Niklas Koep
Niklas Koep
### Describe your issue. I'm not sure if this is expected behavior, but scipy 1.9.0 breaks code that manually vectorizes calls to `scipy.linalg.expm` via `np.vectorize`. I understand support for calling...
The backend functions are currently not tested individually in isolation. This means that bugs in the backend abstractions are only spotted by unit tests if some of the more involved...
By default, the pytorch and tensorflow backends use float32 types, while numpy defaults to 64 bit. This should probably be configurable by user. Numerical tolerances throughout the codebase should also...
This PR adds a provisional JAX backend. The PR supersedes #115 due to changes in the backend API which made it easier to re-implement the backend code. Unlike #115, the...
This (experimental) PR replaces autograd's hvp implementation with a version based on forward-over-reverse mode autodiff (see https://jax.readthedocs.io/en/latest/notebooks/autodiff_cookbook.html?highlight=hvp#hessian-vector-products-using-both-forward-and-reverse-mode). This should generally be more efficient that our handrolled implementation that manually computes...