vi-fine-tuned-f5-tts.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: vin100h-preprocessed-v2  # dataset name
  batch_size_per_gpu: 3200  # 1 GPUs, 1 * 3200 = 3200
  batch_size_type: frame  # frame | 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: 4

optim:
  epochs: 80
  learning_rate: 1e-5
  num_warmup_updates: 2761  # warmup updates
  grad_accumulation_steps: 2  # 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: vi_fine_tuned_t5_tts  # model name
  tokenizer: pinyin  # tokenizer type
  tokenizer_path: null  # if 'custom' tokenizer, define the path want to use (should be vocab.txt)
  backbone: DiT
  arch:
    dim: 1024
    depth: 22
    heads: 16
    ff_mult: 2
    text_dim: 512
    text_mask_padding: False
    conv_layers: 4
    pe_attn_head: 1
    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 | bigvgan
  vocoder:
    is_local: False  # use local offline ckpt or not
    local_path: null  # local vocoder path

ckpts:
  logger: null  # wandb | tensorboard | null
  log_samples: True  # infer random sample per save checkpoint. wip, normal to fail with extra long samples
  save_per_updates: 4000  # save checkpoint per updates
  keep_last_n_checkpoints: 1  # -1 to keep all, 0 to not save intermediate, > 0 to keep last N checkpoints
  last_per_updates: 4000  # save last checkpoint per updates
  save_dir: ckpts/${model.name}_${model.mel_spec.mel_spec_type}_${model.tokenizer}_${datasets.name}