processor(input_tensor, ilens_tensor)
>>> print(output.shape)  # Should output (10, Frames, 20)
>>> print(olens.shape)   # Should output (10,)

NOTE

This module is specifically designed for use in speaker diarization systems, and the choice of window and hop lengths can significantly affect the performance of the aggregation.

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

forward(input: Tensor, ilens: Tensor)

Forward pass for the LabelProcessor module.

This method processes the input tensor containing labels and their respective lengths. It utilizes the label aggregator to produce an output tensor that represents the aggregated labels over the specified window and hop lengths.

• Parameters:
  • input (torch.Tensor) – A tensor of shape (Batch, Nsamples, Label_dim) representing the input labels for each sample in the batch.
  • ilens (torch.Tensor) – A tensor of shape (Batch) containing the lengths of each input sequence in the batch.
• Returns: A tuple containing: : - output (torch.Tensor): A tensor of shape (Batch, Frames, Label_dim) representing the aggregated output labels.
  • olens (torch.Tensor): A tensor of shape (Batch) containing the lengths of the output sequences.
• Return type: Tuple[torch.Tensor, torch.Tensor]

####### Examples

>>> label_processor = LabelProcessor(win_length=512, hop_length=128)
>>> input_tensor = torch.randn(2, 1000, 10)  # Batch of 2, 1000 samples, 10 labels
>>> ilens_tensor = torch.tensor([1000, 800])  # Lengths of each input
>>> output, olens = label_processor(input_tensor, ilens_tensor)
>>> print(output.shape)  # Output shape will depend on aggregation
>>> print(olens)  # Output lengths for each batch

NOTE

Ensure that the input tensor and ilens are correctly formatted according to the expected shapes.

• Raises:
  • ValueError – If the input tensor dimensions do not match the expected
  • shape or if ilens contains invalid lengths. –