nan loss about bbavectors-oriented-object-detection HOT 2 CLOSED

I've met this problem before. It could be the focal loss problem. I think one possible reason is the log (x) part, where the x is extremely approaching 0. I think the output of the heat map would have a very large negative value that after sigmoid it is 0. I think adding a Normalization layer in the second layer of the head part would help. But I didn't have this problem after trying large batch size. Maybe decreasing the learning rate would help too.

The hrsc_evaluation_task1.py is modified from dota_evaluation_task1.py in DOTA_Devkit, I uploaded it to the datasets, the code is similar to the original one.

from bbavectors-oriented-object-detection.

I've met this problem before. It could be the focal loss problem. I think one possible reason is the log (x) part, where the x is extremely approaching 0. I think the output of the heat map would have a very large negative value that after sigmoid it is 0. I think adding a Normalization layer in the second layer of the head part would help. But I didn't have this problem after trying large batch size. Maybe decreasing the learning rate would help too.

The hrsc_evaluation_task1.py is modified from dota_evaluation_task1.py in DOTA_Devkit, I uploaded it to the datasets, the code is similar to the original one.

Thank you very much. I would try.

from bbavectors-oriented-object-detection.