The Std of the dataset was divided by 5 twice in both the train_dataset and val_dataset, and then it happened to equal the Std of the evaluator?

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

The Std of the dataset was divided by 5 twice in both the train_dataset and val_dataset, and then it happened to equal the Std of the evaluator? #50

Closed Dai-Wenxun closed 6 months ago

Dai-Wenxun commented 6 months ago

e096a09368975c39d78a50a8f7c550d 65a1760663a2fc350021cccc2a84e7a 88eb201eaad3535920fa70e3f4aec24 9f95c9828e7d43591e96a120bf3b2f3 f26e0a8d3aa4ca6372be4a63c72e4e1

Dai-Wenxun commented 6 months ago

fc29dfe2aa0196957500b62190d153e

Dai-Wenxun commented 6 months ago

@EricGuo5513 @Murrol

Murrol commented 6 months ago

Yes, it's a bug that remained from the original HumanML3D project. It performs an in-place operation by mistake.

It's not a serious bug and only affects the weight of the feature. So, like all other methods on the HumanML3D benchmark, we follow this setting. https://github.com/EricGuo5513/momask-codes/blob/4e0156ec1f3bd29b8b04e331d840f87d71656a9d/data/t2m_dataset.py#L43

Dai-Wenxun commented 6 months ago

Thanks! bro!