Upload feature extractor
Browse files- feature_extraction_dass.py +242 -0
- preprocessor_config.json +16 -0
feature_extraction_dass.py
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2022 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 DASS.
|
17 |
+
"""
|
18 |
+
# based on https://github.com/huggingface/transformers/blob/v4.49.0/src/
|
19 |
+
# transformers/models/audio_spectrogram_transformer/feature_extraction_audio_spectrogram_transformer.py
|
20 |
+
# added htk_compat=True to mel_filter_bank
|
21 |
+
|
22 |
+
from typing import List, Optional, Union
|
23 |
+
|
24 |
+
import numpy as np
|
25 |
+
|
26 |
+
from transformers.audio_utils import mel_filter_bank, spectrogram, window_function
|
27 |
+
from transformers.feature_extraction_sequence_utils import SequenceFeatureExtractor
|
28 |
+
from transformers.feature_extraction_utils import BatchFeature
|
29 |
+
from transformers.utils import TensorType, is_speech_available, is_torch_available, logging
|
30 |
+
|
31 |
+
|
32 |
+
if is_speech_available():
|
33 |
+
import torchaudio.compliance.kaldi as ta_kaldi
|
34 |
+
|
35 |
+
if is_torch_available():
|
36 |
+
import torch
|
37 |
+
|
38 |
+
|
39 |
+
logger = logging.get_logger(__name__)
|
40 |
+
|
41 |
+
|
42 |
+
class DASSFeatureExtractor(SequenceFeatureExtractor):
|
43 |
+
r"""
|
44 |
+
Constructs a Distilled Audio State-Space (DASS) feature extractor.
|
45 |
+
|
46 |
+
This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
|
47 |
+
most of the main methods. Users should refer to this superclass for more information regarding those methods.
|
48 |
+
|
49 |
+
This class extracts mel-filter bank features from raw speech using TorchAudio if installed or using numpy
|
50 |
+
otherwise, pads/truncates them to a fixed length and normalizes them using a mean and standard deviation.
|
51 |
+
|
52 |
+
Args:
|
53 |
+
feature_size (`int`, *optional*, defaults to 1):
|
54 |
+
The feature dimension of the extracted features.
|
55 |
+
sampling_rate (`int`, *optional*, defaults to 16000):
|
56 |
+
The sampling rate at which the audio files should be digitalized expressed in hertz (Hz).
|
57 |
+
num_mel_bins (`int`, *optional*, defaults to 128):
|
58 |
+
Number of Mel-frequency bins.
|
59 |
+
max_length (`int`, *optional*, defaults to 1024):
|
60 |
+
Maximum length to which to pad/truncate the extracted features.
|
61 |
+
do_normalize (`bool`, *optional*, defaults to `True`):
|
62 |
+
Whether or not to normalize the log-Mel features using `mean` and `std`.
|
63 |
+
mean (`float`, *optional*, defaults to -4.2677393):
|
64 |
+
The mean value used to normalize the log-Mel features. Uses the AudioSet mean by default.
|
65 |
+
std (`float`, *optional*, defaults to 4.5689974):
|
66 |
+
The standard deviation value used to normalize the log-Mel features. Uses the AudioSet standard deviation
|
67 |
+
by default.
|
68 |
+
return_attention_mask (`bool`, *optional*, defaults to `False`):
|
69 |
+
Whether or not [`~ASTFeatureExtractor.__call__`] should return `attention_mask`.
|
70 |
+
"""
|
71 |
+
|
72 |
+
model_input_names = ["input_values", "attention_mask"]
|
73 |
+
|
74 |
+
def __init__(
|
75 |
+
self,
|
76 |
+
feature_size=1,
|
77 |
+
sampling_rate=16000,
|
78 |
+
num_mel_bins=128,
|
79 |
+
max_length=1024,
|
80 |
+
padding_value=0.0,
|
81 |
+
do_normalize=True,
|
82 |
+
mean=-4.2677393,
|
83 |
+
std=4.5689974,
|
84 |
+
return_attention_mask=False,
|
85 |
+
**kwargs,
|
86 |
+
):
|
87 |
+
super().__init__(feature_size=feature_size, sampling_rate=sampling_rate, padding_value=padding_value, **kwargs)
|
88 |
+
self.num_mel_bins = num_mel_bins
|
89 |
+
self.max_length = max_length
|
90 |
+
self.do_normalize = do_normalize
|
91 |
+
self.mean = mean
|
92 |
+
self.std = std
|
93 |
+
self.return_attention_mask = return_attention_mask
|
94 |
+
|
95 |
+
if not is_speech_available():
|
96 |
+
mel_filters = mel_filter_bank(
|
97 |
+
num_frequency_bins=256,
|
98 |
+
num_mel_filters=self.num_mel_bins,
|
99 |
+
min_frequency=20,
|
100 |
+
max_frequency=sampling_rate // 2,
|
101 |
+
sampling_rate=sampling_rate,
|
102 |
+
norm=None,
|
103 |
+
mel_scale="kaldi",
|
104 |
+
triangularize_in_mel_space=True,
|
105 |
+
)
|
106 |
+
|
107 |
+
self.mel_filters = np.pad(mel_filters, ((0, 1), (0, 0)))
|
108 |
+
self.window = window_function(400, "hann", periodic=False)
|
109 |
+
|
110 |
+
def _extract_fbank_features(
|
111 |
+
self,
|
112 |
+
waveform: np.ndarray,
|
113 |
+
max_length: int,
|
114 |
+
) -> np.ndarray:
|
115 |
+
"""
|
116 |
+
Get mel-filter bank features using TorchAudio.
|
117 |
+
"""
|
118 |
+
if is_speech_available():
|
119 |
+
waveform = torch.from_numpy(waveform).unsqueeze(0)
|
120 |
+
waveform = waveform - waveform.mean()
|
121 |
+
fbank = ta_kaldi.fbank(
|
122 |
+
waveform,
|
123 |
+
sample_frequency=self.sampling_rate,
|
124 |
+
window_type="hanning",
|
125 |
+
num_mel_bins=self.num_mel_bins,
|
126 |
+
htk_compat=True,
|
127 |
+
)
|
128 |
+
else:
|
129 |
+
waveform = np.squeeze(waveform)
|
130 |
+
fbank = spectrogram(
|
131 |
+
waveform,
|
132 |
+
self.window,
|
133 |
+
frame_length=400,
|
134 |
+
hop_length=160,
|
135 |
+
fft_length=512,
|
136 |
+
power=2.0,
|
137 |
+
center=False,
|
138 |
+
preemphasis=0.97,
|
139 |
+
mel_filters=self.mel_filters,
|
140 |
+
log_mel="log",
|
141 |
+
mel_floor=1.192092955078125e-07,
|
142 |
+
remove_dc_offset=True,
|
143 |
+
).T
|
144 |
+
|
145 |
+
fbank = torch.from_numpy(fbank)
|
146 |
+
|
147 |
+
n_frames = fbank.shape[0]
|
148 |
+
difference = max_length - n_frames
|
149 |
+
|
150 |
+
# pad or truncate, depending on difference
|
151 |
+
if difference > 0:
|
152 |
+
pad_module = torch.nn.ZeroPad2d((0, 0, 0, difference))
|
153 |
+
fbank = pad_module(fbank)
|
154 |
+
elif difference < 0:
|
155 |
+
fbank = fbank[0:max_length, :]
|
156 |
+
|
157 |
+
fbank = fbank.numpy()
|
158 |
+
|
159 |
+
return fbank
|
160 |
+
|
161 |
+
def normalize(self, input_values: np.ndarray) -> np.ndarray:
|
162 |
+
return (input_values - (self.mean)) / (self.std * 2)
|
163 |
+
|
164 |
+
def __call__(
|
165 |
+
self,
|
166 |
+
raw_speech: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
|
167 |
+
sampling_rate: Optional[int] = None,
|
168 |
+
return_tensors: Optional[Union[str, TensorType]] = None,
|
169 |
+
**kwargs,
|
170 |
+
) -> BatchFeature:
|
171 |
+
"""
|
172 |
+
Main method to featurize and prepare for the model one or several sequence(s).
|
173 |
+
|
174 |
+
Args:
|
175 |
+
raw_speech (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
|
176 |
+
The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
|
177 |
+
values, a list of numpy arrays or a list of list of float values. Must be mono channel audio, not
|
178 |
+
stereo, i.e. single float per timestep.
|
179 |
+
sampling_rate (`int`, *optional*):
|
180 |
+
The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass
|
181 |
+
`sampling_rate` at the forward call to prevent silent errors.
|
182 |
+
return_tensors (`str` or [`~utils.TensorType`], *optional*):
|
183 |
+
If set, will return tensors instead of list of python integers. Acceptable values are:
|
184 |
+
|
185 |
+
- `'tf'`: Return TensorFlow `tf.constant` objects.
|
186 |
+
- `'pt'`: Return PyTorch `torch.Tensor` objects.
|
187 |
+
- `'np'`: Return Numpy `np.ndarray` objects.
|
188 |
+
"""
|
189 |
+
|
190 |
+
if sampling_rate is not None:
|
191 |
+
if sampling_rate != self.sampling_rate:
|
192 |
+
raise ValueError(
|
193 |
+
f"The model corresponding to this feature extractor: {self} was trained using a sampling rate of"
|
194 |
+
f" {self.sampling_rate}. Please make sure that the provided `raw_speech` input was sampled with"
|
195 |
+
f" {self.sampling_rate} and not {sampling_rate}."
|
196 |
+
)
|
197 |
+
else:
|
198 |
+
logger.warning(
|
199 |
+
"It is strongly recommended to pass the `sampling_rate` argument to this function. "
|
200 |
+
"Failing to do so can result in silent errors that might be hard to debug."
|
201 |
+
)
|
202 |
+
|
203 |
+
is_batched_numpy = isinstance(raw_speech, np.ndarray) and len(raw_speech.shape) > 1
|
204 |
+
if is_batched_numpy and len(raw_speech.shape) > 2:
|
205 |
+
raise ValueError(f"Only mono-channel audio is supported for input to {self}")
|
206 |
+
is_batched = is_batched_numpy or (
|
207 |
+
isinstance(raw_speech, (list, tuple)) and (isinstance(raw_speech[0], (np.ndarray, tuple, list)))
|
208 |
+
)
|
209 |
+
|
210 |
+
if is_batched:
|
211 |
+
raw_speech = [np.asarray(speech, dtype=np.float32) for speech in raw_speech]
|
212 |
+
elif not is_batched and not isinstance(raw_speech, np.ndarray):
|
213 |
+
raw_speech = np.asarray(raw_speech, dtype=np.float32)
|
214 |
+
elif isinstance(raw_speech, np.ndarray) and raw_speech.dtype is np.dtype(np.float64):
|
215 |
+
raw_speech = raw_speech.astype(np.float32)
|
216 |
+
|
217 |
+
# always return batch
|
218 |
+
if not is_batched:
|
219 |
+
raw_speech = [raw_speech]
|
220 |
+
|
221 |
+
# extract fbank features and pad/truncate to max_length
|
222 |
+
features = [self._extract_fbank_features(waveform, max_length=self.max_length) for waveform in raw_speech]
|
223 |
+
|
224 |
+
# convert into BatchFeature
|
225 |
+
padded_inputs = BatchFeature({"input_values": features})
|
226 |
+
|
227 |
+
# make sure list is in array format
|
228 |
+
input_values = padded_inputs.get("input_values")
|
229 |
+
if isinstance(input_values[0], list):
|
230 |
+
padded_inputs["input_values"] = [np.asarray(feature, dtype=np.float32) for feature in input_values]
|
231 |
+
|
232 |
+
# normalization
|
233 |
+
if self.do_normalize:
|
234 |
+
padded_inputs["input_values"] = [self.normalize(feature) for feature in input_values]
|
235 |
+
|
236 |
+
if return_tensors is not None:
|
237 |
+
padded_inputs = padded_inputs.convert_to_tensors(return_tensors)
|
238 |
+
|
239 |
+
return padded_inputs
|
240 |
+
|
241 |
+
|
242 |
+
__all__ = ["DASSFeatureExtractor"]
|
preprocessor_config.json
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"auto_map": {
|
3 |
+
"AutoFeatureExtractor": "feature_extraction_dass.DASSFeatureExtractor"
|
4 |
+
},
|
5 |
+
"do_normalize": true,
|
6 |
+
"feature_extractor_type": "DASSFeatureExtractor",
|
7 |
+
"feature_size": 1,
|
8 |
+
"max_length": 1024,
|
9 |
+
"mean": -4.2677393,
|
10 |
+
"num_mel_bins": 128,
|
11 |
+
"padding_side": "right",
|
12 |
+
"padding_value": 0.0,
|
13 |
+
"return_attention_mask": false,
|
14 |
+
"sampling_rate": 16000,
|
15 |
+
"std": 4.5689974
|
16 |
+
}
|