egs/tts/NaturalSpeech2/exp_config.json

{
    "base_config": "egs/tts/NaturalSpeech2/exp_config_base.json",
    "dataset": [
      "libritts"
    ],
    "preprocess": {
      // Specify the output root path to save the processed data 
      "processed_dir": "data",
      "train_file": "train.json",
      "valid_file": "test.json",
      "read_metadata": true,
      "metadata_dir": "metadata"
    },
    // Specify the output root path to save model ckpts and logs
    "log_dir": "ckpts/tts",
    "train": {
        // New trainer and Accelerator
        "gradient_accumulation_step": 1,
        "tracker": ["tensorboard"],
        "max_epoch": 5000,
        "save_checkpoint_stride": [1],
        "keep_last": [1000],
        "run_eval": [true],
        "dataloader": {
          "num_worker": 16,
          "pin_memory": true
        },
        "adam": {
            "lr": 1.0e-4
        },
        "use_dynamic_batchsize": true,
        "batch_size": 8,
        "max_tokens": 7500,
        "max_sentences": 32,
        "lr_warmup_steps": 5000,
        "lr_scheduler": "cosine",
        "num_train_steps": 800000
    }
  }