def check_encoder_decoder_model_generate(self,
config,
decoder_config,
input_values=None,
input_features=None,
**kwargs):
encoder_model, decoder_model = self.get_encoder_decoder_model(
config, decoder_config)
enc_dec_model = SpeechEncoderDecoderModel(encoder=encoder_model,
decoder=decoder_model)
enc_dec_model.to(torch_device)
# make sure EOS token is set to None to prevent early stopping of generation
enc_dec_model.config.eos_token_id = None
if hasattr(enc_dec_model.config, "decoder") and hasattr(
enc_dec_model.config.decoder, "eos_token_id"):
enc_dec_model.config.decoder.eos_token_id = None
inputs = input_values if input_features is None else input_features
# Bert does not have a bos token id, so use pad_token_id instead
generated_output = enc_dec_model.generate(
inputs,
decoder_start_token_id=enc_dec_model.config.decoder.pad_token_id)
self.assertEqual(generated_output.shape,
(inputs.shape[0], ) + (decoder_config.max_length, ))
def check_encoder_decoder_model_output_attentions(self,
config,
attention_mask,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
labels=None,
input_values=None,
input_features=None,
**kwargs):
# make the decoder inputs a different shape from the encoder inputs to harden the test
decoder_input_ids = decoder_input_ids[:, :-1]
decoder_attention_mask = decoder_attention_mask[:, :-1]
encoder_model, decoder_model = self.get_encoder_decoder_model(
config, decoder_config)
enc_dec_model = SpeechEncoderDecoderModel(encoder=encoder_model,
decoder=decoder_model)
enc_dec_model.to(torch_device)
outputs_encoder_decoder = enc_dec_model(
input_values=input_values,
input_features=input_features,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
output_attentions=True,
)
inputs = input_values if input_features is None else input_features
encoder_attentions = outputs_encoder_decoder["encoder_attentions"]
self.assertEqual(len(encoder_attentions), config.num_hidden_layers)
seq_len = enc_dec_model.encoder._get_feat_extract_output_lengths(
inputs.shape[1])
self.assertEqual(encoder_attentions[0].shape[-3:],
(config.num_attention_heads, seq_len, seq_len))
decoder_attentions = outputs_encoder_decoder["decoder_attentions"]
num_decoder_layers = (decoder_config.num_decoder_layers if hasattr(
decoder_config, "num_decoder_layers") else
decoder_config.num_hidden_layers)
self.assertEqual(len(decoder_attentions), num_decoder_layers)
self.assertEqual(
decoder_attentions[0].shape[-3:],
(decoder_config.num_attention_heads, decoder_input_ids.shape[-1],
decoder_input_ids.shape[-1]),
)
cross_attentions = outputs_encoder_decoder["cross_attentions"]
self.assertEqual(len(cross_attentions), num_decoder_layers)
cross_attention_input_seq_len = decoder_input_ids.shape[-1]
self.assertEqual(
cross_attentions[0].shape[-3:],
(decoder_config.num_attention_heads, cross_attention_input_seq_len,
seq_len),
)
def check_equivalence_pt_to_flax(self, config, decoder_config, inputs_dict):
encoder_decoder_config = SpeechEncoderDecoderConfig.from_encoder_decoder_configs(config, decoder_config)
pt_model = SpeechEncoderDecoderModel(encoder_decoder_config)
fx_model = FlaxSpeechEncoderDecoderModel(encoder_decoder_config)
fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
fx_model.params = fx_state
self.check_pt_flax_equivalence(pt_model, fx_model, inputs_dict)
def check_encoder_decoder_model_generate(
self, config, decoder_config, input_values=None, input_features=None, **kwargs
):
encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
enc_dec_model = SpeechEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
enc_dec_model.to(torch_device)
inputs = input_values if input_features is None else input_features
# Bert does not have a bos token id, so use pad_token_id instead
generated_output = enc_dec_model.generate(
inputs, decoder_start_token_id=enc_dec_model.config.decoder.pad_token_id
)
self.assertEqual(generated_output.shape, (inputs.shape[0],) + (decoder_config.max_length,))
def check_encoder_decoder_model_from_pretrained(self,
config,
attention_mask,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
return_dict,
input_values=None,
input_features=None,
**kwargs):
encoder_model, decoder_model = self.get_encoder_decoder_model(
config, decoder_config)
kwargs = {
"encoder_model": encoder_model,
"decoder_model": decoder_model,
"return_dict": return_dict
}
enc_dec_model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
**kwargs)
enc_dec_model.to(torch_device)
outputs_encoder_decoder = enc_dec_model(
input_values=input_values,
input_features=input_features,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
output_hidden_states=True,
return_dict=True,
)
self.assertEqual(outputs_encoder_decoder["logits"].shape,
(decoder_input_ids.shape +
(decoder_config.vocab_size, )))
def test_real_model_save_load_from_pretrained(self):
model_2 = self.get_pretrained_model()
model_2.to(torch_device)
input_name, inputs = self.get_inputs()
decoder_input_ids = ids_tensor([13, 1], model_2.config.encoder.vocab_size)
attention_mask = ids_tensor([13, 5], vocab_size=2)
with torch.no_grad():
outputs = model_2(
**{input_name: inputs},
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
)
out_2 = outputs[0].cpu().numpy()
out_2[np.isnan(out_2)] = 0
with tempfile.TemporaryDirectory() as tmp_dirname:
model_2.save_pretrained(tmp_dirname)
model_1 = SpeechEncoderDecoderModel.from_pretrained(tmp_dirname)
model_1.to(torch_device)
after_outputs = model_1(
**{input_name: inputs},
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
)
out_1 = after_outputs[0].cpu().numpy()
out_1[np.isnan(out_1)] = 0
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def check_pt_flax_equivalence(self, pt_model, fx_model, inputs_dict):
pt_model.to(torch_device)
pt_model.eval()
# prepare inputs
flax_inputs = inputs_dict
pt_inputs = {
k: torch.tensor(v.tolist())
for k, v in flax_inputs.items()
}
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs)
pt_logits = pt_outputs.logits
pt_outputs = pt_outputs.to_tuple()
fx_outputs = fx_model(**inputs_dict)
fx_logits = fx_outputs.logits
fx_outputs = fx_outputs.to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits, pt_logits.numpy(), 4e-2)
# PT -> Flax
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
fx_model_loaded = FlaxSpeechEncoderDecoderModel.from_pretrained(
tmpdirname, from_pt=True)
fx_outputs_loaded = fx_model_loaded(**inputs_dict)
fx_logits_loaded = fx_outputs_loaded.logits
fx_outputs_loaded = fx_outputs_loaded.to_tuple()
self.assertEqual(len(fx_outputs_loaded), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits_loaded, pt_logits.numpy(), 4e-2)
# Flax -> PT
with tempfile.TemporaryDirectory() as tmpdirname:
fx_model.save_pretrained(tmpdirname)
pt_model_loaded = SpeechEncoderDecoderModel.from_pretrained(
tmpdirname, from_flax=True)
pt_model_loaded.to(torch_device)
pt_model_loaded.eval()
with torch.no_grad():
pt_outputs_loaded = pt_model_loaded(**pt_inputs)
pt_logits_loaded = pt_outputs_loaded.logits
pt_outputs_loaded = pt_outputs_loaded.to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs_loaded),
"Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits, pt_logits_loaded.numpy(), 4e-2)
def check_equivalence_flax_to_pt(self, config, decoder_config, inputs_dict):
encoder_decoder_config = SpeechEncoderDecoderConfig.from_encoder_decoder_configs(config, decoder_config)
pt_model = SpeechEncoderDecoderModel(encoder_decoder_config)
fx_model = FlaxSpeechEncoderDecoderModel(encoder_decoder_config)
pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)
self.check_pt_flax_equivalence(pt_model, fx_model, inputs_dict)
def check_encoder_decoder_model(self,
config,
attention_mask,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
input_values=None,
input_features=None,
**kwargs):
encoder_model, decoder_model = self.get_encoder_decoder_model(
config, decoder_config)
enc_dec_model = SpeechEncoderDecoderModel(encoder=encoder_model,
decoder=decoder_model)
self.assertTrue(enc_dec_model.config.decoder.is_decoder)
self.assertTrue(enc_dec_model.config.decoder.add_cross_attention)
self.assertTrue(enc_dec_model.config.is_encoder_decoder)
enc_dec_model.to(torch_device)
outputs_encoder_decoder = enc_dec_model(
input_values=input_values,
input_features=input_features,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
output_hidden_states=True,
)
self.assertEqual(outputs_encoder_decoder["logits"].shape,
(decoder_input_ids.shape +
(decoder_config.vocab_size, )))
encoder_outputs = BaseModelOutput(
last_hidden_state=outputs_encoder_decoder.encoder_hidden_states[-1]
)
outputs_encoder_decoder = enc_dec_model(
encoder_outputs=encoder_outputs,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
)
self.assertEqual(outputs_encoder_decoder["logits"].shape,
(decoder_input_ids.shape +
(decoder_config.vocab_size, )))
def check_encoder_decoder_model_from_pretrained_configs(
self,
config,
attention_mask,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
input_values=None,
input_features=None,
**kwargs):
encoder_decoder_config = SpeechEncoderDecoderConfig.from_encoder_decoder_configs(
config, decoder_config)
self.assertTrue(encoder_decoder_config.decoder.is_decoder)
enc_dec_model = SpeechEncoderDecoderModel(encoder_decoder_config)
enc_dec_model.to(torch_device)
enc_dec_model.eval()
self.assertTrue(enc_dec_model.config.is_encoder_decoder)
outputs_encoder_decoder = enc_dec_model(
input_values=input_values,
input_features=input_features,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
)
self.assertEqual(outputs_encoder_decoder["logits"].shape,
(decoder_input_ids.shape +
(decoder_config.vocab_size, )))
def check_save_and_load_encoder_decoder_model(self,
config,
attention_mask,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
input_values=None,
input_features=None,
**kwargs):
encoder_model, decoder_model = self.get_encoder_decoder_model(
config, decoder_config)
enc_dec_model = SpeechEncoderDecoderModel(encoder=encoder_model,
decoder=decoder_model)
enc_dec_model.to(torch_device)
enc_dec_model.eval()
with torch.no_grad():
outputs = enc_dec_model(
input_values=input_values,
input_features=input_features,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
)
out_2 = outputs[0].cpu().numpy()
out_2[np.isnan(out_2)] = 0
with tempfile.TemporaryDirectory(
) as encoder_tmp_dirname, tempfile.TemporaryDirectory(
) as decoder_tmp_dirname:
enc_dec_model.encoder.save_pretrained(encoder_tmp_dirname)
enc_dec_model.decoder.save_pretrained(decoder_tmp_dirname)
SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
encoder_pretrained_model_name_or_path=encoder_tmp_dirname,
decoder_pretrained_model_name_or_path=decoder_tmp_dirname,
)
after_outputs = enc_dec_model(
input_values=input_values,
input_features=input_features,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
)
out_1 = after_outputs[0].cpu().numpy()
out_1[np.isnan(out_1)] = 0
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def check_encoder_decoder_model_with_inputs(self,
config,
attention_mask,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
input_values=None,
input_features=None,
**kwargs):
inputs = input_values if input_features is None else input_features
encoder_model, decoder_model = self.get_encoder_decoder_model(
config, decoder_config)
enc_dec_model = SpeechEncoderDecoderModel(encoder=encoder_model,
decoder=decoder_model)
enc_dec_model.to(torch_device)
outputs_encoder_decoder = enc_dec_model(
inputs,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
output_hidden_states=True,
)
self.assertEqual(outputs_encoder_decoder["logits"].shape,
(decoder_input_ids.shape +
(decoder_config.vocab_size, )))
outputs_encoder_decoder_kwarg = enc_dec_model(
inputs=inputs,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
output_hidden_states=True,
)
self.assertEqual(outputs_encoder_decoder_kwarg["logits"].shape,
(decoder_input_ids.shape +
(decoder_config.vocab_size, )))
def get_pretrained_model_and_inputs(self):
model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
"facebook/s2t-small-librispeech-asr", "bert-base-cased"
)
batch_size = 13
input_features = floats_tensor([batch_size, 7, 80], scale=1.0)
attention_mask = random_attention_mask([batch_size, 7])
decoder_input_ids = ids_tensor([batch_size, 4], model.decoder.config.vocab_size)
decoder_attention_mask = random_attention_mask([batch_size, 4])
inputs = {
"input_features": input_features,
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
}
return model, inputs
def get_pretrained_model_and_inputs(self):
model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
"facebook/wav2vec2-base-960h", "bert-base-cased"
)
batch_size = 13
input_values = floats_tensor([batch_size, 512], scale=1.0)
attention_mask = random_attention_mask([batch_size, 512])
decoder_input_ids = ids_tensor([batch_size, 4], model.decoder.config.vocab_size)
decoder_attention_mask = random_attention_mask([batch_size, 4])
inputs = {
"input_values": input_values,
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
}
return model, inputs
def test_real_model_save_load_from_pretrained(self):
model_2, inputs = self.get_pretrained_model_and_inputs()
model_2.to(torch_device)
with torch.no_grad():
outputs = model_2(**inputs)
out_2 = outputs[0].cpu().numpy()
out_2[np.isnan(out_2)] = 0
with tempfile.TemporaryDirectory() as tmp_dirname:
model_2.save_pretrained(tmp_dirname)
model_1 = SpeechEncoderDecoderModel.from_pretrained(tmp_dirname)
model_1.to(torch_device)
after_outputs = model_1(**inputs)
out_1 = after_outputs[0].cpu().numpy()
out_1[np.isnan(out_1)] = 0
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def convert_wav2vec2_checkpoint(
checkpoint_path,
pytorch_dump_folder_path,
dict_path,
encoder_config_path,
decoder_config_path,
vocab_size,
num_decoder_layers,
):
"""
Copy/paste/tweak model's weights to transformers design.
"""
encoder_config = Wav2Vec2Config.from_pretrained(encoder_config_path)
decoder_config = Speech2Text2Config.from_pretrained(
decoder_config_path, vocab_size=vocab_size, decoder_layers=num_decoder_layers, do_stable_layer_norm=True
)
feature_extractor = Wav2Vec2FeatureExtractor(
feature_size=1,
sampling_rate=16000,
padding_value=0,
do_normalize=True,
return_attention_mask=True,
)
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path], arg_overrides={"data": "/".join(dict_path.split("/")[:-1])}
)
model = model[0].eval()
# set weights for wav2vec2 encoder
hf_encoder = Wav2Vec2Model(encoder_config)
projection_layer = recursively_load_weights_wav2vec2(model.encoder, hf_encoder)
hf_decoder = Speech2Text2ForCausalLM(decoder_config)
missing_keys, unexpected_keys = hf_decoder.model.decoder.load_state_dict(model.decoder.state_dict(), strict=False)
# set output linear layer
unexpected_keys.remove("embed_out")
hf_decoder.lm_head.weight = nn.Parameter(model.decoder.embed_out.detach())
# layer norm is init to identity matrix so leaving it is fine
logger.warning(f"The following keys are missing when loading the decoder weights: {missing_keys}")
logger.warning(f"The following keys are unexpected when loading the decoder weights: {unexpected_keys}")
hf_wav2vec = SpeechEncoderDecoderModel(encoder=hf_encoder, decoder=hf_decoder)
hf_wav2vec.config.tie_word_embeddings = False
# add projection layer
hf_wav2vec.enc_to_dec_proj.weight = nn.Parameter(projection_layer.weight)
hf_wav2vec.enc_to_dec_proj.bias = nn.Parameter(projection_layer.bias)
vocab_dict = create_vocab_dict(dict_path)
with open(os.path.join(pytorch_dump_folder_path, "vocab.json"), "w") as fp:
json.dump(vocab_dict, fp)
tokenizer = Speech2Text2Tokenizer(os.path.join(pytorch_dump_folder_path, "vocab.json"))
tokenizer.save_pretrained(pytorch_dump_folder_path)
config = hf_wav2vec.config.to_dict()
config["pad_token_id"] = tokenizer.pad_token_id
config["bos_token_id"] = tokenizer.bos_token_id
config["eos_token_id"] = tokenizer.eos_token_id
config["tokenizer_class"] = "speech_to_text_2"
config["feature_extractor_type"] = "wav2vec2"
hf_wav2vec.config = SpeechEncoderDecoderConfig.from_dict(config)
hf_wav2vec.save_pretrained(pytorch_dump_folder_path)
feature_extractor.save_pretrained(pytorch_dump_folder_path)
def get_pretrained_model(self):
return SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
"facebook/wav2vec2-base-960h", "bert-base-cased"
)
def get_pretrained_model(self):
return SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
"facebook/s2t-small-librispeech-asr", "bert-base-cased"
)
def test_flaxwav2vec2bart_pt_flax_equivalence(self):
pt_model = SpeechEncoderDecoderModel.from_pretrained("patrickvonplaten/wav2vec2-2-bart-large")
fx_model = FlaxSpeechEncoderDecoderModel.from_pretrained(
"patrickvonplaten/wav2vec2-2-bart-large", from_pt=True
)
pt_model.to(torch_device)
pt_model.eval()
# prepare inputs
batch_size = 13
input_values = floats_tensor([batch_size, 512], fx_model.config.encoder.vocab_size)
attention_mask = random_attention_mask([batch_size, 512])
decoder_input_ids = ids_tensor([batch_size, 4], fx_model.config.decoder.vocab_size)
decoder_attention_mask = random_attention_mask([batch_size, 4])
inputs_dict = {
"inputs": input_values,
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
}
flax_inputs = inputs_dict
pt_inputs = {k: torch.tensor(v.tolist()) for k, v in flax_inputs.items()}
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs)
pt_logits = pt_outputs.logits
pt_outputs = pt_outputs.to_tuple()
fx_outputs = fx_model(**inputs_dict)
fx_logits = fx_outputs.logits
fx_outputs = fx_outputs.to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits, pt_logits.numpy(), 4e-2)
# PT -> Flax
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
fx_model_loaded = FlaxSpeechEncoderDecoderModel.from_pretrained(tmpdirname, from_pt=True)
fx_outputs_loaded = fx_model_loaded(**inputs_dict)
fx_logits_loaded = fx_outputs_loaded.logits
fx_outputs_loaded = fx_outputs_loaded.to_tuple()
self.assertEqual(len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits_loaded, pt_logits.numpy(), 4e-2)
# Flax -> PT
with tempfile.TemporaryDirectory() as tmpdirname:
fx_model.save_pretrained(tmpdirname)
pt_model_loaded = SpeechEncoderDecoderModel.from_pretrained(tmpdirname, from_flax=True)
pt_model_loaded.to(torch_device)
pt_model_loaded.eval()
with torch.no_grad():
pt_outputs_loaded = pt_model_loaded(**pt_inputs)
pt_logits_loaded = pt_outputs_loaded.logits
pt_outputs_loaded = pt_outputs_loaded.to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits, pt_logits_loaded.numpy(), 4e-2)
def get_pretrained_model(self):
return SpeechEncoderDecoderModel.from_encoder_decoder_pretrained("bert-large-uncased", "facebook/bart-large")
def convert_wav2vec2_checkpoint(
checkpoint_path,
pytorch_dump_folder_path,
dict_path,
config_yaml_path,
encoder_config_path,
decoder_config_path,
add_adapter,
adapter_kernel_size,
adapter_stride,
decoder_start_token_id,
encoder_output_dim,
):
"""
Copy/paste/tweak model's weights to transformers design.
"""
# load configs
encoder_config = Wav2Vec2Config.from_pretrained(
encoder_config_path,
add_adapter=True,
adapter_stride=adapter_stride,
adapter_kernel_size=adapter_kernel_size,
use_auth_token=True,
output_hidden_size=encoder_output_dim,
)
decoder_config = MBartConfig.from_pretrained(decoder_config_path)
# load model
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[checkpoint_path],
arg_overrides={
"config_yaml": config_yaml_path,
"data": "/".join(dict_path.split("/")[:-1]),
"w2v_path": checkpoint_path,
"load_pretrained_decoder_from": None,
},
)
model = model[0].eval()
# load feature extractor
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
encoder_config_path, use_auth_token=True)
# set weights for wav2vec2 encoder
hf_encoder = Wav2Vec2Model(encoder_config)
recursively_load_weights_wav2vec2(model.encoder, hf_encoder)
# load decoder weights
hf_decoder = MBartForCausalLM(decoder_config)
missing_keys, unexpected_keys = hf_decoder.model.decoder.load_state_dict(
model.decoder.state_dict(), strict=False)
logger.warning(
f"The following keys are missing when loading the decoder weights: {missing_keys}"
)
logger.warning(
f"The following keys are unexpected when loading the decoder weights: {unexpected_keys}"
)
hf_wav2vec = SpeechEncoderDecoderModel(encoder=hf_encoder,
decoder=hf_decoder)
hf_wav2vec.config.tie_word_embeddings = False
tokenizer = MBart50Tokenizer(dict_path)
tokenizer.save_pretrained(pytorch_dump_folder_path)
config = hf_wav2vec.config.to_dict()
config["pad_token_id"] = tokenizer.pad_token_id
config["bos_token_id"] = tokenizer.bos_token_id
config["eos_token_id"] = tokenizer.eos_token_id
config["tokenizer_class"] = "mbart50"
config["feature_extractor_type"] = "wav2vec2"
config["decoder_start_token_id"] = tokenizer.eos_token_id
config["forced_bos_token_id"] = 250004
config["forced_eos_token_id"] = tokenizer.eos_token_id
hf_wav2vec.config = SpeechEncoderDecoderConfig.from_dict(config)
hf_wav2vec.save_pretrained(pytorch_dump_folder_path)
feature_extractor.save_pretrained(pytorch_dump_folder_path)
