semantic_detokenizer/ckpt/model.yaml

hydra:
  run:
    # _${datasets.name}/
    dir: ckpts/${model.name}_${model.mel_spec.mel_spec_type}_${model.tokenizer}/${now:%Y-%m-%d}/${now:%H-%M-%S}
    
datasets:
  name: LibriTTS_960_WenetSpeech4TTS_Premium_Standard_TTSDatawithMultiStyleEmotion_Datatang_Haitianruisheng_Qingshu_XiaoyiF143_Bani  # dataset name
  path: /home/ma-user/work/dehua/unit2speech/u2s_training_data/40ms_imedia_45wh_hubert_large_FSQ_8888_CTC_sampled_EN_CH_4000h_wulan_H20_mt_1280k_lr_3e-4_d2v_phase_nomask_20241018_8p_fromDaxin/LibriTTS_WenetSpeech4TTS_TTSDatawithMultiStyleEmotion_XiaoyiF143_Bani/train-960_Premium_Standard_datatang_haitianruisheng_qingshu_freetalk_style1_all-three
  batch_size_per_gpu: 2000 # 38400  # 8 GPUs, 8 * 38400 = 307200
  batch_size_type: frame  # "frame" or "sample"
  max_samples: 4 # 64  # max sequences per batch if use frame-wise batch_size. we set 32 for small models, 64 for base models
  num_workers: 2

optim:
  epochs: 1000 # 15
  learning_rate: 7.5e-5
  num_warmup_updates: 20000  # warmup steps
  grad_accumulation_steps: 1  # note: updates = steps / grad_accumulation_steps
  max_grad_norm: 1.0  # gradient clipping
  bnb_optimizer: False  # use bnb 8bit AdamW optimizer or not

model:
  name: CADiT_Base_train_40ms_imedia_45wh_HuBERT_large_FSQ_8888_CTC_sampled_EN_CH_4000h_CTC_LibriTTS_960_WenetSpeech4TTS_Premium_Standard_TTSDatawithMultiStyleEmotion_Datatang_Haitianruisheng_Qingshu_XiaoyiF143_Bani_Shanghai_V100_5n8g_20250414  # model name
  tokenizer: custom  # pinyin; tokenizer type
  # if tokenizer = 'custom', define the path to the tokenizer you want to use (should be vocab.txt)
  tokenizer_path: /home/ma-user/work/dehua/unit2speech/u2s_training_data/40ms_imedia_45wh_hubert_large_FSQ_8888_CTC_sampled_EN_CH_4000h_wulan_H20_mt_1280k_lr_3e-4_d2v_phase_nomask_20241018_8p_fromDaxin/LibriTTS_WenetSpeech4TTS_TTSDatawithMultiStyleEmotion_XiaoyiF143_Bani/train-960_Premium_Standard_datatang_haitianruisheng_qingshu_freetalk_style1_all-three/vocab.txt
  arch:
    dim: 1024
    depth: 22
    heads: 16
    ff_mult: 2
    text_dim: 512
    should_extend_text: True
    conv_layers: 4
    checkpoint_activations: False  # recompute activations and save memory for extra compute
  mel_spec:
    target_sample_rate: 24000
    n_mel_channels: 100
    hop_length: 256
    win_length: 1024
    n_fft: 1024
    mel_spec_type: vocos  # 'vocos' or 'bigvgan'
  vocoder:
    is_local: True  # use local offline ckpt or not
    local_path: /home/ma-user/work/dehua/unit2speech/CA-F5-TTS/pretrained_vocoder/charactr/vocos-mel-24khz  # local vocoder path

ckpts:
  logger: tensorboard  # wandb | tensorboard | None
  save_per_updates: 50000 # save checkpoint per steps
  last_per_steps: 2000 # 5000 # save last checkpoint per steps
  save_dir: ckpts/${model.name}_${model.mel_spec.mel_spec_type}_${model.tokenizer} # _${datasets.name}