Spaces:
Runtime error
Runtime error
File size: 10,077 Bytes
455a40f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 |
# coding=utf-8
# Copyright 2022 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import itertools
import os
import random
import tempfile
import unittest
import numpy as np
from datasets import load_dataset
from transformers import is_speech_available
from transformers.testing_utils import check_json_file_has_correct_format, require_torch, require_torchaudio
from transformers.utils.import_utils import is_torch_available
from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
if is_speech_available():
from transformers import WhisperFeatureExtractor
if is_torch_available():
import torch
global_rng = random.Random()
def floats_list(shape, scale=1.0, rng=None, name=None):
"""Creates a random float32 tensor"""
if rng is None:
rng = global_rng
values = []
for batch_idx in range(shape[0]):
values.append([])
for _ in range(shape[1]):
values[-1].append(rng.random() * scale)
return values
@require_torch
@require_torchaudio
class WhisperFeatureExtractionTester(unittest.TestCase):
def __init__(
self,
parent,
batch_size=7,
min_seq_length=400,
max_seq_length=2000,
feature_size=10,
hop_length=160,
chunk_length=8,
padding_value=0.0,
sampling_rate=4_000,
return_attention_mask=False,
do_normalize=True,
):
self.parent = parent
self.batch_size = batch_size
self.min_seq_length = min_seq_length
self.max_seq_length = max_seq_length
self.seq_length_diff = (self.max_seq_length - self.min_seq_length) // (self.batch_size - 1)
self.padding_value = padding_value
self.sampling_rate = sampling_rate
self.return_attention_mask = return_attention_mask
self.do_normalize = do_normalize
self.feature_size = feature_size
self.chunk_length = chunk_length
self.hop_length = hop_length
def prepare_feat_extract_dict(self):
return {
"feature_size": self.feature_size,
"hop_length": self.hop_length,
"chunk_length": self.chunk_length,
"padding_value": self.padding_value,
"sampling_rate": self.sampling_rate,
"return_attention_mask": self.return_attention_mask,
"do_normalize": self.do_normalize,
}
def prepare_inputs_for_common(self, equal_length=False, numpify=False):
def _flatten(list_of_lists):
return list(itertools.chain(*list_of_lists))
if equal_length:
speech_inputs = [floats_list((self.max_seq_length, self.feature_size)) for _ in range(self.batch_size)]
else:
# make sure that inputs increase in size
speech_inputs = [
floats_list((x, self.feature_size))
for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff)
]
if numpify:
speech_inputs = [np.asarray(x) for x in speech_inputs]
return speech_inputs
@require_torch
@require_torchaudio
class WhisperFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
feature_extraction_class = WhisperFeatureExtractor if is_speech_available() else None
def setUp(self):
self.feat_extract_tester = WhisperFeatureExtractionTester(self)
def test_feat_extract_from_and_save_pretrained(self):
feat_extract_first = self.feature_extraction_class(**self.feat_extract_dict)
with tempfile.TemporaryDirectory() as tmpdirname:
saved_file = feat_extract_first.save_pretrained(tmpdirname)[0]
check_json_file_has_correct_format(saved_file)
feat_extract_second = self.feature_extraction_class.from_pretrained(tmpdirname)
dict_first = feat_extract_first.to_dict()
dict_second = feat_extract_second.to_dict()
mel_1 = feat_extract_first.mel_filters
mel_2 = feat_extract_second.mel_filters
self.assertTrue(np.allclose(mel_1, mel_2))
self.assertEqual(dict_first, dict_second)
def test_feat_extract_to_json_file(self):
feat_extract_first = self.feature_extraction_class(**self.feat_extract_dict)
with tempfile.TemporaryDirectory() as tmpdirname:
json_file_path = os.path.join(tmpdirname, "feat_extract.json")
feat_extract_first.to_json_file(json_file_path)
feat_extract_second = self.feature_extraction_class.from_json_file(json_file_path)
dict_first = feat_extract_first.to_dict()
dict_second = feat_extract_second.to_dict()
mel_1 = feat_extract_first.mel_filters
mel_2 = feat_extract_second.mel_filters
self.assertTrue(np.allclose(mel_1, mel_2))
self.assertEqual(dict_first, dict_second)
def test_call(self):
# Tests that all call wrap to encode_plus and batch_encode_plus
feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
# create three inputs of length 800, 1000, and 1200
speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs]
# Test feature size
input_features = feature_extractor(np_speech_inputs, padding="max_length", return_tensors="np").input_features
self.assertTrue(input_features.ndim == 3)
self.assertTrue(input_features.shape[-1] == feature_extractor.nb_max_frames)
self.assertTrue(input_features.shape[-2] == feature_extractor.feature_size)
# Test not batched input
encoded_sequences_1 = feature_extractor(speech_inputs[0], return_tensors="np").input_features
encoded_sequences_2 = feature_extractor(np_speech_inputs[0], return_tensors="np").input_features
self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
# Test batched
encoded_sequences_1 = feature_extractor(speech_inputs, return_tensors="np").input_features
encoded_sequences_2 = feature_extractor(np_speech_inputs, return_tensors="np").input_features
for enc_seq_1, enc_seq_2 in zip(encoded_sequences_1, encoded_sequences_2):
self.assertTrue(np.allclose(enc_seq_1, enc_seq_2, atol=1e-3))
# Test truncation required
speech_inputs = [floats_list((1, x))[0] for x in range(200, (feature_extractor.n_samples + 500), 200)]
np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs]
speech_inputs_truncated = [x[: feature_extractor.n_samples] for x in speech_inputs]
np_speech_inputs_truncated = [np.asarray(speech_input) for speech_input in speech_inputs_truncated]
encoded_sequences_1 = feature_extractor(np_speech_inputs, return_tensors="np").input_features
encoded_sequences_2 = feature_extractor(np_speech_inputs_truncated, return_tensors="np").input_features
for enc_seq_1, enc_seq_2 in zip(encoded_sequences_1, encoded_sequences_2):
self.assertTrue(np.allclose(enc_seq_1, enc_seq_2, atol=1e-3))
def test_double_precision_pad(self):
import torch
feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
np_speech_inputs = np.random.rand(100, 32).astype(np.float64)
py_speech_inputs = np_speech_inputs.tolist()
for inputs in [py_speech_inputs, np_speech_inputs]:
np_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="np")
self.assertTrue(np_processed.input_features.dtype == np.float32)
pt_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="pt")
self.assertTrue(pt_processed.input_features.dtype == torch.float32)
def _load_datasamples(self, num_samples):
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
# automatic decoding with librispeech
speech_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"]
return [x["array"] for x in speech_samples]
def test_integration(self):
# fmt: off
EXPECTED_INPUT_FEATURES = torch.tensor(
[
0.1193, -0.0946, -0.1098, -0.0196, 0.0225, -0.0690, -0.1736, 0.0951,
0.0971, -0.0817, -0.0702, 0.0162, 0.0260, 0.0017, -0.0192, -0.1678,
0.0709, -0.1867, -0.0655, -0.0274, -0.0234, -0.1884, -0.0516, -0.0554,
-0.0274, -0.1425, -0.1423, 0.0837, 0.0377, -0.0854
]
)
# fmt: on
input_speech = self._load_datasamples(1)
feaure_extractor = WhisperFeatureExtractor()
input_features = feaure_extractor(input_speech, return_tensors="pt").input_features
self.assertTrue(torch.allclose(input_features[0, 0, :30], EXPECTED_INPUT_FEATURES, atol=1e-4))
def test_zero_mean_unit_variance_normalization_trunc_np_longest(self):
feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
audio = self._load_datasamples(1)[0]
audio = ((audio - audio.min()) / (audio.max() - audio.min())) * 65535 # Rescale to [0, 65535] to show issue
audio = feat_extract.zero_mean_unit_var_norm([audio], attention_mask=None)[0]
self.assertTrue(np.all(np.mean(audio) < 1e-3))
self.assertTrue(np.all(np.abs(np.var(audio) - 1) < 1e-3))
|