def convert_wav2vec2_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 config_path is not None:
config = Wav2Vec2Config.from_pretrained(config_path)
else:
config = Wav2Vec2Config()
if is_finetuned:
hf_wav2vec = Wav2Vec2ForCTC(config)
else:
hf_wav2vec = Wav2Vec2Model(config)
if is_finetuned:
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path], arg_overrides={"data": dict_path})
else:
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path])
model = model[0].eval()
recursively_load_weights(model, hf_wav2vec, is_finetuned)
hf_wav2vec.save_pretrained(pytorch_dump_folder_path)
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])
config = Wav2Vec2Config(
do_stable_layer_norm=self.do_stable_layer_norm,
hidden_size=self.hidden_size,
feat_extract_norm=self.feat_extract_norm,
feat_extract_dropout=self.feat_extract_dropout,
feat_extract_activation=self.feat_extract_activation,
conv_dim=self.conv_dim,
conv_stride=self.conv_stride,
conv_kernel=self.conv_kernel,
conv_bias=self.conv_bias,
num_conv_pos_embeddings=self.num_conv_pos_embeddings,
num_conv_pos_embedding_groups=self.num_conv_pos_embedding_groups,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
hidden_dropout_prob=self.hidden_dropout_prob,
intermediate_size=self.intermediate_size,
layer_norm_eps=self.layer_norm_eps,
hidden_act=self.hidden_act,
initializer_range=self.initializer_range,
vocab_size=self.vocab_size,
)
return config, input_values, attention_mask
def test_processor_from_local_directory_from_repo(self):
with tempfile.TemporaryDirectory() as tmpdirname:
model_config = Wav2Vec2Config()
processor = AutoProcessor.from_pretrained("facebook/wav2vec2-base-960h")
# save in new folder
model_config.save_pretrained(tmpdirname)
processor.save_pretrained(tmpdirname)
processor = AutoProcessor.from_pretrained(tmpdirname)
self.assertIsInstance(processor, Wav2Vec2Processor)
def test_processor_from_local_directory_from_model_config(self):
with tempfile.TemporaryDirectory() as tmpdirname:
model_config = Wav2Vec2Config(processor_class="Wav2Vec2Processor")
model_config.save_pretrained(tmpdirname)
# copy relevant files
copyfile(SAMPLE_VOCAB, os.path.join(tmpdirname, "vocab.json"))
# create emtpy sample processor
with open(os.path.join(tmpdirname, FEATURE_EXTRACTOR_NAME), "w") as f:
f.write("{}")
processor = AutoProcessor.from_pretrained(tmpdirname)
self.assertIsInstance(processor, Wav2Vec2Processor)
def convert_wav2vec2_checkpoint(checkpoint_path, pytorch_dump_folder_path, dict_path=None):
"""
Copy/paste/tweak model's weights to transformers design.
"""
hf_wav2vec = Wav2Vec2ForCTC(Wav2Vec2Config())
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path], arg_overrides={"data": dict_path}
)
model = model[0].eval()
recursively_load_weights(model, hf_wav2vec)
hf_wav2vec.save_pretrained(pytorch_dump_folder_path)
def test_feature_extractor_from_local_directory_from_config(self):
with tempfile.TemporaryDirectory() as tmpdirname:
model_config = Wav2Vec2Config()
# remove feature_extractor_type to make sure config.json alone is enough to load feature processor locally
config_dict = AutoFeatureExtractor.from_pretrained(
SAMPLE_FEATURE_EXTRACTION_CONFIG_DIR).to_dict()
config_dict.pop("feature_extractor_type")
config = Wav2Vec2FeatureExtractor(config_dict)
# save in new folder
model_config.save_pretrained(tmpdirname)
config.save_pretrained(tmpdirname)
config = AutoFeatureExtractor.from_pretrained(tmpdirname)
self.assertIsInstance(config, Wav2Vec2FeatureExtractor)
def get_config(self):
return Wav2Vec2Config(
hidden_size=self.hidden_size,
feat_extract_norm=self.feat_extract_norm,
feat_extract_dropout=self.feat_extract_dropout,
feat_extract_activation=self.feat_extract_activation,
conv_dim=self.conv_dim,
conv_stride=self.conv_stride,
conv_kernel=self.conv_kernel,
conv_bias=self.conv_bias,
num_conv_pos_embeddings=self.num_conv_pos_embeddings,
num_conv_pos_embedding_groups=self.num_conv_pos_embedding_groups,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
hidden_dropout_prob=self.hidden_dropout_prob,
intermediate_size=self.intermediate_size,
layer_norm_eps=self.layer_norm_eps,
hidden_act=self.hidden_act,
initializer_range=self.initializer_range,
vocab_size=self.vocab_size,
)
def test_feature_extractor_from_local_directory_from_config(self):
with tempfile.TemporaryDirectory() as tmpdirname:
model_config = Wav2Vec2Config()
# remove feature_extractor_type to make sure config.json alone is enough to load feature processor locally
config_dict = AutoFeatureExtractor.from_pretrained(SAMPLE_FEATURE_EXTRACTION_CONFIG_DIR).to_dict()
config_dict.pop("feature_extractor_type")
config = Wav2Vec2FeatureExtractor(**config_dict)
# save in new folder
model_config.save_pretrained(tmpdirname)
config.save_pretrained(tmpdirname)
config = AutoFeatureExtractor.from_pretrained(tmpdirname)
# make sure private variable is not incorrectly saved
dict_as_saved = json.loads(config.to_json_string())
self.assertTrue("_processor_class" not in dict_as_saved)
self.assertIsInstance(config, Wav2Vec2FeatureExtractor)
def convert_wav2vec2_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 config_path is not None:
config = Wav2Vec2Config.from_pretrained(config_path)
else:
config = Wav2Vec2Config()
if is_finetuned:
if dict_path:
target_dict = Dictionary.load(dict_path)
config.bos_token_id = target_dict.bos_index
config.eos_token_id = target_dict.eos_index
config.pad_token_id = target_dict.pad_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,
)
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,
)
processor = Wav2Vec2Processor(feature_extractor=feature_extractor,
tokenizer=tokenizer)
processor.save_pretrained(pytorch_dump_folder_path)
hf_wav2vec = Wav2Vec2ForCTC(config)
else:
hf_wav2vec = Wav2Vec2Model(config)
if is_finetuned:
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path], arg_overrides={"data": dict_path})
else:
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path])
model = model[0].eval()
recursively_load_weights(model, hf_wav2vec, is_finetuned)
hf_wav2vec.save_pretrained(pytorch_dump_folder_path)