Dataset difference may exist between your work and official HumanML3D

[qrzou]

Hi, thanks for this attractive work!

I met the NaN problem when evaluating the VQVAE. And I found this is caused by the HumanML3D data that contains some NaN motion data.

However, I didn't see any preprocess in your repo to handle NaN problem. Does this mean the dataset you used is different from the official HumanML3D dataset?

Besides, with the dataset I processed according to the official HumanML and dropping the two NaN raw motion data, your provided pretrained VQVAE FID evaluation result is about 0.090, which is higher than 0.070 reported in your paper.
Does this also imply there is a minor difference in yours and the official HumanML3D dataset?

Thank you!

Below is the NaN data in the official HumanML3D https://github.com/EricGuo5513/HumanML3D/blob/main/cal_mean_variance.ipynb
The outputs in the image of HumanML3D notebook mean that the 007975.npy and M007975.npy contain NaN.

[Chatonz]

i have the same question!