espnet2.enh.loss.criterions.tf_domain.FrequencyDomainCrossEntropy

source

class espnet2.enh.loss.criterions.tf_domain.FrequencyDomainCrossEntropy(compute_on_mask=False, mask_type=None, ignore_id=-100, name=None, only_for_test=False, is_noise_loss=False, is_dereverb_loss=False)

Bases: FrequencyDomainLoss

FrequencyDomainCrossEntropy computes the cross-entropy loss in the frequency

domain for audio signal processing tasks.

This loss is used to evaluate the difference between the predicted and reference distributions in a frequency domain representation, which is useful in applications such as speech enhancement and separation.

compute_on_mask

Indicates whether the loss is computed on the mask.

• Type: bool

mask_type

Type of the mask being used for computation.

• Type: str

ignore_id

ID to ignore in the cross-entropy calculation.

• Type: int

• Parameters:

  • compute_on_mask (bool , optional) – If True, the loss is computed on the mask. Defaults to False.
  • mask_type (str , optional) – Type of mask. Defaults to None.
  • ignore_id (int , optional) – ID to ignore in the loss computation. Defaults to -100.
  • name (str , optional) – Name of the loss instance. Defaults to None.
  • only_for_test (bool , optional) – If True, the loss is only for testing. Defaults to False.
  • is_noise_loss (bool , optional) – If True, the loss is related to noise. Defaults to False.
  • is_dereverb_loss (bool , optional) – If True, the loss is related to dereverberation. Defaults to False.

• Returns: Computed loss for each batch.

• Return type: torch.Tensor

####### Examples

>>> loss_fn = FrequencyDomainCrossEntropy()
>>> ref = torch.tensor([[0, 1, 2], [1, 2, -100]])  # Reference labels
>>> inf = torch.rand(2, 3, 4)  # Predicted logits (Batch, T, nclass)
>>> loss = loss_fn(ref, inf)
>>> print(loss)

• Raises:ValueError – If the input shapes of ref and inf do not match or are invalid.

NOTE

• The input ref should be of shape (Batch, T) or (Batch, T, C) where T is the time dimension.
• The input inf should be of shape (Batch, T, nclass) or (Batch, T, C, nclass).
• The ignore_id can be used to exclude certain labels from the loss computation.

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

property compute_on_mask : bool

forward(ref, inf) → Tensor

Computes the time-frequency cross-entropy loss for audio enhancement tasks.

This loss function is designed to compare a reference tensor with an inference tensor, both of which represent audio data in a time-frequency domain. The cross-entropy loss is useful in tasks such as speech separation and enhancement, where the goal is to classify or predict the presence of certain audio signals.

compute_on_mask

Indicates whether the loss is computed on the mask or the spectrum.

• Type: bool

mask_type

The type of mask being used (e.g., “IBM”, “IRM”).

• Type: str

ignore_id

The label ID to ignore during loss computation.

• Type: int

• Parameters:

  • compute_on_mask (bool) – If True, the loss will be computed on the mask; otherwise, it will be computed on the spectrum.
  • mask_type (str , optional) – The type of mask to use. Defaults to None.
  • ignore_id (int , optional) – The label ID to ignore. Defaults to -100.
  • name (str , optional) – The name of the loss function. Defaults to None.
  • only_for_test (bool , optional) – If True, indicates the loss is only for testing. Defaults to False.
  • is_noise_loss (bool , optional) – If True, indicates the loss is related to noise. Defaults to False.
  • is_dereverb_loss (bool , optional) – If True, indicates the loss is related to dereverberation. Defaults to False.

• Returns: The computed loss for the batch, shape (Batch,).

• Return type: torch.Tensor

• Raises:ValueError – If the shapes of ref and inf are not compatible or if the input dimensions are invalid.

####### Examples

>>> loss_fn = FrequencyDomainCrossEntropy()
>>> ref = torch.tensor([[1, 0], [0, 1]])  # Reference tensor
>>> inf = torch.tensor([[[0.1, 0.9], [0.8, 0.2]],
...                      [[0.7, 0.3], [0.2, 0.8]]])  # Inference tensor
>>> loss = loss_fn(ref, inf)  # Compute loss
>>> print(loss)  # Output the loss value

NOTE

Ensure that the input tensors ref and inf have compatible shapes as per the specifications mentioned in the Args section.

property mask_type : str