princeton-vl/lietorch

ToMatrix() does not have backward operation implemented

lixiny opened this issue · 2 comments

Hi,

I just read through your paper and code and really adore this project on lie group in tangent space.

I hope that the ToMatrix() operation could be fully differentiable for researchers to use.
Do you have a plan to implement the ToMatrix(), or may be FromMatrix() in the future? and a planned date in mind?

Cheers

Lixin

Hi Lixin, I agree this would be good to have. I do have equations for the gradients of ToMatrix so I can add this in the next few days. I will need to think about FromMatrix.

In the future, I am working on trying to make conversions from all different representations differentiable (e.g quaternions, matricies, ect).

Best,
Zach

Best, Zach.
Thanks again for the great contribution you have made.

Lixin