Importing MemoizeJac from SciPy raises a deprecation warning.

[moorepants]

See #157, but that didn't seem to help. See this issue on SciPy: scipy/scipy#17572

[nrontsis]

What is the reason for using (the now seemingly private) MemoizeJac? Does caching function/Jacobian evaluations help noticeably in continuous optimisers like Ipopt? I would assume it has no effect?

[moorepants]

I don't know. I didn't write that code.

[moorepants]

My guess would be that has little to no effect because those functions would rarely be called with the same arguments.

[nrontsis]

Opened a PR to remove it: #161

[jhelgert]

Does caching function/Jacobian evaluations help noticeably in continuous optimisers like Ipopt? I would assume it has no effect?

Depends on the optimization problem.

For instance, for an objective function that contains some expensive-to-evaluate expression tmp which depends on the optimization variables and arises in the gradient as well, you can avoid duplicated computations of tmp by evaluating the objective and the gradient with a single function call and using the jac=True option. Here, the effect would be huge because you save one expensive calculation per iteration. An illustrative example of tmp might be an inverse matrix that depends on the optimization variable.

IIRC, the main intention for adding this feature to cyipopt's scipy interface was because it's in scipy.optimize.minimize too.

[moorepants]

Yes, I think we should leave it in. After watching the discussion over at SciPy I think we will just have to import from scipy.optimize._optimize and risk breaking changes or we can copy MemoizeJac into this library (and the license of SciPy).