[Guide] Using Faceswap on Nvidia RTX 30xx cards

[abigflea]

Just a little note to anyone trying to get the 30 series to work.
This is unsupported. Feels like iron chef type stuff to get the software to work.
Its been said many times before, can't do a whole lot until they fix the upstream dependencies like tensorflow and Nvidia quit changing fiddling with her driver so much.
I'm personally now having trouble even getting Ubuntu to install with a 3060 TI as a primary display
So trust me, all of you that are having trouble, we all feel your frustrations

[torzdf]

You are using legacy extracted faces but have selected 'face' centering which is incompatible. Switching centering to 'legacy'

Could that be the reason for the extremely slow performance?

No.

Most likely there is an issue with the Cuda/cuDNN install.

If you run in "Verbose" mode, it will generate a load of information, which may tell you why it's not working.

[impost3r]

No.

Most likely there is an issue with the Cuda/cuDNN install.

If you run in "Verbose" mode, it will generate a load of information, which may tell you why it's not working.

Thanks for the tip! I checked verbose mode and saw this warning:

2021-01-31 17:00:22.981621: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cusolver64_10.dll'; dlerror: cusolver64_10.dll not found

I had installed cuda 11.1 and it only came with cusolver64_11.dll.

I duplicated this library and renamed the duplicate "cusolver64_10.dll" and faceswap now works!

[torzdf]

I would highly recommend sticking with Cuda 11.0 then. As per original post:

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.

Your solution may work (in which case, great!) but you may also come unstuck at some point.

[Taltos]

For RTX A6000 on Windows:

• Install cuda_11.2.0_460.89_win10

• Install faceswap

Run anaconda prompt

Code: Select all

conda acticate faceswap
conda remove tensorflow
conda install broli
conda install urllib3
pip install tensorflow-gpu==2.4

Download cuDDN compatible with CUDA 11.2 (need to register with nvidia) : cudnn-11.2-windows-x64-v8.1.0.77
Copy extracted files into C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2 (DLL into bin)

You should be fine!

[Megadead]

You are using legacy extracted faces but have selected 'face' centering which is incompatible. Switching centering to 'legacy'

Could that be the reason for the extremely slow performance?

No.

Most likely there is an issue with the Cuda/cuDNN install.

If you run in "Verbose" mode, it will generate a load of information, which may tell you why it's not working.

How do you turn on Verbose Mode?

[torzdf]

Change the "Log Level" option to "Verbose"

[zoji2309]

461.40
cuda_11.0.2 (with Visual Studio 2019)
cudnn-11.0-v8.0.2.39
conda install -c anaconda urllib3(brotli Error Resolution during Tensorflow 2.4 Installation with cmd)

It works!
However, when entering the train process, vram utilization is close to 99%.
And if even a slight movement is detected on the screen, the gpu load increase and throttling is accompanied.
Is there a way to limit vram utilization or is there something else that causes it?

[Yuvix25]

I have an RTX 3070.
I keep getting this error while extracting:

Code: Select all

E tensorflow/stream_executor/cuda/cuda_blas.cc:226] failed to create cublas handle: CUBLAS_STATUS_ALLOC_FAILED

But if I simply ignore this error, the extraction works, but then I get this error:

Code: Select all

tensorflow.python.framework.errors_impl.InternalError:  dnn PoolForward launch failed

And face images are generated but alignments are not generated.
Does anyone know how to fix this?