-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathcache.h
1427 lines (1247 loc) · 48.5 KB
/
cache.h
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
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/**
* \file cache.h
*
* This file contains the ::cache structure, which optimizes MCMC inference for
* the HDP. This structure is specialized for various combinations of
* `BaseDistribution` and `DataDistribution`, and the specializations are also
* implemented in this file.
*
* <!-- Created on: Jul 11, 2016
* Author: asaparov -->
*/
#ifndef CACHE_H_
#define CACHE_H_
#include <math/distributions.h>
/* forward declarations */
#if !defined(DOXYGEN_IGNORE)
template<typename K, typename V> struct node_sampler;
template<typename BaseDistribution, typename DataDistribution, typename K, typename V> struct hdp_sampler;
#endif
template<typename K>
constexpr unsigned int count(const K& observation) {
return 1;
}
template<typename K>
inline unsigned int count(const array_multiset<K>& observations) {
return observations.sum;
}
template<typename V>
inline void cleanup_root_probabilities(V** root_probabilities, unsigned int row_count)
{
for (unsigned int i = 0; i < row_count; i++)
free(root_probabilities[i]);
free(root_probabilities);
}
inline void cleanup_root_probabilities(array<unsigned int>* root_probabilities, unsigned int row_count)
{
for (unsigned int i = 0; i < row_count; i++)
free(root_probabilities[i]);
free(root_probabilities);
}
/* a default implementation that various cache implementations can use */
template<typename BaseDistribution, typename DataDistribution, typename K, typename V>
V** compute_root_probabilities(const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h, const K& observation)
{
/* TODO: there's a potential speedup with a single contiguous array */
/* initialize the root_probabilities array of arrays */
V** root_probabilities = (V**) malloc(sizeof(V*) * h.posterior.length);
if (root_probabilities == NULL) {
fprintf(stderr, "predict ERROR: Insufficient memory for root_probabilities.\n");
return NULL;
}
unsigned int sample_count = 0;
for (unsigned int i = 0; i < h.posterior.length; i++) {
root_probabilities[i] = (V*) malloc(sizeof(V) * h.posterior[i].table_count);
if (root_probabilities[i] == NULL) {
fprintf(stderr, "predict ERROR: Insufficient memory for root_probabilities[%u].\n", i);
cleanup_root_probabilities(root_probabilities, sample_count);
return NULL;
}
sample_count++;
}
/* store the appropriate conditional probabilities in root_probabilities */
for (unsigned int i = 0; i < h.posterior.length; i++) {
for (unsigned int j = 0; j < h.posterior[i].table_count; j++)
/* TODO: test numerical stability */
if (h.posterior[i].descendant_observations[j].counts.size == 0) {
/* theoretically, this should be the prior probability of 'observation',
but since the table is empty, its likelihood will be zero anyway */
root_probabilities[i][j] = 0.0;
} else {
root_probabilities[i][j] = DataDistribution::conditional(
h.pi(), observation, h.posterior[i].descendant_observations[j]);
}
}
return root_probabilities;
}
/* a default implementation that various cache implementations can use */
template<typename BaseDistribution, typename DataDistribution, typename K, typename V>
V** compute_root_probabilities(
const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h,
const K* observations, unsigned int observation_count)
{
V** probabilities = (V**) malloc(sizeof(V*) * observation_count);
if (probabilities == NULL) {
fprintf(stderr, "cache.compute_root_probabilities ERROR: Insufficient memory for matrix.\n");
return NULL;
}
for (unsigned int i = 0; i < observation_count; i++) {
probabilities[i] = (V*) malloc(sizeof(V) * h.table_count);
if (probabilities[i] == NULL) {
fprintf(stderr, "cache.compute_root_probabilities ERROR: Insufficient memory for matrix row.\n");
return NULL;
}
for (unsigned int j = 0; j < h.table_count; j++) {
if (h.descendant_observations[j].counts.size == 0)
/* theoretically, this should be the prior probability, but
since the table is empty, the likelihood will be zero anyway */
probabilities[i][j] = 0.0;
else probabilities[i][j] = DataDistribution::conditional(h.pi(), observations[i], h.descendant_observations[j]);
}
}
return probabilities;
}
/**
* This structure is used to provide optimizations for the Gibbs sampler in
* mcmc.h for specific combinations of `BaseDistribution` and
* `DataDistribution`. This struct is specialized for those particular
* combinations of base distribution and likelihood. These specializations are
* implemented further down in this file. Currently, a default implementation
* is unavailable.
*/
template<typename BaseDistribution, typename DataDistribution, typename K, typename V, class Enable = void>
struct cache {
/* TODO: create a default implementation */
typedef hdp_sampler<BaseDistribution, dense_categorical<V>, unsigned int, V> sampler_root;
/**
* Constructs a cache structure for the HDP sampler hierarchy rooted at `h`.
*/
cache(const sampler_root& h) {
fprintf(stderr, "cache ERROR: Unimplemented. This combination of "
"BaseDistribution and DataDistribution is currently unsupported.\n");
exit(EXIT_FAILURE);
}
/**
* Computes a matrix, where every row corresponds to a
* sample from the posterior, and every column is a table
* at the root node. Each element contains the probability
* of assigning the given observation to each table at the
* root (without taking into account table sizes and the
* alpha parameter).
*/
inline V** compute_root_probabilities(const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h, const K& observation) {
return ::compute_root_probabilities(h, observation);
}
/**
* Constructs a map from observations to a vector, of which
* each element is the probability of assigning that
* observation to each table at the root (without taking
* into account table sizes and the alpha parameter).
*/
inline V** compute_root_probabilities(
const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h,
const K* observations, unsigned int observation_count)
{
return ::compute_root_probabilities(h, observations, observation_count);
}
};
/* this structure caches values of the log rising factorial
log (pi + i)^(n) for various values of i and n */
template<typename V>
struct rising_factorial_cache {
array_map<unsigned int, array<V>> values;
V pi;
rising_factorial_cache(const V& pi) :
values(64), pi(pi)
{ }
~rising_factorial_cache() { free(); }
inline const V* get_rising_factorials(unsigned int exponent, unsigned int largest_table_size) {
if (!values.ensure_capacity((unsigned int) values.size + 1))
return NULL;
bool contains;
array<V>& log_factorials = values.get(exponent, contains);
if (!contains) {
if (!array_init(log_factorials, largest_table_size + 1)) {
fprintf(stderr, "rising_factorial_cache.get_rising_factorials ERROR: Out of memory.\n");
return NULL;
}
values.keys[values.size] = exponent;
values.size++;
} else if (!log_factorials.ensure_capacity(largest_table_size + 1)) {
fprintf(stderr, "rising_factorial_cache.get_rising_factorials"
" ERROR: Unable to expand array of log factorials.\n");
return NULL;
}
for (unsigned int i = (unsigned int) log_factorials.length; i < largest_table_size + 1; i++)
log_factorials[i] = log_rising_factorial(pi + i, exponent);
if (log_factorials.length < largest_table_size + 1)
log_factorials.length = largest_table_size + 1;
const V* internal_array = log_factorials.data;
if (!contains)
sort(values.keys, values.values, (unsigned int) values.size, default_sorter());
return internal_array;
}
bool is_valid() const {
for (unsigned int i = 0; i < values.size; i++)
if (!is_valid(pi, values.keys[i], values.values[i]))
return false;
return true;
}
static bool is_valid(const V& pi, unsigned int exponent, const array<V>& log_factorials) {
for (unsigned int i = 0; i < log_factorials.length; i++) {
if (log_factorials[i] != log_rising_factorial(pi + i, exponent)) {
fprintf(stderr, "rising_factorial_cache.is_valid ERROR: Incorrect "
"log rising factorial for base %lf + %u and exponent %u.\n", pi, i, exponent);
return false;
}
}
return true;
}
template<typename Metric>
static inline long unsigned int size_of(const rising_factorial_cache<V>& cache, const Metric& metric) {
return core::size_of(cache.values, make_key_value_metric(default_metric(), metric)) + core::size_of(cache.pi);
}
static inline void free(rising_factorial_cache<V>& cache) {
cache.free();
core::free(cache.values);
}
private:
inline void free() {
for (unsigned int i = 0; i < values.size; i++)
core::free(values.values[i]);
}
};
template<typename V>
inline bool init(rising_factorial_cache<V>& cache, const V& pi) {
cache.pi = pi;
return array_map_init(cache.values, 64);
}
template<typename V>
struct table_distribution {
/* *denominators* of root-level table assignment likelihoods
(for a fixed size of the observation set) */
array_categorical<V> likelihoods;
/* *denominators* of root-level table assignment probabilities
(the above likelihoods multiplied by the table sizes/alpha) */
V* probabilities;
V sum;
template<typename BaseDistribution, typename DataDistribution, typename K>
inline void set_higher(
const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h,
unsigned int table, const V& log_probability)
{
#if !defined(NDEBUG)
if (log_probability < likelihoods.log_probabilities[table])
fprintf(stderr, "table_distribution.set_higher WARNING: Given probability is smaller than the current value.\n");
#endif
likelihoods.log_probabilities[table] = log_probability;
if (likelihoods.log_probabilities[table] - likelihoods.maximum > CATEGORICAL_MAX_THRESHOLD) {
renormalize(h);
} else {
likelihoods.probabilities[table] = exp(likelihoods.log_probabilities[table] - likelihoods.maximum);
sum -= probabilities[table];
probabilities[table] = h.table_sizes[table] * likelihoods.probabilities[table];
sum += probabilities[table];
}
}
template<typename BaseDistribution, typename DataDistribution, typename K>
inline void set_lower(const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h,
unsigned int table, const V& log_probability)
{
#if !defined(NDEBUG)
if (log_probability > likelihoods.log_probabilities[table])
fprintf(stderr, "table_distribution.set_lower WARNING: Given probability is larger than the current value.\n");
#endif
likelihoods.log_probabilities[table] = log_probability;
likelihoods.probabilities[table] = exp(likelihoods.log_probabilities[table] - likelihoods.maximum);
sum -= probabilities[table];
probabilities[table] = h.table_sizes[table] * likelihoods.probabilities[table];
sum += probabilities[table];
if (sum < CATEGORICAL_MIN_THRESHOLD)
renormalize(h);
}
/* NOTE: this function assumes the new table has already been created in the hdp */
template<typename BaseDistribution, typename DataDistribution, typename K>
inline bool add_new_table(
const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h,
unsigned int count, unsigned int observation_count)
{
if (!likelihoods.resize(h.table_count))
return false;
likelihoods.log_probabilities[h.table_count - 1] =
-log_rising_factorial(h.pi().sum() + count, observation_count);
likelihoods.probabilities[h.table_count - 1] =
exp(likelihoods.log_probabilities[h.table_count - 1] - likelihoods.maximum);
V* new_probabilities = (V*) realloc(probabilities, sizeof(V) * h.table_count);
if (new_probabilities == NULL) {
fprintf(stderr, "table_distribution.add_new_table ERROR: Out of memory.\n");
return false;
}
probabilities = new_probabilities;
probabilities[h.table_count - 1] = h.table_sizes[h.table_count - 1] * likelihoods.probabilities[h.table_count - 1];
sum += probabilities[h.table_count - 1];
return true;
}
template<bool Check, typename BaseDistribution, typename DataDistribution, typename K>
inline void change_descendant_count(
const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h,
unsigned int table, unsigned int new_count, unsigned int observation_count)
{
V new_likelihood = -log_rising_factorial(h.pi().sum() + new_count, observation_count);
if (Check) {
set_lower(h, table, new_likelihood);
} else {
likelihoods.log_probabilities[table] = new_likelihood;
likelihoods.probabilities[table] = exp(likelihoods.log_probabilities[table] - likelihoods.maximum);
sum -= probabilities[table];
probabilities[table] = h.table_sizes[table] * likelihoods.probabilities[table];
sum += probabilities[table];
}
}
inline void change_table_size(unsigned int table, unsigned int new_table_size) {
sum -= probabilities[table];
probabilities[table] = new_table_size * likelihoods.probabilities[table];
sum += probabilities[table];
}
template<typename BaseDistribution, typename DataDistribution, typename K>
bool is_valid(
const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h,
unsigned int observation_count) const
{
V actual_sum = 0.0;
for (unsigned int i = 0; i < h.table_count; i++) {
V expected = -log_rising_factorial(
h.pi().sum() + h.descendant_observations[i].sum, observation_count);
if (h.table_sizes[i] > 0 && fabs(likelihoods.log_probabilities[i] - expected) > fabs(expected) * 1.0e-12) {
fprintf(stderr, "table_distribution.is_valid ERROR: Incorrect log probability for table %u.\n", i);
return false;
}
V diff = fabs(likelihoods.probabilities[i] * h.table_sizes[i] - probabilities[i]);
if (diff > probabilities[i] * 1.0e-12 && diff > 1.0e-12) {
fprintf(stderr, "table_distribution.is_valid ERROR: Incorrect probability for table %u.\n", i);
return false;
}
actual_sum += probabilities[i];
}
if (fabs(actual_sum - sum) > sum * 1.0e-6 && fabs(actual_sum - sum) > 1.0e-6) {
fprintf(stderr, "table_distribution.is_valid ERROR: Incorrect probability sum.\n");
return false;
}
return true;
}
void move_table(unsigned int src, unsigned int dst) {
probabilities[dst] = probabilities[src];
likelihoods.log_probabilities[dst] = likelihoods.log_probabilities[src];
likelihoods.probabilities[dst] = likelihoods.probabilities[src];
}
static inline void move(const table_distribution<V>& src, table_distribution<V>& dst) {
core::move(src.likelihoods, dst.likelihoods);
dst.probabilities = src.probabilities;
dst.sum = src.sum;
}
static inline void swap(table_distribution<V>& first, table_distribution<V>& second) {
core::swap(first.likelihoods, second.likelihoods);
core::swap(first.probabilities, second.probabilities);
core::swap(first.sum, second.sum);
}
static inline long unsigned int size_of(const table_distribution<V>& distribution, unsigned int table_count) {
return core::size_of(distribution.likelihoods, table_count)
+ core::size_of(distribution.sum) + sizeof(V) * table_count;
}
static inline void free(table_distribution<V>& distribution) {
core::free(distribution.probabilities);
core::free(distribution.likelihoods);
}
private:
template<typename BaseDistribution, typename DataDistribution, typename K>
inline void renormalize(
const hdp_sampler<BaseDistribution, DataDistribution, K, V>& h)
{
likelihoods.renormalize(h.table_count);
sum = 0.0;
for (unsigned int i = 0; i < h.table_count; i++) {
probabilities[i] = h.table_sizes[i] * likelihoods.probabilities[i];
sum += probabilities[i];
}
}
};
template<typename V>
inline bool init(table_distribution<V>& distribution, unsigned int table_count) {
distribution.probabilities = (V*) malloc(sizeof(V) * table_count);
if (distribution.probabilities == NULL) {
fprintf(stderr, "init ERROR: Unable to initialize probability array in table_distribution.\n");
return false;
} else if (!init(distribution.likelihoods, table_count)) {
fprintf(stderr, "init ERROR: Unable to initialize array_categorical structure in table_distribution.\n");
core::free(distribution.probabilities);
return false;
}
return true;
}
/**
* A specialization of ::cache where the `BaseDistribution` satisfies
* is_dirichlet and the `DataDistribution` is dense_categorical.
*/
template<typename BaseDistribution, typename V>
struct cache<BaseDistribution, dense_categorical<V>, unsigned int, V,
typename std::enable_if<is_dirichlet<BaseDistribution>::value>::type>
{
typedef hdp_sampler<BaseDistribution, dense_categorical<V>, unsigned int, V> sampler_root;
struct root_distribution {
table_distribution<V>& distribution;
const V* log_rising_factorials;
root_distribution(table_distribution<V>& distribution, const V* log_rising_factorials)
: distribution(distribution), log_rising_factorials(log_rising_factorials) { }
inline V max() const {
return distribution.likelihoods.maximum;
}
inline V sum() const {
return distribution.sum;
}
inline V likelihood(unsigned int index) const {
return distribution.likelihoods.probabilities[index];
}
inline const V* probabilities() const {
return distribution.probabilities;
}
inline void free() { }
};
/* a map from sizes of observation sets to table_distributions */
array_map<unsigned int, table_distribution<V>> distributions;
/* multiset of table sizes over *all tables* in the CRF */
array_multiset<unsigned int> table_histogram;
/* a map from sizes of observation sets to multiset where the keys are root
table identifiers and the values are the number of those observations in
the descendants of that table */
hash_map<unsigned int, array_multiset<unsigned int>> table_counts;
rising_factorial_cache<V> factorial_cache;
/**
* Constructs a cache structure for the HDP sampler hierarchy rooted at `h`.
*/
cache(const sampler_root& h) :
distributions(16), table_histogram(16), table_counts(128), factorial_cache(h.pi().sum()) { }
~cache() { free(); }
bool get_root_probabilities(
const sampler_root& h, table_distribution<V>& distribution,
unsigned int observation_count, unsigned int largest_table_size)
{
if (!init(distribution, h.table_count)) {
fprintf(stderr, "cache.get_root_probabilities"
" ERROR: Unable to initialize new table_distribution.\n");
return false;
}
distribution.sum = 0.0;
const V* log_rising_factorials = factorial_cache.get_rising_factorials(observation_count, largest_table_size);
for (unsigned int i = 0; i < h.table_count; i++)
distribution.likelihoods.place(i, -log_rising_factorials[h.descendant_observations[i].sum]);
distribution.likelihoods.renormalize(h.table_count);
for (unsigned int i = 0; i < h.table_count; i++) {
distribution.probabilities[i] = distribution.likelihoods.probabilities[i] * h.table_sizes[i];
distribution.sum += distribution.probabilities[i];
}
return true;
}
table_distribution<V>& get_root_probabilities(
const sampler_root& h,
unsigned int observation_count)
{
if (!distributions.ensure_capacity((unsigned int) distributions.size + 1)) {
fprintf(stderr, "cache.get_root_probabilities ERROR: Unable to expand array map.\n");
exit(EXIT_FAILURE);
}
bool contains;
table_distribution<V>& distribution = distributions.get(observation_count, contains);
if (contains)
return distribution;
get_root_probabilities(h, distribution, observation_count, largest_table());
distributions.keys[distributions.size] = observation_count;
distributions.size++;
return distribution;
}
inline void increase_descendant_count(unsigned int old_count, unsigned int new_count) {
unsigned int index = 0;
if (old_count > 0)
index = table_histogram.subtract(old_count) + 1;
table_histogram.add(new_count, 1, index);
}
inline void decrease_descendant_count(unsigned int old_count, unsigned int new_count)
{
unsigned int index;
if (new_count > 0)
index = table_histogram.add(new_count);
else index = 0;
table_histogram.subtract(old_count, index);
}
template<bool Check>
inline void increase_descendant_count(
const sampler_root& h, unsigned int table,
unsigned int old_count, unsigned int new_count)
{
for (unsigned int i = 0; i < distributions.size; i++)
distributions.values[i].template change_descendant_count<Check>(h, table, new_count, distributions.keys[i]);
log_cache<V>::instance().ensure_size(new_count);
increase_descendant_count(old_count, new_count);
}
template<bool Check>
inline void decrease_descendant_count(
const sampler_root& h, unsigned int table,
unsigned int old_count, unsigned int new_count)
{
for (unsigned int i = 0; i < distributions.size; i++)
distributions.values[i].template change_descendant_count<Check>(h, table, new_count, distributions.keys[i]);
decrease_descendant_count(old_count, new_count);
}
inline bool add_new_table(unsigned int count) {
if (count > 0)
return table_histogram.add(count);
else return true;
}
inline bool add_new_table(const sampler_root& h, unsigned int count)
{
for (unsigned int i = 0; i < distributions.size; i++)
if (!distributions.values[i].add_new_table(h, count, distributions.keys[i]))
return false;
log_cache<V>::instance().ensure_size(count);
return add_new_table(count);
}
bool is_valid(const sampler_root& h) const {
for (unsigned int i = 0; i < distributions.size; i++) {
if (!distributions.values[i].is_valid(h, distributions.keys[i])) {
fprintf(stderr, "cache.is_valid ERROR: Cache check"
" failed for observation count %u.\n", distributions.keys[i]);
return false;
}
}
if (table_histogram.counts.contains(0)) {
fprintf(stderr, "cache.is_valid ERROR: Table histogram contains the key '1'.");
return false;
}
array_multiset<unsigned int> actual_table_counts =
array_multiset<unsigned int>((unsigned int) table_histogram.counts.size);
compute_table_histogram(actual_table_counts, h);
if (actual_table_counts != table_histogram) {
fprintf(stderr, "cache.is_valid ERROR: Table histogram is incorrect.\n");
return false;
}
if (!factorial_cache.is_valid()) {
fprintf(stderr, "cache.is_valid ERROR: Rising factorial cache is incorrect.\n");
return false;
}
return true;
}
static bool compute_table_histogram(
array_multiset<unsigned int>& histogram,
const node_sampler<unsigned int, V>& n)
{
for (unsigned int i = 0; i < n.child_count(); i++)
if (!compute_table_histogram(histogram, n.children[i]))
return false;
for (unsigned int i = 0; i < n.table_count; i++)
if (n.descendant_observations[i].sum > 0)
histogram.add(n.descendant_observations[i].sum);
return histogram.add(1, n.observation_count());
}
static bool compute_table_histogram(
array_multiset<unsigned int>& histogram,
const sampler_root& h)
{
for (unsigned int i = 0; i < h.child_count(); i++)
if (!compute_table_histogram(histogram, h.children[i]))
return false;
for (unsigned int i = 0; i < h.table_count; i++)
if (h.descendant_observations[i].sum > 0)
histogram.add(h.descendant_observations[i].sum);
histogram.add(1, h.observation_count());
return true;
}
inline bool compute_table_histogram(const sampler_root& h) {
table_histogram.clear();
if (!compute_table_histogram(table_histogram, h))
return false;
log_cache<V>::instance().ensure_size(largest_table());
return true;
}
inline unsigned int largest_table() const {
return table_histogram.counts.keys[table_histogram.counts.size - 1];
}
void clear_cache() {
for (unsigned int i = 0; i < table_histogram.counts.size; i++) {
if (table_histogram.counts.values[i] == 0) {
unsigned int index = distributions.index_of(table_histogram.counts.keys[i]);
if (index < distributions.size) {
core::free(distributions.values[index]);
distributions.remove_at(index);
}
}
}
table_histogram.remove_zeros();
}
bool add_to_cache(const sampler_root& h)
{
for (unsigned int i = 0; i < table_counts.table.capacity; i++)
if (!is_empty(table_counts.table.keys[i]))
core::free(table_counts.values[i]);
table_counts.clear();
for (unsigned int i = 0; i < h.table_count; i++) {
const array_multiset<unsigned int>& observations = h.descendant_observations[i];
for (unsigned int j = 0; j < observations.counts.size; j++) {
unsigned int atom = observations.counts.keys[j];
bool contains;
unsigned int bucket;
if (!table_counts.check_size())
return false;
array_multiset<unsigned int>& histogram = table_counts.get(atom, contains, bucket);
if (!contains) {
if (!init(histogram, 4)) return false;
table_counts.table.keys[bucket] = atom;
table_counts.table.size++;
}
if (!histogram.add_unsorted(i, observations.counts.values[j]))
return false;
}
}
for (unsigned int i = 0; i < table_counts.table.capacity; i++) {
if (!is_empty(table_counts.table.keys[i])) {
if (table_counts.values[i].counts.size > 1) {
sort(table_counts.values[i].counts.keys,
table_counts.values[i].counts.values,
(unsigned int) table_counts.values[i].counts.size);
}
}
}
return true;
}
inline array_multiset<unsigned int>& get_table_counts(unsigned int atom, bool& contains) {
return table_counts.get(atom, contains);
}
inline bool move_to_table(unsigned int atom,
unsigned int src_table, unsigned int dst_table)
{
if (!move_to_table_helper(atom, 1, src_table, dst_table))
return false;
return true;
}
inline bool move_to_table(
const array_multiset<unsigned int>& observations,
unsigned int src_table, unsigned int dst_table)
{
for (unsigned int i = 0; i < observations.counts.size; i++)
if (!move_to_table_helper(observations.counts.keys[i],
observations.counts.values[i], src_table, dst_table))
return false;
return true;
}
inline bool add_to_table(unsigned int atom, unsigned int table)
{
if (!add_to_table_helper(atom, 1, table))
return false;
return true;
}
inline bool move_to_table(
const array_multiset<unsigned int>& observations, unsigned int table)
{
for (unsigned int i = 0; i < observations.counts.size; i++)
if (!add_to_table_helper(observations.counts.keys[i],
observations.counts.values[i], table))
return false;
return true;
}
inline bool remove_from_table(unsigned int atom, unsigned int table)
{
if (!remove_from_table_helper(atom, 1, table))
return false;
return true;
}
inline bool remove_from_table(
const array_multiset<unsigned int>& observations, unsigned int table)
{
for (unsigned int i = 0; i < observations.counts.size; i++)
if (!remove_from_table_helper(observations.counts.keys[i], observations.counts.values[i], table))
return false;
return true;
}
inline void remove_zeros() {
for (unsigned int i = 0; i < table_counts.table.capacity; i++)
if (!is_empty(table_counts.table.keys[i]))
table_counts.values[i].remove_zeros();
}
inline void relabel_tables(const unsigned int* table_map) {
remove_zeros();
for (unsigned int i = 0; i < table_counts.table.capacity; i++) {
if (is_empty(table_counts.table.keys[i]))
continue;
array_map<unsigned int, unsigned int>& map = table_counts.values[i].counts;
for (unsigned int j = 0; j < map.size; j++)
map.keys[j] = table_map[map.keys[j]];
if (map.size > 1)
sort(map.keys, map.values, (unsigned int) map.size);
}
}
inline void on_change_table_size(unsigned int table, unsigned int new_table_size)
{
for (unsigned int i = 0; i < distributions.size; i++)
distributions.values[i].change_table_size(table, new_table_size);
}
inline void on_finished_sampling_hdp() {
remove_zeros();
clear_cache();
}
inline void prepare_sampler(const sampler_root& h) {
compute_table_histogram(h);
get_root_probabilities(h, 1); /* precompute the cache for singleton observations */
add_to_cache(h);
}
inline void on_move_table(unsigned int src, unsigned int dst) {
for (unsigned int i = 0; i < distributions.size; i++)
distributions.values[i].move_table(src, dst);
}
template<typename Observations>
inline void on_move_to_table(
const node_sampler<unsigned int, V>& n,
unsigned int src, unsigned int dst,
const Observations& observations)
{
decrease_descendant_count(n.descendant_observations[src].sum + count(observations), n.descendant_observations[src].sum);
increase_descendant_count(n.descendant_observations[dst].sum - count(observations), n.descendant_observations[dst].sum);
}
template<typename Observations>
inline bool on_move_to_table(
const sampler_root& h,
unsigned int src, unsigned int dst,
const Observations& observations)
{
decrease_descendant_count<false>(h, src,
h.descendant_observations[src].sum + count(observations), h.descendant_observations[src].sum);
increase_descendant_count<true>(h, dst,
h.descendant_observations[dst].sum - count(observations), h.descendant_observations[dst].sum);
return move_to_table(observations, src, dst);
}
template<typename Observations>
inline void on_move_to_new_table(
const node_sampler<unsigned int, V>& n,
unsigned int src, const Observations& observations)
{
decrease_descendant_count(n.descendant_observations[src].sum
+ count(observations), n.descendant_observations[src].sum);
add_new_table(n.descendant_observations[n.table_count - 1].sum);
}
template<typename Observations>
inline bool on_move_to_new_table(
const sampler_root& h,
unsigned int src, const Observations& observations)
{
decrease_descendant_count<false>(h, src,
h.descendant_observations[src].sum + count(observations), h.descendant_observations[src].sum);
add_new_table(h, h.descendant_observations[h.table_count - 1].sum);
return move_to_table(observations, src, h.table_count - 1);
}
template<typename Observations>
inline void on_add_to_table(
const node_sampler<unsigned int, V>& n,
unsigned int table, const Observations& observations)
{
increase_descendant_count(n.descendant_observations[table].sum - count(observations), n.descendant_observations[table].sum);
}
template<typename Observations>
inline bool on_add_to_table(
const sampler_root& h,
unsigned int table,
const Observations& observations)
{
increase_descendant_count<true>(h, table,
h.descendant_observations[table].sum - count(observations), h.descendant_observations[table].sum);
return add_to_table(observations, table);
}
template<typename Observations>
inline void on_add_to_new_table(
const node_sampler<unsigned int, V>& n,
const Observations& observations)
{
add_new_table(n.descendant_observations[n.table_count - 1].sum);
}
template<typename Observations>
inline bool on_add_to_new_table(
const sampler_root& h,
const Observations& observations)
{
add_new_table(h, h.descendant_observations[h.table_count - 1].sum);
return add_to_table(observations, h.table_count - 1);
}
template<typename Observations>
inline void on_remove_from_table(
const node_sampler<unsigned int, V>& n,
unsigned int table, const Observations& observations)
{
decrease_descendant_count(n.descendant_observations[table].sum + count(observations), n.descendant_observations[table].sum);
}
template<typename Observations>
inline bool on_remove_from_table(
const sampler_root& h,
unsigned int table,
const Observations& observations)
{
decrease_descendant_count<false>(h, table,
h.descendant_observations[table].sum + count(observations), h.descendant_observations[table].sum);
return remove_from_table(observations, table);
}
template<typename Observations>
root_distribution compute_root_distribution(
const sampler_root& root,
const Observations& observations)
{
table_distribution<V>& distribution = get_root_probabilities(root, count(observations));
const V* log_rising_factorials =
factorial_cache.get_rising_factorials(count(observations), largest_table());
compute_likelihoods(root, distribution, root.table_count, observations);
return root_distribution(distribution, log_rising_factorials);
}
template<typename Observations>
root_distribution compute_root_distribution(
const sampler_root& root,
const Observations& observations,
unsigned int old_root_assignment)
{
table_distribution<V>& distribution = get_root_probabilities(root, count(observations));
const V* log_rising_factorials =
factorial_cache.get_rising_factorials(count(observations), largest_table());
V new_likelihood = -log_rising_factorials[root.descendant_observations[old_root_assignment].sum - count(observations)];
distribution.set_higher(root, old_root_assignment, new_likelihood);
compute_likelihoods(root, distribution, old_root_assignment, observations);
return root_distribution(distribution, log_rising_factorials);
}
void compute_likelihoods(const sampler_root& h,
table_distribution<V>& root_distribution,
unsigned int old_root_assignment,
unsigned int observation, unsigned int count = 1)
{
bool contains;
array_multiset<unsigned int>& table_counts = get_table_counts(observation, contains);
if (!contains) return;
for (unsigned int i = 0; i < table_counts.counts.size; i++) {
unsigned int table = table_counts.counts.keys[i];
V base = h.pi().get_for_atom(observation) + table_counts.counts.values[i];
if (table == old_root_assignment)
base -= count;
V contribution = log_rising_factorial(base, count);
if (contribution > 0) {
root_distribution.set_higher(
h, table, root_distribution.likelihoods.log_probabilities[table] + contribution);
} else {
root_distribution.set_lower(
h, table, root_distribution.likelihoods.log_probabilities[table] + contribution);
}
}
}
inline void compute_likelihoods(const sampler_root& h,
table_distribution<V>& root_distribution,
unsigned int old_root_assignment,
const array_multiset<unsigned int>& observations)
{
for (unsigned int i = 0; i < observations.counts.size; i++)
compute_likelihoods(h, root_distribution, old_root_assignment,
observations.counts.keys[i], observations.counts.values[i]);
}
void on_sample_table(sampler_root& h,
root_distribution& root_probabilities,
unsigned int old_root_assignment,
unsigned int observation, unsigned int count = 1)
{
bool contains;
array_multiset<unsigned int>& table_counts = get_table_counts(observation, contains);
if (!contains) return;
for (unsigned int i = 0; i < table_counts.counts.size; i++) {
unsigned int table = table_counts.counts.keys[i];
V base = h.pi().get_for_atom(observation) + table_counts.counts.values[i];
if (table == old_root_assignment)
base -= count;
V contribution = log_rising_factorial(base, count);
if (contribution > 0) {
root_probabilities.distribution.set_lower(
h, table, root_probabilities.distribution.likelihoods.log_probabilities[table] - contribution);
} else {
root_probabilities.distribution.set_higher(
h, table, root_probabilities.distribution.likelihoods.log_probabilities[table] - contribution);
}
}
}
inline void on_sample_table(sampler_root& h,
root_distribution& root_probabilities,
unsigned int old_root_assignment,
const array_multiset<unsigned int>& observations)
{
for (unsigned int i = 0; i < observations.counts.size; i++)
on_sample_table(h, root_probabilities, old_root_assignment,
observations.counts.keys[i], observations.counts.values[i]);
}
template<typename Observations>
inline void on_add(const Observations& observations)
{