Pytorch does practically speaking not support computing Jacobians, but does have reasonable support for Jacobian-vector products. We can leverage this to e.g. provide a default implementation of Manifold.inner(). To do this nicely we need some API for determining if 'embed' provides an optional Jacobian or not.
Pytorch does practically speaking not support computing Jacobians, but does have reasonable support for Jacobian-vector products. We can leverage this to e.g. provide a default implementation of Manifold.inner(). To do this nicely we need some API for determining if 'embed' provides an optional Jacobian or not.