train on cocodataset ( AP is so low )

hustvl / SparseInst

[CVPR 2022] SparseInst: Sparse Instance Activation for Real-Time Instance Segmentation

MIT License

586 stars 72 forks source link

train on cocodataset ( AP is so low ) #117

Open Wooho-Moon opened 1 year ago

Wooho-Moon commented 1 year ago

Thanks for nice work! recently, I'm trying to train sparsenetInst on coco2017 dataset. but AP is wired. I read a paper and document	AP	AP50	AP75	APs	APm	APl
17.363	30.127	17.499	2.721	15.622	32.711

my command : CUDA_VISIBLE_DEVICES=0,1 python tools/train_net.py --config-file configs/sparse_inst_cspdarknet53_giam.yaml --num-gpus 2 --dist-url tcp://127.0.0.1:49153

I don't chang any option. What's wrong with me? How should I do?

116022017144 commented 1 year ago

you see the log and check your backbone whether it loaded sucessfully or not!

Wooho-Moon commented 1 year ago

you see the log and check your backbone whether it loaded sucessfully or not! Thansk for reply :) backbone is sucessfully loaded! but, I cannot reproduce this papers' AP. I don't know what I am missing. and It's a another question. I'd like to reduce the resolution to 192 while remain mask's AP. Could you give a favor?? Plese give me some advices for that task!! thanks.