deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 01:49:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #1232 from boji123/bj_dev_feat_len_pad

Browse Source 
a better solution for mismatch of speech feat len and speech token len when trainning
This commit is contained in:
Xiang Lyu
2025-04-30 09:41:50 +08:00
committed by GitHub
parent a442317d17 038ff9f353
commit 88f467a8ac
6 changed files with 59 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
cosyvoice/flow/flow.py
View File

@@ -92,7 +92,6 @@ class MaskedDiffWithXvec(torch.nn.Module):
mask = (~make_pad_mask(feat_len)).to(h)
# NOTE this is unnecessary, feat/h already same shape
feat = F.interpolate(feat.unsqueeze(dim=1), size=h.shape[1:], mode="nearest").squeeze(dim=1)
loss, _ = self.decoder.compute_loss(
feat.transpose(1, 2).contiguous(),
mask.unsqueeze(1),
@@ -214,7 +213,6 @@ class CausalMaskedDiffWithXvec(torch.nn.Module):
h = self.encoder_proj(h)
# get conditions
feat = F.interpolate(feat.unsqueeze(dim=1), size=h.shape[1:], mode="nearest").squeeze(dim=1)
conds = torch.zeros(feat.shape, device=token.device)
for i, j in enumerate(feat_len):
if random.random() < 0.5: