Closed wiseodd closed 3 weeks ago
Before, they were duplicates. The rationale is to make functorch_jacobians accessible even when the subclasses overloaded jacobians. Useful for Bayesian optimization since not all backend's jacobians are backpropable (e.g. ASDL detaches everything).
functorch_jacobians
jacobians
Let's forget about this due to #187. The next plan is to decouple the Hessian (for fit) and Jacobian (for GLM) computations.
Before, they were duplicates. The rationale is to make
functorch_jacobians
accessible even when the subclasses overloadedjacobians
. Useful for Bayesian optimization since not all backend'sjacobians
are backpropable (e.g. ASDL detaches everything).