Pretraining RFBNet on source domain dataset COCO60, it shows loss nan

[LexieYang]

Hello, I really appreciate your work! When I pretrain RFBNet on source domain dataset COCO60, it shows loss nan. Is it correct and why?

[whattoshow]

I have the same problem from iter2940...Is there anybody know why?

[whattoshow]

Hello ,I've asked my classmate, they told me the exploding gradient may cause this problem.

[Ze-Yang]

Hi, do you make any modifications to the code? If so, please first try with the original code.

[whattoshow]

Hi, do you make any modifications to the code? If so, please first try with the original code.

Thanks a lot for your reply~!!! Due to my computer's GPU memory is small , I've got 'CUDA out of memory ' error when running the original code,so I changed the batch size to 32, then pretrain RFBNet on source domain dataset COCO60, and met the same problem of LexieYang's.

[Ze-Yang]

I will verify it and get back to you soon.

[whattoshow]

I will verify it and get back to you soon.

Thank you!!!!:blush:

[jamesrosstwo]

Hello! I am encountering the same problem, have there been any updates regarding this?