The generation of dance movements

[JingYC001]

Hello, I used the Amass dance action dataset and jukebox audio extractor to extract audio features and send them to model training, but the generated results were not very good. May I ask if our model is not suitable for dance audio driven tasks. Is there any other reason?
In addition, I also noticed that when calculating loss, in addition to rot_ Mse, the last few losses were not calculated. Is there any reasonable explanation?

[YoungSeng]

I think it might be more sensitive to the motion representation and the number of iterations, and this EDGE you're talking about should have already done this, it shouldn't have anything to do with the framework, we're all modifying on top of MDM; also about the losses, I just didn't use the other losses in MDM similar to the feet, velocities, etc.; the losses for each joint are still computed, see the code for details.