Whether the calculated Rt is differentiable

[linxcan]

Hello, I observed on your code that when Rt is calculated according to the point correspondence, it is solved by SVD. But is the loss obtained in this way derivable?

https://github.com/mbanani/unsupervisedRR/blob/4ba5b1cc76080f5c6fd99cec155453f43b547bcd/unsupervisedRR/models/alignment.py#L127

[mbanani]

Hi @linxcan,

PyTorch's implementation of SVD is differentiable, as seen here, allowing us to define losses using the estimated Rt and back propagating the loss to the features.

[linxcan]

Thanks for your response!