Feature/344 functorch #378
Conversation
| y = hessian_vp(x) + hessian_perturbation * x | ||
| return cp.from_dlpack(to_dlpack(y)) | ||
|
|
||
| eigen_vals, eigen_vecs = eigsh( |
There was a problem hiding this comment.
There is a lot of repetition and return points in this function. I think it should be simplified by making a single call to eigsh and defining the different variables and functions in the if and else branches, along with an extra function to convert the result to torch tensors.
| @@ -21,7 +21,7 @@ | |||
| ], | |||
| setup_requires=["wheel"], | |||
| tests_require=["pytest"], | |||
| extras_require={"influence": ["torch"]}, | |||
| extras_require={"influence": ["torch>=2.0.0"], "cupy": ["cupy-cuda11x>=12.1.0"]}, | |||
There was a problem hiding this comment.
cupy is only used with influence functions, right? What would be the disadvantage of adding it to the influence extras instead of creating a separate one just for it?
There was a problem hiding this comment.
On mac, there is no wheel cupy-cuda11x. If we include it into the influence extra, then no mac user could install the influence extra
There was a problem hiding this comment.
Alright, we can leave it for now.
I will create a separate issue to consider using environment markers for this.
|
@schroedk could you please merge develop into your branch and mark this PR as ready so that CI can run? |
* remove parameter x0 * simplify if/else condition
|
@AnesBenmerzoug as this PR is not into develop, I would first merge into into to the target feature branch and run the CI there |
Changes
Checklist
- [ ] If notebooks were added/changed, added boilerplate cells are tagged with"nbsphinx":"hidden"