src/f5_tts/configs/E2TTS_Base_train.yaml

hydra:
  run:
    dir: ckpts/${model.name}_${model.mel_spec.mel_spec_type}_${model.tokenizer}_${datasets.name}/${now:%Y-%m-%d}/${now:%H-%M-%S}
    
datasets:
  name: Emilia_ZH_EN  # dataset name
  batch_size_per_gpu: 38400  # 8 GPUs, 8 * 38400 = 307200
  batch_size_type: frame  # "frame" or "sample"
  max_samples: 64  # max sequences per batch if use frame-wise batch_size. we set 32 for small models, 64 for base models
  num_workers: 16

optim:
  epochs: 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: E2TTS_Base
  tokenizer: pinyin
  tokenizer_path: None  # if tokenizer = 'custom', define the path to the tokenizer you want to use (should be vocab.txt)
  arch:
    dim: 1024
    depth: 24
    heads: 16
    ff_mult: 4
  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: False  # use local offline ckpt or not
    local_path: None  # local vocoder path

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