Closed breAchyz closed 3 years ago
I change the loss_bbox to GIoULoss, and it solve my problem. But the reason of loss_bbox=nan while loss_bbox set to L1Loss is not clear yet. If some one have any idea, please leave your comments, thanks.
I only met this problem in the situation where some of the ground truth boxes' widths or heights are zero. So you may also check your data to see whether there are zero-sized boxes.
Thanks for your error report and we appreciate it a lot.
Checklist
Describe the bug When I train using a config file which modified based on faster_rcnn_r50_fpn.py, the most loss_bbox get nan.
Reproduction
What command or script did you run?
Did you make any modifications on the code or config? Did you understand what you have modified? Yes, I change the lr and lr_config. About the datasets, I modified the CLASSES of VOCDatesets, num_classes=2
What dataset did you use? VOCDatasets cloth Environment
Please run
python mmdet/utils/collect_env.py
to collect necessary environment information and paste it here. sys.platform: linux Python: 3.7.7 (default, May 7 2020, 21:25:33) [GCC 7.3.0] CUDA available: True CUDA_HOME: /usr/local/cuda NVCC: Cuda compilation tools, release 9.2, V9.2.148 GPU 0: GeForce GTX 1080 Ti GCC: gcc (Ubuntu 5.4.0-6ubuntu1~16.04.12) 5.4.0 20160609 PyTorch: 1.4.0 PyTorch compiling details: PyTorch built with:TorchVision: 0.5.0 OpenCV: 4.2.0 MMCV: 0.6.2 MMDetection: 2.2.0+741b638 MMDetection Compiler: GCC 5.4 MMDetection CUDA Compiler: 9.2
Error traceback If applicable, paste the error trackback here.
Bug fix If you have already identified the reason, you can provide the information here. If you are willing to create a PR to fix it, please also leave a comment here and that would be much appreciated!