fish-speech/fish_speech/configs/text2semantic_finetune.yaml

defaults:
  - base
  - model@model.model: dual_ar_2_codebook_small
  - _self_

project: text2semantic_finetune_dual_ar
max_length: 2048
ckpt_path: checkpoints/text2semantic-sft-medium-v1.1-4k.pth
resume_weights_only: true

# Lightning Trainer
trainer:
  accumulate_grad_batches: 1
  gradient_clip_val: 1.0
  gradient_clip_algorithm: "norm"
  max_steps: 1000
  precision: bf16-true
  limit_val_batches: 10
  val_check_interval: 100

# Dataset Configuration
tokenizer:
  _target_: transformers.AutoTokenizer.from_pretrained
  pretrained_model_name_or_path: fishaudio/fish-speech-1

# Dataset Configuration
train_dataset:
  _target_: fish_speech.datasets.text.AutoAugTextDataset
  proto_files:
    - data/protos
  tokenizer: ${tokenizer}
  max_length: ${max_length}
  num_codebooks: ${model.model.config.num_codebooks}
  use_speaker: 0.5
  interactive_prob: 0.7

val_dataset:
  _target_: fish_speech.datasets.text.AutoAugTextDataset
  proto_files:
    - data/protos
  tokenizer: ${tokenizer}
  max_length: ${max_length}
  num_codebooks: ${model.model.config.num_codebooks}
  use_speaker: 0.5
  interactive_prob: 0.7

data:
  _target_: fish_speech.datasets.text.TextDataModule
  train_dataset: ${train_dataset}
  val_dataset: ${val_dataset}
  num_workers: 4
  batch_size: 8
  tokenizer: ${tokenizer}
  max_length: ${max_length}

# Model Configuration
model:
  _target_: fish_speech.models.text2semantic.TextToSemantic
  model: {}

  optimizer:
    _target_: torch.optim.AdamW
    _partial_: true
    lr: 1e-5
    weight_decay: 0
    betas: [0.9, 0.95]
    eps: 1e-5

  lr_scheduler:
    _target_: torch.optim.lr_scheduler.LambdaLR
    _partial_: true
    lr_lambda:
      _target_: fish_speech.scheduler.get_cosine_schedule_with_warmup_lr_lambda
      _partial_: true
      num_warmup_steps: 0.1
      num_training_steps: ${trainer.max_steps}

# Callbacks
callbacks:
  model_checkpoint:
    every_n_train_steps: ${trainer.val_check_interval}