Page 7 of 7
Re: [Guide] Using Faceswap on Nvidia RTX 30xx cards
Posted: Sat Oct 09, 2021 11:51 am
by babala
If it still does not works :
- Uninstall FaceSwap
- Uninstall Cuda and Cudnn
- Clean registry keys using softwares like "CCleaner"
- Restart Computer
- Install Cuda 11.4 and Cudnn 8.2.4
- Install FaceSwap
- Force TF-GPU to 2.4.1 using the guide described above (Conda activate...., .....)
Hope it will help you...
welp... I tried yours but no luck for me I guess.
still not using gpu problem, 'train' and 'extract' too.
I'm stuck with TF 2.6
here's my sys info below
============ System Information ============
gpu_cuda: 11.4
gpu_cudnn: 8.2.4
gpu_devices: GPU_0: NVIDIA GeForce RTX 3080
gpu_devices_active: GPU_0
gpu_driver: 472.12
=============== Pip Packages ===============
tensorboard==2.6.0
tensorboard-data-server==0.6.1
tensorboard-plugin-wit==1.8.0
tensorflow-estimator==2.4.0
tensorflow-gpu==2.4.1
edit: and is it okay with just use 'pip install tensorflow-gpu==2.4.1' ?
cuz when I tried to remove tensorflow, conda said
PackagesNotFoundError: The following packages are missing from the target environment:
so I just installed with 'pip install tensorflow-gpu==2.4.1' and automatically remove tf 2.6
I checked error that not using gpu thing, back to tf 2.6 with just type 'pip install tensorflow-gpu==2.6'
now I'm using tf 2.6 again.
so that error is no problematic at all, right?
Re: [Guide] Using Faceswap on Nvidia RTX 30xx cards
Posted: Sat Oct 09, 2021 1:26 pm
by Barnuble
Hi,
Your SysInfo looks correct...
I really don't know why it does'nt work for you.
Just a question : When installing packages, installer asks you to confirm by "Yes" for each package. I have always answered "Yes" except for two AMD Packages where I answered "No".
Other than that I don't know what to tell you ;(
Re: [Guide] Using Faceswap on Nvidia RTX 30xx cards
Posted: Sat Oct 09, 2021 1:50 pm
by babala
Barnuble wrote: ↑Sat Oct 09, 2021 1:26 pm
Just a question : When installing packages, installer asks you to confirm by "Yes" for each package. I have always answered "Yes" except for two AMD Packages where I answered "No".
you mean 'python setup.py' one?
AMD no, docker no to that.
I skipped part this time that torzdf guide.
and started with your guide.
remove all nvidia drive and cuda cudnn, faceswap.
edit : welp, I tried last time to install with cuda 11.0.2 / cudnn 8.0.4 and it's finally working again.
extract and train is now back to last TF 2.4 build.
but train with TF 2.6 sessions are all not working, just working with TF 2.4 sessions that a month ago.
anyway, my problem was cuda / cudnn.
edit2 : I just noticed that GPU memory is 9.5/10.0 GB now with TF 2.4
TF 2.6 was 8.8/10.0 GB and 25% ish slower EGs/sec.
is it relavent to iteration performence?
Re: [Guide] Using Faceswap on Nvidia RTX 30xx cards
Posted: Mon Oct 11, 2021 3:38 pm
by Barnuble
Hi,
My tests were made with max batch size : 256 (My card has 24GB Memory) and i did not notice any different memory usage, only a speed improvement using TF2.4.
You say : "finally working again", is your "EGs/s" speed at the same level as before ?
I only use RealFace Trainer and speed improvement is about 12%.
What trainer are you using ? Phaze-A ?
Your 25% slowdown could mean that this trainer is more GPU dependent than others.
Re: [Guide] Using Faceswap on Nvidia RTX 30xx cards
Posted: Tue Oct 12, 2021 12:10 am
by babala
I'm using Phaze-A model with 'stojo' preset.
when run it with TF 2.4, iteration about 11.512 EGs/s
and TF 2.6 is 7.58 EGs/s
so, yeah torzdf said Phaze-A is heavy one.
according to viewtopic.php?f=27&t=1525