Upload ./RepCodec/examples/hubert_feature_reader.py with huggingface_hub
Browse files
RepCodec/examples/hubert_feature_reader.py
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) ByteDance, Inc. and its affiliates.
|
2 |
+
# Copyright (c) Chutong Meng
|
3 |
+
#
|
4 |
+
# This source code is licensed under the MIT license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
# Based on fairseq (https://github.com/facebookresearch/fairseq)
|
7 |
+
|
8 |
+
import logging
|
9 |
+
|
10 |
+
import fairseq
|
11 |
+
import torch
|
12 |
+
import torch.nn.functional as F
|
13 |
+
|
14 |
+
from fairseq.data.audio.audio_utils import get_features_or_waveform
|
15 |
+
|
16 |
+
logger = logging.getLogger("dump_feature")
|
17 |
+
|
18 |
+
|
19 |
+
class HubertFeatureReader(object):
|
20 |
+
def __init__(self, ckpt_path: str, layer: int, device: str, max_chunk=1600000):
|
21 |
+
(
|
22 |
+
model,
|
23 |
+
cfg,
|
24 |
+
task,
|
25 |
+
) = fairseq.checkpoint_utils.load_model_ensemble_and_task([ckpt_path])
|
26 |
+
|
27 |
+
self.device = device
|
28 |
+
logger.info(f"device = {self.device}")
|
29 |
+
|
30 |
+
self.model = model[0].eval().to(self.device)
|
31 |
+
self.task = task
|
32 |
+
self.layer = layer
|
33 |
+
self.max_chunk = max_chunk
|
34 |
+
logger.info(f"TASK CONFIG:\n{self.task.cfg}")
|
35 |
+
logger.info(f" max_chunk = {self.max_chunk}")
|
36 |
+
|
37 |
+
def read_audio(self, path, ref_len=None):
|
38 |
+
wav = get_features_or_waveform(path, need_waveform=True, use_sample_rate=self.task.cfg.sample_rate)
|
39 |
+
if wav.ndim == 2:
|
40 |
+
wav = wav.mean(-1)
|
41 |
+
assert wav.ndim == 1, wav.ndim
|
42 |
+
if ref_len is not None and abs(ref_len - len(wav)) > 160:
|
43 |
+
logger.warning(f"ref {ref_len} != read {len(wav)} ({path})")
|
44 |
+
return wav
|
45 |
+
|
46 |
+
def get_feats(self, path, ref_len=None):
|
47 |
+
x = self.read_audio(path, ref_len=ref_len)
|
48 |
+
with torch.no_grad():
|
49 |
+
x = torch.from_numpy(x).float().to(self.device)
|
50 |
+
if self.task.cfg.normalize:
|
51 |
+
x = F.layer_norm(x, x.shape)
|
52 |
+
x = x.view(1, -1)
|
53 |
+
|
54 |
+
feat = []
|
55 |
+
for start in range(0, x.size(1), self.max_chunk):
|
56 |
+
x_chunk = x[:, start: start + self.max_chunk]
|
57 |
+
feat_chunk, _ = self.model.extract_features(
|
58 |
+
source=x_chunk,
|
59 |
+
padding_mask=None,
|
60 |
+
mask=False,
|
61 |
+
output_layer=self.layer,
|
62 |
+
)
|
63 |
+
feat.append(feat_chunk)
|
64 |
+
return torch.cat(feat, 1).squeeze(0)
|