run_training_bpe.sh

!/usr/bin/env bash

DOCKER_EXP_DIR="/checkpoints/results"

### Good
CKPT="/checkpoints/streaming/magpie/jason/magpieTTS--val_loss=5.1255-epoch=89-last.ckpt"
HPARAM="/checkpoints/streaming/magpie/jason/magpietts_en_jason_inference.yaml"

PROJECT='magpieTTS_en_newcodec_exps'
CODECMODEL_NAME="C25FPS_Causal_8cb"

# CODEC_PATH="/nemo_codec_checkpoints/21fps_causal_codecmodel.nemo"
CODEC_PATH="/nemo_codec_checkpoints/Low_Frame-rate_25FPS_Speech_Codec++.nemo"
DATASETS="HRLLM"
MODEL_STRUCTURE="MP_CE"
LR=1e-4
PRECISION=bf16
SAMPLE_WEIGHT_LIBRIVOX=0.2
BATCH_SIZE=12
EXP_NAME="${MODEL_STRUCTURE}_${CODECMODEL_NAME}_${DATASETS}_lr${LR}_bs${BATCH_SIZE}_precision${PRECISION}_w${SAMPLE_WEIGHT_LIBRIVOX}_bin_4"

# EXP_DIR="/lustre/fsw/llmservice_nemo_speechlm/users/subhankarg/experiments_slurm_logs/T5TTSMarch2025/magpie2503_causal_codec_causal_enc_attnprior_nolocaltransformer"
DOCKER_EXP_DIR="/checkpoints/local_training/magpie2503_causal_codec_causal_enc_attnprior_nolocaltransformer"

export CUDA_VISIBLE_DEVICES=0

# ++model.train_ds.dataset.include_wait_k=True \
# ++model.train_ds.dataset.max_wait_k=2 \
# ++model.validation_ds.dataset.include_wait_k=True \
# ++model.validation_ds.dataset.max_wait_k=2 \

python examples/tts/magpietts.py \
--config-path=/workspace/NeMo/examples/tts/conf/magpietts \
--config-name=magpietts_en_bpe_25fps \
exp_manager.exp_dir="${DOCKER_EXP_DIR}" \
++exp_manager.name="magpieTTS" \
++exp_manager.create_tensorboard_logger=true \
++exp_manager.create_wandb_logger=false \
+exp_manager.version=0 \
weighted_sampling_steps_per_epoch=1000 \
max_epochs=500 \
batch_size=${BATCH_SIZE} \
~model.text_tokenizers.english_phoneme \
++model.text_tokenizers.qwen._target_=AutoTokenizer \
++model.text_tokenizers.qwen.pretrained_model="Qwen/Qwen3-1.7B" \
++model.use_bpe_char_tokenizer=true \
+train_ds_meta.libri100train.manifest_path="/data/TTS/manifests/libri100__phoneme__nemo_audio_21fps_8codebooks_2kcodes_v2bWithWavLM_simplet5_withContextAudioPaths.json" \
+train_ds_meta.libri100train.audio_dir="/data/TTS/LibriTTS" \
+train_ds_meta.libri100train.feature_dir="/data/TTS/LibriTTS" \
+train_ds_meta.libri100train.sample_weight=1.0 \
+train_ds_meta.libri100train.tokenizer_names="[qwen]" \
+val_ds_meta.libridev.manifest_path="/data/TTS/manifests/dev_clean_withContextAudioPaths.json" \
+val_ds_meta.libridev.audio_dir="/data/TTS/LibriTTS" \
+val_ds_meta.libridev.feature_dir="/data/TTS/LibriTTS" \
+val_ds_meta.libridev.tokenizer_names="[qwen]" \
model.train_ds.dataset.min_duration=0.2 \
model.validation_ds.dataset.min_duration=0.2 \
++model.train_ds.dataset.load_cached_codes_if_available=false \
++model.validation_ds.dataset.load_cached_codes_if_available=false \
++model.train_ds.dataset.include_wait_k=True \
++model.train_ds.dataset.max_wait_k=3 \
++model.train_ds.dataset.wait_k_step_size=2 \
++model.validation_ds.dataset.include_wait_k=True \
++model.validation_ds.dataset.max_wait_k=3 \
++model.validation_ds.dataset.wait_k_step_size=2 \
model.context_duration_min=5.0 \
model.context_duration_max=5.0 \
model.codecmodel_path=${CODEC_PATH} \
model.model_type="decoder_context_tts" \
model.use_text_conditioning_encoder=false \
model.alignment_loss_scale=0.002 \
model.prior_scaling_factor=0.1 \
model.prior_end_step=0 \
model.prior_scaledown_start_step=0 \
model.indefinite_prior_prob=0.5 \
model.use_alignment_encoder=true \
model.binarize_prior_after_step=5 \
model.alignment_encoder_loss_scale=1.0 \
model.use_prior_for_aligner=true \
model.prior_future_context=0 \
model.prior_past_context=1 \
model.prior_future_decay=0.05 \
model.prior_past_decay=0.03 \
model.binarized_prior_epsilon=0.0001 \
model.use_local_transformer=false \
+model.cfg_unconditional_prob=0.1 \
model.encoder.is_causal=true \
model.encoder.kernel_size=3 \
model.context_encoder.is_causal=true \
model.decoder.kernel_size=3 \
model.decoder.xa_n_heads=1 \
model.embedding_dim=768 \
model.train_ds.dataloader_params.num_workers=2 \
model.validation_ds.dataloader_params.num_workers=2 \
trainer.devices=-1 \
trainer.val_check_interval=100 \
+trainer.num_sanity_val_steps=0 \
trainer.precision=${PRECISION} \
model.optim.lr=${LR}

# +trainer.check_val_every_n_epoch=1 \