train loss don't converge

Jeff-sjtu / HybrIK

Official code of "HybrIK: A Hybrid Analytical-Neural Inverse Kinematics Solution for 3D Human Pose and Shape Estimation", CVPR 2021

MIT License

1.17k stars 142 forks source link

train loss don't converge #126

Closed cbyzju closed 1 year ago

cbyzju commented 1 year ago

I use resnet34 as the backbone, after epoch 40, the train loss becomes very large, and don't converge in the remained epoches.

Jeff-sjtu commented 1 year ago

Hi @cbyzju, I have fixed the training code. You can try it again.

Jeff-sjtu commented 1 year ago

Thank you for your quick reply! after removing robust train strategy, model could converge normally. Moreover, I found you add 3d joint regression paradigm with weighted_laplace_loss, is it possible to use RLE loss to get better regression result during train stage?

Yes, I am working on it.