ENH: Jacobian memoization for multivariate minimization #177
Conversation
… L-BFGS-B The `MemoizeJac` decorator added in `optimize` is used for caching the value of the gradient when the latter is calculated along with the objective function.
Add tests for this based on TNC and SLSQP.
| Jacobian of objective function (if None, Jacobian will be | ||
| estimated numerically). Only for CG, BFGS, Newton-CG. | ||
| Jacobian of objective function. Only for CG, BFGS, Newton-CG. | ||
| If None, Jacobian will be estimated numerically. If the `jac==fun`, |
There was a problem hiding this comment.
Changing the semantic role of a parameter based on object identity is IMHO a bit too magical here. I'd prefer to use jac=True or jac="combined" or some other non-callable special value to signal the changed behavior.
I wonder if one should be careful with the comparisons, as the object jac passed in may in principle implement a __eq__ method. Seems rare, though, but if yes, for a string, isinstance(jac, str) and jac == "combined", and for boolean, not callable(jac) and bool(jac) could do it.
There was a problem hiding this comment.
I haven't thought about that, thanks. I choosed the Boolean option and set the default value of jac to False. 4ed75ba
|
Ok, this gets +1 from me. I don't see obvious problems in merging this. |
|
@dlaxalde: you'll have to rebase this on master and test again, because I just merged the L-BFGS-B 3.0 upgrade. If there aren't any problems, I'd say to ahead and push it. |
* enh/optimize/memoize-jac: ENH: jac is either a bool or a callable in minimize ENH: allow objective function to also return the jacobian in minimize ENH: memoize gradient when calculated along with obj. fun. in TNC and L-BFGS-B FIX: drop multiplie imports of optimize in tnc Conflicts: scipy/optimize/lbfgsb.py scipy/optimize/minimize.py scipy/optimize/tnc.py
fun(x, *args) -> f, fprime) inminimizethus making it available for all solvers.I'm quite new to these memoization techniques so any comment very welcome.