Reimplementation results about integral pose regression

[ducongju]

trafficstars

I am using the following command with 4 RTX 2080Ti:

NNODES=1 NODE_RANK=0 PORT=29501 MASTER_ADDR="127.0.0.1" CUDA_VISIBLE_DEVICES=0,1,2,3 bash tools/dist_train.sh configs/body_2d_keypoint/integral_regression/coco/ipr_res50_8xb64-210e_coco-256x256.py 4 --work-dir work_dirs/ipr_res50_8xb64-210e_coco-256x256 --cfg-options train_dataloader.batch_size=8 train_dataloader.num_workers=2

but final results drop about 5% AP compared to the given results 0.633 AP:

Epoch(val) [210][814/814] coco/AP: 0.586075 coco/AP .5: 0.842064 coco/AP .75: 0.655482 coco/AP (M): 0.549865 coco/AP (L): 0.662681 coco/AR: 0.688114 coco/AR .5: 0.906801 coco/AR .75: 0.758974 coco/AR (M): 0.638323 coco/AR (L): 0.758714

Am I missing something?

[Tau-J]

Hi, it seems that your batch_size is too small than ours.

[ducongju]

Hi~what batchsize should I set? 64?