I used PSPNer based on resnet-50 in my own dataset. I trained it on a single RTX2070 with apex amp. During training, the loss declined to 3 (no convergencing). While in evaluation, all metics are nan%. I checked the eval.py and evaluator.py. In evaluator.py the scale_process() function in Evluator, I visualize all score and I found that the line "score=score.permute(1,2,0)" may go wrong. Could you tell me how to fix this problem.
I used PSPNer based on resnet-50 in my own dataset. I trained it on a single RTX2070 with apex amp. During training, the loss declined to 3 (no convergencing). While in evaluation, all metics are nan%. I checked the eval.py and evaluator.py. In evaluator.py the scale_process() function in Evluator, I visualize all score and I found that the line "score=score.permute(1,2,0)" may go wrong. Could you tell me how to fix this problem.