Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

fix reduction mistake in SpectralConvergenceLoss #75

Open
wants to merge 1 commit into
base: main
Choose a base branch
from

Conversation

renared
Copy link

@renared renared commented May 22, 2024

I noticed that when evaluating the STFT loss over my validation dataset, I obtained different results in function of the batch size. I could isolate the cause to be the spectral convergence term, then came across the comment by @egaznep in issue #69. It does not make sense to average the denominator over all dimensions including the batch dimension, so I believe their suggestion should be used instead.

This snippet shows the difference:

import torch
from auraloss.freq import STFTLoss

batches = [(torch.randn(4, 1, 16384), torch.randn(4, 1, 16384)) for i in range(1024)]
batchall = tuple(torch.concat(u, dim=0) for u in zip(*batches))

print("with spectral convergence enabled")
loss = STFTLoss()
print("mean of losses:", torch.mean(torch.tensor(tuple(loss(*batch) for batch in batches))))
print("over full dataset:", loss(*batchall))

print("with spectral convergence disabled")
loss = STFTLoss(w_sc=0)
print("mean of losses:", torch.mean(torch.tensor(tuple(loss(*batch) for batch in batches))))
print("over full dataset:", loss(*batchall))

Before:

with spectral convergence enabled
mean of losses: tensor(1.3511)
over full dataset: tensor(1.3493)
with spectral convergence disabled
mean of losses: tensor(0.6950)
over full dataset: tensor(0.6950)

After:

with spectral convergence enabled
mean of losses: tensor(1.3726)
over full dataset: tensor(1.3726)
with spectral convergence disabled
mean of losses: tensor(0.7095)
over full dataset: tensor(0.7095)

the denominator was averaged over all dimensions including the batch dimension, see comment by @egaznep in csteinmetz1#69
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant