SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[SequenceFeatureExtractor] Rewrite padding logic from pure python to numpy (#13650)

Browse Source 
* Test np padding

* Pass feature extraction tests

* Update type hints

* Fix flaky integration tests

* Try a more stable waveform

* Add to_numpy jax support

* int32 attention masks

* Refactor normalization tests
This commit is contained in:
Anton Lozhkov
2021-09-21 17:10:13 +03:00
committed by GitHub
parent 8d533e6ad6
commit 1417978cd4
8 changed files with 133 additions and 146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
tests/test_feature_extraction_speech_to_text.py
View File

@@ -110,6 +110,10 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
def setUp(self):
self.feat_extract_tester = Speech2TextFeatureExtractionTester(self)
def _check_zero_mean_unit_variance(self, input_vector):
self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3))
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())
@@ -137,17 +141,9 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
feature_extractor = 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)]
# TODO(Patrick, Suraj, Anton) - It's surprising that "non-padded/non-numpified" padding
# results in quite inaccurate variance computation after (see 5e-1 tolerance)
# Issue is filed and PR is underway: https://github.com/huggingface/transformers/issues/13539
# paddings = ["longest", "max_length", "do_not_pad"]
# max_lengths = [None, 16, None]
# var_tolerances = [1e-3, 1e-3, 5e-1]
paddings = ["longest", "max_length"]
max_lengths = [None, 16]
var_tolerances = [1e-3, 1e-3]
for max_length, padding, var_tol in zip(max_lengths, paddings, var_tolerances):
paddings = ["longest", "max_length", "do_not_pad"]
max_lengths = [None, 16, None]
for max_length, padding in zip(max_lengths, paddings):
inputs = feature_extractor(
speech_inputs, padding=padding, max_length=max_length, return_attention_mask=True
)
@@ -155,28 +151,17 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
attention_mask = inputs.attention_mask
fbank_feat_lengths = [np.sum(x) for x in attention_mask]
def _check_zero_mean_unit_variance(input_vector, var_tol=1e-3):
self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < var_tol))
_check_zero_mean_unit_variance(input_features[0][: fbank_feat_lengths[0]], var_tol)
_check_zero_mean_unit_variance(input_features[1][: fbank_feat_lengths[1]], var_tol)
_check_zero_mean_unit_variance(input_features[2][: fbank_feat_lengths[2]], var_tol)
self._check_zero_mean_unit_variance(input_features[0][: fbank_feat_lengths[0]])
self._check_zero_mean_unit_variance(input_features[1][: fbank_feat_lengths[1]])
self._check_zero_mean_unit_variance(input_features[2][: fbank_feat_lengths[2]])
def test_cepstral_mean_and_variance_normalization_np(self):
feature_extractor = 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)]
# TODO(Patrick, Suraj, Anton) - It's surprising that "non-padded/non-numpified" padding
# results in quite inaccurate variance computation after (see 5e-1 tolerance)
# Issue is filed and PR is underway: https://github.com/huggingface/transformers/issues/13539
# paddings = ["longest", "max_length", "do_not_pad"]
# max_lengths = [None, 16, None]
# var_tolerances = [1e-3, 1e-3, 5e-1]
paddings = ["longest", "max_length"]
max_lengths = [None, 16]
var_tolerances = [1e-3, 1e-3]
for max_length, padding, var_tol in zip(max_lengths, paddings, var_tolerances):
paddings = ["longest", "max_length", "do_not_pad"]
max_lengths = [None, 16, None]
for max_length, padding in zip(max_lengths, paddings):
inputs = feature_extractor(
speech_inputs, max_length=max_length, padding=padding, return_tensors="np", return_attention_mask=True
)
@@ -184,15 +169,11 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
attention_mask = inputs.attention_mask
fbank_feat_lengths = [np.sum(x) for x in attention_mask]
def _check_zero_mean_unit_variance(input_vector, var_tol=1e-3):
self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < var_tol))
_check_zero_mean_unit_variance(input_features[0][: fbank_feat_lengths[0]], var_tol)
self._check_zero_mean_unit_variance(input_features[0][: fbank_feat_lengths[0]])
self.assertTrue(input_features[0][fbank_feat_lengths[0] :].sum() < 1e-6)
_check_zero_mean_unit_variance(input_features[1][: fbank_feat_lengths[1]], var_tol)
self._check_zero_mean_unit_variance(input_features[1][: fbank_feat_lengths[1]])
self.assertTrue(input_features[0][fbank_feat_lengths[1] :].sum() < 1e-6)
_check_zero_mean_unit_variance(input_features[2][: fbank_feat_lengths[2]], var_tol)
self._check_zero_mean_unit_variance(input_features[2][: fbank_feat_lengths[2]])
def test_cepstral_mean_and_variance_normalization_trunc_max_length(self):
feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -209,13 +190,9 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
attention_mask = inputs.attention_mask
fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
def _check_zero_mean_unit_variance(input_vector):
self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3))
_check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
_check_zero_mean_unit_variance(input_features[1])
_check_zero_mean_unit_variance(input_features[2])
self._check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
self._check_zero_mean_unit_variance(input_features[1])
self._check_zero_mean_unit_variance(input_features[2])
def test_cepstral_mean_and_variance_normalization_trunc_longest(self):
feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -232,13 +209,9 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
attention_mask = inputs.attention_mask
fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
def _check_zero_mean_unit_variance(input_vector):
self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3))
_check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
_check_zero_mean_unit_variance(input_features[1, : fbank_feat_lengths[1]])
_check_zero_mean_unit_variance(input_features[2])
self._check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
self._check_zero_mean_unit_variance(input_features[1, : fbank_feat_lengths[1]])
self._check_zero_mean_unit_variance(input_features[2])
# make sure that if max_length < longest -> then pad to max_length
self.assertEqual(input_features.shape, (3, 4, 24))
@@ -256,9 +229,9 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
attention_mask = inputs.attention_mask
fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
_check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
_check_zero_mean_unit_variance(input_features[1, : fbank_feat_lengths[1]])
_check_zero_mean_unit_variance(input_features[2])
self._check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
self._check_zero_mean_unit_variance(input_features[1, : fbank_feat_lengths[1]])
self._check_zero_mean_unit_variance(input_features[2])
# make sure that if max_length < longest -> then pad to max_length
self.assertEqual(input_features.shape, (3, 6, 24))