The generated motions are static? about motionclip HOT 7 CLOSED

Hi, I followed the script and code with torch 1.8.1 + cuda 11.1, but still got static motions on my own training on Amass. Official checkpoints work well. Any other clues here? @GuyTevet
Is it possible to share your env info with me? Just wanna align the environment installation. @srph25 Thanks a lot.

from motionclip.

Hi Ling-Hao,
Did you train with our configuration (in readme), or use your own settings?

from motionclip.

Hi Ling-Hao, Did you train with our configuration (in readme), or use your own settings?

I check the training and inference settings just now. The static results are inferenced by 500-epoch pretrained model. Motions are active when I switch to the pretrained model with 100-epoch training.

Are your results the same?

from motionclip.

Yup. We present results on epoch 100. Anyway, please note that the training loop currently has a reproduction issue (detailed in readme). Please reopen if you have any other questions regarding this.

from motionclip.

Hi,
After retraining paper-model with the README configuration, I'm still getting constant movements.

10 epochs generates some movements:

20 epochs makes them all constant:

100 epochs keeps them constant and makes them all the same:

Whereas loading the pretrained weights gives the expected results:

Any ideas what I did wrong? Thank you!

from motionclip.

Solved my previous post:
Turned out conda was loading my site-packages installed pytorch 1.10.2 instead of the repository suggested 1.8.1. Downgraded by uninstalling the more recent version.

Now I get much better loss values and closer results after 100 epochs:

from motionclip.

According to similar problems met by @shunlinlu and @srph25 . I hope to reopen the issue to fix the bug~

from motionclip.