I notebook file that is working in Colab, and the loss is valid, but when I execute it in Ubuntu with a local GPU, I get some accuracy, but I get nan, as Loss. I have no idea why. Have you seen anything like this?
Hard to tell without seeing some code. Maybe start by checking CUDA/CUDNN versions on both platforms?
Colab for me shows:
I have also had problems with old hardware (780ti) a long time ago behaving strangely when memory consumption was close to 100%.
I have the same problem. did you manage to solve it?