[Wav2Vec2] PyCTCDecode Integration to support language model boosted decoding (#14339)
* up * up * up * make it cleaner * correct * make styhahalal * add more tests * finish * small fix * make style * up * tryout to solve cicrle ci * up * fix more tests * fix more tests * apply sylvains suggestions * fix import * correct docs * add pyctcdecode only to speech tests * fix more tests * add tf, flax and pt tests * add pt * fix last tests * fix more tests * Apply suggestions from code review * change lines * Apply suggestions from code review Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com> * correct tests * correct tests * add doc string Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>
This commit is contained in:
Patrick von Platen
committed by
GitHub
GitHub
parent
2e12d90b9e
commit
961732c276
@@ -17,9 +17,19 @@ import math
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
from datasets import load_dataset
|
||||
|
||||
from transformers import Wav2Vec2Config, is_flax_available
|
||||
from transformers.testing_utils import require_datasets, require_flax, require_soundfile, slow
|
||||
from transformers.testing_utils import (
|
||||
is_librosa_available,
|
||||
is_pyctcdecode_available,
|
||||
require_datasets,
|
||||
require_flax,
|
||||
require_librosa,
|
||||
require_pyctcdecode,
|
||||
require_soundfile,
|
||||
slow,
|
||||
)
|
||||
|
||||
from .test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor, random_attention_mask
|
||||
|
||||
@@ -39,6 +49,14 @@ if is_flax_available():
|
||||
)
|
||||
|
||||
|
||||
if is_pyctcdecode_available():
|
||||
from transformers import Wav2Vec2ProcessorWithLM
|
||||
|
||||
|
||||
if is_librosa_available():
|
||||
import librosa
|
||||
|
||||
|
||||
class FlaxWav2Vec2ModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
@@ -354,8 +372,6 @@ class FlaxWav2Vec2UtilsTest(unittest.TestCase):
|
||||
@slow
|
||||
class FlaxWav2Vec2ModelIntegrationTest(unittest.TestCase):
|
||||
def _load_datasamples(self, num_samples):
|
||||
from datasets import load_dataset
|
||||
|
||||
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
|
||||
# automatic decoding with librispeech
|
||||
speech_samples = ds.sort("id").filter(
|
||||
@@ -447,3 +463,22 @@ class FlaxWav2Vec2ModelIntegrationTest(unittest.TestCase):
|
||||
# a random wav2vec2 model has not learned to predict the quantized latent states
|
||||
# => the cosine similarity between quantized states and predicted states is very likely < 0.1
|
||||
self.assertTrue(cosine_sim_masked.mean().item() - 5 * cosine_sim_masked_rand.mean().item() > 0)
|
||||
|
||||
@require_pyctcdecode
|
||||
@require_librosa
|
||||
def test_wav2vec2_with_lm(self):
|
||||
ds = load_dataset("common_voice", "es", split="test", streaming=True)
|
||||
sample = next(iter(ds))
|
||||
|
||||
resampled_audio = librosa.resample(sample["audio"]["array"], 48_000, 16_000)
|
||||
|
||||
model = FlaxWav2Vec2ForCTC.from_pretrained("patrickvonplaten/wav2vec2-large-xlsr-53-spanish-with-lm")
|
||||
processor = Wav2Vec2ProcessorWithLM.from_pretrained("patrickvonplaten/wav2vec2-large-xlsr-53-spanish-with-lm")
|
||||
|
||||
input_values = processor(resampled_audio, return_tensors="np").input_values
|
||||
|
||||
logits = model(input_values).logits
|
||||
|
||||
transcription = processor.batch_decode(np.array(logits)).text
|
||||
|
||||
self.assertEqual(transcription[0], "bien y qué regalo vas a abrir primero")
|
||||
|
||||
@@ -21,9 +21,11 @@ import unittest
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
from datasets import load_dataset
|
||||
|
||||
from transformers import Wav2Vec2Config, is_tf_available
|
||||
from transformers.testing_utils import require_datasets, require_soundfile, require_tf, slow
|
||||
from transformers.file_utils import is_librosa_available, is_pyctcdecode_available
|
||||
from transformers.testing_utils import require_datasets, require_librosa, require_pyctcdecode, require_tf, slow
|
||||
|
||||
from .test_configuration_common import ConfigTester
|
||||
from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
||||
@@ -36,6 +38,14 @@ if is_tf_available():
|
||||
from transformers.models.wav2vec2.modeling_tf_wav2vec2 import _compute_mask_indices
|
||||
|
||||
|
||||
if is_pyctcdecode_available():
|
||||
from transformers import Wav2Vec2ProcessorWithLM
|
||||
|
||||
|
||||
if is_librosa_available():
|
||||
import librosa
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFWav2Vec2ModelTester:
|
||||
def __init__(
|
||||
@@ -474,7 +484,6 @@ class TFWav2Vec2UtilsTest(unittest.TestCase):
|
||||
@require_tf
|
||||
@slow
|
||||
@require_datasets
|
||||
@require_soundfile
|
||||
class TFWav2Vec2ModelIntegrationTest(unittest.TestCase):
|
||||
def _load_datasamples(self, num_samples):
|
||||
from datasets import load_dataset
|
||||
@@ -544,3 +553,22 @@ class TFWav2Vec2ModelIntegrationTest(unittest.TestCase):
|
||||
"his instant panic was followed by a small sharp blow high on his chest",
|
||||
]
|
||||
self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)
|
||||
|
||||
@require_pyctcdecode
|
||||
@require_librosa
|
||||
def test_wav2vec2_with_lm(self):
|
||||
ds = load_dataset("common_voice", "es", split="test", streaming=True)
|
||||
sample = next(iter(ds))
|
||||
|
||||
resampled_audio = librosa.resample(sample["audio"]["array"], 48_000, 16_000)
|
||||
|
||||
model = TFWav2Vec2ForCTC.from_pretrained("patrickvonplaten/wav2vec2-large-xlsr-53-spanish-with-lm")
|
||||
processor = Wav2Vec2ProcessorWithLM.from_pretrained("patrickvonplaten/wav2vec2-large-xlsr-53-spanish-with-lm")
|
||||
|
||||
input_values = processor(resampled_audio, return_tensors="tf").input_values
|
||||
|
||||
logits = model(input_values).logits
|
||||
|
||||
transcription = processor.batch_decode(logits.numpy()).text
|
||||
|
||||
self.assertEqual(transcription[0], "bien y qué regalo vas a abrir primero")
|
||||
|
||||
@@ -18,15 +18,19 @@ import math
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
from datasets import load_dataset
|
||||
|
||||
from tests.test_modeling_common import floats_tensor, ids_tensor, random_attention_mask
|
||||
from transformers import Wav2Vec2Config, is_torch_available
|
||||
from transformers.testing_utils import (
|
||||
is_pt_flax_cross_test,
|
||||
is_pyctcdecode_available,
|
||||
is_torchaudio_available,
|
||||
require_datasets,
|
||||
require_pyctcdecode,
|
||||
require_soundfile,
|
||||
require_torch,
|
||||
require_torchaudio,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@@ -54,6 +58,14 @@ if is_torch_available():
|
||||
)
|
||||
|
||||
|
||||
if is_torchaudio_available():
|
||||
import torchaudio
|
||||
|
||||
|
||||
if is_pyctcdecode_available():
|
||||
from transformers import Wav2Vec2ProcessorWithLM
|
||||
|
||||
|
||||
class Wav2Vec2ModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
@@ -331,7 +343,7 @@ class Wav2Vec2ModelTester:
|
||||
max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths))
|
||||
labels = ids_tensor((input_values.shape[0], max(max_length_labels) - 2), model.config.vocab_size + 100)
|
||||
|
||||
with pytest.raises(ValueError):
|
||||
with self.parent.assertRaises(ValueError):
|
||||
model(input_values, labels=labels)
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
@@ -998,8 +1010,6 @@ class Wav2Vec2UtilsTest(unittest.TestCase):
|
||||
@slow
|
||||
class Wav2Vec2ModelIntegrationTest(unittest.TestCase):
|
||||
def _load_datasamples(self, num_samples):
|
||||
from datasets import load_dataset
|
||||
|
||||
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
|
||||
# automatic decoding with librispeech
|
||||
speech_samples = ds.sort("id").filter(
|
||||
@@ -1009,8 +1019,6 @@ class Wav2Vec2ModelIntegrationTest(unittest.TestCase):
|
||||
return [x["array"] for x in speech_samples]
|
||||
|
||||
def _load_superb(self, task, num_samples):
|
||||
from datasets import load_dataset
|
||||
|
||||
ds = load_dataset("anton-l/superb_dummy", task, split="test")
|
||||
|
||||
return ds[:num_samples]
|
||||
@@ -1337,3 +1345,27 @@ class Wav2Vec2ModelIntegrationTest(unittest.TestCase):
|
||||
|
||||
self.assertListEqual(predicted_ids.tolist(), expected_labels)
|
||||
self.assertTrue(torch.allclose(predicted_logits, expected_logits, atol=1e-2))
|
||||
|
||||
@require_pyctcdecode
|
||||
@require_torchaudio
|
||||
def test_wav2vec2_with_lm(self):
|
||||
ds = load_dataset("common_voice", "es", split="test", streaming=True)
|
||||
sample = next(iter(ds))
|
||||
|
||||
resampled_audio = torchaudio.functional.resample(
|
||||
torch.tensor(sample["audio"]["array"]), 48_000, 16_000
|
||||
).numpy()
|
||||
|
||||
model = Wav2Vec2ForCTC.from_pretrained("patrickvonplaten/wav2vec2-large-xlsr-53-spanish-with-lm").to(
|
||||
torch_device
|
||||
)
|
||||
processor = Wav2Vec2ProcessorWithLM.from_pretrained("patrickvonplaten/wav2vec2-large-xlsr-53-spanish-with-lm")
|
||||
|
||||
input_values = processor(resampled_audio, return_tensors="pt").input_values
|
||||
|
||||
with torch.no_grad():
|
||||
logits = model(input_values.to(torch_device)).logits
|
||||
|
||||
transcription = processor.batch_decode(logits.cpu().numpy()).text
|
||||
|
||||
self.assertEqual(transcription[0], "bien y qué regalo vas a abrir primero")
|
||||
|
||||
236
tests/test_processor_wav2vec2_with_lm.py
Normal file
236
tests/test_processor_wav2vec2_with_lm.py
Normal file
@@ -0,0 +1,236 @@
|
||||
# Copyright 2021 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# 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 json
|
||||
import os
|
||||
import shutil
|
||||
import tempfile
|
||||
import unittest
|
||||
from multiprocessing import Pool
|
||||
|
||||
import numpy as np
|
||||
|
||||
from transformers.file_utils import FEATURE_EXTRACTOR_NAME, is_pyctcdecode_available
|
||||
from transformers.models.wav2vec2 import Wav2Vec2CTCTokenizer, Wav2Vec2FeatureExtractor
|
||||
from transformers.models.wav2vec2.tokenization_wav2vec2 import VOCAB_FILES_NAMES
|
||||
from transformers.testing_utils import require_pyctcdecode
|
||||
|
||||
from .test_feature_extraction_wav2vec2 import floats_list
|
||||
|
||||
|
||||
if is_pyctcdecode_available():
|
||||
from pyctcdecode import BeamSearchDecoderCTC
|
||||
from transformers.models.wav2vec2 import Wav2Vec2ProcessorWithLM
|
||||
|
||||
|
||||
@require_pyctcdecode
|
||||
class Wav2Vec2ProcessorWithLMTest(unittest.TestCase):
|
||||
def setUp(self):
|
||||
vocab = "| <pad> <unk> <s> </s> a b c d e f g h i j k".split()
|
||||
vocab_tokens = dict(zip(vocab, range(len(vocab))))
|
||||
|
||||
self.add_kwargs_tokens_map = {
|
||||
"unk_token": "<unk>",
|
||||
"bos_token": "<s>",
|
||||
"eos_token": "</s>",
|
||||
}
|
||||
feature_extractor_map = {
|
||||
"feature_size": 1,
|
||||
"padding_value": 0.0,
|
||||
"sampling_rate": 16000,
|
||||
"return_attention_mask": False,
|
||||
"do_normalize": True,
|
||||
}
|
||||
|
||||
self.tmpdirname = tempfile.mkdtemp()
|
||||
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
|
||||
self.feature_extraction_file = os.path.join(self.tmpdirname, FEATURE_EXTRACTOR_NAME)
|
||||
with open(self.vocab_file, "w", encoding="utf-8") as fp:
|
||||
fp.write(json.dumps(vocab_tokens) + "\n")
|
||||
|
||||
with open(self.feature_extraction_file, "w", encoding="utf-8") as fp:
|
||||
fp.write(json.dumps(feature_extractor_map) + "\n")
|
||||
|
||||
# load decoder from hub
|
||||
self.decoder_name = "hf-internal-testing/ngram-beam-search-decoder"
|
||||
|
||||
def get_tokenizer(self, **kwargs_init):
|
||||
kwargs = self.add_kwargs_tokens_map.copy()
|
||||
kwargs.update(kwargs_init)
|
||||
return Wav2Vec2CTCTokenizer.from_pretrained(self.tmpdirname, **kwargs)
|
||||
|
||||
def get_feature_extractor(self, **kwargs):
|
||||
return Wav2Vec2FeatureExtractor.from_pretrained(self.tmpdirname, **kwargs)
|
||||
|
||||
def get_decoder(self, **kwargs):
|
||||
return BeamSearchDecoderCTC.load_from_hf_hub(self.decoder_name, **kwargs)
|
||||
|
||||
def tearDown(self):
|
||||
shutil.rmtree(self.tmpdirname)
|
||||
|
||||
def test_save_load_pretrained_default(self):
|
||||
tokenizer = self.get_tokenizer()
|
||||
feature_extractor = self.get_feature_extractor()
|
||||
decoder = self.get_decoder()
|
||||
|
||||
processor = Wav2Vec2ProcessorWithLM(tokenizer=tokenizer, feature_extractor=feature_extractor, decoder=decoder)
|
||||
|
||||
processor.save_pretrained(self.tmpdirname)
|
||||
processor = Wav2Vec2ProcessorWithLM.from_pretrained(self.tmpdirname)
|
||||
|
||||
# tokenizer
|
||||
self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
|
||||
self.assertIsInstance(processor.tokenizer, Wav2Vec2CTCTokenizer)
|
||||
|
||||
# feature extractor
|
||||
self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string())
|
||||
self.assertIsInstance(processor.feature_extractor, Wav2Vec2FeatureExtractor)
|
||||
|
||||
# decoder
|
||||
self.assertEqual(processor.decoder._alphabet.labels, decoder._alphabet.labels)
|
||||
self.assertEqual(
|
||||
processor.decoder.model_container[decoder._model_key]._unigram_set,
|
||||
decoder.model_container[decoder._model_key]._unigram_set,
|
||||
)
|
||||
self.assertIsInstance(processor.decoder, BeamSearchDecoderCTC)
|
||||
|
||||
def test_save_load_pretrained_additional_features(self):
|
||||
processor = Wav2Vec2ProcessorWithLM(
|
||||
tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor(), decoder=self.get_decoder()
|
||||
)
|
||||
processor.save_pretrained(self.tmpdirname)
|
||||
|
||||
# make sure that error is thrown when decoder alphabet doesn't match
|
||||
processor = Wav2Vec2ProcessorWithLM.from_pretrained(
|
||||
self.tmpdirname, alpha=5.0, beta=3.0, score_boundary=-7.0, unk_score_offset=3
|
||||
)
|
||||
|
||||
# decoder
|
||||
self.assertEqual(processor.language_model.alpha, 5.0)
|
||||
self.assertEqual(processor.language_model.beta, 3.0)
|
||||
self.assertEqual(processor.language_model.score_boundary, -7.0)
|
||||
self.assertEqual(processor.language_model.unk_score_offset, 3)
|
||||
|
||||
def test_load_decoder_tokenizer_mismatch_content(self):
|
||||
tokenizer = self.get_tokenizer()
|
||||
# add token to trigger raise
|
||||
tokenizer.add_tokens(["xx"])
|
||||
with self.assertRaisesRegex(ValueError, "include"):
|
||||
Wav2Vec2ProcessorWithLM(
|
||||
tokenizer=tokenizer, feature_extractor=self.get_feature_extractor(), decoder=self.get_decoder()
|
||||
)
|
||||
|
||||
def test_feature_extractor(self):
|
||||
feature_extractor = self.get_feature_extractor()
|
||||
tokenizer = self.get_tokenizer()
|
||||
decoder = self.get_decoder()
|
||||
|
||||
processor = Wav2Vec2ProcessorWithLM(tokenizer=tokenizer, feature_extractor=feature_extractor, decoder=decoder)
|
||||
|
||||
raw_speech = floats_list((3, 1000))
|
||||
|
||||
input_feat_extract = feature_extractor(raw_speech, return_tensors="np")
|
||||
input_processor = processor(raw_speech, return_tensors="np")
|
||||
|
||||
for key in input_feat_extract.keys():
|
||||
self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
|
||||
|
||||
def test_tokenizer(self):
|
||||
feature_extractor = self.get_feature_extractor()
|
||||
tokenizer = self.get_tokenizer()
|
||||
decoder = self.get_decoder()
|
||||
|
||||
processor = Wav2Vec2ProcessorWithLM(tokenizer=tokenizer, feature_extractor=feature_extractor, decoder=decoder)
|
||||
|
||||
input_str = "This is a test string"
|
||||
|
||||
with processor.as_target_processor():
|
||||
encoded_processor = processor(input_str)
|
||||
|
||||
encoded_tok = tokenizer(input_str)
|
||||
|
||||
for key in encoded_tok.keys():
|
||||
self.assertListEqual(encoded_tok[key], encoded_processor[key])
|
||||
|
||||
def _get_dummy_logits(self, shape=(2, 10, 16), seed=77):
|
||||
np.random.seed(seed)
|
||||
return np.random.rand(*shape)
|
||||
|
||||
def test_decoder(self):
|
||||
feature_extractor = self.get_feature_extractor()
|
||||
tokenizer = self.get_tokenizer()
|
||||
decoder = self.get_decoder()
|
||||
|
||||
processor = Wav2Vec2ProcessorWithLM(tokenizer=tokenizer, feature_extractor=feature_extractor, decoder=decoder)
|
||||
|
||||
logits = self._get_dummy_logits(shape=(10, 16), seed=13)
|
||||
|
||||
decoded_processor = processor.decode(logits).text
|
||||
|
||||
decoded_decoder = decoder.decode_beams(logits)[0][0]
|
||||
|
||||
self.assertEqual(decoded_decoder, decoded_processor)
|
||||
self.assertEqual("</s> <s> </s>", decoded_processor)
|
||||
|
||||
def test_decoder_batch(self):
|
||||
feature_extractor = self.get_feature_extractor()
|
||||
tokenizer = self.get_tokenizer()
|
||||
decoder = self.get_decoder()
|
||||
|
||||
processor = Wav2Vec2ProcessorWithLM(tokenizer=tokenizer, feature_extractor=feature_extractor, decoder=decoder)
|
||||
|
||||
logits = self._get_dummy_logits()
|
||||
|
||||
decoded_processor = processor.batch_decode(logits).text
|
||||
|
||||
logits_list = [array for array in logits]
|
||||
decoded_decoder = [d[0][0] for d in decoder.decode_beams_batch(Pool(), logits_list)]
|
||||
|
||||
self.assertListEqual(decoded_decoder, decoded_processor)
|
||||
self.assertListEqual(["<s> <s> </s>", "<s> <s> <s>"], decoded_processor)
|
||||
|
||||
def test_decoder_with_params(self):
|
||||
feature_extractor = self.get_feature_extractor()
|
||||
tokenizer = self.get_tokenizer()
|
||||
decoder = self.get_decoder()
|
||||
|
||||
processor = Wav2Vec2ProcessorWithLM(tokenizer=tokenizer, feature_extractor=feature_extractor, decoder=decoder)
|
||||
|
||||
logits = self._get_dummy_logits()
|
||||
|
||||
beam_width = 20
|
||||
beam_prune_logp = -20.0
|
||||
token_min_logp = -4.0
|
||||
|
||||
decoded_processor_out = processor.batch_decode(
|
||||
logits,
|
||||
beam_width=beam_width,
|
||||
beam_prune_logp=beam_prune_logp,
|
||||
token_min_logp=token_min_logp,
|
||||
)
|
||||
decoded_processor = decoded_processor_out.text
|
||||
|
||||
logits_list = [array for array in logits]
|
||||
decoded_decoder_out = decoder.decode_beams_batch(
|
||||
Pool(),
|
||||
logits_list,
|
||||
beam_width=beam_width,
|
||||
beam_prune_logp=beam_prune_logp,
|
||||
token_min_logp=token_min_logp,
|
||||
)
|
||||
|
||||
decoded_decoder = [d[0][0] for d in decoded_decoder_out]
|
||||
|
||||
self.assertListEqual(decoded_decoder, decoded_processor)
|
||||
self.assertListEqual(["<s> </s> </s>", "<s> <s> </s>"], decoded_processor)
|
||||
Reference in New Issue
Block a user