search

yf19970118 / OPLD-Pytorch

Official implementation of Learning Point-guided Localization for Detection in Remote Sensing Images
MIT License
28 stars 3 forks source link

Training errors #3

Open zhengshangdong opened 3 years ago

zhengshangdong commented 3 years ago

Hello sir! Thank you for your great job! I get the right results in the test_net.py. However, there are some errors when I run your train code. The loss is always Nan when I resume from your model_latest.pth or the original caffe weights resnet101_caffe.pth. Do you have any suggestions? Yours!

yf19970118 commented 3 years ago

Hello sir! Thank you for your great job! I get the right results in the test_net.py. However, there are some errors when I run your train code. The loss is always Nan when I resume from your model_latest.pth or the original caffe weights resnet101_caffe.pth. Do you have any suggestions? Yours!

I haven't met the situation that RPN related loss becomes Nan. When generating data, we should pay extra attention to the situation that the target box is truncated into five points. I made changes in dota_devkit. This repo is separated from my other projects. I have conducted test, but I haven't retrained it. I can't rule out the possibility of bugs. I'm very sorry that I don't have time to investigate the problem in detail.