search

EricGuo5513 / momask-codes

Official implementation of "MoMask: Generative Masked Modeling of 3D Human Motions (CVPR2024)"
https://ericguo5513.github.io/momask/
MIT License
856 stars 73 forks source link

Double division of std #52

Closed seokhyeonhong closed 5 months ago

seokhyeonhong commented 5 months ago

Hello, thanks for the great work and sharing the code. While exploring the training code of VQ-VAE, I found that std is divided by opt.feat_bias, which I can understand. But it seems that std is divided in the training dataset first, and it is divided again in the validation dataset due to the call-by-reference function parameters. And it seems that this double-divided std is saved in the meta_dir, which is further used when training the transformers. As the std is different between training and inference of VQ-VAE, I'm just curious if this was intended or just a bug.

Thanks for the great work again, and I look forward to hearing from you.

seokhyeonhong commented 5 months ago

Oh, I found it in the closed issues. I'll close this!