Pytorch and License
Opened this issue · 2 comments
Do you have a plan for the torch version and any simple inference example on how to test your model on new videos?
I see many source files in the repo are under Apache license. Is the repo under Apache license? Can you eventually add a License file?
In the meantime, is the jax code ready?
Cause I've quickly tried to inference on a new video using a subset of the eval code in supervised_point_prediction.py and the output points are quite random.
It seems that the distributed final checkpoint are not compatible with the restore function in checkpoint utils.
In any case I have a result now but I still see some instabilities also at an higher resolution so I don't know if there is a residual error or it is really a performance limit of the model.
I hope that with the pytorch version we have a more clear path to just run the inference standalone.