Vykstorm / numpy-bool-argmax-ext

Additional helper methods to improve numpy.argmax performance for 1D boolean arrays with strides

Introduction

This library provides an alternative to numpy.argmax to get the maximum value in a 1D boolean array with strides, without performing a copy of the input array, and thus increasing the performance.

The rationale is explained in this stackoverflow question

Requirements

The only dependency is numpy.

For now, only tested with Python>=3.7 and numpy==1.16.4

Installation

Via setup.py script:

git clone https://github.com/Vykstorm/numpy-bool-argmax-ext.git
cd numpy-bool-argmax-ext
python setup.py install

Usage

Now you only need to use the function argmax defined in this library instead of np.argmax

e.g:

import numpy as np
from argmaxext import argmax

a = np.random.randint(0, 2, 10000, np.bool)
print(argmax(a))

Execute the next benchmark to compare both functions when dealing with boolean 1D arrays and -1 as stride value for example:

from timeit import timeit

a = np.zeros([2 ** 18], np.bool)
# Worst case scenario (only the first item is True)
a[0] = True

k = 10000
print("np.argmax(a[::-1]) average time: {:6f} msecs".format(
    1000*timeit(lambda: np.argmax(a[::-1]), number=k) / k
))

print("argmax(a[::-1]) average time: {:6f} msecs".format(
    1000*timeit(lambda: argmax(a[::-1]), number=k) / k
))