File size: 5,793 Bytes
5f89bea |
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 |
# coding=utf-8
# Copyright 2021 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 random
import unittest
import numpy as np
from transformers import WAV_2_VEC_2_PRETRAINED_MODEL_ARCHIVE_LIST, Wav2Vec2Config, Wav2Vec2FeatureExtractor
from transformers.testing_utils import require_torch, slow
from .test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
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
class Wav2Vec2FeatureExtractionTester(unittest.TestCase):
def __init__(
self,
parent,
batch_size=7,
min_seq_length=400,
max_seq_length=2000,
feature_size=1,
padding_value=0.0,
sampling_rate=16000,
return_attention_mask=True,
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.feature_size = feature_size
self.padding_value = padding_value
self.sampling_rate = sampling_rate
self.return_attention_mask = return_attention_mask
self.do_normalize = do_normalize
def prepare_feat_extract_dict(self):
return {
"feature_size": self.feature_size,
"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.batch_size, self.max_seq_length))
else:
speech_inputs = [
_flatten(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
class Wav2Vec2FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
feature_extraction_class = Wav2Vec2FeatureExtractor
def setUp(self):
self.feat_extract_tester = Wav2Vec2FeatureExtractionTester(self)
def test_call(self):
# Tests that all call wrap to encode_plus and batch_encode_plus
feat_extract = 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 not batched input
encoded_sequences_1 = feat_extract(speech_inputs[0], return_tensors="np").input_values
encoded_sequences_2 = feat_extract(np_speech_inputs[0], return_tensors="np").input_values
self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
# Test batched
encoded_sequences_1 = feat_extract(speech_inputs, return_tensors="np").input_values
encoded_sequences_2 = feat_extract(np_speech_inputs, return_tensors="np").input_values
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_zero_mean_unit_variance_normalization(self):
feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
processed = feat_extract(speech_inputs, padding="longest")
input_values = processed.input_values
def _check_zero_mean_unit_variance(input_vector):
self.assertTrue(np.abs(np.mean(input_vector)) < 1e-3)
self.assertTrue(np.abs(np.var(input_vector) - 1) < 1e-3)
_check_zero_mean_unit_variance(input_values[0, :800])
_check_zero_mean_unit_variance(input_values[1, :1000])
_check_zero_mean_unit_variance(input_values[2])
@slow
@require_torch
def test_pretrained_checkpoints_are_set_correctly(self):
# this test makes sure that models that are using
# group norm don't have their feature extractor return the
# attention_mask
for model_id in WAV_2_VEC_2_PRETRAINED_MODEL_ARCHIVE_LIST:
config = Wav2Vec2Config.from_pretrained(model_id)
feat_extract = Wav2Vec2FeatureExtractor.from_pretrained(model_id)
# only "layer" feature extraction norm should make use of
# attention_mask
self.assertEqual(feat_extract.return_attention_mask, config.feat_extract_norm == "layer")
|