Skip to content

从零训练(train from scratch)

RVC-Boss edited this page Jul 23, 2024 · 4 revisions

从零训练(train from scratch)

1、训练推理流程具体可参考这个issue

train&inference pipeline:

https://github.com/RVC-Boss/GPT-SoVITS/issues/67

2、完整的训练流程分为3阶段

The whole training pipeline is divided into three stages:

(1)token编解码(s2)

(1)token encoder and decoder (s2)

(2)GPT生成模型(s1)

(2)GPT model (s1)

(3)冻结s2编码器,微调解码器(s2+)(按需。也可以不做)

(3)fine tune s2 decoder while freezing s2 encoder (additional)

目前开源的2k小时底模只做了前2个阶段

The open sourced base model is only trained on the first two stages.

3、webui微调训练和底模训练(因为代码写得比较丑所以未开源)的代码区别

Difference between webui fine tune training codes and base model training codes

(1)webui微调训练微操降低了SoVITS模块文本编码器的lr,底模训练没有

(1)webui fine tune training decreases the leanring rate of text embedding module, while base model training not

(2)webui微调冻结了SoVITS中的VQ编码器,训练底模是不冻结的,具体在configs/s2.json中的freeze_quantizer。

(2)webui fine tune training freezes the VQ encoder while base model training not. You can modify it in the configs/s2.json (freeze_quantizer)

(3)效果上除了1、2外就没有区别了。代码区别就是webui的写法是环境变量+参数,这个打印下+对着webui代码看看可以很容易得到train-cli的代码

(3)No other diffenece which influences the results. Train-cli codes can easily be reproduced by reading webui-fine-tuning codes (environment variables and sys arguments)

4、训练集的质量要求

The quality requirments of training datasets:

https://github.com/RVC-Boss/GPT-SoVITS/issues/176