WIP: jax eigsh lanczos #359
Conversation
… upstream-master � Conflicts: � tensornetwork/backends/backend_test.py
… upstream-master
… upstream-master
|
Hey @MichaelMarien looks good! I like the idea of having a backend independent lanczos. The reason we decided to put it into the backend is that we want eigsh_lanczos to work under |
What exactly is it we'd like to jit? Is it the (action of the) linear operator we want the eigenvectors of? |
|
Hey @MichaelMarien I think we want to jit both the liner operator as well as the lanczos iteration itself. |
|
implemented in another PR #570 in a jitted way |
This is still very much work in progress, but would like to capture some initial feedback of @Thenerdstation , @mganahl . While taking a first look at #303 I started playing a bit with the code.
It occurs to me that with some additional backend methods, we can have the current numpy code rather generic over the different frameworks. Feedback I would like to capture is whether the inclusion of
backend.dotandbackend.eighmethods is worth making the implementation ofeigsh_lanczosbackend independent (or almost, I'm very close already). There are some other changes already, like refactoring out the checks, but never mind these as this is definitely not a final design.