pytorch_fnet icon indicating copy to clipboard operation
pytorch_fnet copied to clipboard

Published 20 hours ago • verified publisher icon AllenCellModeling

There are many grids on predict results

Open mQvQ opened this issue 3 years ago • 4 comments

I run the download_and_train.py to train a model .But when I run predict.py ,I find many girds on my predict results. why is that?

mQvQ avatar Jul 08 '21 06:07 mQvQ

Try increasing the number of images (default is 40) and the number of iterations (default is 50000).

alxndrkalinin avatar Jul 11 '21 10:07 alxndrkalinin

Try increasing the number of images (default is 40) and the number of iterations (default is 50000).

thanks for your reply ,but when I train 75 images and 80000 iters there are still many grids or even more. 截屏2021-07-13 上午10 25 29

mQvQ avatar Jul 13 '21 02:07 mQvQ

You can also try ad_customization branch that added TTA and changed the prediction tile size. Maybe using non-zero overlap between prediction tiles can help. Finally, I recommend checking out Distill article for better understanding of these artifacts.

In my practice, I've seen these artifacts diminish when using more data and longer training.

alxndrkalinin avatar Jul 13 '21 19:07 alxndrkalinin

Hi, I have the same trouble. Have you solved this circumstances? Could it be caused by the salts-and-peppers noises on the background of target images?

NikLi66 avatar Nov 08 '21 04:11 NikLi66