Two months ago, I trained a segmentation network using Fastai ver.1.0.51, and I got excellent results. However, yesterday I updated Fastai to version 1.0.59, I tried segmenting the same image, and the results were terrible. So, I downgraded Fastai, and the results are back to normal.
Is it the training that gives a worse model or is the same model you had performing worse?
It is the same model I had performing worse!!. The thing that saved me is I used Floydhub website during the training stage. I checked the results in Floydhub and they were as I left them two months ago. I downloaded the trained model from the site on my PC and again I got bad results. Therefore, I knew that the problem is not in the trained model, the problem is either in the installed fastai or some drivers. Luckily, I started with downgrading Fastai to the same version I used in floydhub and it solved the issue.
Ok, then I think it’s because we changed a few things in the unet and the old model might not load properly.
Is there a way to fix this? It is just a pain in the neck starting the training process again.
You just need to map the weights from one state dict to the other. They may have changed names, but they should still match in shapes.