The training loss is contain 1.33, it can not decrease.

[201814476285951]

I used python main.py --task ctdet --exp_id coco_dla --batch_size 32 --master_batch 15 --lr 1.25e-4^C--gpus 0,1
The datasets have not changed.

[xingyizhou]

It does take long to train coco in two GPUs. You can see what happens after training it for a day.

[201814476285951]

It have trained 40 hours,running to about 40 epoch ,and it still contain 1.32

[xingyizhou]

Can you make sure the code is unchanged? Or you can try training pascal voc first to see if you can reproduce the results.

[Holmes-GU]

Hi, when you finish raining, about how larg the hm_loss drops to?