numpy: most efficient frequency counts for unique values in an array

Question or problem about Python programming:

In numpy / scipy, is there an efficient way to get frequency counts for unique values in an array?

Something along these lines:

x = array( [1,1,1,2,2,2,5,25,1,1] )
y = freq_count( x )
print y

>> [[1, 5], [2,3], [5,1], [25,1]]

( For you, R users out there, I’m basically looking for the table() function )

How to solve the problem:

Solution 1:

Take a look at np.bincount:

http://docs.scipy.org/doc/numpy/reference/generated/numpy.bincount.html

import numpy as np
x = np.array([1,1,1,2,2,2,5,25,1,1])
y = np.bincount(x)
ii = np.nonzero(y)[0]

And then:

zip(ii,y[ii]) 
# [(1, 5), (2, 3), (5, 1), (25, 1)]

or:

np.vstack((ii,y[ii])).T
# array([[ 1,  5],
         [ 2,  3],
         [ 5,  1],
         [25,  1]])

or however you want to combine the counts and the unique values.

Solution 2:

As of Numpy 1.9, the easiest and fastest method is to simply use numpy.unique, which now has a return_counts keyword argument:

import numpy as np

x = np.array([1,1,1,2,2,2,5,25,1,1])
unique, counts = np.unique(x, return_counts=True)

print np.asarray((unique, counts)).T

Which gives:

 [[ 1  5]
  [ 2  3]
  [ 5  1]
  [25  1]]

A quick comparison with scipy.stats.itemfreq:

In [4]: x = np.random.random_integers(0,100,1e6)

In [5]: %timeit unique, counts = np.unique(x, return_counts=True)
10 loops, best of 3: 31.5 ms per loop

In [6]: %timeit scipy.stats.itemfreq(x)
10 loops, best of 3: 170 ms per loop

Solution 3:

Update: The method mentioned in the original answer is deprecated, we should use the new way instead:

>>> import numpy as np
>>> x = [1,1,1,2,2,2,5,25,1,1]
>>> np.array(np.unique(x, return_counts=True)).T
    array([[ 1,  5],
           [ 2,  3],
           [ 5,  1],
           [25,  1]])

---

Original answer:

you can use scipy.stats.itemfreq

>>> from scipy.stats import itemfreq
>>> x = [1,1,1,2,2,2,5,25,1,1]
>>> itemfreq(x)
/usr/local/bin/python:1: DeprecationWarning: `itemfreq` is deprecated! `itemfreq` is deprecated and will be removed in a future version. Use instead `np.unique(..., return_counts=True)`
array([[  1.,   5.],
       [  2.,   3.],
       [  5.,   1.],
       [ 25.,   1.]])

Solution 4:

I was also interested in this, so I did a little performance comparison (using perfplot, a pet project of mine). Result:

y = np.bincount(a)
ii = np.nonzero(y)[0]
out = np.vstack((ii, y[ii])).T

is by far the fastest. (Note the log-scaling.)

---

Code to generate the plot:

import numpy as np
import pandas as pd
import perfplot
from scipy.stats import itemfreq


def bincount(a):
    y = np.bincount(a)
    ii = np.nonzero(y)[0]
    return np.vstack((ii, y[ii])).T


def unique(a):
    unique, counts = np.unique(a, return_counts=True)
    return np.asarray((unique, counts)).T


def unique_count(a):
    unique, inverse = np.unique(a, return_inverse=True)
    count = np.zeros(len(unique), np.int)
    np.add.at(count, inverse, 1)
    return np.vstack((unique, count)).T


def pandas_value_counts(a):
    out = pd.value_counts(pd.Series(a))
    out.sort_index(inplace=True)
    out = np.stack([out.keys().values, out.values]).T
    return out


perfplot.show(
    setup=lambda n: np.random.randint(0, 1000, n),
    kernels=[bincount, unique, itemfreq, unique_count, pandas_value_counts],
    n_range=[2 ** k for k in range(26)],
    logx=True,
    logy=True,
    xlabel="len(a)",
)

Solution 5:

Using pandas module:

>>> import pandas as pd
>>> import numpy as np
>>> x = np.array([1,1,1,2,2,2,5,25,1,1])
>>> pd.value_counts(x)
1     5
2     3
25    1
5     1
dtype: int64

Hope this helps!