hopesala

@HMCBSJ thanks you. it solved my same error.

I have vertified that the above problem was caused by too little training time, training for one day is ok to detect another question is below in file test_dota_base.py threshold...

I am training my own dataset the train process loss value is nan the command is : python train.py mot --exp_id crowdhuman_dla34 --gpus 0 --batch_size 1 --load_model '../models/ctdet_coco_dla_2x.pth' --num_epochs 60...

generate odgt file from this guide https://github.com/zengarden/light_head_rcnn/issues/10 and then can train normally

opps nan again...... mot/crowdhuman_dla34 |############################ | train: [1][3308/3766]|Tot: 0:38:24 |ETA: 0:05:18 |loss 18.8693 |hm_loss 1.1229 |wh_loss 0.2640 |off_loss 0.1886 |id_loss 13.351mot/crowdhuman_dla34 |############################ | train: [1][3309/3766]|Tot: 0:38:25 |ETA: 0:05:17 |loss 18.8684...