返回 F5-TTS
README.md
根目录 / src / f5_tts / train / README.md
1 # Training
2
3 Check your FFmpeg installation:
4 ```bash
5 ffmpeg -version
6 ```
7 If not found, install it first (or skip assuming you know of other backends available).
8
9 ## Prepare Dataset
10
11 Example data processing scripts, and you may tailor your own one along with a Dataset class in `src/f5_tts/model/dataset.py`.
12
13 ### 1. Some specific Datasets preparing scripts
14 Download corresponding dataset first, and fill in the path in scripts.
15
16 ```bash
17 # Prepare the Emilia dataset
18 python src/f5_tts/train/datasets/prepare_emilia.py
19
20 # Prepare the Wenetspeech4TTS dataset
21 python src/f5_tts/train/datasets/prepare_wenetspeech4tts.py
22
23 # Prepare the LibriTTS dataset
24 python src/f5_tts/train/datasets/prepare_libritts.py
25
26 # Prepare the LJSpeech dataset
27 python src/f5_tts/train/datasets/prepare_ljspeech.py
28 ```
29
30 ### 2. Create custom dataset with CSV
31 Prepare a CSV with two columns using a required header: `audio_file|text`. Audio paths must be absolute.
32 Use guidance see [#57 here](https://github.com/SWivid/F5-TTS/discussions/57#discussioncomment-10959029).
33
34 ```bash
35 python src/f5_tts/train/datasets/prepare_csv_wavs.py /path/to/metadata.csv /path/to/output
36 ```
37
38 ## Training & Finetuning
39
40 Once your datasets are prepared, you can start the training process.
41
42 ### 1. Training script used for pretrained model
43
44 ```bash
45 # setup accelerate config, e.g. use multi-gpu ddp, fp16
46 # will be to: ~/.cache/huggingface/accelerate/default_config.yaml
47 accelerate config
48
49 # .yaml files are under src/f5_tts/configs directory
50 accelerate launch src/f5_tts/train/train.py --config-name F5TTS_v1_Base.yaml
51
52 # possible to overwrite accelerate and hydra config
53 accelerate launch --mixed_precision=fp16 src/f5_tts/train/train.py --config-name F5TTS_v1_Base.yaml ++datasets.batch_size_per_gpu=19200
54 ```
55
56 ### 2. Finetuning practice
57 Discussion board for Finetuning [#57](https://github.com/SWivid/F5-TTS/discussions/57).
58
59 Gradio UI training/finetuning with `src/f5_tts/train/finetune_gradio.py` see [#143](https://github.com/SWivid/F5-TTS/discussions/143).
60
61 If want to finetune with a variant version e.g. *F5TTS_v1_Base_no_zero_init*, manually download pretrained checkpoint from model weight repository and fill in the path correspondingly on web interface.
62
63 If use tensorboard as logger, install it first with `pip install tensorboard`.
64
65 <ins>The `use_ema = True` might be harmful for early-stage finetuned checkpoints</ins> (which goes just few updates, thus ema weights still dominated by pretrained ones), try turn it off with finetune gradio option or `load_model(..., use_ema=False)`, see if offer better results.
66
67 ### 3. W&B Logging
68
69 The `wandb/` dir will be created under path you run training/finetuning scripts.
70
71 By default, the training script does NOT use logging (assuming you didn't manually log in using `wandb login`).
72
73 To turn on wandb logging, you can either:
74
75 1. Manually login with `wandb login`: Learn more [here](https://docs.wandb.ai/ref/cli/wandb-login)
76 2. Automatically login programmatically by setting an environment variable: Get an API KEY at https://wandb.ai/authorize and set the environment variable as follows:
77
78 On Mac & Linux:
79
80 ```
81 export WANDB_API_KEY=<YOUR WANDB API KEY>
82 ```
83
84 On Windows:
85
86 ```
87 set WANDB_API_KEY=<YOUR WANDB API KEY>
88 ```
89 Moreover, if you couldn't access W&B and want to log metrics offline, you can set the environment variable as follows:
90
91 ```
92 export WANDB_MODE=offline
93 ```
94
94 lines MARKDOWN