def test_save_load_pretrained_additional_features(self):
processor = MCTCTProcessor(
tokenizer=self.get_tokenizer(),
feature_extractor=self.get_feature_extractor())
processor.save_pretrained(self.tmpdirname)
tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)",
eos_token="(EOS)")
feature_extractor_add_kwargs = self.get_feature_extractor(
do_normalize=False, padding_value=1.0)
processor = MCTCTProcessor.from_pretrained(self.tmpdirname,
bos_token="(BOS)",
eos_token="(EOS)",
do_normalize=False,
padding_value=1.0)
self.assertEqual(processor.tokenizer.get_vocab(),
tokenizer_add_kwargs.get_vocab())
self.assertIsInstance(processor.tokenizer, Wav2Vec2CTCTokenizer)
self.assertEqual(processor.feature_extractor.to_json_string(),
feature_extractor_add_kwargs.to_json_string())
self.assertIsInstance(processor.feature_extractor,
MCTCTFeatureExtractor)
def test_tokenizer_decode(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = MCTCTProcessor(tokenizer=tokenizer,
feature_extractor=feature_extractor)
predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
decoded_processor = processor.batch_decode(predicted_ids)
decoded_tok = tokenizer.batch_decode(predicted_ids)
self.assertListEqual(decoded_tok, decoded_processor)
def test_inference_ctc_robust_batched(self):
model = MCTCTForCTC.from_pretrained("speechbrain/m-ctc-t-large").to(
torch_device)
processor = MCTCTProcessor.from_pretrained("speechbrain/m-ctc-t-large",
do_lower_case=True)
input_speech = self._load_datasamples(4)
inputs = processor(input_speech,
return_tensors="pt",
padding=True,
return_attention_mask=True)
input_features = inputs.input_features.to(torch_device)
attention_mask = inputs.attention_mask.to(torch_device)
with torch.no_grad():
logits = model(input_features,
attention_mask=attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
predicted_trans = processor.batch_decode(predicted_ids)
EXPECTED_TRANSCRIPTIONS = [
"a man said to the universe, sir, i exist.",
'"sweat-covered brion\'s body, trickling into the tight-lowing clossa was the only germent huor." "',
"\"the cadona's chest still-dripping bloodthe acofis overstrained eyes, even the soring arena around him"
" with thousands of spectators retrivialities not worth-thinking about.",
"his instant panic was followed by a small sharp blow high on his chestr.",
]
self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)
def test_inference_ctc_normal_batched(self):
model = MCTCTForCTC.from_pretrained("speechbrain/m-ctc-t-large")
model.to(torch_device)
processor = MCTCTProcessor.from_pretrained("speechbrain/m-ctc-t-large",
do_lower_case=True)
input_speech = self._load_datasamples(2)
inputs = processor(input_speech, return_tensors="pt", padding=True)
input_features = inputs.input_features.to(torch_device)
attention_mask = inputs.attention_mask.to(torch_device)
with torch.no_grad():
logits = model(input_features,
attention_mask=attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
predicted_trans = processor.batch_decode(predicted_ids)
EXPECTED_TRANSCRIPTIONS = [
"a man said to the universe, sir, i exist.",
'"sweat-covered brion\'s body, trickling into the tight-lowing clossa was the only germent huor."',
]
self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)
def test_tokenizer(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = MCTCTProcessor(tokenizer=tokenizer,
feature_extractor=feature_extractor)
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 test_save_load_pretrained_default(self):
tokenizer = self.get_tokenizer()
feature_extractor = self.get_feature_extractor()
processor = MCTCTProcessor(tokenizer=tokenizer,
feature_extractor=feature_extractor)
processor.save_pretrained(self.tmpdirname)
processor = MCTCTProcessor.from_pretrained(self.tmpdirname)
self.assertEqual(processor.tokenizer.get_vocab(),
tokenizer.get_vocab())
self.assertIsInstance(processor.tokenizer, Wav2Vec2CTCTokenizer)
self.assertEqual(processor.feature_extractor.to_json_string(),
feature_extractor.to_json_string())
self.assertIsInstance(processor.feature_extractor,
MCTCTFeatureExtractor)
def test_feature_extractor(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = MCTCTProcessor(tokenizer=tokenizer,
feature_extractor=feature_extractor)
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_inference_ctc_normal(self):
model = MCTCTForCTC.from_pretrained("speechbrain/m-ctc-t-large")
model.to(torch_device)
processor = MCTCTProcessor.from_pretrained("speechbrain/m-ctc-t-large",
do_lower_case=True)
input_speech = self._load_datasamples(1)
input_features = processor(
input_speech, return_tensors="pt").input_features.to(torch_device)
with torch.no_grad():
logits = model(input_features).logits
predicted_ids = torch.argmax(logits, dim=-1)
predicted_trans = processor.batch_decode(predicted_ids)
EXPECTED_TRANSCRIPTIONS = ["a man said to the universe, sir, i exist."]
self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)