search

open-mmlab / mmskeleton

A OpenMMLAB toolbox for human pose estimation, skeleton-based action recognition, and action synthesis.
Apache License 2.0
2.92k stars 1.03k forks source link

vanishing gradient #197

Open HenryDo147 opened 5 years ago

HenryDo147 commented 5 years ago

Code version (Git Hash) and PyTorch version

Dataset used

I used NTU‘s RGB-D dataset same as you, but I used ColorX and ColorY information, not 3 channels that you used.

Expected behavior

Actual behavior

I have faced vanishing gradient problem, have you ever faced this problem.

Steps to reproduce the behavior

Other comments

yysijie commented 5 years ago

Since we add BN layers before all convolution layers, the gradient should be stable in normal case. In this codebase, we simply add a BN layer at the begin of networks for data normalization. It works well on two datasets. I think the value scale is also not the reason for causing this problem. Is possible that ColorX and ColorY you mentioned are not effective in action recognition?