242

I have a NumPy array 'boolarr' of boolean type. I want to count the number of elements whose values are True. Is there a NumPy or Python routine dedicated for this task? Or, do I need to iterate over the elements in my script?

cs95
  • 379,657
  • 97
  • 704
  • 746
norio
  • 3,652
  • 3
  • 25
  • 33
  • 4
    For pandas: http://stackoverflow.com/questions/26053849/counting-non-zero-values-in-each-column-of-a-dataframe-in-python – Private Apr 03 '17 at 08:16

6 Answers6

343

You have multiple options. Two options are the following.

boolarr.sum()
numpy.count_nonzero(boolarr)

Here's an example:

>>> import numpy as np
>>> boolarr = np.array([[0, 0, 1], [1, 0, 1], [1, 0, 1]], dtype=np.bool)
>>> boolarr
array([[False, False,  True],
       [ True, False,  True],
       [ True, False,  True]], dtype=bool)

>>> boolarr.sum()
5

Of course, that is a bool-specific answer. More generally, you can use numpy.count_nonzero.

>>> np.count_nonzero(boolarr)
5
Eric O. Lebigot
  • 91,433
  • 48
  • 218
  • 260
David Alber
  • 17,624
  • 6
  • 65
  • 71
  • 3
    Thanks, David. They look neat. About the method with sum(..), is True always equal to 1 in python (or at least in numpy)? If it is not guaranteed, I will add a check, 'if True==1:' beforehand. About count_nonzero(..), unfortunately, it seems not implemented in my numpy module at version 1.5.1, but I may have a chance to use it in the future. – norio Dec 03 '11 at 01:52
  • 5
    @norio Regarding `bool`: boolean values are treated as 1 and 0 in arithmetic operations. See "[Boolean Values](http://docs.python.org/library/stdtypes.html#boolean-values)" in the Python Standard Library documentation. Note that NumPy's `bool` and Python `bool` are not the same, but they are compatible (see [here](http://docs.scipy.org/doc/numpy/reference/arrays.scalars.html#built-in-scalar-types) for more information). – David Alber Dec 03 '11 at 04:39
  • 1
    @norio Regarding `numpy.count_nonzero` not being in NumPy v1.5.1: you are right. According to this [release announcement](http://mail.scipy.org/pipermail/numpy-discussion/2011-May/056295.html), it was added in NumPy v1.6.0. – David Alber Dec 03 '11 at 04:41
  • 1
    Thank you very much for the replies with the links! – norio Dec 03 '11 at 08:34
  • 31
    FWIW, `numpy.count_nonzero` is about a thousand times faster, in my Python interpreter, at least. `python -m timeit -s "import numpy as np; bools = np.random.uniform(size=1000) >= 0.5" "np.count_nonzero(bools)"` vs. `python -m timeit -s "import numpy as np; bools = np.random.uniform(size=1000) >= 0.5" "sum(bools)"` – chbrown Nov 19 '13 at 21:10
  • 10
    @chbrown you are right. But you should compare to `np.sum(bools)` instead! However, `np.count_nonzero(bools)` is still ~12x faster. – mab Nov 23 '15 at 18:15
  • If I try either of those, it works as long as my answer is non-zero. But if I get 0 and I'm doing it in a pivot table, my answer is always False. – Elliptica Aug 03 '16 at 23:46
  • If you intend to check if there are more 1 or more elements in the array after true values have been counted, you can do this with `np.any(bools)` – Zikoat Nov 21 '20 at 13:27
  • @DavidAlber `numpy.count_nonzero` returns wrong results for masked array. If masked array has some mask values and all True values in other cells, it's different from `np.sum()`. Is it bug or expected result? – discover Sep 28 '21 at 07:23
33

That question solved a quite similar question for me and I thought I should share :

In raw python you can use sum() to count True values in a list :

>>> sum([True,True,True,False,False])
3

But this won't work :

>>> sum([[False, False, True], [True, False, True]])
TypeError...
wuerfelfreak
  • 2,363
  • 1
  • 14
  • 29
Guillaume Gendre
  • 2,504
  • 28
  • 17
  • 2
    You should "flatten" the array of arrays first. unfortunately, there's no builtin method, see http://stackoverflow.com/questions/2158395/flatten-an-irregular-list-of-lists-in-python – tommy chheng Dec 07 '12 at 23:32
  • 2
    Thanks Guillaume! Works with Pandas dataframes as well. – JJFord3 Dec 01 '16 at 19:11
  • The raw built-in `sum` is much slower for Pandas `DataFrame`s and numpy arrays than their respective `sum` methods. – Elias Hasle Dec 08 '22 at 09:00
5

In terms of comparing two numpy arrays and counting the number of matches (e.g. correct class prediction in machine learning), I found the below example for two dimensions useful:

import numpy as np
result = np.random.randint(3,size=(5,2)) # 5x2 random integer array
target = np.random.randint(3,size=(5,2)) # 5x2 random integer array

res = np.equal(result,target)
print result
print target
print np.sum(res[:,0])
print np.sum(res[:,1])

which can be extended to D dimensions.

The results are:

Prediction:

[[1 2]
 [2 0]
 [2 0]
 [1 2]
 [1 2]]

Target:

[[0 1]
 [1 0]
 [2 0]
 [0 0]
 [2 1]]

Count of correct prediction for D=1: 1

Count of correct prediction for D=2: 2

salehinejad
  • 7,258
  • 3
  • 18
  • 26
2
b[b].size

where b is the Boolean ndarray in question. It filters b for True, and then count the length of the filtered array.

This probably isn't as efficient np.count_nonzero() mentioned previously, but is useful if you forget the other syntax. Plus, this shorter syntax saves programmer time.

Demo:

In [1]: a = np.array([0,1,3])

In [2]: a
Out[2]: array([0, 1, 3])

In [3]: a[a>=1].size
Out[3]: 2

In [5]: b=a>=1

In [6]: b
Out[6]: array([False,  True,  True])

In [7]: b[b].size
Out[7]: 2
flow2k
  • 3,999
  • 40
  • 55
0
boolarr.sum(axis=1 or axis=0)

axis = 1 will output number of trues in a row and axis = 0 will count number of trues in columns so

boolarr[[true,true,true],[false,false,true]]
print(boolarr.sum(axis=1))

will be (3,1)

Roohullah Kazmi
  • 337
  • 3
  • 14
0

For 1D array, this is what worked for me:

import numpy as np
numbers= np.array([3, 1, 5, 2, 5, 1, 1, 5, 1, 4, 2, 1, 4, 5, 3, 4, 
                  5, 2, 4, 2, 6, 6, 3, 6, 2, 3, 5, 6, 5])

numbersGreaterThan2= np.count_nonzero(numbers> 2)
jose pablo solano
  • 1