search

megvii-research / Iter-E2EDET

Official implementation of the paper "Progressive End-to-End Object Detection in Crowded Scenes"
MIT License
88 stars 8 forks source link

AssertionError Boxes #29

Open mdaehl opened 12 months ago

mdaehl commented 12 months ago

Hello, I like your work and wanted to try your code, but I am constantly encountering AssertionErrors (see image). I use the config you recommend for 1 GPU: "--config-file configs/50e.6h.500pro.ignore.yaml --num-gpus 1 SOLVER.IMS_PER_BATCH 2 SOLVER.BASE_LR 0.0025" and crowdhuman data. After debugging I realized that the network predicts nan boxes that cause the error. However, I do not know how to cirumvent this. In your code I found the comment "degenerate boxes gives inf / nan results", but there is no indication how to fix this. Do you have an idea how to solve this problem or how did you handle such boxes? (Note: The loss is decreasing during the training.)

Thanks in advance!

crash
yexiguafuqihao commented 12 months ago

Try to decrease the learning rate. One case is to replace the learning rate with 0 and observe whether the mistake happens or not.