前言：

本教程仅适用于A卡，使用N卡的您应当前往soVITS3.0炼丹教程。本人并不是专业人员，仅业余爱好者，有错误欢迎指正。本人参考过的所有文章链接都放在文章末尾。您应当知道炼丹对显卡的伤害不小于挖矿！

前期准备：

Ubuntu22.04、Python3.7.0

推荐使用Anaconda管理虚拟环境。系统盘不应小于50G。您不应该使用虚拟机炼丹。

#打上显卡驱动

wget https://repo.radeon.com/amdgpu-install/22.40/ubuntu/jammy/amdgpu-install_5.4.50401-1_all.deb

sudo dpkg -i amdgpu-install_5.4.50401-1_all.deb sudo apt update


sudo apt upgrade -y


sudo amdgpu-install --no-dkms


sudo apt install rocm-dev


sudo apt install radeontop（此项非必选）

sudo apt-get install libbz2-dev

sudo reboot

#配置环境


ls -l /dev/dri/render*


sudo usermod -a -G render $LOGNAME


sudo usermod -a -G video $LOGNAME


sudo reboot

#测试环境


rocm-smi


/opt/rocm/bin/rocminfo


/opt/rocm/opencl/bin/clinfo

没报错即可


PIP安装依赖（事故多发地）

#请先执行

python -m pip install --upgrade pip

pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple

sudo apt-get install libasound-dev portaudio19-dev libportaudio2 libportaudiocpp0

sudo apt-get install libbz2-dev

#防止后面安装pyaudio报错，同时加快pip的安装速度

（不推荐）你对自己处理依赖库有信心可以直接复制下面的指令安装。

pip install -r requirements.txt -i https://pypi.mirrors.ustc.edu.cn/simple/

（推荐)自己动手一个一个安装

pip install librosa==0.8.1

pip install Flask==2.1.2


pip install Flask_Cors==3.0.10


pip install gradio==3.4.1

pip install numpy==1.21.6

pip install playsound==1.3.0


pip install PyAudio==0.2.12

pip install pydub==0.25.1

pip install pyworld==0.3.0

pip install requests==2.28.1

pip install scipy==1.7.3

pip install sounddevice==0.4.5

pip install SoundFile==0.10.3.post1

pip install starlette==0.19.1

pip install tqdm==4.63.0

pip install scikit-maad

pip install praat-parselmouth

pip install tensorboard

pip install onnx

pip install onnxsim


pip install onnxoptimizer

安装MIopen

安装pytorch

RDNA2架构的显卡：pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm5.2

RX580：pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm3.7

到这里前期工作基本完成！

部署项目

#安装预训练模型

wget -P hubert/ https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt

wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth

wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth

#数据预处理

重采样至 32khz

python resample.py

自动划分训练集 验证集 测试集 以及自动生成配置文件

HSA_OVERRIDE_GFX_VERSION=10.3.0 python preprocess_flist_config.py

生成hubert与f0

HSA_OVERRIDE_GFX_VERSION=10.3.0 python preprocess_hubert_f0.py

#训练

HSA_OVERRIDE_GFX_VERSION=10.3.0 python train.py -c configs/config.json -m 32k

#推理

HSA_OVERRIDE_GFX_VERSION=10.3.0 python inference_main.py

#训练监控

tensorboard --logdir=logs/32k
浏览器打开http://127.0.0.1:6006即可查看训练情况
rocm-smi
终端输入即可查看显卡运行情况

本文参考的连接：

1. https://blog.csdn.net/qq_44948500/article/details/127346390?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522167708346516800213013835%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.%2522%257D&request_id=167708346516800213013835&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~all~first_rank_ecpm_v1~rank_v31_ecpm-3-127346390-null-null.142^v73^wechat_v2,201^v4^add_ask,239^v2^insert_chatgpt&utm_term=xiaozhou&spm=1018.2226.3001.4187
   

2. https://www.bilibili.com/read/cv20500632?spm_id_from=333.999.0.0

3. https://www.bilibili.com/read/cv20997087?spm_id_from=333.999.0.0

4. https://github.com/innnky/so-vits-svc

5. https://www.bilibili.com/read/cv20533940?spm_id_from=333.999.0.0

6. https://www.bilibili.com/video/BV1fD4y1w7KX/?spm_id_from=333.999.0.0&;vd_source=305e43fe4b1fa2a34139fec29c9becdf

so-vits-svc4.0的部署方法基本相同，制作不易，求个赞～