Most efficient way to find mode in numpy array

Check scipy.stats.mode() (inspired by @tom10's comment):

import numpy as np
from scipy import stats

a = np.array([[1, 3, 4, 2, 2, 7],
              [5, 2, 2, 1, 4, 1],
              [3, 3, 2, 2, 1, 1]])

m = stats.mode(a)
print(m)

Output:

ModeResult(mode=array([[1, 3, 2, 2, 1, 1]]), count=array([[1, 2, 2, 2, 1, 2]]))

As you can see, it returns both the mode as well as the counts. You can select the modes directly via m[0]:

print(m[0])

Output:

[[1 3 2 2 1 1]]

Update

The scipy.stats.mode function has been significantly optimized since this post, and would be the recommended method

Old answer

This is a tricky problem, since there is not much out there to calculate mode along an axis. The solution is straight forward for 1-D arrays, where numpy.bincount is handy, along with numpy.unique with the return_counts arg as True. The most common n-dimensional function I see is scipy.stats.mode, although it is prohibitively slow- especially for large arrays with many unique values. As a solution, I've developed this function, and use it heavily:

import numpy

def mode(ndarray, axis=0):
    # Check inputs
    ndarray = numpy.asarray(ndarray)
    ndim = ndarray.ndim
    if ndarray.size == 1:
        return (ndarray[0], 1)
    elif ndarray.size == 0:
        raise Exception('Cannot compute mode on empty array')
    try:
        axis = range(ndarray.ndim)[axis]
    except:
        raise Exception('Axis "{}" incompatible with the {}-dimension array'.format(axis, ndim))

    # If array is 1-D and numpy version is > 1.9 numpy.unique will suffice
    if all([ndim == 1,
            int(numpy.__version__.split('.')[0]) >= 1,
            int(numpy.__version__.split('.')[1]) >= 9]):
        modals, counts = numpy.unique(ndarray, return_counts=True)
        index = numpy.argmax(counts)
        return modals[index], counts[index]

    # Sort array
    sort = numpy.sort(ndarray, axis=axis)
    # Create array to transpose along the axis and get padding shape
    transpose = numpy.roll(numpy.arange(ndim)[::-1], axis)
    shape = list(sort.shape)
    shape[axis] = 1
    # Create a boolean array along strides of unique values
    strides = numpy.concatenate([numpy.zeros(shape=shape, dtype='bool'),
                                 numpy.diff(sort, axis=axis) == 0,
                                 numpy.zeros(shape=shape, dtype='bool')],
                                axis=axis).transpose(transpose).ravel()
    # Count the stride lengths
    counts = numpy.cumsum(strides)
    counts[~strides] = numpy.concatenate([[0], numpy.diff(counts[~strides])])
    counts[strides] = 0
    # Get shape of padded counts and slice to return to the original shape
    shape = numpy.array(sort.shape)
    shape[axis] += 1
    shape = shape[transpose]
    slices = [slice(None)] * ndim
    slices[axis] = slice(1, None)
    # Reshape and compute final counts
    counts = counts.reshape(shape).transpose(transpose)[slices] + 1

    # Find maximum counts and return modals/counts
    slices = [slice(None, i) for i in sort.shape]
    del slices[axis]
    index = numpy.ogrid[slices]
    index.insert(axis, numpy.argmax(counts, axis=axis))
    return sort[index], counts[index]

Result:

In [2]: a = numpy.array([[1, 3, 4, 2, 2, 7],
                         [5, 2, 2, 1, 4, 1],
                         [3, 3, 2, 2, 1, 1]])

In [3]: mode(a)
Out[3]: (array([1, 3, 2, 2, 1, 1]), array([1, 2, 2, 2, 1, 2]))

Some benchmarks:

In [4]: import scipy.stats

In [5]: a = numpy.random.randint(1,10,(1000,1000))

In [6]: %timeit scipy.stats.mode(a)
10 loops, best of 3: 41.6 ms per loop

In [7]: %timeit mode(a)
10 loops, best of 3: 46.7 ms per loop

In [8]: a = numpy.random.randint(1,500,(1000,1000))

In [9]: %timeit scipy.stats.mode(a)
1 loops, best of 3: 1.01 s per loop

In [10]: %timeit mode(a)
10 loops, best of 3: 80 ms per loop

In [11]: a = numpy.random.random((200,200))

In [12]: %timeit scipy.stats.mode(a)
1 loops, best of 3: 3.26 s per loop

In [13]: %timeit mode(a)
1000 loops, best of 3: 1.75 ms per loop

EDIT: Provided more of a background and modified the approach to be more memory-efficient