espnet2.enh.loss.criterions.tf_domain.FrequencyDomainLoss

source

class espnet2.enh.loss.criterions.tf_domain.FrequencyDomainLoss(name, only_for_test=False, is_noise_loss=False, is_dereverb_loss=False)

Bases: AbsEnhLoss, ABC

Base class for all frequency-domain enhancement loss modules.

This abstract class provides a structure for defining various types of frequency-domain loss functions used in audio enhancement tasks. Derived classes must implement specific loss computations and define the mask type used for those computations.

compute_on_mask

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

• Type: bool

mask_type

The type of mask used in loss computation.

• Type: str

name

The name of the loss function.

• Type: str

only_for_test

If True, this loss is only used during testing.

• Type: bool

is_noise_loss

If True, this loss is related to noise.

• Type: bool

is_dereverb_loss

If True, this loss is related to dereverberation.

• Type: bool

• Parameters:

  • name (str) – The name of the loss function.
  • only_for_test (bool , optional) – Whether the loss is only for testing. Defaults to False.
  • is_noise_loss (bool , optional) – Whether the loss is related to noise. Defaults to False.
  • is_dereverb_loss (bool , optional) – Whether the loss is related to dereverberation. Defaults to False.

• Raises:ValueError – If both is_noise_loss and is_dereverb_loss are True.

####### Examples

>>> loss = FrequencyDomainLoss(name="MyLoss", is_noise_loss=True)
>>> print(loss.name)
MyLoss

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

abstract property compute_on_mask : bool

create_mask_label(mix_spec, ref_spec, noise_spec=None)

Create a mask label based on the provided spectrograms.

This method generates a mask label for the input mixed spectrogram and reference spectrograms based on the specified mask type. It utilizes the _create_mask_label helper function to compute the mask.

• Parameters:
  • mix_spec (ComplexTensor) – The mixed spectrogram of shape (B, T, [C,] F).
  • ref_spec (List *[*ComplexTensor ]) – A list of reference spectrograms of shape (B, T, [C,] F) for each reference signal.
  • noise_spec (ComplexTensor , optional) – The noise spectrogram of shape (B, T, [C,] F). This is only used for IBM and IRM masks. Defaults to None.
• Returns: A list of masks of shape : (B, T, [C,] F) or (B, T, F) depending on the mask type.
• Return type: List[Tensor] or List[ComplexTensor]

####### Examples

>>> mix = torch.randn(4, 256, 1, 512)  # Mixed spectrogram
>>> ref = [torch.randn(4, 256, 1, 512) for _ in range(2)]  # Two refs
>>> masks = create_mask_label(mix, ref)
>>> len(masks)  # Should return 2 (one for each reference)

NOTE

Ensure that the mask_type attribute is set correctly in the class to use the desired masking method.

• Raises:AssertionError – If the mask_type is not supported.

property is_dereverb_loss : bool

property is_noise_loss : bool

abstract property mask_type : str

property name : str

property only_for_test : bool