This should be faster than your current approach.
Rather than searching through mass looking for pairs of matching numbers we pair each number in mass with its index and sort those pairs. We can then use groupby to find groups of equal numbers. If there are more than 2 of the same number we use the first and last, since they will have the greatest sum between them.
from operator import itemgetter
from itertools import groupby
raw = '3 5 6 3 5 4'
mass = [int(u) for u in raw.split()]
result = []
a = sorted((u, i) for i, u in enumerate(mass))
for _, g in groupby(a, itemgetter(0)):
g = list(g)
if len(g) > 1:
u, v = g[0][1], g[-1][1]
result.append((sum(mass[u:v+1]), u+1, v+1))
print(max(result))
output
(19, 2, 5)
Note that this code will not necessarily give the maximum sum between equal elements in the list if the list contains negative numbers. It will still work correctly with negative numbers if no group of equal numbers has more than two members. If that's not the case, we need to use a slower algorithm that tests every pair within a group of equal numbers.
Here's a more efficient version. Instead of using the sum function we build a list of the cumulative sums of the whole list. This doesn't make much of a difference for small lists, but it's much faster when the list size is large. Eg, for a list of 10,000 elements this approach is about 10 times faster. To test it, I create an array of random positive integers.
from operator import itemgetter
from itertools import groupby
from random import seed, randrange
seed(42)
def maxsum(seq):
total = 0
sums = [0]
for u in seq:
total += u
sums.append(total)
result = []
a = sorted((u, i) for i, u in enumerate(seq))
for _, g in groupby(a, itemgetter(0)):
g = list(g)
if len(g) > 1:
u, v = g[0][1], g[-1][1]
result.append((sums[v+1] - sums[u], u+1, v+1))
return max(result)
num = 25000
hi = num // 2
mass = [randrange(1, hi) for _ in range(num)]
print(maxsum(mass))
output
(155821402, 21, 24831)
If you're using a recent version of Python you can use itertools.accumulate to build the list of cumulative sums. This is around 10% faster.
from itertools import accumulate
def maxsum(seq):
sums = [0] + list(accumulate(seq))
result = []
a = sorted((u, i) for i, u in enumerate(seq))
for _, g in groupby(a, itemgetter(0)):
g = list(g)
if len(g) > 1:
u, v = g[0][1], g[-1][1]
result.append((sums[v+1] - sums[u], u+1, v+1))
return max(result)
Here's a faster version, derived from code by Stefan Pochmann, which uses a dict, instead of sorting & groupby. Thanks, Stefan!
def maxsum(seq):
total = 0
sums = [0]
for u in seq:
total += u
sums.append(total)
where = {}
for i, x in enumerate(seq, 1):
where.setdefault(x, [i, i])[1] = i
return max((sums[j] - sums[i - 1], i, j)
for i, j in where.values())
If the list contains no duplicate items (and hence no subsequences bound by duplicate items) it returns the maximum item in the list.
Here are two more variations. These can handle negative items correctly, and if there are no duplicate items they return None. In Python 3 that could be handled elegantly by passing default=None to max, but that option isn't available in Python 2, so instead I catch the ValueError exception that's raised when you attempt to find the max of an empty iterable.
The first version, maxsum_combo, uses itertools.combinations to generate all combinations of a group of equal numbers and thence finds the combination that gives the maximum sum. The second version, maxsum_kadane uses a variation of Kadane's algorithm to find the maximum subsequence within a group.
If there aren't many duplicates in the original sequence, so the average group size is small, maxsum_combo is generally faster. But if the groups are large, then maxsum_kadane is much faster than maxsum_combo. The code below tests these functions on random sequences of 15000 items, firstly on sequences with few duplicates (and hence small mean group size) and then on sequences with lots of duplicates. It verifies that both versions give the same results, and then it performs timeit tests.
from __future__ import print_function
from itertools import groupby, combinations
from random import seed, randrange
from timeit import Timer
seed(42)
def maxsum_combo(seq):
total = 0
sums = [0]
for u in seq:
total += u
sums.append(total)
where = {}
for i, x in enumerate(seq, 1):
where.setdefault(x, []).append(i)
try:
return max((sums[j] - sums[i - 1], i, j)
for v in where.values() for i, j in combinations(v, 2))
except ValueError:
return None
def maxsum_kadane(seq):
total = 0
sums = [0]
for u in seq:
total += u
sums.append(total)
where = {}
for i, x in enumerate(seq, 1):
where.setdefault(x, []).append(i)
try:
return max(max_sublist([(sums[j] - sums[i-1], i, j)
for i, j in zip(v, v[1:])], k)
for k, v in where.items() if len(v) > 1)
except ValueError:
return None
# Kadane's Algorithm to find maximum sublist
# From https://en.wikipedia.org/wiki/Maximum_subarray_problem
def max_sublist(seq, k):
max_ending_here = max_so_far = seq[0]
for x in seq[1:]:
y = max_ending_here[0] + x[0] - k, max_ending_here[1], x[2]
max_ending_here = max(x, y)
max_so_far = max(max_so_far, max_ending_here)
return max_so_far
def test(num, hi, loops):
print('\nnum = {0}, hi = {1}, loops = {2}'.format(num, hi, loops))
print('Verifying...')
for k in range(5):
mass = [randrange(-hi // 2, hi) for _ in range(num)]
a = maxsum_combo(mass)
b = maxsum_kadane(mass)
print(a, b, a==b)
print('\nTiming...')
for func in maxsum_combo, maxsum_kadane:
t = Timer(lambda: func(mass))
result = sorted(t.repeat(3, loops))
result = ', '.join([format(u, '.5f') for u in result])
print('{0:14} : {1}'.format(func.__name__, result))
loops = 20
num = 15000
hi = num // 4
test(num, hi, loops)
loops = 10
hi = num // 100
test(num, hi, loops)
output
num = 15000, hi = 3750, loops = 20
Verifying...
(13983131, 44, 14940) (13983131, 44, 14940) True
(13928837, 27, 14985) (13928837, 27, 14985) True
(14057416, 40, 14995) (14057416, 40, 14995) True
(13997395, 65, 14996) (13997395, 65, 14996) True
(14050007, 12, 14972) (14050007, 12, 14972) True
Timing...
maxsum_combo : 1.72903, 1.73780, 1.81138
maxsum_kadane : 2.17738, 2.22108, 2.22394
num = 15000, hi = 150, loops = 10
Verifying...
(553789, 21, 14996) (553789, 21, 14996) True
(550174, 1, 14992) (550174, 1, 14992) True
(551017, 13, 14991) (551017, 13, 14991) True
(554317, 2, 14986) (554317, 2, 14986) True
(558663, 15, 14988) (558663, 15, 14988) True
Timing...
maxsum_combo : 7.29226, 7.34213, 7.36688
maxsum_kadane : 1.07532, 1.07695, 1.10525
This code runs on both Python 2 and Python 3. The above results were generated on an old 32 bit 2GHz machine running Python 2.6.6 on a Debian derivative of Linux. The speeds for Python 3.6.0 are similar.
If you want to include groups that consist of a single non-repeated number, and also want to include the numbers that are in groups as a "subsequence" of length 1, you can use this version:
def maxsum_kadane(seq):
if not seq:
return None
total = 0
sums = [0]
for u in seq:
total += u
sums.append(total)
where = {}
for i, x in enumerate(seq, 1):
where.setdefault(x, []).append(i)
# Find the maximum of the single items
m_single = max((k, v[0], v[0]) for k, v in where.items())
# Find the maximum of the subsequences
try:
m_subseq = max(max_sublist([(sums[j] - sums[i-1], i, j)
for i, j in zip(v, v[1:])], k)
for k, v in where.items() if len(v) > 1)
return max(m_single, m_subseq)
except ValueError:
# No subsequences
return m_single
I haven't tested it extensively, but it should work. ;)
