Update: TF2.4 support has now been rolled into mainline. The tf2.4 branch has been removed, so you should now just stay on the Master branch
First up, this is entirely unsupported, and you are on your own. These instructions are here as a courtesy whilst we wait for upstream (Conda mainly) to update it's packages for the latest GPU support.
You will not receive official support for this kind of setup, so please don't ask for it. Official support will come when upstream catches up. You should keep all comments, feedback, bugs and suggestions to this thread in the hope that others will be able to help you and the devs can pick up on bugs.
This solution is not automatic, there will be manual and command line work involved. These instructions are untested and just reeled off from the top of my head. If you have suggestions for changes please post them.
Another thing to bear in mind is that officially the minimum Cuda version required for 30xx cards is 11.1 (https://www.phoronix.com/scan.php?page= ... 1-Released), however the latest Tensorflow version has been compiled for 11.0 (https://www.tensorflow.org/install/source#gpu). These instructions will allow you to run the 11.0 version of Tensorflow (as per the official release). This may or may not bring with it issues for 30xx cards. Please do report them here.
To use a version for Cuda 11.1 then you will need to compile Tensorflow yourself, which is well outside of the scope of these instructions. (Compiling Tensorflow from source: https://www.tensorflow.org/install/source). Alternatively you could try some pre-compiled wheels (Google is your friend. your mileage may vary), although we cannot vouch for 3rd party compiled versions of Tensorflow.
- Remove any and all versions of Cuda and cuDNN from your PC.
- Install Cuda Tool Kit 11.0 from https://developer.nvidia.com/cuda-11.0-download-archive.
- Install cuDNN 8.0 for Cuda Tool Kit 11.0 from https://developer.nvidia.com/cudnn-download-survey. (NB: I cannot provide a direct link as you need to log in, make sure you download the correct cuDNN version (8.0) for the correct Cuda version (11.0)).
- Download the latest Faceswap installer (do not try to upgrade your existing Faceswap install).
- Select "CPU" when asked which version of Faceswap you want to install. This is very important as it will ensure we use the Cuda version we installed in the last step.
- Run the installer and let it complete.
- Delete the file
<faceswap folder>/config/.faceswap(alternatively you can edit this file and change the word
We have to remove the CPU version of Tensorflow we installed and replace it with a version compatible with 30xx cards. I will assume Linux users are comfortable with entering a Conda environment, so these instructions are for Windows:
- Enter the following commands to activate the faceswap environment, remove Tensorflow cpu and install Tensorflow-GPU. (Change the name
faceswapin the first command if you selected a different environment name when installing):
Code: Select all
conda activate faceswap conda remove tensorflow pip install tensorflow-gpu==2.4
We now have Faceswap, Cuda and Tensorflow installed. Opening Faceswap should see you running tf.2.4 for 30xx cards.
Please post issues and bugs here. We will not provide setup support, but bugs and issues will help us to remedy asap.