Bug in image reconstruction with model.py

[ariskoutris]

Hi,

I'm trying to set up the environment and test it by executing model.py. Even though no errors are returned, the reconstructed image seems very wrong.

My setup is:
WSL Ubuntu 18.04
Anaconda Python 3.6
tensorflow-gpu=1.8.0, horovod=0.13.8
GPU Compute Capability 8.6 (3060Ti)
Cuda 9.0, CuDNN 7.6.3

Has anyone dealt with something similar?

Thank you in advance.

[ariskoutris]

Created the same environment on a machine with a GTX 1080 and the same experiment produces correct results.

I believe that there is a compatibility issue with the newer 3060Ti but I cannot really say what it is. Please note that I tried multiple compatible combinations of CUDA {9,10,11} and CUDNN with no success.