1 # -*- coding: latin-1 -*- 2 3 """Heap queue algorithm (a.k.a. priority queue). 4 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. 9 10 Usage: 11 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 19 20 Our API differs from textbook heap algorithms as follows: 21 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. 25 26 - Our heappop() method returns the smallest item, not the largest. 27 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! 31 """ 32 33 # Original code by Kevin O'Connor, augmented by Tim Peters and Raymond Hettinger 34 35 __about__ = """Heap queues 36 37 [explanation by Franois Pinard] 38 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. 43 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]: 46 47 0 48 49 1 2 50 51 3 4 5 6 52 53 7 8 9 10 11 12 13 14 54 55 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 56 57 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. 67 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. 75 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 :-). 85 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. 91 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. 102 103 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 110 effective! 111 112 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 114 around. :-) 115 116 -------------------- 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! :-) 127 """ 128 129 __all__ = ['heappush', 'heappop', 'heapify', 'heapreplace', 'merge', 130 'nlargest', 'nsmallest', 'heappushpop'] 131 132 from itertools import islice, repeat, count, imap, izip, tee, chain 133 from operator import itemgetter 134 import bisect 135 136 def cmp_lt(x, y): 137 # Use __lt__ if available; otherwise, try __le__. 138 # In Py3.x, only __lt__ will be called. 139 return (x < y) if hasattr(x, '__lt__') else (not y <= x) 140 141 def heappush(heap, item): 142 """Push item onto heap, maintaining the heap invariant.""" 143 heap.append(item) 144 _siftdown(heap, 0, len(heap)-1) 145 146 def heappop(heap): 147 """Pop the smallest item off the heap, maintaining the heap invariant.""" 148 lastelt = heap.pop() # raises appropriate IndexError if heap is empty 149 if heap: 150 returnitem = heap[0] 151 heap[0] = lastelt 152 _siftup(heap, 0) 153 else: 154 returnitem = lastelt 155 return returnitem 156 157 def heapreplace(heap, item): 158 """Pop and return the current smallest value, and add the new item. 159 160 This is more efficient than heappop() followed by heappush(), and can be 161 more appropriate when using a fixed-size heap. Note that the value 162 returned may be larger than item! That constrains reasonable uses of 163 this routine unless written as part of a conditional replacement: 164 165 if item > heap[0]: 166 item = heapreplace(heap, item) 167 """ 168 returnitem = heap[0] # raises appropriate IndexError if heap is empty 169 heap[0] = item 170 _siftup(heap, 0) 171 return returnitem 172 173 def heappushpop(heap, item): 174 """Fast version of a heappush followed by a heappop.""" 175 if heap and cmp_lt(heap[0], item): 176 item, heap[0] = heap[0], item 177 _siftup(heap, 0) 178 return item 179 180 def heapify(x): 181 """Transform list into a heap, in-place, in O(len(x)) time.""" 182 n = len(x) 183 # Transform bottom-up. The largest index there's any point to looking at 184 # is the largest with a child index in-range, so must have 2*i + 1 < n, 185 # or i < (n-1)/2. If n is even = 2*j, this is (2*j-1)/2 = j-1/2 so 186 # j-1 is the largest, which is n//2 - 1. If n is odd = 2*j+1, this is 187 # (2*j+1-1)/2 = j so j-1 is the largest, and that's again n//2-1. 188 for i in reversed(xrange(n//2)): 189 _siftup(x, i) 190 191 def nlargest(n, iterable): 192 """Find the n largest elements in a dataset. 193 194 Equivalent to: sorted(iterable, reverse=True)[:n] 195 """ 196 it = iter(iterable) 197 result = list(islice(it, n)) 198 if not result: 199 return result 200 heapify(result) 201 _heappushpop = heappushpop 202 for elem in it: 203 _heappushpop(result, elem) 204 result.sort(reverse=True) 205 return result 206 207 def nsmallest(n, iterable): 208 """Find the n smallest elements in a dataset. 209 210 Equivalent to: sorted(iterable)[:n] 211 """ 212 if hasattr(iterable, '__len__') and n * 10 <= len(iterable): 213 # For smaller values of n, the bisect method is faster than a minheap. 214 # It is also memory efficient, consuming only n elements of space. 215 it = iter(iterable) 216 result = sorted(islice(it, 0, n)) 217 if not result: 218 return result 219 insort = bisect.insort 220 pop = result.pop 221 los = result[-1] # los --> Largest of the nsmallest 222 for elem in it: 223 if cmp_lt(elem, los): 224 insort(result, elem) 225 pop() 226 los = result[-1] 227 return result 228 # An alternative approach manifests the whole iterable in memory but 229 # saves comparisons by heapifying all at once. Also, saves time 230 # over bisect.insort() which has O(n) data movement time for every 231 # insertion. Finding the n smallest of an m length iterable requires 232 # O(m) + O(n log m) comparisons. 233 h = list(iterable) 234 heapify(h) 235 return map(heappop, repeat(h, min(n, len(h)))) 236 237 # 'heap' is a heap at all indices >= startpos, except possibly for pos. pos 238 # is the index of a leaf with a possibly out-of-order value. Restore the 239 # heap invariant. 240 def _siftdown(heap, startpos, pos): 241 newitem = heap[pos] 242 # Follow the path to the root, moving parents down until finding a place 243 # newitem fits. 244 while pos > startpos: 245 parentpos = (pos - 1) >> 1 246 parent = heap[parentpos] 247 if cmp_lt(newitem, parent): 248 heap[pos] = parent 249 pos = parentpos 250 continue 251 break 252 heap[pos] = newitem 253 254 # The child indices of heap index pos are already heaps, and we want to make 255 # a heap at index pos too. We do this by bubbling the smaller child of 256 # pos up (and so on with that child's children, etc) until hitting a leaf, 257 # then using _siftdown to move the oddball originally at index pos into place. 258 # 259 # We *could* break out of the loop as soon as we find a pos where newitem <= 260 # both its children, but turns out that's not a good idea, and despite that 261 # many books write the algorithm that way. During a heap pop, the last array 262 # element is sifted in, and that tends to be large, so that comparing it 263 # against values starting from the root usually doesn't pay (= usually doesn't 264 # get us out of the loop early). See Knuth, Volume 3, where this is 265 # explained and quantified in an exercise. 266 # 267 # Cutting the # of comparisons is important, since these routines have no 268 # way to extract "the priority" from an array element, so that intelligence 269 # is likely to be hiding in custom __cmp__ methods, or in array elements 270 # storing (priority, record) tuples. Comparisons are thus potentially 271 # expensive. 272 # 273 # On random arrays of length 1000, making this change cut the number of 274 # comparisons made by heapify() a little, and those made by exhaustive 275 # heappop() a lot, in accord with theory. Here are typical results from 3 276 # runs (3 just to demonstrate how small the variance is): 277 # 278 # Compares needed by heapify Compares needed by 1000 heappops 279 # -------------------------- -------------------------------- 280 # 1837 cut to 1663 14996 cut to 8680 281 # 1855 cut to 1659 14966 cut to 8678 282 # 1847 cut to 1660 15024 cut to 8703 283 # 284 # Building the heap by using heappush() 1000 times instead required 285 # 2198, 2148, and 2219 compares: heapify() is more efficient, when 286 # you can use it. 287 # 288 # The total compares needed by list.sort() on the same lists were 8627, 289 # 8627, and 8632 (this should be compared to the sum of heapify() and 290 # heappop() compares): list.sort() is (unsurprisingly!) more efficient 291 # for sorting. 292 293 def _siftup(heap, pos): 294 endpos = len(heap) 295 startpos = pos 296 newitem = heap[pos] 297 # Bubble up the smaller child until hitting a leaf. 298 childpos = 2*pos + 1 # leftmost child position 299 while childpos < endpos: 300 # Set childpos to index of smaller child. 301 rightpos = childpos + 1 302 if rightpos < endpos and not cmp_lt(heap[childpos], heap[rightpos]): 303 childpos = rightpos 304 # Move the smaller child up. 305 heap[pos] = heap[childpos] 306 pos = childpos 307 childpos = 2*pos + 1 308 # The leaf at pos is empty now. Put newitem there, and bubble it up 309 # to its final resting place (by sifting its parents down). 310 heap[pos] = newitem 311 _siftdown(heap, startpos, pos) 312 313 # If available, use C implementation 314 try: 315 from _heapq import * 316 except ImportError: 317 pass 318 319 def merge(*iterables): 320 '''Merge multiple sorted inputs into a single sorted output. 321 322 Similar to sorted(itertools.chain(*iterables)) but returns a generator, 323 does not pull the data into memory all at once, and assumes that each of 324 the input streams is already sorted (smallest to largest). 325 326 >>> list(merge([1,3,5,7], [0,2,4,8], [5,10,15,20], [], [25])) 327 [0, 1, 2, 3, 4, 5, 5, 7, 8, 10, 15, 20, 25] 328 329 ''' 330 _heappop, _heapreplace, _StopIteration = heappop, heapreplace, StopIteration 331 332 h = [] 333 h_append = h.append 334 for itnum, it in enumerate(map(iter, iterables)): 335 try: 336 next = it.next 337 h_append([next(), itnum, next]) 338 except _StopIteration: 339 pass 340 heapify(h) 341 342 while 1: 343 try: 344 while 1: 345 v, itnum, next = s = h[0] # raises IndexError when h is empty 346 yield v 347 s[0] = next() # raises StopIteration when exhausted 348 _heapreplace(h, s) # restore heap condition 349 except _StopIteration: 350 _heappop(h) # remove empty iterator 351 except IndexError: 352 return 353 354 # Extend the implementations of nsmallest and nlargest to use a key= argument 355 _nsmallest = nsmallest 356 def nsmallest(n, iterable, key=None): 357 """Find the n smallest elements in a dataset. 358 359 Equivalent to: sorted(iterable, key=key)[:n] 360 """ 361 # Short-cut for n==1 is to use min() when len(iterable)>0 362 if n == 1: 363 it = iter(iterable) 364 head = list(islice(it, 1)) 365 if not head: 366 return [] 367 if key is None: 368 return [min(chain(head, it))] 369 return [min(chain(head, it), key=key)] 370 371 # When n>=size, it's faster to use sorted() 372 try: 373 size = len(iterable) 374 except (TypeError, AttributeError): 375 pass 376 else: 377 if n >= size: 378 return sorted(iterable, key=key)[:n] 379 380 # When key is none, use simpler decoration 381 if key is None: 382 it = izip(iterable, count()) # decorate 383 result = _nsmallest(n, it) 384 return map(itemgetter(0), result) # undecorate 385 386 # General case, slowest method 387 in1, in2 = tee(iterable) 388 it = izip(imap(key, in1), count(), in2) # decorate 389 result = _nsmallest(n, it) 390 return map(itemgetter(2), result) # undecorate 391 392 _nlargest = nlargest 393 def nlargest(n, iterable, key=None): 394 """Find the n largest elements in a dataset. 395 396 Equivalent to: sorted(iterable, key=key, reverse=True)[:n] 397 """ 398 399 # Short-cut for n==1 is to use max() when len(iterable)>0 400 if n == 1: 401 it = iter(iterable) 402 head = list(islice(it, 1)) 403 if not head: 404 return [] 405 if key is None: 406 return [max(chain(head, it))] 407 return [max(chain(head, it), key=key)] 408 409 # When n>=size, it's faster to use sorted() 410 try: 411 size = len(iterable) 412 except (TypeError, AttributeError): 413 pass 414 else: 415 if n >= size: 416 return sorted(iterable, key=key, reverse=True)[:n] 417 418 # When key is none, use simpler decoration 419 if key is None: 420 it = izip(iterable, count(0,-1)) # decorate 421 result = _nlargest(n, it) 422 return map(itemgetter(0), result) # undecorate 423 424 # General case, slowest method 425 in1, in2 = tee(iterable) 426 it = izip(imap(key, in1), count(0,-1), in2) # decorate 427 result = _nlargest(n, it) 428 return map(itemgetter(2), result) # undecorate 429 430 if __name__ == "__main__": 431 # Simple sanity test 432 heap = [] 433 data = [1, 3, 5, 7, 9, 2, 4, 6, 8, 0] 434 for item in data: 435 heappush(heap, item) 436 sort = [] 437 while heap: 438 sort.append(heappop(heap)) 439 print sort 440 441 import doctest 442 doctest.testmod() 443