forked from kswersky/CaRBM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sum_cardinality.py
972 lines (740 loc) · 32.7 KB
/
sum_cardinality.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
"""
Code provided by Kevin Swersky, Danny Tarlow, Ilya Sutskever, Ruslan Salakhutdinov, Rich Zemel and Ryan Adams.
Permission is granted for anyone to copy, use, modify, or distribute this
program and accompanying programs and documents for any purpose, provided
this copyright notice is retained and prominently displayed, along with
a note saying that the original programs are available from our
web page.
The programs and documents are distributed without any warranty, express or
implied. As the programs were written for research purposes only, they have
not been tested to the degree that would be advisable in any important
application. All use of these programs is entirely at the user's own risk.
This code implements the methods described in the paper:
Cardinality Restricted Boltzmann Machines. NIPS 2012.
"""
import numpy as np
import sys
import scipy.signal as sig
import scipy.cluster.hierarchy as hac
from scipy.spatial.distance import pdist, squareform
from time import clock, time
try:
import matplotlib
import matplotlib.pylab as plt
except:
pass
FFT_CROSSOVER_POINT = 32
def neg_energy(assn, node_potentials, count_potential):
nen = np.dot(assn, node_potentials) + count_potential[np.sum(assn)]
return nen
def idx_to_assn(idx, D):
""" convert 1D index to setting of binary variables """
return np.array([(idx / 2**d) % 2 for d in range(D)])
def random_categorical(probs):
return probs.cumsum().searchsorted(np.random.rand())
def hierarchical_cluster(trainx):
""" See the scipy.cluster.hierarchy documentation for the
meanings of entries in T.
The result can be plotted by calling hac.dendrogram(T). """
T = hac.complete(pdist(trainx.T) + .1)
return T
def make_balanced_binary_tree(D):
""" Make a tree structure to be used by the convolution tree
algorithm, where the tree is as balanced as possible.
The result can be plotted by calling hac.dendrogram(T). """
if D == 1: return np.zeros((0,4))
T = np.zeros((D-1,4))
ctr = 0
top_level = np.arange(D)
new_top_level = np.zeros(np.ceil(D/2.))
lvl = 1
while new_top_level.shape[0] > 1:
for d in range(top_level.shape[0]/2):
new_top_level[d] = ctr+D
T[ctr,0] = top_level[2*d]
T[ctr,1] = top_level[2*d+1]
T[ctr,2] = lvl
T[ctr,3] = 0
if T[ctr,0] < D: T[ctr,3] += 1
else: T[ctr,3] += T[T[ctr,0]-D,3]
if T[ctr,1] < D: T[ctr,3] += 1
else: T[ctr,3] += T[T[ctr,1]-D,3]
ctr += 1
if top_level.shape[0] % 2 == 1: new_top_level[-1] = top_level[-1]
top_level = new_top_level
new_top_level = np.zeros(np.ceil(top_level.shape[0]/2.))
lvl += 1
T[ctr,0] = top_level[0]
T[ctr,1] = top_level[1]
T[ctr,2] = lvl
if T[ctr,0] < D: T[ctr,3] += 1
else: T[ctr,3] += T[T[ctr,0]-D,3]
if T[ctr,1] < D: T[ctr,3] += 1
else: T[ctr,3] += T[T[ctr,1]-D,3]
return T
def runtime_test(Ds):
results = np.zeros((len(Ds), 4))
for d, D in enumerate(Ds):
print D
tree, fft, chain = test_our_algs(D, 1)
results[d,0] = D
results[d,1:] = [tree, fft, chain]
return results
def plot_runtime_results(results):
plt.rcParams["figure.figsize"] = 7,7
plt.rcParams["font.size"] = 22
matplotlib.rc("xtick", labelsize=24)
matplotlib.rc("ytick", labelsize=24)
params = {"text.fontsize" : 32,
"font.size" : 32,
"legend.fontsize" : 30,
"axes.labelsize" : 32,
"text.usetex" : False
}
plt.rcParams.update(params)
#plt.semilogx(results[:,0], results[:,3], 'r-x', lw=3)
#plt.semilogx(results[:,0], results[:,1], 'g-D', lw=3)
#plt.semilogx(results[:,0], results[:,2], 'b-s', lw=3)
plt.plot(results[:,0], results[:,3], 'r-x', lw=3, ms=10)
plt.plot(results[:,0], results[:,1], 'g-D', lw=3, ms=10)
plt.plot(results[:,0], results[:,2], 'b-s', lw=3, ms=10)
plt.legend(["Chain", "Tree", "FFT Tree"], loc="upper left")
plt.xticks([1e5, 2e5, 3e5])
plt.yticks([0, 60, 120, 180])
plt.xlabel("Problem Size")
plt.ylabel("Runtime (sec)")
return results
def test_our_algs(D, num_runs, count_cap=None):
from time import time, clock
# need a tree structure for the convolution tree algorithm.
T = make_balanced_binary_tree(D)
tree_time = 0
fftree_time = 0
chain_time = 0
for r in range(num_runs):
exp_node_pots = np.exp(1 * np.random.randn(D))
exp_count_pots = np.exp(0 * np.random.randn(D+1))
if count_cap is not None:
exp_count_pots[count_cap:] = -np.inf
exp_count_pot_dict = {} # dictionary mapping internal node indexes to exp count potentials
root_idx = D + T.shape[0]-1
exp_count_pot_dict[root_idx] = exp_count_pots
start = clock()
nm_conv, cm_conv, Z_conv = conv_tree(exp_node_pots, exp_count_pot_dict, T,
use_fft=False)
cm_conv = cm_conv[root_idx]
tree_time += (clock() - start)
start = clock()
nm_fconv, cm_fconv, Z_fconv = conv_tree(exp_node_pots, exp_count_pot_dict, T,
use_fft=True)
cm_fconv = cm_fconv[root_idx]
fftree_time += (clock() - start)
if D <= 20000:
start = clock()
nm_chain, cm_chain = pass_all_messages(exp_node_pots, exp_count_pots)
chain_time += (clock() - start)
# Make sure marginals agree to within eps
if False:
print cm_chain, cm_conv, cm_fconv
assert np.sum((nm_conv-nm_chain)**2) < 1e-8
assert np.sum((cm_conv-cm_chain)**2) < 1e-8
assert np.sum((nm_fconv-nm_chain)**2) < 1e-8
assert np.sum((cm_fconv-cm_chain)**2) < 1e-8
else:
chain_time = 0
avg_tree_time = tree_time / num_runs
avg_fftree_time = fftree_time / num_runs
avg_chain_time = chain_time / num_runs
print " Tree time \t%s" % avg_tree_time
print "FFTree time \t%s" % avg_fftree_time
print " Chain time \t%s" % avg_chain_time
return avg_tree_time, avg_fftree_time, avg_chain_time
def test_convolution_speeds(D, mode="full"):
""" fftconvolve should be *much* faster than the others.
if it's not (as it wasn't for me originally), you probably
need to upgrade your scipy version -- there was a bug in
previous versions that caused it to be very slow for
some inputs. """
from time import time, clock
a = np.random.rand(D)
b = np.random.rand(D)
functions = [np.convolve, sig.convolve, sig.fftconvolve]
times = np.zeros(len(functions))
for f, fn in enumerate(functions):
start = clock()
c = fn(a,b,mode=mode)
times[f] += (clock() - start)
return times
def conv_tree(exp_node_pots, exp_count_pot_dict, T,
use_fft=True, VERBOSE=False, count_cap=None):
if count_cap is None:
D = exp_node_pots.shape[0]
root_idx = D + T.shape[0]-1
count_cap = np.max(np.nonzero(exp_count_pot_dict[root_idx] > 0))
if VERBOSE: print "Setting count_cap = ", count_cap
# easier to special case D=1 than make code below work for it
if exp_node_pots.shape[0] == 1:
p0 = 1. / (1. + exp_node_pots[0])
p1 = exp_node_pots[0] / (1. + exp_node_pots[0])
if 1 in exp_count_pot_dict:
p0 *= exp_count_pot_dict[1][0]
p1 *= exp_count_pot_dict[1][1]
node_margs = np.array([p1])
count_margs = {}
count_margs[0] = np.array([p0, p1])
log_Z = np.log(1. + exp_node_pots[0])
return node_margs, count_margs, log_Z
TIME_up = 0
#TIME_uconv = 0
TIME_down = 0
#TIME_dconv = 0
D = exp_node_pots.shape[0]
cards = np.zeros(2*D-1, dtype=np.int)
cards[:D] = 1
if False: # slow way
# traverse the tree upwards to compute node cardinalities -- slow
for merge in range(T.shape[0]):
cards[D+merge] = cards[T[merge,0]] + cards[T[merge,1]]
cards[D+merge] = np.minimum(count_cap, cards[D+merge])
cards = np.int32(cards)
else: # fast way
cards[D:] = np.minimum(count_cap, T[:,3].astype(np.int))
# all messages will be stored in a single array. this
# array lets us know where to find them.
start_idxs = np.cumsum(np.hstack([0, cards + 1])).astype(np.int)
up_messages = np.zeros(np.sum(cards + 1))
down_messages = np.zeros(np.sum(cards + 1))
if VERBOSE: print "allocating message arrays of size", up_messages.shape[0]
# fill in unary potentials at leaves
if True: # slow way
for d in range(D):
start = start_idxs[d]
end = start + cards[d] + 1
if np.isinf(exp_node_pots[d]):
up_messages[start:end] = [0, 1]
else:
up_messages[start:end] = [1, exp_node_pots[d]]
else: # fast way
up_messages[:2*D:2] = 1
up_messages[1:2*D:2] = exp_node_pots
if VERBOSE:
print cards.shape, D, T.shape
print "Cards, starts"
print cards
print start_idxs
print "Initial msgs"
print up_messages
log_Z = 0
# Upward pass
start_up = time()
for m in range(T.shape[0]):
# merging T[m,0] and T[m,1] to get node dd
dd = D + m # index of parent node
ch1, ch2 = int(T[m,0]), int(T[m,1]) # indices of children nodes
start_ch1 = start_idxs[ch1]; end_ch1 = start_ch1 + cards[ch1] + 1
start_ch2 = start_idxs[ch2]; end_ch2 = start_ch2 + cards[ch2] + 1
start_p = start_idxs[dd]; end_p = start_p + cards[dd] + 1
use_fft_here = use_fft and np.minimum(cards[ch1], cards[ch2]) > FFT_CROSSOVER_POINT
ch1_msg = up_messages[start_ch1:end_ch1]
ch2_msg = up_messages[start_ch2:end_ch2]
# multiply in any subset count potentials
if ch1 in exp_count_pot_dict: ch1_msg *= exp_count_pot_dict[ch1]
if ch2 in exp_count_pot_dict: ch2_msg *= exp_count_pot_dict[ch2]
#start_conv = time()
if use_fft_here:
up_messages[start_p:end_p] = sig.fftconvolve(ch1_msg, ch2_msg, mode="full")[:count_cap+1]
else:
up_messages[start_p:end_p] = np.convolve(ch1_msg, ch2_msg, mode="full")[:count_cap+1]
#TIME_uconv += (time() - start_conv)
# normalize messages for numerical reasons, but store constants
# so we can compute the partition function (Z)
Z_m = np.sum(up_messages[start_p:end_p])
assert Z_m != 0, "Partition function is 0!"
up_messages[start_p:end_p] /= Z_m
log_Z += np.log(Z_m)
TIME_up += (time() - start_up)
# Last contribution to the partition function
count_beliefs = up_messages[start_p:end_p].copy()
root_idx = D + T.shape[0]-1
if root_idx in exp_count_pot_dict:
count_beliefs *= exp_count_pot_dict[D + T.shape[0]-1][:count_cap+1]
Z_ct = np.sum(count_beliefs)
count_beliefs /= Z_ct
if Z_ct < 1e-10:
#print "Warning Z_ct=%s (size=%s, ct[0]=%s)" %(Z_ct, D, exp_count_pot_dict[D+T.shape[0]-1][0])
#print np.prod(1. / (1.0 + exp_node_pots))
pass
log_Z += np.log(Z_ct)
if False:
print "End upward pass up messages"
print start_p, end_p, up_messages[start_p:end_p]
print "count beliefs"
print count_beliefs
print
# Downward pass
# set count potential to be uniform at root. will get multiplied after down_p_msg
# is created from root to children
start_down = time()
down_messages[start_p:end_p] = 1 #np.ones(count_cap+1)
count_margs = {}
for m in reversed(range(T.shape[0])):
# merged T[m,0] and T[m,1] to get node dd.
# now need to send message from parent down to children
dd = D + m # index of parent node
ch1, ch2 = int(T[m,0]), int(T[m,1]) # indices of children nodes
start_ch1 = start_idxs[ch1]; end_ch1 = start_ch1 + cards[ch1] + 1
start_ch2 = start_idxs[ch2]; end_ch2 = start_ch2 + cards[ch2] + 1
start_p = start_idxs[dd]; end_p = start_p + cards[dd] + 1
# Add just enough padding of 0's so that convolutions can be
# done with a simple call to convolve with mode="valid"
down_p_msg = np.zeros(cards[ch1] + cards[ch2] + 1)
down_p_msg[:end_p-start_p] = down_messages[start_p:end_p]
if dd in exp_count_pot_dict:
#print dd, "down_p before", down_p_msg
down_p_msg *= exp_count_pot_dict[dd][:down_p_msg.shape[0]]
#print dd, "down_p after", down_p_msg
count_margs[dd] = down_p_msg[:cards[dd]+1] * up_messages[start_p:end_p]
count_margs[dd] /= np.sum(count_margs[dd])
# could reverse child messages like this...
ch1_msg = up_messages[start_ch1:end_ch1]
ch2_msg = up_messages[start_ch2:end_ch2]
ch1_rev_msg = ch1_msg[::-1] # reversed!
ch2_rev_msg = ch2_msg[::-1] # reversed!
# ... but it's faster to do it all in one go
#if start_ch1 == 0: ch1_rev_msg = up_messages[end_ch1::-1]
#else: ch1_rev_msg = up_messages[end_ch1:start_ch1-1:-1]
#if start_ch2 == 0: ch2_rev_msg = up_messages[end_ch2::-1]
#else: ch2_rev_msg = up_messages[end_ch2:start_ch2-1:-1]
use_fft_here = use_fft and np.minimum(cards[ch1], cards[ch2]) > FFT_CROSSOVER_POINT
#start_conv = time()
if use_fft_here:
down1 = sig.fftconvolve(down_p_msg, ch2_rev_msg, mode="valid")
down2 = sig.fftconvolve(down_p_msg, ch1_rev_msg, mode="valid")
down1 = np.maximum(down1, 0)
down2 = np.maximum(down2, 0)
else:
down1 = np.convolve(down_p_msg, ch2_rev_msg, mode="valid")
down2 = np.convolve(down_p_msg, ch1_rev_msg, mode="valid")
#TIME_dconv += (time() - start_conv)
if False:
full_down = np.zeros(T[m,3]+1)
full_down[:down_p_msg.shape[0]] = down_p_msg
if ch1 < D: real_ch1_card = 1
else: real_ch1_card = T[ch1-D,3]
if ch2 < D: real_ch2_card = 1
else: real_ch2_card = T[ch2-D,3]
print T[m,3], real_ch1_card, real_ch2_card
print cards[dd], cards[ch1], cards[ch2]
print down_p_msg.shape[0], ch1_rev_msg.shape[0], ch2_rev_msg.shape[0]
full_ch1 = np.zeros(real_ch1_card+1)
full_ch1[:ch1_rev_msg.shape[0]] = ch1_rev_msg[::-1]
full_ch2 = np.zeros(real_ch2_card+1)
full_ch2[:ch2_rev_msg.shape[0]] = ch2_rev_msg[::-1]
full_ch1 = full_ch1[::-1]
full_ch2 = full_ch2[::-1]
print "Fulls"
print full_down
print full_ch1
print full_ch2
# Pull out the right part of down1/down2 to use as the downward message.
cstart_ch1 = down1.shape[0] - cards[ch1]-1
cend_ch1 = down1.shape[0]
cstart_ch2 = down2.shape[0] - cards[ch2]-1
cend_ch2 = down2.shape[0]
#print dd, "d1", down1, down1[cstart_ch1:cend_ch1], start_ch1, end_ch1
#print dd, "d2", down2, down2[cstart_ch2:cend_ch2], start_ch2, end_ch2
down_messages[start_ch1:end_ch1] = down1[cstart_ch1:cend_ch1]
down_messages[start_ch2:end_ch2] = down2[cstart_ch2:cend_ch2]
# to combat over/under-flow -- could be less aggressive here to make it faster
Z1 = np.sum(down_messages[start_ch1:end_ch1])
Z2 = np.sum(down_messages[start_ch2:end_ch2])
down_messages[start_ch1:end_ch1] /= Z1
down_messages[start_ch2:end_ch2] /= Z2
TIME_down += (time() - start_down)
node_beliefs = down_messages[:(2*D)] * up_messages[:(2*D)]
b0 = node_beliefs[::2]
b1 = node_beliefs[1::2]
node_margs = b1 / (b0 + b1)
if VERBOSE:
#print "Node marginals"
#print node_margs
print "Times"
print " UP", TIME_up
#print "uCONV", TIME_uconv
print " DOWN", TIME_down
#print "dCONV", TIME_dconv
return node_margs, count_margs, log_Z
def up_conv_tree_down_sample(exp_node_pots, exp_count_pot_dict, T,
use_fft=True, VERBOSE=False):
D = exp_node_pots.shape[0]
cards = np.zeros(2*D-1)
cards[:D] = 1
# traverse the tree upwards to compute node cardinalities
for merge in range(T.shape[0]):
cards[D+merge] = cards[T[merge,0]] + cards[T[merge,1]]
cards = np.int32(cards)
# all messages will be stored in a single array. this
# array lets us know where to find them.
start_idxs = np.cumsum(np.hstack([0, cards + 1]))
up_messages = np.zeros(np.sum(cards + 1))
down_messages = np.zeros(np.sum(cards + 1))
# fill in unary potentials at leaves
for d in range(D):
start = start_idxs[d]
end = start + cards[d] + 1
if np.isinf(exp_node_pots[d]):
up_messages[start:end] = [0, 1]
else:
up_messages[start:end] = [1, exp_node_pots[d]]
# Upward pass -- copied from conv_tree
#start_up = time()
for m in range(T.shape[0]):
# merging T[m,0] and T[m,1] to get node dd
dd = D + m # index of parent node
ch1, ch2 = int(T[m,0]), int(T[m,1]) # indices of children nodes
start_ch1 = start_idxs[ch1]; end_ch1 = start_ch1 + cards[ch1] + 1
start_ch2 = start_idxs[ch2]; end_ch2 = start_ch2 + cards[ch2] + 1
start_p = start_idxs[dd]; end_p = start_p + cards[dd] + 1
ch1_msg = up_messages[start_ch1:end_ch1]
ch2_msg = up_messages[start_ch2:end_ch2]
# multiply in any subset count potentials
if ch1 in exp_count_pot_dict: ch1_msg *= exp_count_pot_dict[ch1]
if ch2 in exp_count_pot_dict: ch2_msg *= exp_count_pot_dict[ch2]
use_fft_here = use_fft and np.minimum(cards[ch1], cards[ch2]) > FFT_CROSSOVER_POINT
if use_fft_here:
up_messages[start_p:end_p] = sig.fftconvolve(ch1_msg, ch2_msg, mode="full")
else:
up_messages[start_p:end_p] = np.convolve(ch1_msg, ch2_msg, mode="full")
# normalize messages for numerical reasons, but store constants
# so we can compute the partition function (Z)
Z_m = np.sum(up_messages[start_p:end_p])
assert Z_m != 0, "Partition function is 0!"
up_messages[start_p:end_p] /= Z_m
# Downward sampling pass
count_margs = {}
num_on = {}
root_idx = D + T.shape[0]-1
root_start = start_idxs[root_idx]; root_end = root_start + cards[root_idx] + 1
root_count_dist = up_messages[root_start:root_end]
if root_idx in exp_count_pot_dict: root_count_dist *= exp_count_pot_dict[root_idx]
root_count_dist /= np.sum(root_count_dist)
num_on[root_idx] = random_categorical(root_count_dist)
#print "root: %s on" % num_on[root_idx]
for m in reversed(range(T.shape[0])):
# merged T[m,0] and T[m,1] to get node dd.
# now need to send message from parent down to children
dd = D + m # index of parent node
ch1, ch2 = int(T[m,0]), int(T[m,1]) # indices of children nodes
start_ch1 = start_idxs[ch1]
start_ch2 = start_idxs[ch2]
parent_ct = num_on[dd]
# now need to sample children values given that their sum is p_ct
ch1_dist = np.zeros(cards[ch1]+1)
for ch1_ct in range(cards[ch1]+1):
ch2_ct = parent_ct - ch1_ct
#print "cts", parent_ct, "%s/%s" % (ch1_ct, cards[ch1]), "%s/%s" % (ch2_ct, cards[ch2])
if ch2_ct < 0 or ch2_ct > cards[ch2]: continue # assign this 0 prob
ch1_dist[ch1_ct] = up_messages[start_ch1+ch1_ct] * up_messages[start_ch2+ch2_ct]
if ch1 in exp_count_pot_dict: ch1_dist[ch1_ct] *= exp_count_pot_dict[ch1][ch1_ct]
if ch2 in exp_count_pot_dict: ch1_dist[ch1_ct] *= exp_count_pot_dict[ch2][ch2_ct]
#print ch1_dist
ch1_dist /= np.sum(ch1_dist)
num_on[ch1] = random_categorical(ch1_dist)
num_on[ch2] = parent_ct - num_on[ch1]
#print "(parent %s: %s)" % (dd, parent_ct)
#print "node %s <-- %s" % (ch1, num_on[ch1])
#print "node %s <-- %s" % (ch2, num_on[ch2])
return np.array([num_on[d] for d in range(D)])
def avg_logprob_from_empirical_margs(node_potentials, count_pot_dict,
empirical_margs, empirical_count_dict, log_Z):
logprob = np.sum(node_potentials * empirical_margs)
for tt in count_pot_dict:
logprob += np.sum(count_pot_dict[tt] * empirical_count_dict[tt])
logprob -= log_Z
return logprob
def parent_to_child_message(d, u, method="brute_force"):
""" A little test showing how to compute downward messages
using convolutions. """
# parent = ch1 + ch2, so ch1 = parent - ch2.
# this leads to messages to ch1 of the form:
#
# m(k) = \sum_i d(k+i)u(i) for i=0 to |u|
if method == "brute_force":
D_ch2 = u.shape[0]
D_ch1 = d.shape[0]-D_ch2
m = np.zeros(D_ch1+1)
for k in range(D_ch1+1):
for i in range(D_ch2):
m[k] += d[k+i]*u[i]
return m
else:
urev = u[::-1]
return sig.convolve(d, urev, mode="valid")
def marginals(node_potentials, count_potential, brute_force=False, VERBOSE=False,
print_messages=False):
Z = 0
D = node_potentials.shape[0]
marginals = np.zeros(D)
ct_marginals = np.zeros(D+1)
if brute_force:
for idx in range(2**D):
assn = idx_to_assn(idx, D)
nen = neg_energy(assn, node_potentials, count_potential)
for d in range(D):
if assn[d] == 1: marginals[d] += np.exp(nen)
ct_marginals[np.sum(assn)] += np.exp(nen)
Z += np.exp(nen)
print "Z=", Z
ct_marginals /= Z
marginals /= Z
else:
marginals, ct_marginals = pass_all_messages(np.exp(node_potentials), np.exp(count_potential),
print_messages=print_messages)
if VERBOSE:
print "Z =", Z
for idx in range(2**D):
assn = idx_to_assn(idx, D)
nen = neg_energy(assn, node_potentials, count_potential)
print "%s\t%4f" % (assn, np.exp(nen)/Z)
print "marginals", marginals
return marginals, ct_marginals
def independent_sample(node_potentials, count_potential, brute_force=True):
""" Draw samples from p(hj|v) for each j independently. """
D = node_potentials.shape[0]
qs, ct_margs = marginals(node_potentials, count_potential, brute_force=brute_force)
return np.int32(np.random.rand(D) < qs)
def joint_sample(node_potentials, count_potential, brute_force=True):
""" Draw a joint sample from p(h|v). """
if brute_force:
Z = 0
D = node_potentials.shape[0]
joint_probs = np.zeros(2**D)
for idx in range(2**D):
assn = idx_to_assn(idx, D)
nen = neg_energy(assn, node_potentials, count_potential)
joint_probs[idx] = np.exp(nen)
Z += np.exp(nen)
joint_probs /= Z
idx_arr = np.nonzero(np.random.multinomial(1, joint_probs) == 1)[0]
return idx_to_assn(idx_arr[0], D)
else:
return backward_messages_forward_sample(np.exp(node_potentials),
np.exp(count_potential), n=1)
def compute_bmessage(exp_node_potential, h, result_matrix, result_row, count_cap=None):
""" Compute backward message and put result in result_matrix[result_row,:].
Do it this way so that we don't allocate new memory.
Assumes z_d was defined as z_d = y_d + z_{d+1}. """
result_matrix[result_row,:] = h
result_matrix[result_row,1:] += exp_node_potential * h[:-1]
#result_matrix[result_row,count_cap+1:] = 0
# normalize for numerical stability
result_matrix[result_row,:] /= np.sum(result_matrix[result_row,:])
def compute_fmessage(exp_node_potential, h, result_matrix, result_row, count_cap=None):
""" Compute forward message and put result in result_matrix[result_row,:].
Do it this way so that we don't allocate new memory.
Assumes z_d was defined as z_d = y_d + z_{d+1}. """
result_matrix[result_row,:] = exp_node_potential * np.hstack([h[1:], 0])
#result_matrix[result_row,:-result_row] += h[:-result_row]
result_matrix[result_row,:] += h
# normalize for numerical stability
result_matrix[result_row,:] /= np.sum(result_matrix[result_row,:])
def pass_all_messages(exp_node_potentials, exp_count_potential, count_cap=None,
print_messages=False):
D = exp_node_potentials.shape[0]
if count_cap is None:
count_cap = np.max(np.nonzero(exp_count_potential > 1e-20))
#count_cap = (np.cumsum(exp_count_potential)*(exp_count_potential>1e-20)).argmax()
#print "Count cap =", count_cap
#fmsgs = np.zeros((D+1,D+1)) # forward messages
#bmsgs = np.zeros((D+1,D+1)) # backward messages
fmsgs = np.zeros((D+1,count_cap+1)) # forward messages
bmsgs = np.zeros((D+1,count_cap+1)) # backward messages
fmsgs[0, :] = exp_count_potential[:count_cap+1]
for d in range(D-1):
compute_fmessage(exp_node_potentials[d], fmsgs[d,:], fmsgs, d+1, count_cap=count_cap)
bmsgs[D,:2] = [1, exp_node_potentials[D-1]]
for d in reversed(range(1,D)):
compute_bmessage(exp_node_potentials[d-1], bmsgs[d+1,:], bmsgs, d, count_cap=count_cap)
if print_messages:
print "forward messages"
print fmsgs
print "backward messages"
print bmsgs
count_beliefs = exp_count_potential[:count_cap+1] * bmsgs[1,:]
count_marginals = np.zeros(D+1)
count_marginals[:count_cap+1] = count_beliefs / np.sum(count_beliefs)
# construct pairwise beliefs (without explicitly instantiating the D^2
# size matrices), then sum the diagonal to get b0, and the off-diagonal
# to get b1. b1/(b0+b1) gives marginal for original y_d for all except
# the last variable, y_D. we need to special case it, because there is
# no pairwise potential that represents \theta_D -- it's just a unary in
# the transformed model.
bb = bmsgs[2:,:]
ff = fmsgs[:-2,:]
b0 = np.sum(bb*ff,axis=1)
b1 = np.sum(bb[:,:-1]*ff[:,1:], axis=1) * exp_node_potentials[:-1]
marginals = np.zeros(D)
marginals[:-1] = b1/(b0+b1)
# could probably structure things so the Dth var doesn't need to be
# special-cased. but this will do for now. rather than computing
# a belief at a pairwise potential, we do it at the variable.
b0_D = fmsgs[D-1,0]*bmsgs[D,0]
b1_D = fmsgs[D-1,1]*bmsgs[D,1]
marginals[D-1] = b1_D / (b0_D+b1_D)
VERBOSE = False
if VERBOSE:
print "BP marginals"
print marginals
print
print "forward/backward msgs"
print fmsgs
print bmsgs
print
print "BP count marginals", count_beliefs / np.sum(count_beliefs)
print "BPZ", np.sum(count_beliefs)
print
return marginals, count_marginals
def backward_messages_forward_sample(exp_node_potentials, exp_count_potential, n=50000, count_cap=None):
D = exp_node_potentials.shape[0]
if count_cap is None:
count_cap = np.max(np.nonzero(exp_count_potential > 1e-20))
#print "Count cap =", count_cap
bmsgs = np.zeros((D+1,count_cap+1))
bmsgs[D,:2] = [1, exp_node_potentials[D-1]]
for d in reversed(range(1,D)):
compute_bmessage(exp_node_potentials[d-1], bmsgs[d+1,:], bmsgs, d, count_cap=count_cap)
count_beliefs = exp_count_potential[:count_cap+1] * bmsgs[1,:]
# could repeat the following many times if desired
count_freq = np.zeros(D+1)
marginals = np.zeros(D)
NUM_SAMPLES = n
for i in range(NUM_SAMPLES):
sample = np.zeros(D)
D_on = random_categorical(count_beliefs/np.sum(count_beliefs))
z_d = D_on
for d in range(D-1):
if z_d == 0: break
# unnormalized conditional probs p(z_d-1|z_d) = p(y_d=1|z_d) using backwards
# messages to compute the reparameterization.
p1 = exp_node_potentials[d] * bmsgs[d+2,z_d-1]
p0 = bmsgs[d+2,z_d]
sample[d] = np.random.rand() < p1 / (p0 + p1)
z_d -= sample[d]
sample[D-1] = z_d
assert D_on == np.sum(sample)
if n == 1: return sample
count_freq[np.sum(sample)] += 1
marginals += sample
print "Drew %s samples..." % n
print "Back-msgs, forward-sample count marginals"
print count_freq / np.sum(count_freq)
print "Back-msgs, forward-sample marginals"
print marginals / NUM_SAMPLES
def assignment_to_hard_node_pots(assn):
""" assn is assumed binary """
D = assn.shape[0]
exp_node_pots = np.zeros(D)
for d in range(D):
if assn[d] == 0: exp_node_pots[d] = 0
else: exp_node_pots[d] = np.inf
return exp_node_pots
if __name__ == "__main__":
D = int(sys.argv[1])
np.random.seed(0)
node_potentials = np.log((.0001+np.arange(D))/D) #np.random.randn(D)
count_potential = -np.inf * np.ones(D+1)
count_potential[2:5] = 0
#count_potential = np.ones(D+1)
#count_potential = np.random.randn(D+1) #np.zeros(D+1)
#count_potential[3:] = -np.inf
#count_potential[2] = np.log(10000)
if False:
A = np.zeros(7)
B = np.zeros(5)
C = np.zeros(3)
a = np.array([1,10])
b = np.array([1,2,3])
c = np.array([3,4])
A[:a.shape[0]] = a
B[:b.shape[0]] = b
C[:c.shape[0]] = c
res1 = np.convolve(A, B[::-1], mode="valid")
res2 = np.convolve(A, C[::-1], mode="valid")
print 80 * "*"
print A, B
print res1
print A, C
print res2
print 80 * "*"
print
res1 = np.convolve(a, b[::-1], mode="full")
res2 = np.convolve(a, c[::-1], mode="full")
print 80 * "*"
print a, b
print res1
print a, c
print res2
print 80 * "*"
print
sys.exit(0)
print
print 80 * "*"
print
print "NP", node_potentials
print "CP", count_potential
print
print 80 * "*"
print
CAN_BRUTE_FORCE = D <= 16
if CAN_BRUTE_FORCE:
qs, ct_margs = marginals(node_potentials, count_potential)
print "Exact marginals (brute force)"
print qs
print "Exact count marginals (brute force)"
print ct_margs
print
qs, ct_margs = marginals(node_potentials, count_potential, brute_force=False, print_messages=True)
print "Exact marginals (chain BP)"
print qs
print "Exact count marginals (chain BP)"
print ct_margs
print
T = make_balanced_binary_tree(D)
# view tree with: scipy.cluster.hierarchy.dendrogram(T)
# count potentials are stored in a dictionary, where internal node indexes map
# to count potentials
exp_count_pot_dict = {} # dictionary mapping internal node indexes to exp count potentials
root_idx = D + T.shape[0]-1
exp_count_pot_dict[root_idx] = np.exp(count_potential)
fft_margs, fft_ct_margs, fft_log_Z = conv_tree(np.exp(node_potentials), exp_count_pot_dict, T,
use_fft=False)
print "Exact marginals (FFT)"
print fft_margs
print "Exact count marginals (FFT)"
print fft_ct_margs[root_idx]
fft_ct_at_root = fft_ct_margs[root_idx]
print "FFT max error:", np.max(np.sqrt((fft_ct_at_root - ct_margs[:fft_ct_at_root.shape[0]])**2))
print "Test backward-messages, forward-sample"
backward_messages_forward_sample(np.exp(node_potentials), np.exp(count_potential),
n=1000)
sys.exit(0)
if False:
print "Independent samples"
for k in range(0):
sample = independent_sample(node_potentials, count_potential)
print np.sum(sample), sample
print
if CAN_BRUTE_FORCE:
print "Exact joint samples (brute force)"
for k in range(10):
sample = joint_sample(node_potentials, count_potential)
print np.sum(sample), sample
print
print "Exact joint samples (BP)"
for k in range(10):
sample = joint_sample(node_potentials, count_potential, brute_force=False)
print np.sum(sample), sample
print
print "Test backward-messages, forward-sample"
backward_messages_forward_sample(np.exp(node_potentials), np.exp(count_potential),
n=10)