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

is_batched fix for remaining 2-D numpy arrays (#23309)

Browse Source 
* Fix is_batched code to allow 2-D numpy arrays for audio

* Tests

* Fix typo

* Incorporate comments from PR #23223
This commit is contained in:
LWprogramming
2023-05-23 11:37:35 -07:00
committed by GitHub
parent 6b7d6f848b
commit 3d57404464
14 changed files with 106 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tests/models/audio_spectrogram_transformer/test_feature_extraction_audio_spectrogram_transformer.py
View File

@@ -125,6 +125,14 @@ class ASTFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.Test
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 2-D numpy arrays are batched.
speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
np_speech_inputs = np.asarray(speech_inputs)
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))
@require_torch
def test_double_precision_pad(self):
import torch

8
tests/models/clap/test_feature_extraction_clap.py
View File

@@ -139,6 +139,14 @@ class ClapFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.Tes
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 2-D numpy arrays are batched.
speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
np_speech_inputs = np.asarray(speech_inputs)
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))
def test_double_precision_pad(self):
import torch

8
tests/models/mctct/test_feature_extraction_mctct.py
View File

@@ -134,6 +134,14 @@ class MCTCTFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.Te
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 2-D numpy arrays are batched.
speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
np_speech_inputs = np.asarray(speech_inputs)
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))
def test_cepstral_mean_and_variance_normalization(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(8000, 14000, 2000)]

8
tests/models/speech_to_text/test_feature_extraction_speech_to_text.py
View File

@@ -136,6 +136,14 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
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 2-D numpy arrays are batched.
speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
np_speech_inputs = np.asarray(speech_inputs)
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))
def test_cepstral_mean_and_variance_normalization(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)]

8
tests/models/speecht5/test_feature_extraction_speecht5.py
View File

@@ -275,6 +275,14 @@ class SpeechT5FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest
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 2-D numpy arrays are batched.
speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
np_speech_inputs = np.asarray(speech_inputs)
encoded_sequences_1 = feature_extractor(speech_inputs, return_tensors="np").input_values
encoded_sequences_2 = feature_extractor(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_batch_feature_target(self):
speech_inputs = self.feat_extract_tester.prepare_inputs_for_target()
feat_extract = self.feature_extraction_class(**self.feat_extract_dict)

9
tests/models/tvlt/test_feature_extraction_tvlt.py
View File

@@ -189,6 +189,15 @@ class TvltFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.Tes
self.assertTrue(encoded_audios.shape[-2] <= feature_extractor.spectrogram_length)
self.assertTrue(encoded_audios.shape[-3] == feature_extractor.num_channels)
# Test 2-D numpy arrays are batched.
speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
np_speech_inputs = np.asarray(speech_inputs)
encoded_audios = feature_extractor(np_speech_inputs, return_tensors="np", sampling_rate=44100).audio_values
self.assertTrue(encoded_audios.ndim == 4)
self.assertTrue(encoded_audios.shape[-1] == feature_extractor.feature_size)
self.assertTrue(encoded_audios.shape[-2] <= feature_extractor.spectrogram_length)
self.assertTrue(encoded_audios.shape[-3] == feature_extractor.num_channels)
def _load_datasamples(self, num_samples):
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
# automatic decoding with librispeech

8
tests/models/whisper/test_feature_extraction_whisper.py
View File

@@ -173,6 +173,14 @@ class WhisperFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.
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 2-D numpy arrays are batched.
speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
np_speech_inputs = np.asarray(speech_inputs)
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]