fix(loss): use raw mel magnitude for the mag_weight term in MultiScal…

[andrewwhitecdw]

…eMelSpectrogramLoss

Summary

In MultiScaleMelSpectrogramLoss.forward() (loss.py), the magnitude (mag_weight) portion of the multi-scale mel loss is computed between the log-mel spectrograms instead of the raw mel spectrograms. As a result, setting mag_weight > 0 does not add a magnitude-domain loss at all — it just adds a second, differently-weighted copy of the log-mel loss.

This is latent with the default mag_weight=0.0 (term is multiplied by zero), so default BigVGAN training is unaffected; any user enabling mag_weight gets the wrong objective.

Root cause

Copy-paste error duplicating the log-mel line:

loss += self.log_weight * self.loss_fn(x_logmels, y_logmels)
loss += self.mag_weight * self.loss_fn(x_logmels, y_logmels)  # wrong operands

The class docstring states it is copied from descriptinc/descript-audio-codec dac/nn/loss.py, where the upstream MelSpectrogramLoss.forward() reads:

loss += self.log_weight * self.loss_fn(x_logmels, y_logmels)
loss += self.mag_weight * self.loss_fn(x_mels, y_mels)

(and DAC's MultiScaleSTFTLoss likewise compares raw magnitudes for the mag_weight term).

Fix

One-line change: compare x_mels/y_mels (raw) for the magnitude term:

loss += self.mag_weight * self.loss_fn(x_mels, y_mels)

Testing

⚠️ The repo's own tests (tests/test_activation*.py, tests/test_cuda_vs_torch_model.py) require a CUDA GPU and the fused CUDA kernel build; no GPU is available on this machine, so they could not be run here — please rely on CI.

Ran a standalone regression script (torch CPU, loss.py imported directly): instantiated MultiScaleMelSpectrogramLoss(log_weight=0, mag_weight=1) and compared its output against a hand-computed L1 between raw mel spectrograms:

• with fix: loss = 0.061000 == expected 0.061000 → PASS
• with fix stashed (unmodified code): loss = 0.193965 (= L1 of log-mels) → FAIL, confirming the bug is real and pre-existing on main

Also verified the default path is unchanged: with mag_weight=0.0 the term is multiplied by zero, so default training behavior is identical.

Why existing tests missed it

The test suite only covers the Snake activations and CUDA-vs-torch kernel parity; nothing exercises MultiScaleMelSpectrogramLoss, and the default mag_weight=0.0 masks the bug even if it did.