Ethanfel/ComfyUI-SelVA
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: exclude GAFilter params from L2-SP regularization

Browse Source 
L2-SP anchors trainable params to their pretrained values. GAFilter is a
newly initialized module (identity FIR filter) with no pretrained values —
anchoring it to identity initialization would resist learning. Exclude
gafilter params from the L2-SP loss so they train freely.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Ethanfel
2026-04-09 16:19:52 +02:00
parent db112394e8
commit 45fced55bc
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nodes/selva_bigvgan_trainer.py
+3 -1
View File
@@ -812,7 +812,9 @@ def _do_train(vocoder, mel_converter, clips,
l2sp_loss = torch.zeros(1, device=device) l2sp_loss = torch.zeros(1, device=device)
if lambda_l2sp > 0.0 and ref_params: if lambda_l2sp > 0.0 and ref_params:
for name, param in vocoder.named_parameters(): for name, param in vocoder.named_parameters():
if name in ref_params and param.requires_grad: # Skip GAFilter — newly initialized, not pretrained; L2-SP
# anchoring to identity would fight against learning.
if name in ref_params and param.requires_grad and "gafilter" not in name:
l2sp_loss = l2sp_loss + F.mse_loss( l2sp_loss = l2sp_loss + F.mse_loss(
param, ref_params[name], reduction="sum" param, ref_params[name], reduction="sum"
) )