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

Metric changes resulting from changing the training mode #74

Closed erlingling closed 2 months ago

erlingling commented 2 months ago

Hello and thank you for your work! I would like to ask if when training the mask transformer, I set the conditional mode to uncond and changed the mask ratio from random to fixed 20% ZERHGO ~`E %7%0~K~WQ D9 , which results in a large fid (around 17), do you know what this is asking?

Murrol commented 2 months ago

The fixed mask ratio would cause inconsistency between training and inference.