Why calculating the std during training?

[XingliangJin]

I find that in the class Text2MotionDataset, the code calculates the std of the motions. Why don't you use the provided value in HumanML3D? What is feat_bias used for?

[h-y1heng]

I find that in the class Text2MotionDataset, the code calculates the std of the motions. Why don't you use the provided value in HumanML3D? What is feat_bias used for?

I followed the dataset code provided by T2M, which is a benchmark work on HumanML3D. In the original benchmark, the std value provided in HumanML3D was mainly used for the evaluator.

Regarding the feat_bias, I personally think it is used to adjust the feature values. Specifically, it helps to shift the feature distribution to a more appropriate range. I've actually tried omitting this step, and it didn't seem to have a significant impact on the results. But since I noticed that other code implementations included it, I remain it as well.