training loss is strange,for openpose body_25 model training about realtime_multi-person_pose_estimation HOT 9 OPEN

hello, have you solved this question? my architecture is the same as you

from realtime_multi-person_pose_estimation.

@SyGoing no , my finall loss is too big ,i still try to change lr .but ~~~~~,i have no idea ~~~~, how do u think about this?

from realtime_multi-person_pose_estimation.

maybe the real architecture is only one L1 loss and L2 loss. I am trying it again !

from realtime_multi-person_pose_estimation.

@SyGoing i think should have multi stage , i add normlize for net, my last problem is solved. but ~~~~finall loss is still too big~~~,and result is not good

from realtime_multi-person_pose_estimation.

@SyGoing sorry not normlize, should be initialize

from realtime_multi-person_pose_estimation.

yean ,like your architecture(4 stages L2 loss and 2 stages L1 loss, I use the vgg19 pretrained model to initialize the pose model , and the loss is changing smaller during training.

from realtime_multi-person_pose_estimation.

@oujieww, hello. I want to train body_25 on my dataset too. And my question, how did you obtain your train model?

from realtime_multi-person_pose_estimation.

Hi, @oujieww I wanna train body 25 model by this rep. I do not need feet keypoints, so I only modify the prototxt file like yours but with the 17-points coco data (download from this rep, 188G lmdb file). However, my new trained model can not detect any points. Can you give me some advice?
I set the outputs of 'Mconv7_stage0_L1' and 'Mconv7_stage1_L1' as 19, I wonder why you set them to 21? I will get an error when I set them to 21.

from realtime_multi-person_pose_estimation.

How can I find models and papers about body25?

from realtime_multi-person_pose_estimation.