search

microsoft / RepPoints

Represent Visual Objects by Point Sets
MIT License
589 stars 75 forks source link

Why did I get 37.5 AP with "reppoints_minmax_r50_fpn_1x.py" on coco dataset? #3

Open yangli18 opened 5 years ago

yangli18 commented 5 years ago

Thanks for sharing your work! I used the following command to train the model on coco dataset. 

./mmdetection/tools/dist_train.sh ./configs/reppoints_minmax_r50_fpn_1x.py 8 --validate

I have tried using 4 gpus and 8 gpus for model training, but all the results are about 37.5 AP. I wonder what I got wrong or missed? Thanks very much!

yangze0930 commented 5 years ago

Hi, have you decreased the learning rate to 0.005 when you tried using 4 gpus?

yangli18 commented 5 years ago

No, I used the same initial learning rate 0.01 when training with 8 gpus and 4 gpus. I will try training on 4 gpus with the decreased learning rate and check the results again. Thanks for your suggestion!

yangli18 commented 5 years ago

@yangze0930 Hi, I got 38.0 AP / 58.6 AP50 / 40.8 AP75 this time (0.005 lr, 4 gpus), and it's still different from 38.2 AP / 59.7 AP50 / 40.7 AP75 in your paper. Is this a normal deviation?

yangze0930 commented 5 years ago

@yangli18 Sure. In fact, the minmax method seems not as stable as moment method, as only four points receive localization supervision in the refinement.

yangli18 commented 5 years ago

@yangze0930 I see. Thanks for explaining!