| 1 | hydra: |
| 2 | run: |
| 3 | dir: ckpts/${model.name}_${model.mel_spec.mel_spec_type}_${model.tokenizer}_${datasets.name}/${now:%Y-%m-%d}/${now:%H-%M-%S} |
| 4 | |
| 5 | datasets: |
| 6 | name: Emilia_ZH_EN # dataset name |
| 7 | batch_size_per_gpu: 38400 # 8 GPUs, 8 * 38400 = 307200 |
| 8 | batch_size_type: frame # frame | sample |
| 9 | max_samples: 64 # max sequences per batch if use frame-wise batch_size. we set 32 for small models, 64 for base models |
| 10 | num_workers: 16 |
| 11 | |
| 12 | optim: |
| 13 | epochs: 11 |
| 14 | learning_rate: 7.5e-5 |
| 15 | num_warmup_updates: 20000 # warmup updates |
| 16 | grad_accumulation_steps: 1 # note: updates = steps / grad_accumulation_steps |
| 17 | max_grad_norm: 1.0 # gradient clipping |
| 18 | bnb_optimizer: False # use bnb 8bit AdamW optimizer or not |
| 19 | |
| 20 | model: |
| 21 | name: F5TTS_v1_Base # model name |
| 22 | tokenizer: pinyin # tokenizer type |
| 23 | tokenizer_path: null # if 'custom' tokenizer, define the path want to use (should be vocab.txt) |
| 24 | backbone: DiT |
| 25 | arch: |
| 26 | dim: 1024 |
| 27 | depth: 22 |
| 28 | heads: 16 |
| 29 | ff_mult: 2 |
| 30 | text_dim: 512 |
| 31 | text_mask_padding: True |
| 32 | qk_norm: null # null | rms_norm |
| 33 | conv_layers: 4 |
| 34 | pe_attn_head: null |
| 35 | attn_backend: torch # torch | flash_attn |
| 36 | attn_mask_enabled: False |
| 37 | checkpoint_activations: False # recompute activations and save memory for extra compute |
| 38 | mel_spec: |
| 39 | target_sample_rate: 24000 |
| 40 | n_mel_channels: 100 |
| 41 | hop_length: 256 |
| 42 | win_length: 1024 |
| 43 | n_fft: 1024 |
| 44 | mel_spec_type: vocos # vocos | bigvgan |
| 45 | vocoder: |
| 46 | is_local: False # use local offline ckpt or not |
| 47 | local_path: null # local vocoder path |
| 48 | |
| 49 | ckpts: |
| 50 | logger: wandb # wandb | tensorboard | null |
| 51 | wandb_project: CFM-TTS # wandb project name |
| 52 | wandb_run_name: ${model.name}_${model.mel_spec.mel_spec_type}_${model.tokenizer}_${datasets.name} # wandb run name |
| 53 | wandb_resume_id: null # wandb run id for resuming, null to auto-detect from checkpoint |
| 54 | log_samples: True # infer random sample per save checkpoint. wip, normal to fail with extra long samples |
| 55 | save_per_updates: 50000 # save checkpoint per updates |
| 56 | keep_last_n_checkpoints: -1 # -1 to keep all, 0 to not save intermediate, > 0 to keep last N checkpoints |
| 57 | last_per_updates: 5000 # save last checkpoint per updates |
| 58 | save_dir: ckpts/${model.name}_${model.mel_spec.mel_spec_type}_${model.tokenizer}_${datasets.name} |