Use interpolant caching for performance improvements

In the ROQLikelihood and calibration, interp1d instances are repeatedly instantiated. Minor improvements are made by made by caching the interpolant and only updating what is required.

changed milestone to %0.5.3

To show why this improves things,

>>> x = np.linspace(0, 1, 1000)                                             
>>> y = np.sin(x) 
>>> %timeit I = interp1d(x, y, kind='cubic'); I(x)                         
454 µs ± 15 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
>>> %timeit I(x)                                                           
20.7 µs ± 292 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
>>> %timeit I.x = x; I.y = y; I(x)                                         
20.9 µs ± 280 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)

The difference is more apparent with cubic (with linear it is about the same).

I've also checked that this does in fact update the interpolant.

added 1 commit

• 9375117c - Fix missing argument

Compare with previous version

Fascinating! I think the property names of the CubicSpline have an extra get_, but otherwise this looks great.

added 1 commit

• 2872807e - Fix bugs in tests

Compare with previous version

@colm.talbot I love your enthusiasm.

approved this merge request

approved this merge request

mentioned in commit b0fb571f

merged

mentioned in commit be76cf72

mentioned in merge request !557 (merged)

mentioned in commit 2e944741