DeepFaceLab CentOS Linux

Note: This installation method works for CentOS 7.x - your milage may vary.

This is a simple addition to DeepFaceLab to allow the latest code to be run directly on Linux.

Install requirements

Download the NVIDIA cuDNN RPM. note: You need to register and make sure you download the specific version libcudnn7-7.6.5.32-1.cuda10.0

Manual method

Install the rest of the requirements as your favourite user (root is used by default and as example):

yum -y install epel-release http://li.nux.ro/download/nux/dextop/el7/x86_64/nux-dextop-release-0-5.el7.nux.noarch.rpm
wget -O /etc/yum.repos.d/cuda.repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
yum -y update
yum -y install python3-3.6.8 git cuda-toolkit-10-0.x86_64 ffmpeg ffmpeg-devel libSM libXext libXrender-devel libXrender gcc gcc-c++ 
yum -y install kernel-devel make cuda-libraries-10-0.x86_64 nvidia-driver-cuda.x86_64 nvidia-driver-cuda-libs.x86_64 

cd ~ && git clone https://github.com/iperov/DeepFaceLab.git
cd ~ && git clone https://github.com/elemantalcode/dfl.git

cd ~/DeepFaceLab && cp -advpR ../dfl/scripts .
cd ~/DeepFaceLab && python3.6 -m pip install -r requirements-cuda.txt
yum -y install /root/libcudnn7-7.6.5.32-1.cuda10.0.x86_64.rpm

Scripted method

Run the setup.bash script

sudo setup.bash

Post-install check

Once installed, you should be able to run DeepFaceLab as per usual:

cd ~/DeepFaceLab/scripts
./1_clear_workspace.sh

Quick FAQ

Q: Why CentOS?

A: It's stable, close to RHEL and doesn't suffer the same fate as CoreOS ... yet.

Q: Why not docker?

A: Why abstract yet again when you don't have to?

Q: Why not use Anaconda3 like DeepFaceLab for Linux

A: Different approaches to solve the same challange. I personally don't like adding frameworks/structures that I don't intend on making good use of. For python developers, Anaconda makes sense - for me, I just want to run DeepFaceLab under Linux with minimal hassle.

Q: Anything else particular noteworthy?

A: You need some Linux knowledge.