Repository navigation
Expand file tree
/
Copy pathfastcluster_python.cpp
More file actions
1248 lines (1123 loc) · 35.2 KB
/
Copy pathfastcluster_python.cpp
File metadata and controls
1248 lines (1123 loc) · 35.2 KB
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
/*
fastcluster: Fast hierarchical clustering routines for R and Python
Copyright © 2011 Daniel Müllner
<http://danifold.net>
*/
// for INT32_MAX in fastcluster.cpp
// This must be defined here since Python.h loads the header file pyport.h,
// and from this stdint.h. INT32_MAX is defined in stdint.h, but only if
// __STDC_LIMIT_MACROS is defined.
#define __STDC_LIMIT_MACROS
#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
#if __GNUC__ > 4 || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 6))
#define HAVE_DIAGNOSTIC 1
#endif
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wswitch-default"
#pragma GCC diagnostic ignored "-Wpadded"
#pragma GCC diagnostic ignored "-Wlong-long"
#pragma GCC diagnostic ignored "-Wformat"
#endif
#include <Python.h>
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wlong-long"
#pragma GCC diagnostic ignored "-Wpedantic"
#pragma GCC diagnostic ignored "-Wpadded"
#pragma GCC diagnostic ignored "-Wcast-qual"
#endif
#include <numpy/arrayobject.h>
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
/* It's complicated, but if I do not include the C++ math headers, GCC
will complain about conversions from 'double' to 'float', whenever 'isnan'
is called in a templated function (but not outside templates).
The '#include <cmath>' seems to cure the problem.
*/
//#include <cmath>
#define fc_isnan(X) ((X)!=(X))
// There is Py_IS_NAN but it is so much slower on my x86_64 system with GCC!
#include <cstddef> // for std::ptrdiff_t
#include <limits> // for std::numeric_limits<...>::infinity()
#include <algorithm> // for std::stable_sort
#include <new> // for std::bad_alloc
#include <exception> // for std::exception
#include "fastcluster.cpp"
// backwards compatibility
#ifndef NPY_ARRAY_CARRAY_RO
#define NPY_ARRAY_CARRAY_RO NPY_CARRAY_RO
#endif
/* Since the public interface is given by the Python respectively R interface,
* we do not want other symbols than the interface initalization routines to be
* visible in the shared object file. The "visibility" switch is a GCC concept.
* Hiding symbols keeps the relocation table small and decreases startup time.
* See http://gcc.gnu.org/wiki/Visibility
*/
#if HAVE_VISIBILITY
#pragma GCC visibility push(hidden)
#endif
/*
Convenience class for the output array: automatic counter.
*/
class linkage_output {
private:
t_float * Z;
public:
linkage_output(t_float * const Z_)
: Z(Z_)
{}
void append(const t_index node1, const t_index node2, const t_float dist,
const t_float size) {
if (node1<node2) {
*(Z++) = static_cast<t_float>(node1);
*(Z++) = static_cast<t_float>(node2);
}
else {
*(Z++) = static_cast<t_float>(node2);
*(Z++) = static_cast<t_float>(node1);
}
*(Z++) = dist;
*(Z++) = size;
}
};
/*
Generate the SciPy-specific output format for a dendrogram from the
clustering output.
The list of merging steps can be sorted or unsorted.
*/
// The size of a node is either 1 (a single point) or is looked up from
// one of the clusters.
#define size_(r_) ( ((r_<N) ? 1 : Z_(r_-N,3)) )
template <const bool sorted>
static void generate_SciPy_dendrogram(t_float * const Z, cluster_result & Z2, const t_index N) {
// The array "nodes" is a union-find data structure for the cluster
// identities (only needed for unsorted cluster_result input).
union_find nodes(sorted ? 0 : N);
if (!sorted) {
std::stable_sort(Z2[0], Z2[N-1]);
}
linkage_output output(Z);
t_index node1, node2;
for (node const * NN=Z2[0]; NN!=Z2[N-1]; ++NN) {
// Get two data points whose clusters are merged in step i.
if (sorted) {
node1 = NN->node1;
node2 = NN->node2;
}
else {
// Find the cluster identifiers for these points.
node1 = nodes.Find(NN->node1);
node2 = nodes.Find(NN->node2);
// Merge the nodes in the union-find data structure by making them
// children of a new node.
nodes.Union(node1, node2);
}
output.append(node1, node2, NN->dist, size_(node1)+size_(node2));
}
}
/*
Python interface code
*/
static PyObject * linkage_wrap(PyObject * const self, PyObject * const args);
static PyObject * linkage_vector_wrap(PyObject * const self, PyObject * const args);
// List the C++ methods that this extension provides.
static PyMethodDef _fastclusterWrapMethods[] = {
{"linkage_wrap", linkage_wrap, METH_VARARGS, NULL},
{"linkage_vector_wrap", linkage_vector_wrap, METH_VARARGS, NULL},
{NULL, NULL, 0, NULL} /* Sentinel - marks the end of this structure */
};
/* Tell Python about these methods.
Python 2.x and 3.x differ in their C APIs for this part.
*/
#if PY_VERSION_HEX >= 0x03000000
static struct PyModuleDef fastclustermodule = {
PyModuleDef_HEAD_INIT,
"_fastcluster",
NULL, // no module documentation
-1, /* size of per-interpreter state of the module,
or -1 if the module keeps state in global variables. */
_fastclusterWrapMethods,
NULL, NULL, NULL, NULL
};
/* Make the interface initalization routines visible in the shared object
* file.
*/
#if HAVE_VISIBILITY
#pragma GCC visibility push(default)
#endif
PyMODINIT_FUNC PyInit__fastcluster(void) {
PyObject * m;
m = PyModule_Create(&fastclustermodule);
if (!m) {
return NULL;
}
import_array(); // Must be present for NumPy. Called first after above line.
return m;
}
#if HAVE_VISIBILITY
#pragma GCC visibility pop
#endif
# else // Python 2.x
#if HAVE_VISIBILITY
#pragma GCC visibility push(default)
#endif
PyMODINIT_FUNC init_fastcluster(void) {
(void) Py_InitModule("_fastcluster", _fastclusterWrapMethods);
import_array(); // Must be present for NumPy. Called first after above line.
}
#if HAVE_VISIBILITY
#pragma GCC visibility pop
#endif
#endif // PY_VERSION
class GIL_release
{
private:
// noncopyable
GIL_release(GIL_release const &);
GIL_release & operator=(GIL_release const &);
public:
inline
GIL_release(bool really = true)
: _save(really ? PyEval_SaveThread() : NULL)
{
}
inline
~GIL_release()
{
if (_save)
PyEval_RestoreThread(_save);
}
private:
PyThreadState * _save;
};
/*
Interface to Python, part 1:
The input is a dissimilarity matrix.
*/
static PyObject *linkage_wrap(PyObject * const, PyObject * const args) {
PyArrayObject * D, * Z;
long int N_ = 0;
unsigned char method;
try{
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wold-style-cast"
#endif
// Parse the input arguments
if (!PyArg_ParseTuple(args, "lO!O!b",
&N_, // signed long integer
&PyArray_Type, &D, // NumPy array
&PyArray_Type, &Z, // NumPy array
&method)) { // unsigned char
return NULL; // Error if the arguments have the wrong type.
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
if (N_ < 1 ) {
// N must be at least 1.
PyErr_SetString(PyExc_ValueError,
"At least one element is needed for clustering.");
return NULL;
}
/*
(1)
The biggest index used below is 4*(N-2)+3, as an index to Z. This must
fit into the data type used for indices.
(2)
The largest representable integer, without loss of precision, by a
floating point number of type t_float is 2^T_FLOAT_MANT_DIG. Here, we
make sure that all cluster labels from 0 to 2N-2 in the output can be
accurately represented by a floating point number.
Conversion of N to 64 bits below is not really necessary but it prevents
a warning ("shift count >= width of type") on systems where "long int"
is 32 bits wide.
*/
if (N_ > MAX_INDEX/4 ||
static_cast<int64_t>(N_-1)>>(T_FLOAT_MANT_DIG-1) > 0) {
PyErr_SetString(PyExc_ValueError,
"Data is too big, index overflow.");
return NULL;
}
t_index N = static_cast<t_index>(N_);
// Allow threads!
GIL_release G;
t_float * const D_ = reinterpret_cast<t_float *>(PyArray_DATA(D));
cluster_result Z2(N-1);
auto_array_ptr<t_index> members;
// For these methods, the distance update formula needs the number of
// data points in a cluster.
if (method==METHOD_METR_AVERAGE ||
method==METHOD_METR_WARD ||
method==METHOD_METR_CENTROID) {
members.init(N, 1);
}
// Operate on squared distances for these methods.
if (method==METHOD_METR_WARD ||
method==METHOD_METR_CENTROID ||
method==METHOD_METR_MEDIAN) {
for (t_float * DD = D_; DD!=D_+static_cast<std::ptrdiff_t>(N)*(N-1)/2;
++DD)
*DD *= *DD;
}
switch (method) {
case METHOD_METR_SINGLE:
MST_linkage_core(N, D_, Z2);
break;
case METHOD_METR_COMPLETE:
NN_chain_core<METHOD_METR_COMPLETE, t_index>(N, D_, NULL, Z2);
break;
case METHOD_METR_AVERAGE:
NN_chain_core<METHOD_METR_AVERAGE, t_index>(N, D_, members, Z2);
break;
case METHOD_METR_WEIGHTED:
NN_chain_core<METHOD_METR_WEIGHTED, t_index>(N, D_, NULL, Z2);
break;
case METHOD_METR_WARD:
NN_chain_core<METHOD_METR_WARD, t_index>(N, D_, members, Z2);
break;
case METHOD_METR_CENTROID:
generic_linkage<METHOD_METR_CENTROID, t_index>(N, D_, members, Z2);
break;
case METHOD_METR_MEDIAN:
generic_linkage<METHOD_METR_MEDIAN, t_index>(N, D_, NULL, Z2);
break;
default:
throw std::runtime_error(std::string("Invalid method index."));
}
if (method==METHOD_METR_WARD ||
method==METHOD_METR_CENTROID ||
method==METHOD_METR_MEDIAN) {
Z2.sqrt();
}
t_float * const Z_ = reinterpret_cast<t_float *>(PyArray_DATA(Z));
if (method==METHOD_METR_CENTROID ||
method==METHOD_METR_MEDIAN) {
generate_SciPy_dendrogram<true>(Z_, Z2, N);
}
else {
generate_SciPy_dendrogram<false>(Z_, Z2, N);
}
} // try
catch (const std::bad_alloc&) {
return PyErr_NoMemory();
}
catch(const std::exception& e){
PyErr_SetString(PyExc_EnvironmentError, e.what());
return NULL;
}
catch(const nan_error&){
PyErr_SetString(PyExc_FloatingPointError, "NaN dissimilarity value.");
return NULL;
}
#ifdef FE_INVALID
catch(const fenv_error&){
PyErr_SetString(PyExc_FloatingPointError,
"NaN dissimilarity value in intermediate results.");
return NULL;
}
#endif
catch(...){
PyErr_SetString(PyExc_EnvironmentError,
"C++ exception (unknown reason). Please send a bug report.");
return NULL;
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wold-style-cast"
#endif
Py_RETURN_NONE;
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
}
/*
Part 2: Clustering on vector data
*/
/* Metric codes.
These codes must agree with the dictionary mtridx in fastcluster.py.
*/
enum metric_codes {
// metrics
METRIC_EUCLIDEAN = 0,
METRIC_MINKOWSKI = 1,
METRIC_CITYBLOCK = 2,
METRIC_SEUCLIDEAN = 3,
METRIC_SQEUCLIDEAN = 4,
METRIC_COSINE = 5,
METRIC_HAMMING = 6,
METRIC_JACCARD = 7,
METRIC_CHEBYCHEV = 8,
METRIC_CANBERRA = 9,
METRIC_BRAYCURTIS = 10,
METRIC_MAHALANOBIS = 11,
METRIC_YULE = 12,
METRIC_MATCHING = 13,
METRIC_DICE = 14,
METRIC_ROGERSTANIMOTO = 15,
METRIC_RUSSELLRAO = 16,
METRIC_SOKALSNEATH = 17,
METRIC_KULSINSKI = 18,
METRIC_USER = 19,
METRIC_INVALID = 20, // sentinel
METRIC_JACCARD_BOOL = 21, // separate function for Jaccard metric on
}; // Boolean input data
/*
Helper class: Throw this if calling the Python interpreter from within
C returned an error.
*/
class pythonerror {};
/*
This class handles all the information about the dissimilarity
computation.
*/
class python_dissimilarity {
private:
t_float * Xa;
std::ptrdiff_t dim; // size_t saves many statis_cast<> in products
t_index N;
auto_array_ptr<t_float> Xnew;
t_index * members;
void (cluster_result::*postprocessfn) (const t_float) const;
t_float postprocessarg;
t_float (python_dissimilarity::*distfn) (const t_index, const t_index) const;
// for user-defined metrics
PyObject * X_Python;
PyObject * userfn;
auto_array_ptr<t_float> precomputed;
t_float * precomputed2;
PyArrayObject * V;
const t_float * V_data;
// noncopyable
python_dissimilarity();
python_dissimilarity(python_dissimilarity const &);
python_dissimilarity & operator=(python_dissimilarity const &);
public:
// Ignore warning about uninitialized member variables. I know what I am
// doing here, and some member variables are only used for certain metrics.
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Weffc++"
#endif
python_dissimilarity (PyArrayObject * const Xarg,
t_index * const members_,
const method_codes method,
const metric_codes metric,
PyObject * const extraarg,
bool temp_point_array)
: Xa(reinterpret_cast<t_float *>(PyArray_DATA(Xarg))),
dim(PyArray_DIM(Xarg, 1)),
N(static_cast<t_index>(PyArray_DIM(Xarg, 0))),
Xnew(temp_point_array ? (N-1)*dim : 0),
members(members_),
postprocessfn(NULL),
V(NULL)
{
switch (method) {
case METHOD_METR_SINGLE:
postprocessfn = NULL; // default
switch (metric) {
case METRIC_EUCLIDEAN:
set_euclidean();
break;
case METRIC_SEUCLIDEAN:
if (extraarg==NULL) {
PyErr_SetString(PyExc_TypeError,
"The 'seuclidean' metric needs a variance parameter.");
throw pythonerror();
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wold-style-cast"
#endif
V = reinterpret_cast<PyArrayObject *>(PyArray_FromAny(extraarg,
PyArray_DescrFromType(NPY_DOUBLE),
1, 1,
NPY_ARRAY_CARRAY_RO,
NULL));
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
if (PyErr_Occurred()) {
throw pythonerror();
}
if (PyArray_DIM(V, 0)!=dim) {
PyErr_SetString(PyExc_ValueError,
"The variance vector must have the same dimensionality as the data.");
throw pythonerror();
}
V_data = reinterpret_cast<t_float *>(PyArray_DATA(V));
distfn = &python_dissimilarity::seuclidean;
postprocessfn = &cluster_result::sqrt;
break;
case METRIC_SQEUCLIDEAN:
distfn = &python_dissimilarity::sqeuclidean<false>;
break;
case METRIC_CITYBLOCK:
set_cityblock();
break;
case METRIC_CHEBYCHEV:
set_chebychev();
break;
case METRIC_MINKOWSKI:
set_minkowski(extraarg);
break;
case METRIC_COSINE:
distfn = &python_dissimilarity::cosine;
postprocessfn = &cluster_result::plusone;
// precompute norms
precomputed.init(N);
for (t_index i=0; i<N; ++i) {
t_float sum=0;
for (t_index k=0; k<dim; ++k) {
sum += X(i,k)*X(i,k);
}
precomputed[i] = 1/sqrt(sum);
}
break;
case METRIC_HAMMING:
distfn = &python_dissimilarity::hamming;
postprocessfn = &cluster_result::divide;
postprocessarg = static_cast<t_float>(dim);
break;
case METRIC_JACCARD:
distfn = &python_dissimilarity::jaccard;
break;
case METRIC_CANBERRA:
distfn = &python_dissimilarity::canberra;
break;
case METRIC_BRAYCURTIS:
distfn = &python_dissimilarity::braycurtis;
break;
case METRIC_MAHALANOBIS:
if (extraarg==NULL) {
PyErr_SetString(PyExc_TypeError,
"The 'mahalanobis' metric needs a parameter for the inverse covariance.");
throw pythonerror();
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wold-style-cast"
#endif
V = reinterpret_cast<PyArrayObject *>(PyArray_FromAny(extraarg,
PyArray_DescrFromType(NPY_DOUBLE),
2, 2,
NPY_ARRAY_CARRAY_RO,
NULL));
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
if (PyErr_Occurred()) {
throw pythonerror();
}
if (PyArray_DIM(V, 0)!=N || PyArray_DIM(V, 1)!=dim) {
PyErr_SetString(PyExc_ValueError,
"The inverse covariance matrix has the wrong size.");
throw pythonerror();
}
V_data = reinterpret_cast<t_float *>(PyArray_DATA(V));
distfn = &python_dissimilarity::mahalanobis;
postprocessfn = &cluster_result::sqrt;
break;
case METRIC_YULE:
distfn = &python_dissimilarity::yule;
break;
case METRIC_MATCHING:
distfn = &python_dissimilarity::matching;
postprocessfn = &cluster_result::divide;
postprocessarg = static_cast<t_float>(dim);
break;
case METRIC_DICE:
distfn = &python_dissimilarity::dice;
break;
case METRIC_ROGERSTANIMOTO:
distfn = &python_dissimilarity::rogerstanimoto;
break;
case METRIC_RUSSELLRAO:
distfn = &python_dissimilarity::russellrao;
postprocessfn = &cluster_result::divide;
postprocessarg = static_cast<t_float>(dim);
break;
case METRIC_SOKALSNEATH:
distfn = &python_dissimilarity::sokalsneath;
break;
case METRIC_KULSINSKI:
distfn = &python_dissimilarity::kulsinski;
postprocessfn = &cluster_result::plusone;
precomputed.init(N);
for (t_index i=0; i<N; ++i) {
t_index sum=0;
for (t_index k=0; k<dim; ++k) {
sum += Xb(i,k);
}
precomputed[i] = -.5/static_cast<t_float>(sum);
}
break;
case METRIC_USER:
X_Python = reinterpret_cast<PyObject *>(Xarg);
this->userfn = extraarg;
distfn = &python_dissimilarity::user;
break;
default: // case METRIC_JACCARD_BOOL:
distfn = &python_dissimilarity::jaccard_bool;
}
break;
case METHOD_METR_WARD:
postprocessfn = &cluster_result::sqrtdouble;
break;
default:
postprocessfn = &cluster_result::sqrt;
}
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
~python_dissimilarity() {
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wold-style-cast"
#endif
Py_XDECREF(V);
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
}
inline t_float operator () (const t_index i, const t_index j) const {
return (this->*distfn)(i,j);
}
inline t_float X (const t_index i, const t_index j) const {
return Xa[i*dim+j];
}
inline bool Xb (const t_index i, const t_index j) const {
return reinterpret_cast<bool *>(Xa)[i*dim+j];
}
inline t_float * Xptr(const t_index i, const t_index j) const {
return Xa+i*dim+j;
}
void merge(const t_index i, const t_index j, const t_index newnode) const {
t_float const * const Pi = i<N ? Xa+i*dim : Xnew+(i-N)*dim;
t_float const * const Pj = j<N ? Xa+j*dim : Xnew+(j-N)*dim;
for(t_index k=0; k<dim; ++k) {
Xnew[(newnode-N)*dim+k] = (Pi[k]*static_cast<t_float>(members[i]) +
Pj[k]*static_cast<t_float>(members[j])) /
static_cast<t_float>(members[i]+members[j]);
}
members[newnode] = members[i]+members[j];
}
void merge_weighted(const t_index i, const t_index j, const t_index newnode)
const {
t_float const * const Pi = i<N ? Xa+i*dim : Xnew+(i-N)*dim;
t_float const * const Pj = j<N ? Xa+j*dim : Xnew+(j-N)*dim;
for(t_index k=0; k<dim; ++k) {
Xnew[(newnode-N)*dim+k] = (Pi[k]+Pj[k])*.5;
}
}
void merge_inplace(const t_index i, const t_index j) const {
t_float const * const Pi = Xa+i*dim;
t_float * const Pj = Xa+j*dim;
for(t_index k=0; k<dim; ++k) {
Pj[k] = (Pi[k]*static_cast<t_float>(members[i]) +
Pj[k]*static_cast<t_float>(members[j])) /
static_cast<t_float>(members[i]+members[j]);
}
members[j] += members[i];
}
void merge_inplace_weighted(const t_index i, const t_index j) const {
t_float const * const Pi = Xa+i*dim;
t_float * const Pj = Xa+j*dim;
for(t_index k=0; k<dim; ++k) {
Pj[k] = (Pi[k]+Pj[k])*.5;
}
}
void postprocess(cluster_result & Z2) const {
if (postprocessfn!=NULL) {
(Z2.*postprocessfn)(postprocessarg);
}
}
inline t_float ward(const t_index i, const t_index j) const {
t_float mi = static_cast<t_float>(members[i]);
t_float mj = static_cast<t_float>(members[j]);
return sqeuclidean<true>(i,j)*mi*mj/(mi+mj);
}
inline t_float ward_initial(const t_index i, const t_index j) const {
// alias for sqeuclidean
// Factor 2!!!
return sqeuclidean<true>(i,j);
}
// This method must not produce NaN if the input is non-NaN.
inline static t_float ward_initial_conversion(const t_float min) {
return min*.5;
}
inline t_float ward_extended(const t_index i, const t_index j) const {
t_float mi = static_cast<t_float>(members[i]);
t_float mj = static_cast<t_float>(members[j]);
return sqeuclidean_extended(i,j)*mi*mj/(mi+mj);
}
/* We need two variants of the Euclidean metric: one that does not check
for a NaN result, which is used for the initial distances, and one which
does, for the updated distances during the clustering procedure.
*/
template <const bool check_NaN>
t_float sqeuclidean(const t_index i, const t_index j) const {
t_float sum = 0;
/*
for (t_index k=0; k<dim; ++k) {
t_float diff = X(i,k) - X(j,k);
sum += diff*diff;
}
*/
// faster
t_float const * Pi = Xa+i*dim;
t_float const * Pj = Xa+j*dim;
for (t_index k=0; k<dim; ++k) {
t_float diff = Pi[k] - Pj[k];
sum += diff*diff;
}
if (check_NaN) {
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
if (fc_isnan(sum))
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
throw(nan_error());
}
return sum;
}
t_float sqeuclidean_extended(const t_index i, const t_index j) const {
t_float sum = 0;
t_float const * Pi = i<N ? Xa+i*dim : Xnew+(i-N)*dim; // TBD
t_float const * Pj = j<N ? Xa+j*dim : Xnew+(j-N)*dim;
for (t_index k=0; k<dim; ++k) {
t_float diff = Pi[k] - Pj[k];
sum += diff*diff;
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
if (fc_isnan(sum))
throw(nan_error());
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
return sum;
}
private:
void set_minkowski(PyObject * extraarg) {
if (extraarg==NULL) {
PyErr_SetString(PyExc_TypeError,
"The Minkowski metric needs a parameter.");
throw pythonerror();
}
postprocessarg = PyFloat_AsDouble(extraarg);
if (PyErr_Occurred()) {
throw pythonerror();
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
if (postprocessarg==std::numeric_limits<t_float>::infinity()) {
set_chebychev();
}
else if (postprocessarg==1.0){
set_cityblock();
}
else if (postprocessarg==2.0){
set_euclidean();
}
else {
distfn = &python_dissimilarity::minkowski;
postprocessfn = &cluster_result::power;
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
}
void set_euclidean() {
distfn = &python_dissimilarity::sqeuclidean<false>;
postprocessfn = &cluster_result::sqrt;
}
void set_cityblock() {
distfn = &python_dissimilarity::cityblock;
}
void set_chebychev() {
distfn = &python_dissimilarity::chebychev;
}
t_float seuclidean(const t_index i, const t_index j) const {
t_float sum = 0;
for (t_index k=0; k<dim; ++k) {
t_float diff = X(i,k)-X(j,k);
sum += diff*diff/V_data[k];
}
return sum;
}
t_float cityblock(const t_index i, const t_index j) const {
t_float sum = 0;
for (t_index k=0; k<dim; ++k) {
sum += fabs(X(i,k)-X(j,k));
}
return sum;
}
t_float minkowski(const t_index i, const t_index j) const {
t_float sum = 0;
for (t_index k=0; k<dim; ++k) {
sum += pow(fabs(X(i,k)-X(j,k)),postprocessarg);
}
return sum;
}
t_float chebychev(const t_index i, const t_index j) const {
t_float max = 0;
for (t_index k=0; k<dim; ++k) {
t_float diff = fabs(X(i,k)-X(j,k));
if (diff>max) {
max = diff;
}
}
return max;
}
t_float cosine(const t_index i, const t_index j) const {
t_float sum = 0;
for (t_index k=0; k<dim; ++k) {
sum -= X(i,k)*X(j,k);
}
return sum*precomputed[i]*precomputed[j];
}
t_float hamming(const t_index i, const t_index j) const {
t_float sum = 0;
for (t_index k=0; k<dim; ++k) {
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
sum += (X(i,k)!=X(j,k));
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
}
return sum;
}
// Differs from scipy.spatial.distance: equal vectors correctly
// return distance 0.
t_float jaccard(const t_index i, const t_index j) const {
t_index sum1 = 0;
t_index sum2 = 0;
for (t_index k=0; k<dim; ++k) {
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
sum1 += (X(i,k)!=X(j,k));
sum2 += ((X(i,k)!=0) || (X(j,k)!=0));
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
}
return sum1==0 ? 0 : static_cast<t_float>(sum1) / static_cast<t_float>(sum2);
}
t_float canberra(const t_index i, const t_index j) const {
t_float sum = 0;
for (t_index k=0; k<dim; ++k) {
t_float numerator = fabs(X(i,k)-X(j,k));
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
sum += numerator==0 ? 0 : numerator / (fabs(X(i,k)) + fabs(X(j,k)));
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
}
return sum;
}
t_float user(const t_index i, const t_index j) const {
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wold-style-cast"
#endif
PyObject * u = PySequence_ITEM(X_Python, i);
PyObject * v = PySequence_ITEM(X_Python, j);
PyObject * result = PyObject_CallFunctionObjArgs(userfn, u, v, NULL);
Py_DECREF(u);
Py_DECREF(v);
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
if (result==NULL) {
throw pythonerror();
}
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wold-style-cast"
#endif
const t_float C_result = PyFloat_AsDouble(result);
Py_DECREF(result);
#if HAVE_DIAGNOSTIC
#pragma GCC diagnostic pop
#endif
if (PyErr_Occurred()) {
throw pythonerror();
}
return C_result;
}
t_float braycurtis(const t_index i, const t_index j) const {
t_float sum1 = 0;
t_float sum2 = 0;
for (t_index k=0; k<dim; ++k) {
sum1 += fabs(X(i,k)-X(j,k));
sum2 += fabs(X(i,k)+X(j,k));
}
return sum1/sum2;
}
t_float mahalanobis(const t_index i, const t_index j) const {
// V_data contains the product X*VI
t_float sum = 0;
for (t_index k=0; k<dim; ++k) {
sum += (V_data[i*dim+k]-V_data[j*dim+k])*(X(i,k)-X(j,k));
}
return sum;
}
t_index mutable NTT; // 'local' variables
t_index mutable NXO;
t_index mutable NTF;
#define NTFFT NTF
#define NFFTT NTT
void nbool_correspond(const t_index i, const t_index j) const {
NTT = 0;
NXO = 0;
for (t_index k=0; k<dim; ++k) {
NTT += (Xb(i,k) & Xb(j,k)) ;
NXO += (Xb(i,k) ^ Xb(j,k)) ;
}
}
void nbool_correspond_tfft(const t_index i, const t_index j) const {
NTT = 0;
NXO = 0;
NTF = 0;
for (t_index k=0; k<dim; ++k) {