Source-Free-Object-Detection-by-Learning-to-Overlook-Domain-Style How large was the loss after 160000 iterations for a typical Enhance task?

How large was the loss after 160000 iterations for a typical Enhance task?

Open QiuJueqin opened this issue 2 years ago • 1 comments

Hi, I tried to reproduce Enhance stage in my own datasets (autonomous city scenarios, not special), but after 160000 iterations, both loss_c + loss_const and loss_s_1 + loss_s_2 are still very large (about 1E3 and 1E2 respectively), and the enhanced images looked very abnormal, is this a normal phenomenon?

BTW, I changed RandomCrop(128) to RandomCrop(512) in data transform, and kept other settings unchanged.

Could you share loss curves or something to let me verify if my training went well? Thanks.

Jul 04 '22 11:07 QiuJueqin

Yes, your losses are normal. I guess you may have added too much style to the original image. You can change the alpha during testing. Or you can also set a small style_weight during training.

Jul 06 '22 10:07 Flashkong

Source-Free-Object-Detection-by-Learning-to-Overlook-Domain-Style Source-Free-Object-Detection-by-Learning-to-Overlook-Domain-Style copied to clipboard

How large was the loss after 160000 iterations for a typical Enhance task?

Source-Free-Object-Detection-by-Learning-to-Overlook-Domain-Style
Source-Free-Object-Detection-by-Learning-to-Overlook-Domain-Style copied to clipboard