Delete feature_extraction_granite_speech.py
Browse files
feature_extraction_granite_speech.py
DELETED
@@ -1,118 +0,0 @@
|
|
1 |
-
# coding=utf-8
|
2 |
-
# Copyright 2025 The HuggingFace Inc. team.
|
3 |
-
#
|
4 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
-
# you may not use this file except in compliance with the License.
|
6 |
-
# You may obtain a copy of the License at
|
7 |
-
#
|
8 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
-
#
|
10 |
-
# Unless required by applicable law or agreed to in writing, software
|
11 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
-
# See the License for the specific language governing permissions and
|
14 |
-
# limitations under the License.
|
15 |
-
"""
|
16 |
-
Feature extractor class for Speech Granite
|
17 |
-
"""
|
18 |
-
|
19 |
-
import math
|
20 |
-
from typing import List, Optional
|
21 |
-
|
22 |
-
from transformers.feature_extraction_utils import BatchFeature, FeatureExtractionMixin
|
23 |
-
from transformers.utils import is_torch_available, is_torchaudio_available, logging
|
24 |
-
|
25 |
-
|
26 |
-
logger = logging.get_logger(__name__)
|
27 |
-
|
28 |
-
if is_torch_available():
|
29 |
-
import torch
|
30 |
-
|
31 |
-
if is_torchaudio_available():
|
32 |
-
import torchaudio
|
33 |
-
|
34 |
-
|
35 |
-
class GraniteSpeechFeatureExtractor(FeatureExtractionMixin):
|
36 |
-
model_input_names = ["input_features"]
|
37 |
-
|
38 |
-
def __init__(
|
39 |
-
self,
|
40 |
-
sampling_rate=16000,
|
41 |
-
n_fft=512,
|
42 |
-
win_length=400,
|
43 |
-
hop_length=160,
|
44 |
-
n_mels=80,
|
45 |
-
projector_window_size=15,
|
46 |
-
projector_downsample_rate=5,
|
47 |
-
**kwargs,
|
48 |
-
):
|
49 |
-
super().__init__(**kwargs)
|
50 |
-
self.melspec_kwargs = {
|
51 |
-
"sample_rate": sampling_rate,
|
52 |
-
"n_fft": n_fft,
|
53 |
-
"win_length": win_length,
|
54 |
-
"hop_length": hop_length,
|
55 |
-
"n_mels": n_mels,
|
56 |
-
}
|
57 |
-
# HACK - for now, lazily initialize the mel spectrogram transform;
|
58 |
-
# the feature extractor mixin explodes otherwise because
|
59 |
-
# it tries to log the feature extractor, and the melspectrogram
|
60 |
-
# transform isn't json serializable...
|
61 |
-
self.melspec = None
|
62 |
-
self.projector_window_size = projector_window_size
|
63 |
-
self.projector_downsample_rate = projector_downsample_rate
|
64 |
-
|
65 |
-
def _ensure_melspec_transform_is_initialized(self):
|
66 |
-
if self.melspec is None:
|
67 |
-
self.melspec = torchaudio.transforms.MelSpectrogram(**self.melspec_kwargs)
|
68 |
-
|
69 |
-
def __call__(
|
70 |
-
self,
|
71 |
-
x: torch.Tensor,
|
72 |
-
device: Optional[str] = "cpu",
|
73 |
-
) -> BatchFeature:
|
74 |
-
# TODO there is probably a better way to do both of these things...
|
75 |
-
self._ensure_melspec_transform_is_initialized()
|
76 |
-
if device is not None:
|
77 |
-
melspec = self.melspec.to(device)
|
78 |
-
x = x.to(device)
|
79 |
-
else:
|
80 |
-
melspec = self.melspec
|
81 |
-
|
82 |
-
B, _ = x.shape
|
83 |
-
with torch.no_grad():
|
84 |
-
mel = melspec(x.float())
|
85 |
-
logmel = mel.transpose(-1, -2).clip_(min=1e-10).log10_()
|
86 |
-
mx = logmel.amax(dim=(-2, -1), keepdim=True)
|
87 |
-
logmel = torch.maximum(logmel, mx - 8.0).div_(4).add_(1)
|
88 |
-
if logmel.shape[1] % 2 == 1:
|
89 |
-
logmel = logmel[:, :-1] # remove last frame if odd
|
90 |
-
x = logmel.reshape(B, -1, 2 * logmel.shape[-1]) # stacking and skipping by 2
|
91 |
-
|
92 |
-
if x.device != "cpu":
|
93 |
-
return x.detach().cpu()
|
94 |
-
return x
|
95 |
-
|
96 |
-
def _get_num_audio_features(self, audio_lengths: List[int]) -> List[int]:
|
97 |
-
"""
|
98 |
-
Gets the (variable length) variable length number of features
|
99 |
-
(i.e., projector output) for the sequences being considered.
|
100 |
-
"""
|
101 |
-
hop_length = self.melspec_kwargs["hop_length"]
|
102 |
-
effective_window_size = self.projector_window_size // self.projector_downsample_rate
|
103 |
-
|
104 |
-
projector_lengths = []
|
105 |
-
for raw_length in audio_lengths:
|
106 |
-
# mel sequence length computation
|
107 |
-
mel_length = raw_length // hop_length + 1
|
108 |
-
# encoder frame takes two mel features
|
109 |
-
encoder_length = mel_length // 2
|
110 |
-
nblocks = math.ceil(encoder_length / self.projector_window_size)
|
111 |
-
# projector output length
|
112 |
-
projector_length = nblocks * effective_window_size
|
113 |
-
projector_lengths.append(projector_length)
|
114 |
-
|
115 |
-
return projector_lengths
|
116 |
-
|
117 |
-
|
118 |
-
__all__ = ["GraniteSpeechFeatureExtractor"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|