create_model.py

import jax
import jax.numpy as jnp
from transformers import AutoFeatureExtractor, AutoTokenizer
from models.modeling_flax_speech_encoder_decoder import FlaxSpeechEncoderDecoderModel

encoder_id = "facebook/wav2vec2-large-lv60"
decoder_id = "facebook/bart-large"

model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True, decoder_from_pt=True)

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.decoder_start_token_id = model.config.decoder.bos_token_id
model.config.pad_token_id = model.config.decoder.pad_token_id
model.config.eos_token_id = model.config.decoder.eos_token_id
model.config.max_length = 40
model.config.num_beams = 1
model.config.encoder.layerdrop = 0.0
model.config.use_cache = False
model.config.processor_class = "Wav2Vec2Processor"

# need to upcast bart-large weights from float16 to float32
model.params = jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype != jnp.float32 else x, model.params)

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

model.save_pretrained("./")

feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
feature_extractor.save_pretrained("./")
tokenizer = AutoTokenizer.from_pretrained(decoder_id)
tokenizer.save_pretrained("./")