A few tweaks needed in core/fitting.py

[micahfolsom]

While implementing the new model, I noticed a few things:

1. The num args of ConstantModel.guess() and LineModel.guess() are not used
2. The divides in expgauss() (in the same file) could be multiplies for potentially significant performance improvements (I would want to demonstrate this to be sure, though)
3. Almost none of these functions/models have docstrings

I'll probably address these myself when I've got time

[jvavrek]

I took a quick look at changing expgauss with some toy data and found no significant speedup:

>>> %timeit -n 1000 -r 7 fitter.fit()  # with original /2
# 8.48 ms ± 31.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
>>> %timeit -n 1000 -r 7 fitter.fit()  # with new *0.5
# 8.54 ms ± 94.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

[micahfolsom]

Great, thanks for checking!

[cosama]

I think in general using multiplies is a good choice, particular with 2 (0.5), but you probably need a lot of these cases for it to really make a significant difference. It also looks like numpy is probably not as sensitive to this (maybe they do some optimizations here):

import numpy as np

a = np.random.rand(int(1e7))

%timeit a / 2

19.4 ms ± 1.01 ms per loop (mean ± std. dev. of 7 runs, 100 loops each)

%timeit a * 0.5

18.9 ms ± 386 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

Also googling this a bit, it looks like there might be also the issue with 0.5 not actually being exactly 0.5 in it's floating point representation, hence you introduce some additional floating point rounding errors. Not sure how that works with division by 2. It probably is all hardware dependent.

I would use 0.5, but I feel it is a bit nitpicking.

[jvavrek]

@cosama 0.5 should be exactly 0.5 in floating point