search

EricGuo5513 / momask-codes

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

What is the role of "opt.feat_bias"? #61

Closed Run542968 closed 1 month ago

Run542968 commented 4 months ago

Hi, thanks for this great work!

When I was reading the code, I found that each time the dataset was loaded, a new mean.npy and std.npy would be calculated based on opt.feat_bias. May I ask what role does opt.feat_bias play? And how to determine this hyperparameter?

Looking forward to your reply, thx!

Murrol commented 4 months ago

Hi, thanks for your interest.

The mean and std are calculated during dataset loading only when training the autoencoder. The rest of the time, they are loaded from the meta files.

The feat_bias is a weight for the feature, like a weight in the loss function for each feature.

Run542968 commented 4 months ago

Thx! I've noticed that this feat_bias is mainly applied to root and foot, so perhaps I can understand that the purpose of this is to reduce AE's attention to these joints?

vadeli commented 4 months ago

Hi, this was my question too. @Run542968 Did you figure it out?

Run542968 commented 4 months ago

Ah, I am also waiting for the author's further response. Perhaps, as he said, you can consider it as a hyperparameter. @vadeli