We are trying to upgrade our target system from Ubuntu 18.04 to Ubuntu 24.04. In this process, torch, torchvision, and torchaudio versions are upgraded, among other packages.
We are seeing mostly similar model output. But every once in a while, the label and confidences completely change. I’m not looking for a solution, but rather a place to start debugging. Is it possible that the change in torch.save/load methods could cause this? This was my first thought. And is my first path to debugging.
Are there other things that could contribute to such large changes in model outputs?
Thank you for your time.