so-vits-svc/vencoder/ContentVec768L12.py

34 lines
1.2 KiB
Python
Raw Normal View History

2023-05-13 15:45:56 +00:00
from vencoder.encoder import SpeechEncoder
import torch
2023-05-14 06:39:07 +00:00
from fairseq import checkpoint_utils
2023-05-13 15:45:56 +00:00
class ContentVec768L12(SpeechEncoder):
2023-05-14 06:39:07 +00:00
def __init__(self,vec_path = "pretrain/checkpoint_best_legacy_500.pt",device=None):
2023-05-13 15:45:56 +00:00
print("load model(s) from {}".format(vec_path))
2023-05-14 06:39:07 +00:00
self.hidden_dim = 768
2023-05-13 15:45:56 +00:00
models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
[vec_path],
suffix="",
)
2023-05-14 06:39:07 +00:00
if device is None:
self.dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
else:
self.dev = torch.device(device)
self.model = models[0].to(self.dev)
2023-05-13 15:45:56 +00:00
self.model.eval()
def encoder(self, wav):
feats = wav
if feats.dim() == 2: # double channels
feats = feats.mean(-1)
assert feats.dim() == 1, feats.dim()
feats = feats.view(1, -1)
padding_mask = torch.BoolTensor(feats.shape).fill_(False)
inputs = {
"source": feats.to(wav.device),
"padding_mask": padding_mask.to(wav.device),
"output_layer": 12, # layer 12
}
with torch.no_grad():
logits = self.model.extract_features(**inputs)
return logits[0].transpose(1, 2)