index_timing.py

"""
snippet to time different indexing strategies

based on:
http://wesmckinney.com/blog/?p=215
http://stackoverflow.com/questions/11800075/faster-numpy-fancy-indexing-and-reduction/11813040#11813040
http://stackoverflow.com/questions/14386822/fast-numpy-fancy-indexing?rq=1
"""

import numpy as np
import random

indexer = np.arange(5000)
random.shuffle(indexer)

arr = np.array(np.random.randn(10000, 10000), order="F")
%timeit arr[:,indexer]
%timeit arr.take(indexer, axis=1)

arr = np.array(np.random.randn(10000, 10000), order="C")
%timeit arr[:,indexer]
%timeit arr.take(indexer, axis=1)


"""
In [46]: indexer = np.arange(5000)

In [47]: random.shuffle(indexer)

In [48]:

In [48]: arr = np.array(np.random.randn(10000, 10000), order="F")

In [49]: %timeit arr[:,indexer]
1 loops, best of 3: 1.41 s per loop

In [50]: %timeit arr.take(indexer, axis=1)
1 loops, best of 3: 2.35 s per loop

In [51]:

In [51]: arr = np.array(np.random.randn(10000, 10000), order="C")

In [52]: %timeit arr[:,indexer]
1 loops, best of 3: 3.08 s per loop

In [53]: %timeit arr.take(indexer, axis=1)
1 loops, best of 3: 438 ms per loop

"""

# verdict: use C-order with take