Adam Lewis
Adam Lewis
I'd recommend we put something like num_krylov_vecs = min(num_krylov_vecs, n-1) somewhere far enough down the execution path that n-1 is available.
It's possible I misdiagnosed this - I'll try and see what's up tomorrow On Fri, Aug 7, 2020, 01:25 Chase Roberts wrote: > @alewis I tried the code you wrote...
I'm still getting it from eigsh_lanczos though - I'm trying to pinpoint exactly when
``` def test_eigsh_lanczos(): """ Compares linalg.krylov.eigsh_lanczos with backend.eigsh_lanczos. """ n = 2 n_kry = 4 shape = (n, n) dtype = np.float32 A = jax.numpy.ones(shape, dtype=dtype) A = 0.5 *...
Ignore the docstring. Does that code pass for you?
It's actually new information to me that this algorithm is expected to fail for singular matrices. Why does it only fail in Jax?
I just checked and when n = 8, n_kry = n - 1 still gives NaN. So indeed it is probably to do with the matrix and not with the...
What a weird coincidence that implicit Arnoldi had some apparently unrelated bug
This one I haven't touched yet. I wanted to clarify what dtypes we needed to focus on.
I'm doing it. vuMPS is done but I want to write it in terms of tn.Tensor, which I'm almost ready to do On Wed, Jul 1, 2020, 09:56 Martin Ganahl...