#!/bin/bash

# Train Canary ASR model
HYDRA_FULL_ERROR=1 python scripts/speech_to_text_aed.py \
  --config-path="/home/ubuntu/NeMo/config" \
  --config-name="fast-conformer_aed.yaml" \
  name="canary-small" \
  model.prompt_format="canary2" \
  model.train_ds.manifest_filepath="/home/ubuntu/NeMo/data/tsukasa_train.json" \
  model.validation_ds.manifest_filepath="/home/ubuntu/NeMo/data/tsukasa_val.json" \
  model.test_ds.manifest_filepath="/home/ubuntu/NeMo/data/tsukasa_test.json" \
  model.tokenizer.langs.ja.dir="/home/ubuntu/NeMo/tokenizers/jp_TSUKA_1024/tokenizer_spe_bpe_v1024" \
  model.tokenizer.langs.spl_tokens.dir="/home/ubuntu/NeMo/tokenizers/spl_tokens" \
  spl_tokens.model_dir="/home/ubuntu/NeMo/tokenizers/spl_tokens" \
  model.encoder.n_layers=17 \
  model.transf_decoder.config_dict.num_layers=4 \
  model.transf_decoder.config_dict.max_sequence_length=512 \
  model.model_defaults.asr_enc_hidden=512 \
  model.model_defaults.lm_dec_hidden=1024 \
  exp_manager.exp_dir="canary_results" \
  exp_manager.resume_ignore_no_checkpoint=true \
  trainer.max_steps=200_000 \
  trainer.log_every_n_steps=50