Reproducing KIT results

[nefeliandreou]

Hi, I trained from scratch the VAE and diffusion modules on the KIT dataset. I followed your instructions and:

• downloaded the processed data from HumanML3D
• changed only the dataset params on the config files.

I observed that the AVE and APE metrics achieve very high values (see image below).

It seems to me that the two datasets (KIT and HumanML3D) have different scales, so I was wondering whether other hyperparameters (e.g. LOSS.REC_JOINTS, LOSS.LAMBDA_REC) should be changed when training on the KIT dataset. Do you have any insights on this?

Thanks a lot!