Open ChenyanWu opened 4 years ago
my result is also 90.1.hava you changed the parameter and got the result of the paper
My result is also 90.1, have you solved this problem?
you can get 90.03 while yot run python tools/test.py \ --cfg experiments/mpii/hrnet/w32_256x256_adam_lr1e-3.yaml \ TEST.MODEL_FILE models/pytorch/pose_mpii/pose_hrnet_w32_256x256.pth
.
I finally get 90.3. I change the learning strategy. The total epoch is 140, and I change the learning rate at 90 and 120 epoch.
I finally get 90.3. I change the learning strategy. The total epoch is 140, and I change the learning rate at 90 and 120 epoch.
How many GPUs and batchsize you use?
I finally get 90.3. I change the learning strategy. The total epoch is 140, and I change the learning rate at 90 and 120 epoch.
May I ask if you have solved this problem? I changed the learning strategy at 90 and 120epoch according to your method, but the accuracy of 140epoch in total is still less than 90.3 and only 89.3
Recently I reproduced the experiments of HRnet and trained COCO and MPII dataset by myself. I find that the val accuracy of COCO dataset can exactly reach the accuracy in paper. But for MPII dataset my best result of val is 90.1, which is worse than the accuracy in paper 90.3. Is the parameter setting of this github repository exactly same with the parameter setting in paper?