Skip to content

stellanoxUN/RVC-backup

Repository files navigation

RVC 巷口星尘 部署备份

环境

  • Windows 10/11
  • Python 3.8 (conda env: rvc38)
  • CUDA 11.7 + PyTorch 2.0.1
  • fairseq 0.12.2 (源码编译)
  • GPU: RTX 3050 4GB

恢复步骤

1. 安装 Python 3.8

conda create -n rvc38 python=3.8
conda activate rvc38

2. 安装 Visual C++ Build Tools

下载: https://visualstudio.microsoft.com/visual-cpp-build-tools/ 安装时勾选 "C++ 桌面开发" 工作负载

3. 安装依赖

pip install -r requirements_rvc38.txt

4. 编译 fairseq

git clone https://github.com/facebookresearch/fairseq --branch v0.12.2
# 修补 setup.py: 将 hydra-core==1.0.7 改为 hydra-core>=1.0.7
# 将 omegaconf<2.1 改为 omegaconf<3.0
# 然后编译:
set DISTUTILS_USE_SDK=1
pip install -e . --no-build-isolation --no-deps
# 补充依赖:
pip install bitarray sacrebleu Cython

5. 下载 RVC

git clone https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI.git

将本仓库中的 infer-web.py / train.py / preprocess.py 覆盖到对应位置

6. 下载预训练模型

从 hf-mirror.com/lj1995/VoiceConversionWebUI 下载:

  • assets/hubert/hubert_base.pt
  • assets/rmvpe/rmvpe.pt
  • assets/pretrained/*.pth (12个)
  • assets/pretrained_v2/*.pth (12个)
  • ffmpeg.exe / ffprobe.exe (放根目录)

7. 准备训练数据

将巷口星尘人声干音放入 train_data/ 目录

8. 启动

python infer-web.py --port 7930

训练命令

python train_final.py

修改记录

  • train.py: num_workers 改为 0, prefetch_factor 改为 None (适配4GB显存)
  • preprocess.py: argv 访问改为有默认值的防御写法
  • infer-web.py: server_name="127.0.0.1", share=True

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors