Speedy

[atarashansky]

Hello!

I've upgraded the xicor implementation to be much faster. The results are identical compared to the original implementation.

A couple notes:

1. I removed all the properties. Properties are not implicitly cached and so every time you reference a property, it is recalculated. As a result, you are unecessarily re-ranking the data many, many times.

2. I am using numpy.unique with return_inverse and return_counts flags to quickly calculate f and g, which means I don't need to explicitly rank any of the data.

3. The ties argument in the p-value function was inverted. I fixed the if statement.

Let me know what you think.

[PetterS]

This is quite nice but I think it is possible to make it even faster with a single call to unique. E.g. something like this:

def _ranks(x):
    _, inverse, counts = np.unique(
        x, return_index=False, return_inverse=True, return_counts=True
    )
    r = np.cumsum(counts)[inverse]
    # Compute l the same way but flip the counts first.
    counts = np.flip(counts)
    inverse = len(counts) - inverse - 1
    l = np.cumsum(counts)[inverse]
    return r, l

It is what I did when I tried computing this correlation coefficient.

[zbarry]

Hey @atarashansky - cached_property https://docs.python.org/3/library/functools.html#functools.cached_property might be useful for taking care of the multiple calls to properties.