espnet2.enh.loss.wrappers.fixed_order.FixedOrderSolver

source

class espnet2.enh.loss.wrappers.fixed_order.FixedOrderSolver(criterion: AbsEnhLoss, weight=1.0)

Bases: AbsLossWrapper

FixedOrderSolver is a wrapper class that implements a naive fixed-order

solver for computing the loss between reference and inferred signals.

This class extends the AbsLossWrapper and utilizes a specified loss criterion to calculate the minimum loss with the best permutation of the input signals.

criterion

The loss criterion used for calculating the loss between reference and inferred signals.

• Type:AbsEnhLoss

weight

A scaling factor for the computed loss.

• Type: float

• Parameters:

  • criterion (AbsEnhLoss) – The loss criterion to be used for the solver.
  • weight (float , optional) – A weight factor for the loss. Defaults to 1.0.

• Returns: A tuple containing: : - loss (torch.Tensor): The minimum loss with the best permutation.

  • stats (dict): A dictionary containing training status statistics.
  • others (dict): Reserved for additional information, includes the permutation used.

• Return type: Tuple[torch.Tensor, dict, dict]

• Raises:AssertionError – If the length of reference and inferred lists do not match.

####### Examples

>>> solver = FixedOrderSolver(criterion=my_loss_criterion, weight=1.0)
>>> ref_signals = [torch.randn(10, 1) for _ in range(2)]  # Two speakers
>>> inf_signals = [torch.randn(10, 1) for _ in range(2)]
>>> loss, stats, others = solver.forward(ref_signals, inf_signals)
>>> print(loss)  # Outputs the computed loss
>>> print(stats)  # Outputs training statistics
>>> print(others)  # Outputs the permutation used

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

forward(ref, inf, others={})

An implementation of a naive fixed-order solver for loss computation in

speech enhancement.

This class inherits from AbsLossWrapper and implements the forward method, which computes the loss based on the reference and inferred signals, considering a fixed order of speakers.

criterion

The loss criterion used for computing the loss.

• Type:AbsEnhLoss

weight

Weighting factor for the loss.

• Type: float

• Parameters:

  • ref (List *[*torch.Tensor ]) – A list of reference tensors, each of shape (batch, …), where n_spk is the number of speakers.
  • inf (List *[*torch.Tensor ]) – A list of inferred tensors, each of shape (batch, …).
  • others (dict , optional) – Reserved for additional parameters (default: {}).

• Returns: A tuple containing: : - loss (torch.Tensor): The minimum loss with the best permutation.

  • stats (dict): A dictionary collecting training statistics.
  • others (dict): A dictionary containing reserved information, including the speaker permutation.

• Return type: Tuple[torch.Tensor, dict, dict]

• Raises:AssertionError – If the lengths of ref and inf do not match.

####### Examples

>>> solver = FixedOrderSolver(criterion=my_criterion)
>>> ref_signals = [torch.randn(10, 2) for _ in range(3)]
>>> inf_signals = [torch.randn(10, 2) for _ in range(3)]
>>> loss, stats, others = solver.forward(ref_signals, inf_signals)

NOTE

This method assumes that the number of reference signals matches the number of inferred signals and that they are arranged in a specific order corresponding to speakers.