通过一组值过滤 numpy 数组的最快方法

Question

I am pretty new to numpy, I also am using pypy 2.2 which has limited numpy support (see http://buildbot.pypy.org/numpy-status/latest.html ) but what I'm trying to do is filter an array by a set of values (ie keep subarray if it contains a value in a set). I can do with a list comprehension but I'd rather do without the intermediate list as on longer arrays it isn't fast and I can't help but think numpy filtering will be faster.

>> a = np.array([[   368,    322, 175238,      2],
       [   430,    382, 121486,      2],
       [   451,    412, 153521,      2],
       [   480,    442, 121468,      2],
       [   517,    475, 109543,      2],
       [   543,    503, 121471,      2],
       [   576,    537, 100566,      2],
       [   607,    567, 121473,      2],
       [   640,    597, 153561,      2]])

>> b = {121486, 153521, 121473}

>> np.array([x for x in a if x[2] in b])

>> array([[   430,    382, 121486,      2],
   [   451,    412, 153521,      2],
   [   607,    567, 121473,      2]])

Answer 1

You can do it in one line, but you have to use list(b) , so it might not actually be any faster:

>>> a[np.in1d(a[:,2], list(b))]
array([[   430,    382, 121486,      2],
       [   451,    412, 153521,      2],
       [   607,    567, 121473,      2]])

It works because np.in1d tells you which of the first item are in the second:

>>> np.in1d(a[:,2], list(b))
array([False,  True,  True, False, False, False, False,  True, False], dtype=bool)

---

For large a and b , this is probably faster than your solution, as it still uses b as a set, but builds only boolean array instead of rebuilding the entire array one line at a time. For large a and small b , I think np.in1d might be faster.

ainb = np.array([x in b for x in a[:,2]])
a[ainb]

For small a and large b , your own solution is probably fastest.

Answer 2

For relatively small inputs like those in your question, the fastest method is by far and large the naïve one:

np.array([x for x in a if x[2] in b])

This would be true especially for PyPy.

For larger inputs, @askewchan solution with NumPy may be faster:

a[np.in1d(a[:,2], list(b))]

However, when using CPython, a Numba-based implementation would be even faster (at all scale):

import numpy as np
import numba as nb


@nb.jit
def custom_filter(arr, values):
    values = set(values)
    n, m = arr.shape
    result = np.empty((n, m), dtype=arr.dtype)
    k = 0
    for i in range(n):
        if arr[i, 2] in values:
            result[k, :] = arr[i, :]
            k += 1
    return result[:k, :].copy()


@nb.jit
def custom_filter2(arr, values):
    values = set(values)
    n, m = arr.shape
    k = 0
    for i in range(n):
        if arr[i, 2] in values:
            k += 1
    result = np.empty((k, m), dtype=arr.dtype)
    k = 0
    for i in range(n):
        if arr[i, 2] in values:
            result[k, :] = arr[i, :]
            k += 1
    return result

A quick glimpse into benchmarks:

aa = np.tile(a, (1000, 1))
bb = set(list(range(121000, 122000)))


%timeit np.array([x for x in a if x[2] in b])
# 8.54 µs ± 110 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
%timeit custom_filter(a, tuple(b))
# 1.59 µs ± 179 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
%timeit custom_filter2(a, tuple(b))
# 1.45 µs ± 21.8 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)
%timeit a[np.in1d(a[:,2], tuple(b))]
# 25.2 µs ± 1.43 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

%timeit np.array([x for x in aa if x[2] in b])
# 6.76 ms ± 313 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
%timeit custom_filter(aa, tuple(b))
# 90.6 µs ± 3.91 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
%timeit custom_filter2(aa, tuple(b))
# 135 µs ± 5.84 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
%timeit aa[np.in1d(aa[:, 2], tuple(b))]
# 147 µs ± 9.29 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

%timeit np.array([x for x in aa if x[2] in bb])
# 7.26 ms ± 176 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
%timeit custom_filter(aa, tuple(bb))
# 226 µs ± 5.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
%timeit custom_filter2(aa, tuple(bb))
# 278 µs ± 10.4 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
%timeit aa[np.in1d(aa[:, 2], tuple(bb))]
# 756 µs ± 62.2 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)