asteroid.losses.multi_scale_spectral module

class asteroid.losses.multi_scale_spectral.SingleSrcMultiScaleSpectral(n_filters=None, windows_size=None, hops_size=None, alpha=1.0)

Bases: sphinx.ext.autodoc.importer._MockObject

Measure multi-scale spectral loss as described in [1]

Parameters:

• n_filters (list) – list containing the number of filter desired for each STFT
• windows_size (list) – list containing the size of the window desired for each STFT
• hops_size (list) – list containing the size of the hop desired for each STFT

Shape:

est_targets (torch.Tensor): Expected shape [batch, time].

Batch of target estimates.

targets (torch.Tensor): Expected shape [batch, time].

Batch of training targets.

alpha (float) : Weighting factor for the log term

Returns:torch.Tensor – with shape [batch]

Examples

>>> import torch
>>> targets = torch.randn(10, 32000)
>>> est_targets = torch.randn(10, 32000)
>>> # Using it by itself on a pair of source/estimate
>>> loss_func = SingleSrcMultiScaleSpectral()
>>> loss = loss_func(est_targets, targets)

>>> import torch
>>> from asteroid.losses import PITLossWrapper
>>> targets = torch.randn(10, 2, 32000)
>>> est_targets = torch.randn(10, 2, 32000)
>>> # Using it with PITLossWrapper with sets of source/estimates
>>> loss_func = PITLossWrapper(SingleSrcMultiScaleSpectral(),
>>>                            pit_from='pw_pt')
>>> loss = loss_func(est_targets, targets)

References

[1] Jesse Engel and Lamtharn (Hanoi) Hantrakul and Chenjie Gu and Adam Roberts DDSP: Differentiable Digital Signal Processing International Conference on Learning Representations ICLR 2020 $

compute_spectral_loss(encoder, est_target, target)

forward(est_target, target)

static norm1(a)