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

[Wav2Vec2] Add New Wav2Vec2 Translation (#14392)

Browse Source 
* add new wav2vec2 translation

* correct

* up

* add tests

* correct end copy

* correct more

* up

* correct unispeech sat

* finish

* finalize

* finish

* up
This commit is contained in:
Patrick von Platen
2021-11-17 14:38:56 +01:00
committed by GitHub
parent b567510cff
commit 700a748fe6
8 changed files with 544 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
tests/test_modeling_wav2vec2.py
View File

@@ -82,6 +82,8 @@ class Wav2Vec2ModelTester:
mask_time_length=2,
vocab_size=32,
do_stable_layer_norm=False,
num_adapter_layers=1,
adapter_stride=2,
scope=None,
):
self.parent = parent
@@ -107,6 +109,8 @@ class Wav2Vec2ModelTester:
self.initializer_range = initializer_range
self.vocab_size = vocab_size
self.do_stable_layer_norm = do_stable_layer_norm
self.num_adapter_layers = num_adapter_layers
self.adapter_stride = adapter_stride
self.mask_time_prob = mask_time_prob
self.mask_time_length = mask_time_length
self.scope = scope
@@ -117,6 +121,8 @@ class Wav2Vec2ModelTester:
self.output_seq_length = int(math.ceil(output_seq_length))
self.encoder_seq_length = self.output_seq_length
self.adapter_output_seq_length = (self.output_seq_length - 1) // adapter_stride + 1
def prepare_config_and_inputs(self):
input_values = floats_tensor([self.batch_size, self.seq_length], self.vocab_size)
attention_mask = random_attention_mask([self.batch_size, self.seq_length])
@@ -148,6 +154,8 @@ class Wav2Vec2ModelTester:
hidden_act=self.hidden_act,
initializer_range=self.initializer_range,
vocab_size=self.vocab_size,
num_adapter_layers=self.num_adapter_layers,
adapter_stride=self.adapter_stride,
)
def create_and_check_model(self, config, input_values, attention_mask):
@@ -159,6 +167,28 @@ class Wav2Vec2ModelTester:
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
)
def create_and_check_model_with_adapter(self, config, input_values, attention_mask):
config.add_adapter = True
model = Wav2Vec2Model(config=config)
model.to(torch_device)
model.eval()
result = model(input_values, attention_mask=attention_mask)
self.parent.assertEqual(
result.last_hidden_state.shape, (self.batch_size, self.adapter_output_seq_length, self.hidden_size)
)
def create_and_check_model_with_adapter_proj_dim(self, config, input_values, attention_mask):
config.add_adapter = True
config.output_hidden_size = 8
model = Wav2Vec2Model(config=config)
model.to(torch_device)
model.eval()
result = model(input_values, attention_mask=attention_mask)
self.parent.assertEqual(
result.last_hidden_state.shape,
(self.batch_size, self.adapter_output_seq_length, config.output_hidden_size),
)
def create_and_check_batch_inference(self, config, input_values, *args):
# test does not pass for models making use of `group_norm`
# check: https://github.com/pytorch/fairseq/issues/3227
@@ -332,6 +362,14 @@ class Wav2Vec2ModelTest(ModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_with_adapter(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model_with_adapter(*config_and_inputs)
def test_model_with_adapter_proj_dim(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model_with_adapter_proj_dim(*config_and_inputs)
def test_ctc_loss_inference(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.check_ctc_loss(*config_and_inputs)
@@ -544,6 +582,14 @@ class Wav2Vec2RobustModelTest(ModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_with_adapter(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model_with_adapter(*config_and_inputs)
def test_model_with_adapter_proj_dim(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model_with_adapter_proj_dim(*config_and_inputs)
def test_batched_inference(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_batch_inference(*config_and_inputs)

38
tests/test_pipelines_automatic_speech_recognition.py
View File

@@ -203,3 +203,41 @@ class AutomaticSpeechRecognitionPipelineTests(unittest.TestCase, metaclass=Pipel
data = f.read()
output = asr(data)
self.assertEqual(output, {"text": "Un uomo disse all'universo: \"Signore, io esisto."})
@slow
@require_torch
@require_torchaudio
@require_datasets
def test_xls_r_to_en(self):
speech_recognizer = pipeline(
task="automatic-speech-recognition",
model="facebook/wav2vec2-xls-r-1b-21-to-en",
feature_extractor="facebook/wav2vec2-xls-r-1b-21-to-en",
framework="pt",
)
from datasets import load_dataset
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation").sort("id")
filename = ds[40]["file"]
output = speech_recognizer(filename)
self.assertEqual(output, {"text": "A man said to the universe: “Sir, I exist."})
@slow
@require_torch
@require_torchaudio
@require_datasets
def test_xls_r_from_en(self):
speech_recognizer = pipeline(
task="automatic-speech-recognition",
model="facebook/wav2vec2-xls-r-1b-en-to-15",
feature_extractor="facebook/wav2vec2-xls-r-1b-en-to-15",
framework="pt",
)
from datasets import load_dataset
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation").sort("id")
filename = ds[40]["file"]
output = speech_recognizer(filename)
self.assertEqual(output, {"text": "Ein Mann sagte zu dem Universum, Sir, ich bin da."})