sanchit-gandhi
/

wav2vec2-2-gpt2-regularisation

Automatic Speech Recognition Transformers PyTorch TensorBoard

speech-encoder-decoder generated_from_trainer Inference Endpoints

Model card Files Files and versions Metrics Training metrics Community

wav2vec2-2-gpt2-regularisation / create_model.py

sanchit-gandhi's picture

sanchit-gandhi HF staff

Training in progress, step 2500

cf12caf about 2 years ago

raw history blame contribute delete

No virus

1.61 kB

	from transformers import SpeechEncoderDecoderModel, AutoFeatureExtractor, GPT2Tokenizer
	import torch

	# checkpoints to leverage
	encoder_id = "facebook/wav2vec2-large-lv60"
	decoder_id = "gpt2-medium"

	model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True)

	# force GPT2 to append EOS to begin and end of seq
	def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
	outputs = [self.bos_token_id] + token_ids_0 + [self.eos_token_id]
	return outputs

	GPT2Tokenizer.build_inputs_with_special_tokens = build_inputs_with_special_tokens
	gpt2_tokenizer = GPT2Tokenizer.from_pretrained(decoder_id)
	# set pad_token_id to unk_token_id, note: unk_token_id == eos_token_id == bos_token_id
	gpt2_tokenizer.pad_token = gpt2_tokenizer.unk_token
	gpt2_tokenizer.save_pretrained("./")

	model.config.pad_token_id = gpt2_tokenizer.pad_token_id
	model.config.decoder_start_token_id = model.decoder.config.bos_token_id
	model.config.eos_token_id = model.decoder.config.eos_token_id
	model.config.max_length = 50
	model.config.num_beams = 1

	# same enc regularisation as wav2vec2-2-bart
	model.config.encoder.feat_proj_dropout = 0.0
	model.config.encoder.final_dropout = 0.0
	model.config.encoder.mask_time_prob = 0.1

	model.config.use_cache = False
	model.config.decoder.use_cache = False
	model.config.processor_class = "Wav2Vec2Processor"

	# check if generation works
	out = model.generate(torch.ones((1, 2000)))

	model.save_pretrained("./")

	feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
	feature_extractor.save_pretrained("./")