crlandsc / torch-log-wmse

logWMSE, an audio quality metric & loss function with support for digital silence target. Useful for training and evaluating audio source separation systems.

Hi there 👋 my name is Chris!

I am an audio machine learning engineer and researcher working on advancing audio AI/ML and spatial audio capabilities.

My recent work has focused on binaural externalization, audio waveform diffusion for generative audio, and audio source separation for music "demixing".

I also make music under the name 🎶After August.

Please reach out if you have any questions, or if you are interested in chatting about audio, music, AI/ML, spatial audio, or all of the above!

Follow my work, writing, and music on:

My Website | Medium | LinkedIn
YouTube | Spotify | Facebook | Instagram

torch-log-wmse's Issues

Detailed introduction document

Dear author, hello! Your work is excellent. May I ask if there are detailed introduction materials?

Add an option to disable the frequency curve weigthing

It would be great to have a option to easily disable the frequency curve weighting, when working on bass heavy audio (bass/drums/kick stems), it can lead to kinda unexpected results (high score while there are obvious audible issues in really low frequencies).

Returns nan when one of the entries in the batch is a digital silence triplet

Here is some code that reproduces the issue:

import torch
from torch_log_wmse import LogWMSE

loss_function = LogWMSE(audio_length=1, return_as_loss=True)
torch.manual_seed(0)
raw = torch.randn(2, 1, 44100, dtype=torch.float32)
raw[0] = 0.0
est = torch.randn(2, 1, 44100, dtype=torch.float32)
est[0] = 0.0
gt = torch.randn(2, 1, 44100, dtype=torch.float32)
gt[0] = 0.0
loss0 = loss_function(raw[0:1], est[0:1].unsqueeze(2), gt[0:1].unsqueeze(2))  # -73.6827
print(loss0)
assert not torch.isnan(loss0)
loss1 = loss_function(raw[1:2], est[1:2].unsqueeze(2), gt[1:2].unsqueeze(2))  # 2.7946
print(loss1)
assert not torch.isnan(loss1)
loss_combined = loss_function(raw, est.unsqueeze(2), gt.unsqueeze(2))  # nan
print(loss_combined)
assert not torch.isnan(loss_combined)

It looks like this part of the code does not add EPS to the batch entry that is all digital zero:

torch-log-wmse/torch_log_wmse/metric.py

Line 99 in de47dbc

if input_rms.sum() < ERROR_TOLERANCE_THRESHOLD:

That leads to the problematic division by zero here:

torch-log-wmse/torch_log_wmse/metric.py

Line 103 in de47dbc

scaling_factor = 1 / input_rms