tuna2134/sbv2-v2
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

WIP: split vocoder #2

Draft
tuna2134 wants to merge 3 commits from patch into master
pull from: patch
merge into: tuna2134:master
Conversation 0 Commits 3 Files Changed 8 +13 -4
1 changed files with 13 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit db37175b87 - Show all commits

17
style_bert_vits2/models/infer.py
View File

@@ -224,14 +224,23 @@ def infer(
en_bert = en_bert[:, :-2] en_bert = en_bert[:, :-2]
with torch.no_grad(): with torch.no_grad():
# BERT may run in fp16 (for example on Colab) while a loaded TTS model
# remains fp32. Conv1d/Linear require inputs and parameters to use the
# same dtype outside autocast, so normalize all floating conditioning
# tensors to the TTS model dtype at this boundary.
model_dtype = next(net_g.parameters()).dtype
x_tst = phones.to(device).unsqueeze(0) x_tst = phones.to(device).unsqueeze(0)
tones = tones.to(device).unsqueeze(0) tones = tones.to(device).unsqueeze(0)
lang_ids = lang_ids.to(device).unsqueeze(0) lang_ids = lang_ids.to(device).unsqueeze(0)
bert = bert.to(device).unsqueeze(0) bert = bert.to(device=device, dtype=model_dtype).unsqueeze(0)
ja_bert = ja_bert.to(device).unsqueeze(0) ja_bert = ja_bert.to(device=device, dtype=model_dtype).unsqueeze(0)
en_bert = en_bert.to(device).unsqueeze(0) en_bert = en_bert.to(device=device, dtype=model_dtype).unsqueeze(0)
x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device) x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
style_vec_tensor = torch.from_numpy(style_vec).to(device).unsqueeze(0) style_vec_tensor = (
torch.from_numpy(style_vec)
.to(device=device, dtype=model_dtype)
.unsqueeze(0)
)
del phones del phones
sid_tensor = torch.LongTensor([sid]).to(device) sid_tensor = torch.LongTensor([sid]).to(device)