argrelmin/argrelmax fail with smooth data and plateaus

[Chris7]

There are certain cases that argrelmin/max will fail due to the comparator being used. For instance:

from scipy.signal import argrelmin, argrelmax
import numpy as np
x = np.array([1,2,3,3,2,1])
print(argrelmax(x))

returns (array([], dtype=int64),)

This happens because the comparator is np.greater. Should this be fixed to provide some resolution or is this the desired behavior?

[person142]

Essentially the same issue as #3749.

[Chris7]

Here is some code that is one possible fix for _boolrelextrema:

def _boolrelextrema(data, comparator, axis=0, order=1, mode='clip'):
    if((int(order) != order) or (order < 1)):
        raise ValueError('Order must be an int >= 1')

    datalen = data.shape[axis]
    locs = np.arange(0, datalen)
    data = np.ma.masked_array(data, mask=np.hstack(([1], np.diff(data)))==0)
    if np.ma.is_masked(data):
        locs = locs[np.ma.getmask(data)==False]
        main = data.take(locs, axis=axis, mode=mode)
        results = np.zeros(data.shape, dtype=bool)
        for index, result in enumerate(_boolrelextrema(main, comparator, axis=axis, order=order, mode=mode)):
            results[locs[index]] = result
        return results
    else:
        locs = locs[np.ma.getmask(data)==False]
        results = np.ones(data.shape, dtype=bool)
        main = data.take(locs, axis=axis, mode=mode)
        for shift in xrange(1, order + 1):
            plus = data.take(locs + shift, axis=axis, mode=mode)
            minus = data.take(locs - shift, axis=axis, mode=mode)
            results &= comparator(main, plus)
            results &= comparator(main, minus)
            if(~results.any()):
                return results
        return results
_boolrelextrema(x, np.greater)

It makes assumptions about how you want to report the duplicates, but it's workable for me atm.

[Chris7]

Bah, sorry for the dupe @person142. Feel free to close/reference this from there if you wish.

[ilayn]

@Chris7 Since it is referenced already, you can close this one yourself anyways.

[person142]

The other doesn't mention argrelmax or argrelmin, so IMO it's not unreasonable to leave both open so long as they are linked (even though fixing _boolrelextrema fixes both). @Chris7 re

Here is some code that is one possible fix

any interest in turning that into a PR?

[Chris7]

I can try. I don't know how I can really translate it to the multi-dimensional case. Since what I am in effect doing is popping out the duplicates, that will quickly fail once I am working with anything > 1d.

[lagru]

There is now an alternative to _boolextrema for the 1D case: _argmaxima1d introduced in #8350. It approximates the center of flat maxima. I think that function could cover the >1D case with a few small changes as well.