def test_inference_pretrained_batched(self):
model = SEWDModel.from_pretrained("asapp/sew-d-tiny-100k").to(torch_device)
processor = Wav2Vec2FeatureExtractor.from_pretrained("asapp/sew-d-tiny-100k")
input_speech = self._load_datasamples(2)
inputs = processor(input_speech, return_tensors="pt", padding=True)
input_values = inputs.input_values.to(torch_device)
with torch.no_grad():
outputs = model(input_values).last_hidden_state
# expected outputs taken from the original SEW-D implementation
expected_outputs_first = torch.tensor(
[
[
[-0.1619, 0.6995, 0.4062, -0.1014],
[-0.1364, 0.5960, 0.0952, -0.0873],
[-0.1572, 0.5718, 0.4228, -0.0864],
[-0.1325, 0.6823, 0.1387, -0.0871],
],
[
[-0.1296, 0.4008, 0.4952, -0.1450],
[-0.1152, 0.3693, 0.3037, -0.1290],
[-0.1194, 0.6074, 0.3531, -0.1466],
[-0.1113, 0.3135, 0.2224, -0.1338],
],
],
device=torch_device,
)
expected_outputs_last = torch.tensor(
[
[
[-0.1577, 0.5108, 0.8553, 0.2550],
[-0.1530, 0.3580, 0.6143, 0.2672],
[-0.1535, 0.4954, 0.8503, 0.1387],
[-0.1572, 0.3363, 0.6217, 0.1490],
],
[
[-0.1338, 0.5459, 0.9607, -0.1133],
[-0.1502, 0.3738, 0.7313, -0.0986],
[-0.0953, 0.4708, 1.0821, -0.0944],
[-0.1474, 0.3598, 0.7248, -0.0748],
],
],
device=torch_device,
)
expected_output_sum = 54201.0469
self.assertTrue(torch.allclose(outputs[:, :4, :4], expected_outputs_first, atol=1e-3))
self.assertTrue(torch.allclose(outputs[:, -4:, -4:], expected_outputs_last, atol=1e-3))
self.assertTrue(abs(outputs.sum() - expected_output_sum) < 1)
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
model = SEWDModel(config=config)
model.to(torch_device)
model.eval()
input_values = input_values[:3]
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
# pad input
for i in range(len(input_lengths)):
input_values[i, input_lengths[i] :] = 0.0
attention_mask[i, input_lengths[i] :] = 0.0
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
for i in range(input_values.shape[0]):
input_slice = input_values[i : i + 1, : input_lengths[i]]
output = model(input_slice).last_hidden_state
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
def create_and_check_model(self, config, input_values, attention_mask):
model = SEWDModel(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.output_seq_length, self.hidden_size)
)
def test_model_from_pretrained(self):
model = SEWDModel.from_pretrained("asapp/sew-d-tiny-100k")
self.assertIsNotNone(model)
def convert_sew_checkpoint(checkpoint_path,
pytorch_dump_folder_path,
config_path=None,
dict_path=None,
is_finetuned=True):
"""
Copy/paste/tweak model's weights to transformers design.
"""
if is_finetuned:
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path],
arg_overrides={"data": "/".join(dict_path.split("/")[:-1])})
else:
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path])
if config_path is not None:
config = SEWDConfig.from_pretrained(config_path)
else:
config = convert_config(model[0])
model = model[0].eval()
return_attention_mask = True if config.feat_extract_norm == "layer" else False
feature_extractor = Wav2Vec2FeatureExtractor(
feature_size=1,
sampling_rate=16000,
padding_value=0,
do_normalize=True,
return_attention_mask=return_attention_mask,
)
if is_finetuned:
if dict_path:
target_dict = Dictionary.load(dict_path)
# important change bos & pad token id since CTC symbol is <pad> and
# not <s> as in fairseq
config.bos_token_id = target_dict.pad_index
config.pad_token_id = target_dict.bos_index
config.eos_token_id = target_dict.eos_index
config.vocab_size = len(target_dict.symbols)
vocab_path = os.path.join(pytorch_dump_folder_path, "vocab.json")
if not os.path.isdir(pytorch_dump_folder_path):
logger.error(
"--pytorch_dump_folder_path ({}) should be a directory".
format(pytorch_dump_folder_path))
return
os.makedirs(pytorch_dump_folder_path, exist_ok=True)
with open(vocab_path, "w", encoding="utf-8") as vocab_handle:
json.dump(target_dict.indices, vocab_handle)
tokenizer = Wav2Vec2CTCTokenizer(
vocab_path,
unk_token=target_dict.unk_word,
pad_token=target_dict.pad_word,
bos_token=target_dict.bos_word,
eos_token=target_dict.eos_word,
word_delimiter_token="|",
do_lower_case=False,
)
processor = Wav2Vec2Processor(feature_extractor=feature_extractor,
tokenizer=tokenizer)
processor.save_pretrained(pytorch_dump_folder_path)
hf_model = SEWDForCTC(config)
else:
hf_model = SEWDModel(config)
feature_extractor.save_pretrained(pytorch_dump_folder_path)
recursively_load_weights(model, hf_model, is_finetuned)
hf_model.save_pretrained(pytorch_dump_folder_path)