HRNet / DEKR

This is an official implementation of our CVPR 2021 paper "Bottom-Up Human Pose Estimation Via Disentangled Keypoint Regression" (https://arxiv.org/abs/2104.02300)
MIT License
437 stars 76 forks source link

about training loss #12

Closed shanyun123456 closed 3 years ago

shanyun123456 commented 3 years ago

The loss of training is very small and decreases slowly when i finetune model on cocodatasets. Is't normal? and how many epoches do you set?

shanyun123456 commented 3 years ago

taining from scatch on cocodatasets, not finetune

Gengzigang commented 3 years ago

We used 140 epochs. You can find the settings that we used exactly in the experiments folder.

shanyun123456 commented 3 years ago

what about the loss of training

Gengzigang commented 3 years ago

The heatmap loss is usually on the order of 1e-4 and the offset loss is usually on the order of 1e-5.

shanyun123456 commented 3 years ago

Thanks, By the way, What the loss of training did you get when you finish the 140 epochs of training?

Gengzigang commented 3 years ago

For the model with HRNet-W48 backbone, the AP on COCO validation with multi-scale testing is 72.3, the heatmap loss is 1.752e-4 and the offset loss is 4.764e-5.

shanyun123456 commented 3 years ago

ok , thank you!