NumPy grouping using itertools.groupby performance

Sorting is theta(NlogN), I’d go for amortized O(N) provided by Python’s hashtable implementation. Just use defaultdict(int) for keeping counts of the integers and just iterate over the array once:

counts = collections.defaultdict(int)
for v in values:
    counts[v] += 1

This is theoretically faster, unfortunately I have no way to check now. Allocating the additional memory might make it actually slower than your solution, which is in-place.

Edit: If you need to save memory try radix sort, which is much faster on integers than quicksort (which I believe is what numpy uses).

This is a numpy solution:

def group():
    import numpy as np
    values = np.array(np.random.randint(0,1<<32,size=35000000),dtype='u4')

    # we sort in place
    values.sort()

    # when sorted the number of occurences for a unique element is the index of 
    # the first occurence when searching from the right - the index of the first
    # occurence when searching from the left.
    #
    # np.dstack() is the numpy equivalent to Python's zip()

    l = np.dstack((values, values.searchsorted(values, side='right') - 
                   values.searchsorted(values, side='left')))

    index = np.fromiter(l, dtype='u4,u2')

if __name__=='__main__':
    from timeit import Timer
    t = Timer("group()","from __main__ import group")
    print t.timeit(number=1)

Runs in about 25 seconds on my machine compared to about 96 for your initial solution (which is a nice improvement).

There might be still room for improvement, I don’t use numpy that often.

Edit: added some comments in code.

Replacing len(list(g)) with sum(1 for i in g) gives a 2x speedup

By request, here is a Cython version of this. I did two passes through the array. The first one finds out how many unique elements there are so that can my arrays for the unique values and counts of the appropriate size.

import numpy as np
cimport numpy as np
cimport cython

@cython.boundscheck(False)
def dogroup():
    cdef unsigned long tot = 1
    cdef np.ndarray[np.uint32_t, ndim=1] values = np.array(np.random.randint(35000000,size=35000000),dtype=np.uint32)
    cdef unsigned long i, ind, lastval
    values.sort()
    for i in xrange(1,len(values)):
        if values[i] != values[i-1]:
            tot += 1
    cdef np.ndarray[np.uint32_t, ndim=1] vals = np.empty(tot,dtype=np.uint32)
    cdef np.ndarray[np.uint32_t, ndim=1] count = np.empty(tot,dtype=np.uint32)
    vals[0] = values[0]
    ind = 1
    lastval = 0
    for i in xrange(1,len(values)):
        if values[i] != values[i-1]:
            vals[ind] = values[i]
            count[ind-1] = i - lastval
            lastval = i
            ind += 1
    count[ind-1] = len(values) - lastval

The sorting is actually taking the most time here by far. Using the values array given in my code, the sorting is taking 4.75 seconds and the actual finding of the unique values and counts takes .67 seconds. With the pure Numpy code using Paul’s code (but with the same form of the values array) with the fix I suggested in a comment, finding the unique values and counts takes 1.9 seconds (sorting still takes the same amount of time of course).

It makes sense for most of the time to be taken up by the sorting because it is O(N log N) and the counting is O(N). You can speed up the sort a little bit over Numpy’s (which uses C’s qsort if I remember correctly), but you have to really know what you are doing and it probably isn’t worthwhile. Also, there might be some way to speed up my Cython code a little bit more, but it probably isn’t worthwhile.

This is a fairly old thread, but I thought I’d mention that there’s a small improvement to be made on the currently-accepted solution:

def group_by_edge():
    import numpy as np
    values = np.array(np.random.randint(0,1<<32,size=35000000),dtype='u4')
    values.sort()
    edges = (values[1:] != values[:-1]).nonzero()[0] - 1
    idx = np.concatenate(([0], edges, [len(values)]))
    index = np.empty(len(idx) - 1, dtype= 'u4, u2')
    index['f0'] = values[idx[:-1]]
    index['f1'] = np.diff(idx)

This tested as about a half-second faster on my machine; not a huge improvement, but worth something. Additionally, I think it’s clearer what’s happening here; the two step diff approach is a bit opaque at first glance.

I guess the most obvious and still not mentioned approach is, to simply use collections.Counter. Instead of building a huge amount of temporarily used lists with groupby, it just upcounts integers. It’s a oneliner and a 2-fold speedup, but still slower than the pure numpy solutions.

def group():
    import sys
    import numpy as np
    from collections import Counter
    values = np.array(np.random.randint(0,sys.maxint,size=35000000),dtype='u4')
    c = Counter(values)

if __name__=='__main__':
    from timeit import Timer
    t = Timer("group()","from __main__ import group")
    print t.timeit(number=1)

I get a speedup from 136 s to 62 s for my machine, compared to the initial solution.

You could try the following (ab)use of scipy.sparse:

from scipy import sparse
def sparse_bincount(values):
    M = sparse.csr_matrix((np.ones(len(values)), values.astype(int), [0, len(values)]))
    M.sum_duplicates()
    index = np.empty(len(M.indices),dtype='u4,u2')
    index['f0'] = M.indices
    index['f1']= M.data
    return index

This is slower than the winning answer, perhaps because scipy currently doesn’t support unsigned as indices types…

More than 5 years have passed since Paul’s answer was accepted. Interestingly,
the sort() is still the bottleneck in the accepted solution.

Line #      Hits         Time  Per Hit   % Time  Line Contents
==============================================================
     3                                           @profile
     4                                           def group_paul():
     5         1        99040  99040.0      2.4      import numpy as np
     6         1       305651 305651.0      7.4      values = np.array(np.random.randint(0, 2**32,size=35000000),dtype='u4')
     7         1      2928204 2928204.0    71.3      values.sort()
     8         1        78268  78268.0      1.9      diff = np.concatenate(([1],np.diff(values)))
     9         1       215774 215774.0      5.3      idx = np.concatenate((np.where(diff)[0],[len(values)]))
    10         1           95     95.0      0.0      index = np.empty(len(idx)-1,dtype='u4,u2')
    11         1       386673 386673.0      9.4      index['f0'] = values[idx[:-1]]
    12         1        91492  91492.0      2.2      index['f1'] = np.diff(idx)

The accepted solution runs for 4.0 s on my machine, with radix sort it
drops down to 1.7 s.

Just by switching to radix sort, I get an overall 2.35x speedup. The radix sort is more than 4x faster than quicksort in this case.

See How to sort an array of integers faster than quicksort? that was motivated by your question.

For the profiling I used line_profiler and kernprof (the @profile comes from there).

In the latest version of numpy, we have this.

import numpy as np
frequency = np.unique(values, return_counts=True)