Repository navigation
Allow np.ufunc to make new compound models #3866
Description
Activity
It would be nice to get a lot of analytical functions for free as models instead of having to write wrappers for each one. (Assuming this can be implemented ...) What are the limitations? I guess these models couldn't have
fit_deriv? But that's often not important. Anything else obvious that's hard or impossible?fit_derivwould not be hard -- see https://github.com/astropy/astropy/pull/3715/files#diff-2a3560744629608434d0eff93c50cf2aR13
The main issue, I think, is that there is as far as I know no way to make this work without__numpy_ufunc__, i.e., those with numpy <1.10 will have to import the mathematical functions from somewhere.In any case
fit_derivis not currently implemented for compound models in general, though I think there's an open issue for that wherein I've described what needs to be done to implement it. For the most part I think it's straightforward, but I haven't actually tried yet.In numpy/numpy#5844 (comment), @argriffing pointed to
algopy, which defines functions on models as well, and might well have gradient-estimating code, etc., that is useful here (or for modeling more generally).
This issue is mostly to remind us/@embray of an idea that mentioned in #3861 (comment): could one not just support operations like
+on model classes but alsonp.ufunc? This would seem possible if one defined__numpy_ufunc__appropriately, and would mean one would nearly for free get allufuncs covered as models, ie., one could havenp.sin(Const1D+Scale)construct a new model class.Possibly somewhat unrelatedly, by way of similarity, it might then be good to have
F | Gbeing equivalent toG(F)-- obviously, this needs some trickery inG.__new__...