SegDiff icon indicating copy to clipboard operation
SegDiff copied to clipboard

Errors in training

Open 170744039 opened this issue 1 year ago • 1 comments

Hello, thank you for sharing. When I trained on my own binary semantic segmentation dataset and used Monu's data loading, I found that the test images saved during the training were all ineffective noise images. After the training was completed, loading the ema.pt file for testing still yielded the same result, but the model.pt file was normal. May I know which step I might have made an error in?

170744039 avatar May 15 '23 07:05 170744039

Hmm, I don't know the reason, in my work I used the ema.pt model in all datasets and in all experiments. I would start by training the binary segmentation monuseg dataset (that is in the repo) and validate that the ema.pt is working as well, because maybe it is related to changed you made or to the dataset (or maybe not long enough training)

tomeramit avatar Jun 03 '23 07:06 tomeramit