Now I am trying to reproduce papers result(MPJPE: 40.6). I use 'integral_human_pose/experiments/hm36/resnet50v1_ft/d-mh_ps-256_deconv256x3_min-int-l1_adam_bs32-4gpus_x300-270-290/lr1e-3.yaml' and HM36_train_cache file downloaded from your dropbox. And I put h36m images in the right position.
Now i have trained for 50 epochs, but I found the MPJPE in h36m_validmin_dataset stayed around 80mm and stopped reducing.
could you please help me?
Hi,50 epochs is not enough You'd better complete 300 epochs, and besides try other learning rates to see accuracy varying. What's more, could you post the learning curve?
Now I am trying to reproduce papers result(MPJPE: 40.6). I use 'integral_human_pose/experiments/hm36/resnet50v1_ft/d-mh_ps-256_deconv256x3_min-int-l1_adam_bs32-4gpus_x300-270-290/lr1e-3.yaml' and HM36_train_cache file downloaded from your dropbox. And I put h36m images in the right position.
Now i have trained for 50 epochs, but I found the MPJPE in h36m_validmin_dataset stayed around 80mm and stopped reducing. could you please help me?