blob: 0e3555cf9118eaabb5033bbccbcf2f3607716221 [file] [log] [blame]
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +00001"""Heap queue algorithm (a.k.a. priority queue).
2
3Heaps are arrays for which a[k] <= a[2*k+1] and a[k] <= a[2*k+2] for
4all k, counting elements from 0. For the sake of comparison,
5non-existing elements are considered to be infinite. The interesting
6property of a heap is that a[0] is always its smallest element.
7
8Usage:
9
10heap = [] # creates an empty heap
11heappush(heap, item) # pushes a new item on the heap
12item = heappop(heap) # pops the smallest item from the heap
13item = heap[0] # smallest item on the heap without popping it
14heapify(x) # transforms list into a heap, in-place, in linear time
15item = heapreplace(heap, item) # pops and returns smallest item, and adds
16 # new item; the heap size is unchanged
17
18Our API differs from textbook heap algorithms as follows:
19
20- We use 0-based indexing. This makes the relationship between the
21 index for a node and the indexes for its children slightly less
22 obvious, but is more suitable since Python uses 0-based indexing.
23
24- Our heappop() method returns the smallest item, not the largest.
25
26These two make it possible to view the heap as a regular Python list
27without surprises: heap[0] is the smallest item, and heap.sort()
28maintains the heap invariant!
29"""
30
Raymond Hettinger33ecffb2004-06-10 05:03:17 +000031# Original code by Kevin O'Connor, augmented by Tim Peters and Raymond Hettinger
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +000032
33__about__ = """Heap queues
34
Mark Dickinsonb4a17a82010-07-04 19:23:49 +000035[explanation by François Pinard]
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +000036
37Heaps are arrays for which a[k] <= a[2*k+1] and a[k] <= a[2*k+2] for
38all k, counting elements from 0. For the sake of comparison,
39non-existing elements are considered to be infinite. The interesting
40property of a heap is that a[0] is always its smallest element.
41
42The strange invariant above is meant to be an efficient memory
43representation for a tournament. The numbers below are `k', not a[k]:
44
45 0
46
47 1 2
48
49 3 4 5 6
50
51 7 8 9 10 11 12 13 14
52
53 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
54
55
56In the tree above, each cell `k' is topping `2*k+1' and `2*k+2'. In
Martin Panter6245cb32016-04-15 02:14:19 +000057a usual binary tournament we see in sports, each cell is the winner
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +000058over the two cells it tops, and we can trace the winner down the tree
59to see all opponents s/he had. However, in many computer applications
60of such tournaments, we do not need to trace the history of a winner.
61To be more memory efficient, when a winner is promoted, we try to
62replace it by something else at a lower level, and the rule becomes
63that a cell and the two cells it tops contain three different items,
64but the top cell "wins" over the two topped cells.
65
66If this heap invariant is protected at all time, index 0 is clearly
67the overall winner. The simplest algorithmic way to remove it and
68find the "next" winner is to move some loser (let's say cell 30 in the
69diagram above) into the 0 position, and then percolate this new 0 down
70the tree, exchanging values, until the invariant is re-established.
71This is clearly logarithmic on the total number of items in the tree.
72By iterating over all items, you get an O(n ln n) sort.
73
74A nice feature of this sort is that you can efficiently insert new
75items while the sort is going on, provided that the inserted items are
76not "better" than the last 0'th element you extracted. This is
77especially useful in simulation contexts, where the tree holds all
78incoming events, and the "win" condition means the smallest scheduled
79time. When an event schedule other events for execution, they are
80scheduled into the future, so they can easily go into the heap. So, a
81heap is a good structure for implementing schedulers (this is what I
82used for my MIDI sequencer :-).
83
84Various structures for implementing schedulers have been extensively
85studied, and heaps are good for this, as they are reasonably speedy,
86the speed is almost constant, and the worst case is not much different
87than the average case. However, there are other representations which
88are more efficient overall, yet the worst cases might be terrible.
89
90Heaps are also very useful in big disk sorts. You most probably all
91know that a big sort implies producing "runs" (which are pre-sorted
92sequences, which size is usually related to the amount of CPU memory),
93followed by a merging passes for these runs, which merging is often
94very cleverly organised[1]. It is very important that the initial
95sort produces the longest runs possible. Tournaments are a good way
96to that. If, using all the memory available to hold a tournament, you
97replace and percolate items that happen to fit the current run, you'll
98produce runs which are twice the size of the memory for random input,
99and much better for input fuzzily ordered.
100
101Moreover, if you output the 0'th item on disk and get an input which
102may not fit in the current tournament (because the value "wins" over
103the last output value), it cannot fit in the heap, so the size of the
104heap decreases. The freed memory could be cleverly reused immediately
105for progressively building a second heap, which grows at exactly the
106same rate the first heap is melting. When the first heap completely
107vanishes, you switch heaps and start a new run. Clever and quite
108effective!
109
110In a word, heaps are useful memory structures to know. I use them in
111a few applications, and I think it is good to keep a `heap' module
112around. :-)
113
114--------------------
115[1] The disk balancing algorithms which are current, nowadays, are
116more annoying than clever, and this is a consequence of the seeking
117capabilities of the disks. On devices which cannot seek, like big
118tape drives, the story was quite different, and one had to be very
119clever to ensure (far in advance) that each tape movement will be the
120most effective possible (that is, will best participate at
121"progressing" the merge). Some tapes were even able to read
122backwards, and this was also used to avoid the rewinding time.
123Believe me, real good tape sorts were quite spectacular to watch!
124From all times, sorting has always been a Great Art! :-)
125"""
126
Thomas Wouterscf297e42007-02-23 15:07:44 +0000127__all__ = ['heappush', 'heappop', 'heapify', 'heapreplace', 'merge',
Christian Heimesdd15f6c2008-03-16 00:07:10 +0000128 'nlargest', 'nsmallest', 'heappushpop']
Raymond Hettinger33ecffb2004-06-10 05:03:17 +0000129
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000130def heappush(heap, item):
131 """Push item onto heap, maintaining the heap invariant."""
132 heap.append(item)
133 _siftdown(heap, 0, len(heap)-1)
134
135def heappop(heap):
136 """Pop the smallest item off the heap, maintaining the heap invariant."""
137 lastelt = heap.pop() # raises appropriate IndexError if heap is empty
138 if heap:
139 returnitem = heap[0]
140 heap[0] = lastelt
141 _siftup(heap, 0)
Raymond Hettinger356902d2014-05-19 22:13:45 +0100142 return returnitem
143 return lastelt
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000144
145def heapreplace(heap, item):
146 """Pop and return the current smallest value, and add the new item.
147
148 This is more efficient than heappop() followed by heappush(), and can be
149 more appropriate when using a fixed-size heap. Note that the value
150 returned may be larger than item! That constrains reasonable uses of
Raymond Hettinger8158e842004-09-06 07:04:09 +0000151 this routine unless written as part of a conditional replacement:
Raymond Hettinger28224f82004-06-20 09:07:53 +0000152
Raymond Hettinger8158e842004-09-06 07:04:09 +0000153 if item > heap[0]:
154 item = heapreplace(heap, item)
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000155 """
156 returnitem = heap[0] # raises appropriate IndexError if heap is empty
157 heap[0] = item
158 _siftup(heap, 0)
159 return returnitem
160
Christian Heimesdd15f6c2008-03-16 00:07:10 +0000161def heappushpop(heap, item):
162 """Fast version of a heappush followed by a heappop."""
Georg Brandlf78e02b2008-06-10 17:40:04 +0000163 if heap and heap[0] < item:
Christian Heimesdd15f6c2008-03-16 00:07:10 +0000164 item, heap[0] = heap[0], item
165 _siftup(heap, 0)
166 return item
167
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000168def heapify(x):
Éric Araujo395ba352011-04-15 23:34:31 +0200169 """Transform list into a heap, in-place, in O(len(x)) time."""
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000170 n = len(x)
171 # Transform bottom-up. The largest index there's any point to looking at
172 # is the largest with a child index in-range, so must have 2*i + 1 < n,
173 # or i < (n-1)/2. If n is even = 2*j, this is (2*j-1)/2 = j-1/2 so
174 # j-1 is the largest, which is n//2 - 1. If n is odd = 2*j+1, this is
175 # (2*j+1-1)/2 = j so j-1 is the largest, and that's again n//2-1.
Guido van Rossum805365e2007-05-07 22:24:25 +0000176 for i in reversed(range(n//2)):
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000177 _siftup(x, i)
178
Raymond Hettinger35db4392014-05-30 02:28:36 -0700179def _heappop_max(heap):
180 """Maxheap version of a heappop."""
181 lastelt = heap.pop() # raises appropriate IndexError if heap is empty
182 if heap:
183 returnitem = heap[0]
184 heap[0] = lastelt
185 _siftup_max(heap, 0)
186 return returnitem
187 return lastelt
188
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700189def _heapreplace_max(heap, item):
190 """Maxheap version of a heappop followed by a heappush."""
191 returnitem = heap[0] # raises appropriate IndexError if heap is empty
192 heap[0] = item
193 _siftup_max(heap, 0)
194 return returnitem
Raymond Hettingerf6b26672013-03-05 01:36:30 -0500195
196def _heapify_max(x):
197 """Transform list into a maxheap, in-place, in O(len(x)) time."""
198 n = len(x)
199 for i in reversed(range(n//2)):
200 _siftup_max(x, i)
201
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000202# 'heap' is a heap at all indices >= startpos, except possibly for pos. pos
203# is the index of a leaf with a possibly out-of-order value. Restore the
204# heap invariant.
205def _siftdown(heap, startpos, pos):
206 newitem = heap[pos]
207 # Follow the path to the root, moving parents down until finding a place
208 # newitem fits.
209 while pos > startpos:
210 parentpos = (pos - 1) >> 1
211 parent = heap[parentpos]
Georg Brandlf78e02b2008-06-10 17:40:04 +0000212 if newitem < parent:
213 heap[pos] = parent
214 pos = parentpos
215 continue
216 break
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000217 heap[pos] = newitem
218
219# The child indices of heap index pos are already heaps, and we want to make
220# a heap at index pos too. We do this by bubbling the smaller child of
221# pos up (and so on with that child's children, etc) until hitting a leaf,
222# then using _siftdown to move the oddball originally at index pos into place.
223#
224# We *could* break out of the loop as soon as we find a pos where newitem <=
225# both its children, but turns out that's not a good idea, and despite that
226# many books write the algorithm that way. During a heap pop, the last array
227# element is sifted in, and that tends to be large, so that comparing it
228# against values starting from the root usually doesn't pay (= usually doesn't
229# get us out of the loop early). See Knuth, Volume 3, where this is
230# explained and quantified in an exercise.
231#
232# Cutting the # of comparisons is important, since these routines have no
233# way to extract "the priority" from an array element, so that intelligence
Mark Dickinsona56c4672009-01-27 18:17:45 +0000234# is likely to be hiding in custom comparison methods, or in array elements
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000235# storing (priority, record) tuples. Comparisons are thus potentially
236# expensive.
237#
238# On random arrays of length 1000, making this change cut the number of
239# comparisons made by heapify() a little, and those made by exhaustive
240# heappop() a lot, in accord with theory. Here are typical results from 3
241# runs (3 just to demonstrate how small the variance is):
242#
243# Compares needed by heapify Compares needed by 1000 heappops
244# -------------------------- --------------------------------
245# 1837 cut to 1663 14996 cut to 8680
246# 1855 cut to 1659 14966 cut to 8678
247# 1847 cut to 1660 15024 cut to 8703
248#
249# Building the heap by using heappush() 1000 times instead required
250# 2198, 2148, and 2219 compares: heapify() is more efficient, when
251# you can use it.
252#
253# The total compares needed by list.sort() on the same lists were 8627,
254# 8627, and 8632 (this should be compared to the sum of heapify() and
255# heappop() compares): list.sort() is (unsurprisingly!) more efficient
256# for sorting.
257
258def _siftup(heap, pos):
259 endpos = len(heap)
260 startpos = pos
261 newitem = heap[pos]
262 # Bubble up the smaller child until hitting a leaf.
263 childpos = 2*pos + 1 # leftmost child position
264 while childpos < endpos:
265 # Set childpos to index of smaller child.
266 rightpos = childpos + 1
Georg Brandlf78e02b2008-06-10 17:40:04 +0000267 if rightpos < endpos and not heap[childpos] < heap[rightpos]:
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000268 childpos = rightpos
269 # Move the smaller child up.
270 heap[pos] = heap[childpos]
271 pos = childpos
272 childpos = 2*pos + 1
273 # The leaf at pos is empty now. Put newitem there, and bubble it up
274 # to its final resting place (by sifting its parents down).
275 heap[pos] = newitem
276 _siftdown(heap, startpos, pos)
277
Raymond Hettingerf6b26672013-03-05 01:36:30 -0500278def _siftdown_max(heap, startpos, pos):
279 'Maxheap variant of _siftdown'
280 newitem = heap[pos]
281 # Follow the path to the root, moving parents down until finding a place
282 # newitem fits.
283 while pos > startpos:
284 parentpos = (pos - 1) >> 1
285 parent = heap[parentpos]
286 if parent < newitem:
287 heap[pos] = parent
288 pos = parentpos
289 continue
290 break
291 heap[pos] = newitem
292
293def _siftup_max(heap, pos):
Raymond Hettinger2e8d9a72013-03-05 02:11:10 -0500294 'Maxheap variant of _siftup'
Raymond Hettingerf6b26672013-03-05 01:36:30 -0500295 endpos = len(heap)
296 startpos = pos
297 newitem = heap[pos]
298 # Bubble up the larger child until hitting a leaf.
299 childpos = 2*pos + 1 # leftmost child position
300 while childpos < endpos:
301 # Set childpos to index of larger child.
302 rightpos = childpos + 1
303 if rightpos < endpos and not heap[rightpos] < heap[childpos]:
304 childpos = rightpos
305 # Move the larger child up.
306 heap[pos] = heap[childpos]
307 pos = childpos
308 childpos = 2*pos + 1
309 # The leaf at pos is empty now. Put newitem there, and bubble it up
310 # to its final resting place (by sifting its parents down).
311 heap[pos] = newitem
312 _siftdown_max(heap, startpos, pos)
313
Raymond Hettinger35db4392014-05-30 02:28:36 -0700314def merge(*iterables, key=None, reverse=False):
Thomas Wouterscf297e42007-02-23 15:07:44 +0000315 '''Merge multiple sorted inputs into a single sorted output.
316
Guido van Rossumd8faa362007-04-27 19:54:29 +0000317 Similar to sorted(itertools.chain(*iterables)) but returns a generator,
Thomas Wouterscf297e42007-02-23 15:07:44 +0000318 does not pull the data into memory all at once, and assumes that each of
319 the input streams is already sorted (smallest to largest).
320
321 >>> list(merge([1,3,5,7], [0,2,4,8], [5,10,15,20], [], [25]))
322 [0, 1, 2, 3, 4, 5, 5, 7, 8, 10, 15, 20, 25]
323
Raymond Hettinger35db4392014-05-30 02:28:36 -0700324 If *key* is not None, applies a key function to each element to determine
325 its sort order.
326
327 >>> list(merge(['dog', 'horse'], ['cat', 'fish', 'kangaroo'], key=len))
328 ['dog', 'cat', 'fish', 'horse', 'kangaroo']
329
Thomas Wouterscf297e42007-02-23 15:07:44 +0000330 '''
Thomas Wouterscf297e42007-02-23 15:07:44 +0000331
332 h = []
333 h_append = h.append
Raymond Hettinger35db4392014-05-30 02:28:36 -0700334
335 if reverse:
336 _heapify = _heapify_max
337 _heappop = _heappop_max
338 _heapreplace = _heapreplace_max
339 direction = -1
340 else:
341 _heapify = heapify
342 _heappop = heappop
343 _heapreplace = heapreplace
344 direction = 1
345
346 if key is None:
347 for order, it in enumerate(map(iter, iterables)):
348 try:
349 next = it.__next__
350 h_append([next(), order * direction, next])
351 except StopIteration:
352 pass
353 _heapify(h)
354 while len(h) > 1:
355 try:
356 while True:
357 value, order, next = s = h[0]
358 yield value
359 s[0] = next() # raises StopIteration when exhausted
360 _heapreplace(h, s) # restore heap condition
361 except StopIteration:
362 _heappop(h) # remove empty iterator
363 if h:
364 # fast case when only a single iterator remains
365 value, order, next = h[0]
366 yield value
367 yield from next.__self__
368 return
369
Raymond Hettingerab09c132014-05-26 17:08:27 -0700370 for order, it in enumerate(map(iter, iterables)):
Thomas Wouterscf297e42007-02-23 15:07:44 +0000371 try:
Georg Brandla18af4e2007-04-21 15:47:16 +0000372 next = it.__next__
Raymond Hettinger35db4392014-05-30 02:28:36 -0700373 value = next()
374 h_append([key(value), order * direction, value, next])
Raymond Hettingerab09c132014-05-26 17:08:27 -0700375 except StopIteration:
Thomas Wouterscf297e42007-02-23 15:07:44 +0000376 pass
Raymond Hettinger35db4392014-05-30 02:28:36 -0700377 _heapify(h)
Raymond Hettingerab09c132014-05-26 17:08:27 -0700378 while len(h) > 1:
Thomas Wouterscf297e42007-02-23 15:07:44 +0000379 try:
Raymond Hettingerf2762322013-09-11 01:15:40 -0500380 while True:
Raymond Hettinger35db4392014-05-30 02:28:36 -0700381 key_value, order, value, next = s = h[0]
Raymond Hettingerab09c132014-05-26 17:08:27 -0700382 yield value
Raymond Hettinger35db4392014-05-30 02:28:36 -0700383 value = next()
384 s[0] = key(value)
385 s[2] = value
386 _heapreplace(h, s)
Raymond Hettingerab09c132014-05-26 17:08:27 -0700387 except StopIteration:
Raymond Hettinger35db4392014-05-30 02:28:36 -0700388 _heappop(h)
Raymond Hettingerf2762322013-09-11 01:15:40 -0500389 if h:
Raymond Hettinger450ed102014-06-01 23:40:01 -0700390 key_value, order, value, next = h[0]
Raymond Hettingerab09c132014-05-26 17:08:27 -0700391 yield value
Raymond Hettingerf2762322013-09-11 01:15:40 -0500392 yield from next.__self__
Thomas Wouterscf297e42007-02-23 15:07:44 +0000393
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700394
395# Algorithm notes for nlargest() and nsmallest()
396# ==============================================
397#
Raymond Hettinger356902d2014-05-19 22:13:45 +0100398# Make a single pass over the data while keeping the k most extreme values
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700399# in a heap. Memory consumption is limited to keeping k values in a list.
400#
401# Measured performance for random inputs:
402#
403# number of comparisons
404# n inputs k-extreme values (average of 5 trials) % more than min()
Raymond Hettinger356902d2014-05-19 22:13:45 +0100405# ------------- ---------------- --------------------- -----------------
Raymond Hettinger7df3c152014-06-02 01:32:23 -0700406# 1,000 100 3,317 231.7%
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700407# 10,000 100 14,046 40.5%
408# 100,000 100 105,749 5.7%
409# 1,000,000 100 1,007,751 0.8%
410# 10,000,000 100 10,009,401 0.1%
411#
412# Theoretical number of comparisons for k smallest of n random inputs:
413#
414# Step Comparisons Action
415# ---- -------------------------- ---------------------------
416# 1 1.66 * k heapify the first k-inputs
417# 2 n - k compare remaining elements to top of heap
418# 3 k * (1 + lg2(k)) * ln(n/k) replace the topmost value on the heap
419# 4 k * lg2(k) - (k/2) final sort of the k most extreme values
Raymond Hettinger41331e82014-05-26 00:58:56 -0700420#
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700421# Combining and simplifying for a rough estimate gives:
Raymond Hettinger41331e82014-05-26 00:58:56 -0700422#
Raymond Hettingerb321e792014-06-14 00:03:28 -0700423# comparisons = n + k * (log(k, 2) * log(n/k) + log(k, 2) + log(n/k))
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700424#
425# Computing the number of comparisons for step 3:
426# -----------------------------------------------
427# * For the i-th new value from the iterable, the probability of being in the
428# k most extreme values is k/i. For example, the probability of the 101st
429# value seen being in the 100 most extreme values is 100/101.
430# * If the value is a new extreme value, the cost of inserting it into the
431# heap is 1 + log(k, 2).
Berker Peksag1ed2e692014-10-19 18:07:05 +0300432# * The probability times the cost gives:
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700433# (k/i) * (1 + log(k, 2))
434# * Summing across the remaining n-k elements gives:
Raymond Hettinger41331e82014-05-26 00:58:56 -0700435# sum((k/i) * (1 + log(k, 2)) for i in range(k+1, n+1))
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700436# * This reduces to:
437# (H(n) - H(k)) * k * (1 + log(k, 2))
438# * Where H(n) is the n-th harmonic number estimated by:
439# gamma = 0.5772156649
Raymond Hettinger41331e82014-05-26 00:58:56 -0700440# H(n) = log(n, e) + gamma + 1 / (2 * n)
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700441# http://en.wikipedia.org/wiki/Harmonic_series_(mathematics)#Rate_of_divergence
442# * Substituting the H(n) formula:
443# comparisons = k * (1 + log(k, 2)) * (log(n/k, e) + (1/n - 1/k) / 2)
444#
445# Worst-case for step 3:
446# ----------------------
447# In the worst case, the input data is reversed sorted so that every new element
448# must be inserted in the heap:
449#
450# comparisons = 1.66 * k + log(k, 2) * (n - k)
451#
452# Alternative Algorithms
453# ----------------------
454# Other algorithms were not used because they:
455# 1) Took much more auxiliary memory,
456# 2) Made multiple passes over the data.
457# 3) Made more comparisons in common cases (small k, large n, semi-random input).
458# See the more detailed comparison of approach at:
459# http://code.activestate.com/recipes/577573-compare-algorithms-for-heapqsmallest
460
Raymond Hettinger4901a1f2004-12-02 08:59:14 +0000461def nsmallest(n, iterable, key=None):
462 """Find the n smallest elements in a dataset.
463
464 Equivalent to: sorted(iterable, key=key)[:n]
465 """
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700466
Raymond Hettinger354bffa2014-06-15 14:40:18 -0700467 # Short-cut for n==1 is to use min()
Benjamin Peterson18e95122009-01-18 22:46:33 +0000468 if n == 1:
469 it = iter(iterable)
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700470 sentinel = object()
Alexander Marshalove22072f2018-07-24 10:58:21 +0700471 result = min(it, default=sentinel, key=key)
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700472 return [] if result is sentinel else [result]
Benjamin Peterson18e95122009-01-18 22:46:33 +0000473
Éric Araujo395ba352011-04-15 23:34:31 +0200474 # When n>=size, it's faster to use sorted()
Benjamin Peterson18e95122009-01-18 22:46:33 +0000475 try:
476 size = len(iterable)
477 except (TypeError, AttributeError):
478 pass
479 else:
480 if n >= size:
481 return sorted(iterable, key=key)[:n]
482
483 # When key is none, use simpler decoration
Georg Brandl3a9b0622009-01-03 22:07:57 +0000484 if key is None:
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700485 it = iter(iterable)
Raymond Hettinger450ed102014-06-01 23:40:01 -0700486 # put the range(n) first so that zip() doesn't
487 # consume one too many elements from the iterator
Raymond Hettinger41331e82014-05-26 00:58:56 -0700488 result = [(elem, i) for i, elem in zip(range(n), it)]
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700489 if not result:
490 return result
491 _heapify_max(result)
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700492 top = result[0][0]
Raymond Hettinger41331e82014-05-26 00:58:56 -0700493 order = n
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700494 _heapreplace = _heapreplace_max
495 for elem in it:
496 if elem < top:
497 _heapreplace(result, (elem, order))
Raymond Hettinger1bfbe782017-09-04 11:47:58 -0700498 top, _order = result[0]
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700499 order += 1
500 result.sort()
Raymond Hettinger1bfbe782017-09-04 11:47:58 -0700501 return [elem for (elem, order) in result]
Benjamin Peterson18e95122009-01-18 22:46:33 +0000502
503 # General case, slowest method
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700504 it = iter(iterable)
505 result = [(key(elem), i, elem) for i, elem in zip(range(n), it)]
506 if not result:
507 return result
508 _heapify_max(result)
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700509 top = result[0][0]
Raymond Hettinger41331e82014-05-26 00:58:56 -0700510 order = n
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700511 _heapreplace = _heapreplace_max
512 for elem in it:
513 k = key(elem)
514 if k < top:
515 _heapreplace(result, (k, order, elem))
Raymond Hettinger1bfbe782017-09-04 11:47:58 -0700516 top, _order, _elem = result[0]
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700517 order += 1
518 result.sort()
Raymond Hettinger1bfbe782017-09-04 11:47:58 -0700519 return [elem for (k, order, elem) in result]
Raymond Hettinger4901a1f2004-12-02 08:59:14 +0000520
Raymond Hettinger4901a1f2004-12-02 08:59:14 +0000521def nlargest(n, iterable, key=None):
522 """Find the n largest elements in a dataset.
523
524 Equivalent to: sorted(iterable, key=key, reverse=True)[:n]
525 """
Benjamin Peterson18e95122009-01-18 22:46:33 +0000526
Raymond Hettinger354bffa2014-06-15 14:40:18 -0700527 # Short-cut for n==1 is to use max()
Benjamin Peterson18e95122009-01-18 22:46:33 +0000528 if n == 1:
529 it = iter(iterable)
Raymond Hettinger277842e2014-05-11 01:55:46 -0700530 sentinel = object()
Alexander Marshalove22072f2018-07-24 10:58:21 +0700531 result = max(it, default=sentinel, key=key)
Raymond Hettinger277842e2014-05-11 01:55:46 -0700532 return [] if result is sentinel else [result]
Benjamin Peterson18e95122009-01-18 22:46:33 +0000533
Éric Araujo395ba352011-04-15 23:34:31 +0200534 # When n>=size, it's faster to use sorted()
Benjamin Peterson18e95122009-01-18 22:46:33 +0000535 try:
536 size = len(iterable)
537 except (TypeError, AttributeError):
538 pass
539 else:
540 if n >= size:
541 return sorted(iterable, key=key, reverse=True)[:n]
542
543 # When key is none, use simpler decoration
Georg Brandl3a9b0622009-01-03 22:07:57 +0000544 if key is None:
Raymond Hettinger277842e2014-05-11 01:55:46 -0700545 it = iter(iterable)
Raymond Hettinger41331e82014-05-26 00:58:56 -0700546 result = [(elem, i) for i, elem in zip(range(0, -n, -1), it)]
Raymond Hettinger277842e2014-05-11 01:55:46 -0700547 if not result:
548 return result
549 heapify(result)
Raymond Hettinger277842e2014-05-11 01:55:46 -0700550 top = result[0][0]
Raymond Hettinger41331e82014-05-26 00:58:56 -0700551 order = -n
Raymond Hettinger277842e2014-05-11 01:55:46 -0700552 _heapreplace = heapreplace
553 for elem in it:
554 if top < elem:
Raymond Hettinger277842e2014-05-11 01:55:46 -0700555 _heapreplace(result, (elem, order))
Raymond Hettinger1bfbe782017-09-04 11:47:58 -0700556 top, _order = result[0]
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700557 order -= 1
Raymond Hettinger277842e2014-05-11 01:55:46 -0700558 result.sort(reverse=True)
Raymond Hettinger1bfbe782017-09-04 11:47:58 -0700559 return [elem for (elem, order) in result]
Benjamin Peterson18e95122009-01-18 22:46:33 +0000560
561 # General case, slowest method
Raymond Hettinger277842e2014-05-11 01:55:46 -0700562 it = iter(iterable)
563 result = [(key(elem), i, elem) for i, elem in zip(range(0, -n, -1), it)]
564 if not result:
565 return result
566 heapify(result)
Raymond Hettinger277842e2014-05-11 01:55:46 -0700567 top = result[0][0]
Raymond Hettinger41331e82014-05-26 00:58:56 -0700568 order = -n
Raymond Hettinger277842e2014-05-11 01:55:46 -0700569 _heapreplace = heapreplace
570 for elem in it:
571 k = key(elem)
572 if top < k:
Raymond Hettinger277842e2014-05-11 01:55:46 -0700573 _heapreplace(result, (k, order, elem))
Raymond Hettinger1bfbe782017-09-04 11:47:58 -0700574 top, _order, _elem = result[0]
Raymond Hettinger234fb2d2014-05-11 14:21:23 -0700575 order -= 1
Raymond Hettinger277842e2014-05-11 01:55:46 -0700576 result.sort(reverse=True)
Raymond Hettinger1bfbe782017-09-04 11:47:58 -0700577 return [elem for (k, order, elem) in result]
Raymond Hettinger277842e2014-05-11 01:55:46 -0700578
Raymond Hettinger48f68d02014-06-14 16:43:35 -0700579# If available, use C implementation
580try:
581 from _heapq import *
582except ImportError:
583 pass
584try:
585 from _heapq import _heapreplace_max
586except ImportError:
587 pass
588try:
589 from _heapq import _heapify_max
590except ImportError:
591 pass
592try:
593 from _heapq import _heappop_max
594except ImportError:
595 pass
596
Raymond Hettinger4901a1f2004-12-02 08:59:14 +0000597
Raymond Hettingerc46cb2a2004-04-19 19:06:21 +0000598if __name__ == "__main__":
Thomas Wouterscf297e42007-02-23 15:07:44 +0000599
600 import doctest
Raymond Hettinger450ed102014-06-01 23:40:01 -0700601 print(doctest.testmod())