1
# -*- coding: Latin-1 -*-
3
"""Heap queue algorithm (a.k.a. priority queue).
5
Heaps are arrays for which a[k] <= a[2*k+1] and a[k] <= a[2*k+2] for
6
all k, counting elements from 0. For the sake of comparison,
7
non-existing elements are considered to be infinite. The interesting
8
property of a heap is that a[0] is always its smallest element.
12
heap = [] # creates an empty heap
13
heappush(heap, item) # pushes a new item on the heap
14
item = heappop(heap) # pops the smallest item from the heap
15
item = heap[0] # smallest item on the heap without popping it
16
heapify(x) # transforms list into a heap, in-place, in linear time
17
item = heapreplace(heap, item) # pops and returns smallest item, and adds
18
# new item; the heap size is unchanged
20
Our API differs from textbook heap algorithms as follows:
22
- We use 0-based indexing. This makes the relationship between the
23
index for a node and the indexes for its children slightly less
24
obvious, but is more suitable since Python uses 0-based indexing.
26
- Our heappop() method returns the smallest item, not the largest.
28
These two make it possible to view the heap as a regular Python list
29
without surprises: heap[0] is the smallest item, and heap.sort()
30
maintains the heap invariant!
33
# Original code by Kevin O'Connor, augmented by Tim Peters and Raymond Hettinger
35
__about__ = """Heap queues
37
[explanation by Franļæ½ois Pinard]
39
Heaps are arrays for which a[k] <= a[2*k+1] and a[k] <= a[2*k+2] for
40
all k, counting elements from 0. For the sake of comparison,
41
non-existing elements are considered to be infinite. The interesting
42
property of a heap is that a[0] is always its smallest element.
44
The strange invariant above is meant to be an efficient memory
45
representation for a tournament. The numbers below are `k', not a[k]:
55
15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
58
In the tree above, each cell `k' is topping `2*k+1' and `2*k+2'. In
59
an usual binary tournament we see in sports, each cell is the winner
60
over the two cells it tops, and we can trace the winner down the tree
61
to see all opponents s/he had. However, in many computer applications
62
of such tournaments, we do not need to trace the history of a winner.
63
To be more memory efficient, when a winner is promoted, we try to
64
replace it by something else at a lower level, and the rule becomes
65
that a cell and the two cells it tops contain three different items,
66
but the top cell "wins" over the two topped cells.
68
If this heap invariant is protected at all time, index 0 is clearly
69
the overall winner. The simplest algorithmic way to remove it and
70
find the "next" winner is to move some loser (let's say cell 30 in the
71
diagram above) into the 0 position, and then percolate this new 0 down
72
the tree, exchanging values, until the invariant is re-established.
73
This is clearly logarithmic on the total number of items in the tree.
74
By iterating over all items, you get an O(n ln n) sort.
76
A nice feature of this sort is that you can efficiently insert new
77
items while the sort is going on, provided that the inserted items are
78
not "better" than the last 0'th element you extracted. This is
79
especially useful in simulation contexts, where the tree holds all
80
incoming events, and the "win" condition means the smallest scheduled
81
time. When an event schedule other events for execution, they are
82
scheduled into the future, so they can easily go into the heap. So, a
83
heap is a good structure for implementing schedulers (this is what I
84
used for my MIDI sequencer :-).
86
Various structures for implementing schedulers have been extensively
87
studied, and heaps are good for this, as they are reasonably speedy,
88
the speed is almost constant, and the worst case is not much different
89
than the average case. However, there are other representations which
90
are more efficient overall, yet the worst cases might be terrible.
92
Heaps are also very useful in big disk sorts. You most probably all
93
know that a big sort implies producing "runs" (which are pre-sorted
94
sequences, which size is usually related to the amount of CPU memory),
95
followed by a merging passes for these runs, which merging is often
96
very cleverly organised[1]. It is very important that the initial
97
sort produces the longest runs possible. Tournaments are a good way
98
to that. If, using all the memory available to hold a tournament, you
99
replace and percolate items that happen to fit the current run, you'll
100
produce runs which are twice the size of the memory for random input,
101
and much better for input fuzzily ordered.
103
Moreover, if you output the 0'th item on disk and get an input which
104
may not fit in the current tournament (because the value "wins" over
105
the last output value), it cannot fit in the heap, so the size of the
106
heap decreases. The freed memory could be cleverly reused immediately
107
for progressively building a second heap, which grows at exactly the
108
same rate the first heap is melting. When the first heap completely
109
vanishes, you switch heaps and start a new run. Clever and quite
112
In a word, heaps are useful memory structures to know. I use them in
113
a few applications, and I think it is good to keep a `heap' module
117
[1] The disk balancing algorithms which are current, nowadays, are
118
more annoying than clever, and this is a consequence of the seeking
119
capabilities of the disks. On devices which cannot seek, like big
120
tape drives, the story was quite different, and one had to be very
121
clever to ensure (far in advance) that each tape movement will be the
122
most effective possible (that is, will best participate at
123
"progressing" the merge). Some tapes were even able to read
124
backwards, and this was also used to avoid the rewinding time.
125
Believe me, real good tape sorts were quite spectacular to watch!
126
From all times, sorting has always been a Great Art! :-)
129
__all__ = ['heappush', 'heappop', 'heapify', 'heapreplace', 'nlargest',
132
from itertools import islice, repeat
135
def heappush(heap, item):
136
"""Push item onto heap, maintaining the heap invariant."""
138
_siftdown(heap, 0, len(heap)-1)
141
"""Pop the smallest item off the heap, maintaining the heap invariant."""
142
lastelt = heap.pop() # raises appropriate IndexError if heap is empty
151
def heapreplace(heap, item):
152
"""Pop and return the current smallest value, and add the new item.
154
This is more efficient than heappop() followed by heappush(), and can be
155
more appropriate when using a fixed-size heap. Note that the value
156
returned may be larger than item! That constrains reasonable uses of
157
this routine unless written as part of a conditional replacement:
160
item = heapreplace(heap, item)
162
returnitem = heap[0] # raises appropriate IndexError if heap is empty
168
"""Transform list into a heap, in-place, in O(len(heap)) time."""
170
# Transform bottom-up. The largest index there's any point to looking at
171
# is the largest with a child index in-range, so must have 2*i + 1 < n,
172
# or i < (n-1)/2. If n is even = 2*j, this is (2*j-1)/2 = j-1/2 so
173
# j-1 is the largest, which is n//2 - 1. If n is odd = 2*j+1, this is
174
# (2*j+1-1)/2 = j so j-1 is the largest, and that's again n//2-1.
175
for i in reversed(xrange(n//2)):
178
def nlargest(n, iterable):
179
"""Find the n largest elements in a dataset.
181
Equivalent to: sorted(iterable, reverse=True)[:n]
184
result = list(islice(it, n))
188
_heapreplace = heapreplace
189
sol = result[0] # sol --> smallest of the nlargest
193
_heapreplace(result, elem)
195
result.sort(reverse=True)
198
def nsmallest(n, iterable):
199
"""Find the n smallest elements in a dataset.
201
Equivalent to: sorted(iterable)[:n]
203
if hasattr(iterable, '__len__') and n * 10 <= len(iterable):
204
# For smaller values of n, the bisect method is faster than a minheap.
205
# It is also memory efficient, consuming only n elements of space.
207
result = sorted(islice(it, 0, n))
210
insort = bisect.insort
212
los = result[-1] # los --> Largest of the nsmallest
220
# An alternative approach manifests the whole iterable in memory but
221
# saves comparisons by heapifying all at once. Also, saves time
222
# over bisect.insort() which has O(n) data movement time for every
223
# insertion. Finding the n smallest of an m length iterable requires
224
# O(m) + O(n log m) comparisons.
227
return map(heappop, repeat(h, min(n, len(h))))
229
# 'heap' is a heap at all indices >= startpos, except possibly for pos. pos
230
# is the index of a leaf with a possibly out-of-order value. Restore the
232
def _siftdown(heap, startpos, pos):
234
# Follow the path to the root, moving parents down until finding a place
236
while pos > startpos:
237
parentpos = (pos - 1) >> 1
238
parent = heap[parentpos]
239
if parent <= newitem:
245
# The child indices of heap index pos are already heaps, and we want to make
246
# a heap at index pos too. We do this by bubbling the smaller child of
247
# pos up (and so on with that child's children, etc) until hitting a leaf,
248
# then using _siftdown to move the oddball originally at index pos into place.
250
# We *could* break out of the loop as soon as we find a pos where newitem <=
251
# both its children, but turns out that's not a good idea, and despite that
252
# many books write the algorithm that way. During a heap pop, the last array
253
# element is sifted in, and that tends to be large, so that comparing it
254
# against values starting from the root usually doesn't pay (= usually doesn't
255
# get us out of the loop early). See Knuth, Volume 3, where this is
256
# explained and quantified in an exercise.
258
# Cutting the # of comparisons is important, since these routines have no
259
# way to extract "the priority" from an array element, so that intelligence
260
# is likely to be hiding in custom __cmp__ methods, or in array elements
261
# storing (priority, record) tuples. Comparisons are thus potentially
264
# On random arrays of length 1000, making this change cut the number of
265
# comparisons made by heapify() a little, and those made by exhaustive
266
# heappop() a lot, in accord with theory. Here are typical results from 3
267
# runs (3 just to demonstrate how small the variance is):
269
# Compares needed by heapify Compares needed by 1000 heappops
270
# -------------------------- --------------------------------
271
# 1837 cut to 1663 14996 cut to 8680
272
# 1855 cut to 1659 14966 cut to 8678
273
# 1847 cut to 1660 15024 cut to 8703
275
# Building the heap by using heappush() 1000 times instead required
276
# 2198, 2148, and 2219 compares: heapify() is more efficient, when
279
# The total compares needed by list.sort() on the same lists were 8627,
280
# 8627, and 8632 (this should be compared to the sum of heapify() and
281
# heappop() compares): list.sort() is (unsurprisingly!) more efficient
284
def _siftup(heap, pos):
288
# Bubble up the smaller child until hitting a leaf.
289
childpos = 2*pos + 1 # leftmost child position
290
while childpos < endpos:
291
# Set childpos to index of smaller child.
292
rightpos = childpos + 1
293
if rightpos < endpos and heap[rightpos] <= heap[childpos]:
295
# Move the smaller child up.
296
heap[pos] = heap[childpos]
299
# The leaf at pos is empty now. Put newitem there, and bubble it up
300
# to its final resting place (by sifting its parents down).
302
_siftdown(heap, startpos, pos)
304
# If available, use C implementation
306
from _heapq import heappush, heappop, heapify, heapreplace, nlargest, nsmallest
310
if __name__ == "__main__":
313
data = [1, 3, 5, 7, 9, 2, 4, 6, 8, 0]
318
sort.append(heappop(heap))