isotree 0.2.0 → 0.3.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +16 -1
- data/LICENSE.txt +2 -2
- data/README.md +41 -23
- data/ext/isotree/ext.cpp +144 -31
- data/ext/isotree/extconf.rb +7 -7
- data/lib/isotree/dataset.rb +0 -1
- data/lib/isotree/isolation_forest.rb +114 -30
- data/lib/isotree/version.rb +1 -1
- data/vendor/isotree/LICENSE +1 -1
- data/vendor/isotree/README.md +165 -27
- data/vendor/isotree/include/isotree.hpp +2111 -0
- data/vendor/isotree/include/isotree_oop.hpp +394 -0
- data/vendor/isotree/inst/COPYRIGHTS +62 -0
- data/vendor/isotree/src/RcppExports.cpp +525 -52
- data/vendor/isotree/src/Rwrapper.cpp +1931 -268
- data/vendor/isotree/src/c_interface.cpp +953 -0
- data/vendor/isotree/src/crit.hpp +4232 -0
- data/vendor/isotree/src/dist.hpp +1886 -0
- data/vendor/isotree/src/exp_depth_table.hpp +134 -0
- data/vendor/isotree/src/extended.hpp +1444 -0
- data/vendor/isotree/src/external_facing_generic.hpp +399 -0
- data/vendor/isotree/src/fit_model.hpp +2401 -0
- data/vendor/isotree/src/{dealloc.cpp → headers_joined.hpp} +38 -22
- data/vendor/isotree/src/helpers_iforest.hpp +813 -0
- data/vendor/isotree/src/{impute.cpp → impute.hpp} +353 -122
- data/vendor/isotree/src/indexer.cpp +515 -0
- data/vendor/isotree/src/instantiate_template_headers.cpp +118 -0
- data/vendor/isotree/src/instantiate_template_headers.hpp +240 -0
- data/vendor/isotree/src/isoforest.hpp +1659 -0
- data/vendor/isotree/src/isotree.hpp +1804 -392
- data/vendor/isotree/src/isotree_exportable.hpp +99 -0
- data/vendor/isotree/src/merge_models.cpp +159 -16
- data/vendor/isotree/src/mult.hpp +1321 -0
- data/vendor/isotree/src/oop_interface.cpp +842 -0
- data/vendor/isotree/src/oop_interface.hpp +278 -0
- data/vendor/isotree/src/other_helpers.hpp +219 -0
- data/vendor/isotree/src/predict.hpp +1932 -0
- data/vendor/isotree/src/python_helpers.hpp +134 -0
- data/vendor/isotree/src/ref_indexer.hpp +154 -0
- data/vendor/isotree/src/robinmap/LICENSE +21 -0
- data/vendor/isotree/src/robinmap/README.md +483 -0
- data/vendor/isotree/src/robinmap/include/tsl/robin_growth_policy.h +406 -0
- data/vendor/isotree/src/robinmap/include/tsl/robin_hash.h +1620 -0
- data/vendor/isotree/src/robinmap/include/tsl/robin_map.h +807 -0
- data/vendor/isotree/src/robinmap/include/tsl/robin_set.h +660 -0
- data/vendor/isotree/src/serialize.cpp +4300 -139
- data/vendor/isotree/src/sql.cpp +141 -59
- data/vendor/isotree/src/subset_models.cpp +174 -0
- data/vendor/isotree/src/utils.hpp +3808 -0
- data/vendor/isotree/src/xoshiro.hpp +467 -0
- data/vendor/isotree/src/ziggurat.hpp +405 -0
- metadata +40 -106
- data/vendor/cereal/LICENSE +0 -24
- data/vendor/cereal/README.md +0 -85
- data/vendor/cereal/include/cereal/access.hpp +0 -351
- data/vendor/cereal/include/cereal/archives/adapters.hpp +0 -163
- data/vendor/cereal/include/cereal/archives/binary.hpp +0 -169
- data/vendor/cereal/include/cereal/archives/json.hpp +0 -1019
- data/vendor/cereal/include/cereal/archives/portable_binary.hpp +0 -334
- data/vendor/cereal/include/cereal/archives/xml.hpp +0 -956
- data/vendor/cereal/include/cereal/cereal.hpp +0 -1089
- data/vendor/cereal/include/cereal/details/helpers.hpp +0 -422
- data/vendor/cereal/include/cereal/details/polymorphic_impl.hpp +0 -796
- data/vendor/cereal/include/cereal/details/polymorphic_impl_fwd.hpp +0 -65
- data/vendor/cereal/include/cereal/details/static_object.hpp +0 -127
- data/vendor/cereal/include/cereal/details/traits.hpp +0 -1411
- data/vendor/cereal/include/cereal/details/util.hpp +0 -84
- data/vendor/cereal/include/cereal/external/base64.hpp +0 -134
- data/vendor/cereal/include/cereal/external/rapidjson/allocators.h +0 -284
- data/vendor/cereal/include/cereal/external/rapidjson/cursorstreamwrapper.h +0 -78
- data/vendor/cereal/include/cereal/external/rapidjson/document.h +0 -2652
- data/vendor/cereal/include/cereal/external/rapidjson/encodedstream.h +0 -299
- data/vendor/cereal/include/cereal/external/rapidjson/encodings.h +0 -716
- data/vendor/cereal/include/cereal/external/rapidjson/error/en.h +0 -74
- data/vendor/cereal/include/cereal/external/rapidjson/error/error.h +0 -161
- data/vendor/cereal/include/cereal/external/rapidjson/filereadstream.h +0 -99
- data/vendor/cereal/include/cereal/external/rapidjson/filewritestream.h +0 -104
- data/vendor/cereal/include/cereal/external/rapidjson/fwd.h +0 -151
- data/vendor/cereal/include/cereal/external/rapidjson/internal/biginteger.h +0 -290
- data/vendor/cereal/include/cereal/external/rapidjson/internal/diyfp.h +0 -271
- data/vendor/cereal/include/cereal/external/rapidjson/internal/dtoa.h +0 -245
- data/vendor/cereal/include/cereal/external/rapidjson/internal/ieee754.h +0 -78
- data/vendor/cereal/include/cereal/external/rapidjson/internal/itoa.h +0 -308
- data/vendor/cereal/include/cereal/external/rapidjson/internal/meta.h +0 -186
- data/vendor/cereal/include/cereal/external/rapidjson/internal/pow10.h +0 -55
- data/vendor/cereal/include/cereal/external/rapidjson/internal/regex.h +0 -740
- data/vendor/cereal/include/cereal/external/rapidjson/internal/stack.h +0 -232
- data/vendor/cereal/include/cereal/external/rapidjson/internal/strfunc.h +0 -69
- data/vendor/cereal/include/cereal/external/rapidjson/internal/strtod.h +0 -290
- data/vendor/cereal/include/cereal/external/rapidjson/internal/swap.h +0 -46
- data/vendor/cereal/include/cereal/external/rapidjson/istreamwrapper.h +0 -128
- data/vendor/cereal/include/cereal/external/rapidjson/memorybuffer.h +0 -70
- data/vendor/cereal/include/cereal/external/rapidjson/memorystream.h +0 -71
- data/vendor/cereal/include/cereal/external/rapidjson/msinttypes/inttypes.h +0 -316
- data/vendor/cereal/include/cereal/external/rapidjson/msinttypes/stdint.h +0 -300
- data/vendor/cereal/include/cereal/external/rapidjson/ostreamwrapper.h +0 -81
- data/vendor/cereal/include/cereal/external/rapidjson/pointer.h +0 -1414
- data/vendor/cereal/include/cereal/external/rapidjson/prettywriter.h +0 -277
- data/vendor/cereal/include/cereal/external/rapidjson/rapidjson.h +0 -656
- data/vendor/cereal/include/cereal/external/rapidjson/reader.h +0 -2230
- data/vendor/cereal/include/cereal/external/rapidjson/schema.h +0 -2497
- data/vendor/cereal/include/cereal/external/rapidjson/stream.h +0 -223
- data/vendor/cereal/include/cereal/external/rapidjson/stringbuffer.h +0 -121
- data/vendor/cereal/include/cereal/external/rapidjson/writer.h +0 -709
- data/vendor/cereal/include/cereal/external/rapidxml/license.txt +0 -52
- data/vendor/cereal/include/cereal/external/rapidxml/manual.html +0 -406
- data/vendor/cereal/include/cereal/external/rapidxml/rapidxml.hpp +0 -2624
- data/vendor/cereal/include/cereal/external/rapidxml/rapidxml_iterators.hpp +0 -175
- data/vendor/cereal/include/cereal/external/rapidxml/rapidxml_print.hpp +0 -428
- data/vendor/cereal/include/cereal/external/rapidxml/rapidxml_utils.hpp +0 -123
- data/vendor/cereal/include/cereal/macros.hpp +0 -154
- data/vendor/cereal/include/cereal/specialize.hpp +0 -139
- data/vendor/cereal/include/cereal/types/array.hpp +0 -79
- data/vendor/cereal/include/cereal/types/atomic.hpp +0 -55
- data/vendor/cereal/include/cereal/types/base_class.hpp +0 -203
- data/vendor/cereal/include/cereal/types/bitset.hpp +0 -176
- data/vendor/cereal/include/cereal/types/boost_variant.hpp +0 -164
- data/vendor/cereal/include/cereal/types/chrono.hpp +0 -72
- data/vendor/cereal/include/cereal/types/common.hpp +0 -129
- data/vendor/cereal/include/cereal/types/complex.hpp +0 -56
- data/vendor/cereal/include/cereal/types/concepts/pair_associative_container.hpp +0 -73
- data/vendor/cereal/include/cereal/types/deque.hpp +0 -62
- data/vendor/cereal/include/cereal/types/forward_list.hpp +0 -68
- data/vendor/cereal/include/cereal/types/functional.hpp +0 -43
- data/vendor/cereal/include/cereal/types/list.hpp +0 -62
- data/vendor/cereal/include/cereal/types/map.hpp +0 -36
- data/vendor/cereal/include/cereal/types/memory.hpp +0 -425
- data/vendor/cereal/include/cereal/types/optional.hpp +0 -66
- data/vendor/cereal/include/cereal/types/polymorphic.hpp +0 -483
- data/vendor/cereal/include/cereal/types/queue.hpp +0 -132
- data/vendor/cereal/include/cereal/types/set.hpp +0 -103
- data/vendor/cereal/include/cereal/types/stack.hpp +0 -76
- data/vendor/cereal/include/cereal/types/string.hpp +0 -61
- data/vendor/cereal/include/cereal/types/tuple.hpp +0 -123
- data/vendor/cereal/include/cereal/types/unordered_map.hpp +0 -36
- data/vendor/cereal/include/cereal/types/unordered_set.hpp +0 -99
- data/vendor/cereal/include/cereal/types/utility.hpp +0 -47
- data/vendor/cereal/include/cereal/types/valarray.hpp +0 -89
- data/vendor/cereal/include/cereal/types/variant.hpp +0 -109
- data/vendor/cereal/include/cereal/types/vector.hpp +0 -112
- data/vendor/cereal/include/cereal/version.hpp +0 -52
- data/vendor/isotree/src/Makevars +0 -4
- data/vendor/isotree/src/crit.cpp +0 -912
- data/vendor/isotree/src/dist.cpp +0 -749
- data/vendor/isotree/src/extended.cpp +0 -790
- data/vendor/isotree/src/fit_model.cpp +0 -1090
- data/vendor/isotree/src/helpers_iforest.cpp +0 -324
- data/vendor/isotree/src/isoforest.cpp +0 -771
- data/vendor/isotree/src/mult.cpp +0 -607
- data/vendor/isotree/src/predict.cpp +0 -853
- data/vendor/isotree/src/utils.cpp +0 -1566
@@ -0,0 +1,3808 @@
|
|
1
|
+
/* Isolation forests and variations thereof, with adjustments for incorporation
|
2
|
+
* of categorical variables and missing values.
|
3
|
+
* Writen for C++11 standard and aimed at being used in R and Python.
|
4
|
+
*
|
5
|
+
* This library is based on the following works:
|
6
|
+
* [1] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
|
7
|
+
* "Isolation forest."
|
8
|
+
* 2008 Eighth IEEE International Conference on Data Mining. IEEE, 2008.
|
9
|
+
* [2] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
|
10
|
+
* "Isolation-based anomaly detection."
|
11
|
+
* ACM Transactions on Knowledge Discovery from Data (TKDD) 6.1 (2012): 3.
|
12
|
+
* [3] Hariri, Sahand, Matias Carrasco Kind, and Robert J. Brunner.
|
13
|
+
* "Extended Isolation Forest."
|
14
|
+
* arXiv preprint arXiv:1811.02141 (2018).
|
15
|
+
* [4] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
|
16
|
+
* "On detecting clustered anomalies using SCiForest."
|
17
|
+
* Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, Berlin, Heidelberg, 2010.
|
18
|
+
* [5] https://sourceforge.net/projects/iforest/
|
19
|
+
* [6] https://math.stackexchange.com/questions/3388518/expected-number-of-paths-required-to-separate-elements-in-a-binary-tree
|
20
|
+
* [7] Quinlan, J. Ross. C4. 5: programs for machine learning. Elsevier, 2014.
|
21
|
+
* [8] Cortes, David.
|
22
|
+
* "Distance approximation using Isolation Forests."
|
23
|
+
* arXiv preprint arXiv:1910.12362 (2019).
|
24
|
+
* [9] Cortes, David.
|
25
|
+
* "Imputing missing values with unsupervised random trees."
|
26
|
+
* arXiv preprint arXiv:1911.06646 (2019).
|
27
|
+
* [10] https://math.stackexchange.com/questions/3333220/expected-average-depth-in-random-binary-tree-constructed-top-to-bottom
|
28
|
+
* [11] Cortes, David.
|
29
|
+
* "Revisiting randomized choices in isolation forests."
|
30
|
+
* arXiv preprint arXiv:2110.13402 (2021).
|
31
|
+
* [12] Guha, Sudipto, et al.
|
32
|
+
* "Robust random cut forest based anomaly detection on streams."
|
33
|
+
* International conference on machine learning. PMLR, 2016.
|
34
|
+
* [13] Cortes, David.
|
35
|
+
* "Isolation forests: looking beyond tree depth."
|
36
|
+
* arXiv preprint arXiv:2111.11639 (2021).
|
37
|
+
* [14] Ting, Kai Ming, Yue Zhu, and Zhi-Hua Zhou.
|
38
|
+
* "Isolation kernel and its effect on SVM"
|
39
|
+
* Proceedings of the 24th ACM SIGKDD
|
40
|
+
* International Conference on Knowledge Discovery & Data Mining. 2018.
|
41
|
+
*
|
42
|
+
* BSD 2-Clause License
|
43
|
+
* Copyright (c) 2019-2022, David Cortes
|
44
|
+
* All rights reserved.
|
45
|
+
* Redistribution and use in source and binary forms, with or without
|
46
|
+
* modification, are permitted provided that the following conditions are met:
|
47
|
+
* * Redistributions of source code must retain the above copyright notice, this
|
48
|
+
* list of conditions and the following disclaimer.
|
49
|
+
* * Redistributions in binary form must reproduce the above copyright notice,
|
50
|
+
* this list of conditions and the following disclaimer in the documentation
|
51
|
+
* and/or other materials provided with the distribution.
|
52
|
+
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
53
|
+
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
54
|
+
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
55
|
+
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
56
|
+
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
57
|
+
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
58
|
+
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
59
|
+
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
60
|
+
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
61
|
+
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
62
|
+
*/
|
63
|
+
#include "isotree.hpp"
|
64
|
+
|
65
|
+
/* ceil(log2(x)) done with bit-wise operations ensures perfect precision (and it's faster too)
|
66
|
+
https://stackoverflow.com/questions/2589096/find-most-significant-bit-left-most-that-is-set-in-a-bit-array
|
67
|
+
https://stackoverflow.com/questions/11376288/fast-computing-of-log2-for-64-bit-integers */
|
68
|
+
#if SIZE_MAX == UINT32_MAX /* 32-bit systems */
|
69
|
+
constexpr static const uint32_t MultiplyDeBruijnBitPosition[32] =
|
70
|
+
{
|
71
|
+
0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30,
|
72
|
+
8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31
|
73
|
+
};
|
74
|
+
size_t log2ceil( size_t v )
|
75
|
+
{
|
76
|
+
v--;
|
77
|
+
v |= v >> 1; // first round down to one less than a power of 2
|
78
|
+
v |= v >> 2;
|
79
|
+
v |= v >> 4;
|
80
|
+
v |= v >> 8;
|
81
|
+
v |= v >> 16;
|
82
|
+
|
83
|
+
return MultiplyDeBruijnBitPosition[( uint32_t )( v * 0x07C4ACDDU ) >> 27] + 1;
|
84
|
+
}
|
85
|
+
#elif SIZE_MAX == UINT64_MAX /* 64-bit systems */
|
86
|
+
constexpr static const uint64_t tab64[64] = {
|
87
|
+
63, 0, 58, 1, 59, 47, 53, 2,
|
88
|
+
60, 39, 48, 27, 54, 33, 42, 3,
|
89
|
+
61, 51, 37, 40, 49, 18, 28, 20,
|
90
|
+
55, 30, 34, 11, 43, 14, 22, 4,
|
91
|
+
62, 57, 46, 52, 38, 26, 32, 41,
|
92
|
+
50, 36, 17, 19, 29, 10, 13, 21,
|
93
|
+
56, 45, 25, 31, 35, 16, 9, 12,
|
94
|
+
44, 24, 15, 8, 23, 7, 6, 5};
|
95
|
+
|
96
|
+
size_t log2ceil(size_t value)
|
97
|
+
{
|
98
|
+
value--;
|
99
|
+
value |= value >> 1;
|
100
|
+
value |= value >> 2;
|
101
|
+
value |= value >> 4;
|
102
|
+
value |= value >> 8;
|
103
|
+
value |= value >> 16;
|
104
|
+
value |= value >> 32;
|
105
|
+
return tab64[((uint64_t)((value - (value >> 1))*0x07EDD5E59A4E28C2)) >> 58] + 1;
|
106
|
+
}
|
107
|
+
#else /* other architectures - might be much slower */
|
108
|
+
#if (__cplusplus >= 202002L)
|
109
|
+
#include <bit>
|
110
|
+
size_t log2ceil(size_t value)
|
111
|
+
{
|
112
|
+
size_t out = std::numeric_limits<size_t>::digits - std::countl_zero(value);
|
113
|
+
out -= (value == ((size_t)1 << (out-1)));
|
114
|
+
return out;
|
115
|
+
}
|
116
|
+
#else
|
117
|
+
size_t log2ceil(size_t value)
|
118
|
+
{
|
119
|
+
size_t value_ = value;
|
120
|
+
size_t out = 0;
|
121
|
+
while (value >= 1) {
|
122
|
+
value = value >> 1;
|
123
|
+
out++;
|
124
|
+
}
|
125
|
+
out -= (value_ == ((size_t)1 << (out-1)));
|
126
|
+
return out;
|
127
|
+
}
|
128
|
+
#endif
|
129
|
+
#endif
|
130
|
+
|
131
|
+
/* adapted from cephes */
|
132
|
+
#define EULERS_GAMMA 0.577215664901532860606512
|
133
|
+
double digamma(double x)
|
134
|
+
{
|
135
|
+
double y, z, z2;
|
136
|
+
|
137
|
+
/* check for positive integer up to 128 */
|
138
|
+
if (unlikely((x <= 64) && (x == std::floor(x)))) {
|
139
|
+
return harmonic_recursive(1.0, (double)x) - EULERS_GAMMA;
|
140
|
+
}
|
141
|
+
|
142
|
+
if (likely(x < 1.0e17 ))
|
143
|
+
{
|
144
|
+
z = 1.0/(x * x);
|
145
|
+
z2 = square(z);
|
146
|
+
y = z * ( 8.33333333333333333333E-2
|
147
|
+
-8.33333333333333333333E-3*z
|
148
|
+
+3.96825396825396825397E-3*z2
|
149
|
+
-4.16666666666666666667E-3*z2*z
|
150
|
+
+7.57575757575757575758E-3*square(z2)
|
151
|
+
-2.10927960927960927961E-2*square(z2)*z
|
152
|
+
+8.33333333333333333333E-2*square(z2)*z2);
|
153
|
+
}
|
154
|
+
else {
|
155
|
+
y = 0.0;
|
156
|
+
}
|
157
|
+
|
158
|
+
y = ((-0.5/x) - y) + std::log(x);
|
159
|
+
return y;
|
160
|
+
}
|
161
|
+
|
162
|
+
/* http://fredrik-j.blogspot.com/2009/02/how-not-to-compute-harmonic-numbers.html
|
163
|
+
https://en.wikipedia.org/wiki/Harmonic_number
|
164
|
+
https://github.com/scikit-learn/scikit-learn/pull/19087 */
|
165
|
+
template <class ldouble_safe>
|
166
|
+
double harmonic(size_t n)
|
167
|
+
{
|
168
|
+
ldouble_safe temp = (ldouble_safe)1 / square((ldouble_safe)n);
|
169
|
+
return - (ldouble_safe)0.5 * temp * ( (ldouble_safe)1/(ldouble_safe)6 - temp * ((ldouble_safe)1/(ldouble_safe)60 - ((ldouble_safe)1/(ldouble_safe)126)*temp) )
|
170
|
+
+ (ldouble_safe)0.5 * ((ldouble_safe)1/(ldouble_safe)n)
|
171
|
+
+ std::log((ldouble_safe)n) + (ldouble_safe)EULERS_GAMMA;
|
172
|
+
}
|
173
|
+
|
174
|
+
/* usage for getting harmonic(n) is like this: harmonic_recursive((double)1, (double)(n + 1)); */
|
175
|
+
double harmonic_recursive(double a, double b)
|
176
|
+
{
|
177
|
+
if (b == a + 1) return 1. / a;
|
178
|
+
double m = std::floor((a + b) / 2.);
|
179
|
+
return harmonic_recursive(a, m) + harmonic_recursive(m, b);
|
180
|
+
}
|
181
|
+
|
182
|
+
/* https://stats.stackexchange.com/questions/423542/isolation-forest-and-average-expected-depth-formula
|
183
|
+
https://math.stackexchange.com/questions/3333220/expected-average-depth-in-random-binary-tree-constructed-top-to-bottom */
|
184
|
+
#include "exp_depth_table.hpp"
|
185
|
+
template <class ldouble_safe>
|
186
|
+
double expected_avg_depth(size_t sample_size)
|
187
|
+
{
|
188
|
+
if (likely(sample_size <= N_PRECALC_EXP_DEPTH)) {
|
189
|
+
return exp_depth_table[sample_size - 1];
|
190
|
+
}
|
191
|
+
return 2. * (harmonic<ldouble_safe>(sample_size) - 1.);
|
192
|
+
}
|
193
|
+
|
194
|
+
/* Note: H(x) = psi(x+1) + gamma */
|
195
|
+
template <class ldouble_safe>
|
196
|
+
double expected_avg_depth(ldouble_safe approx_sample_size)
|
197
|
+
{
|
198
|
+
if (approx_sample_size <= 1)
|
199
|
+
return 0;
|
200
|
+
else if (approx_sample_size < (ldouble_safe)INT32_MAX)
|
201
|
+
return 2. * (digamma(approx_sample_size + 1.) + EULERS_GAMMA - 1.);
|
202
|
+
else {
|
203
|
+
ldouble_safe temp = (ldouble_safe)1 / square(approx_sample_size);
|
204
|
+
return (ldouble_safe)2 * std::log(approx_sample_size) + (ldouble_safe)2*((ldouble_safe)EULERS_GAMMA - (ldouble_safe)1)
|
205
|
+
+ ((ldouble_safe)1/approx_sample_size)
|
206
|
+
- temp * ( (ldouble_safe)1/(ldouble_safe)6 - temp * ((ldouble_safe)1/(ldouble_safe)60 - ((ldouble_safe)1/(ldouble_safe)126)*temp) );
|
207
|
+
}
|
208
|
+
}
|
209
|
+
|
210
|
+
/* https://math.stackexchange.com/questions/3388518/expected-number-of-paths-required-to-separate-elements-in-a-binary-tree */
|
211
|
+
#define THRESHOLD_EXACT_S 87670 /* difference is <5e-4 */
|
212
|
+
double expected_separation_depth(size_t n)
|
213
|
+
{
|
214
|
+
switch(n)
|
215
|
+
{
|
216
|
+
case 0: return 0.;
|
217
|
+
case 1: return 0.;
|
218
|
+
case 2: return 1.;
|
219
|
+
case 3: return 1. + (1./3.);
|
220
|
+
case 4: return 1. + (1./3.) + (2./9.);
|
221
|
+
case 5: return 1.71666666667;
|
222
|
+
case 6: return 1.84;
|
223
|
+
case 7: return 1.93809524;
|
224
|
+
case 8: return 2.01836735;
|
225
|
+
case 9: return 2.08551587;
|
226
|
+
case 10: return 2.14268078;
|
227
|
+
default:
|
228
|
+
{
|
229
|
+
if (n >= THRESHOLD_EXACT_S)
|
230
|
+
return 3;
|
231
|
+
else
|
232
|
+
return expected_separation_depth_hotstart((double)2.14268078, (size_t)10, n);
|
233
|
+
}
|
234
|
+
}
|
235
|
+
}
|
236
|
+
|
237
|
+
double expected_separation_depth_hotstart(double curr, size_t n_curr, size_t n_final)
|
238
|
+
{
|
239
|
+
if (n_final >= 1360)
|
240
|
+
{
|
241
|
+
if (n_final >= THRESHOLD_EXACT_S)
|
242
|
+
return 3;
|
243
|
+
else if (n_final >= 40774)
|
244
|
+
return 2.999;
|
245
|
+
else if (n_final >= 18844)
|
246
|
+
return 2.998;
|
247
|
+
else if (n_final >= 11956)
|
248
|
+
return 2.997;
|
249
|
+
else if (n_final >= 8643)
|
250
|
+
return 2.996;
|
251
|
+
else if (n_final >= 6713)
|
252
|
+
return 2.995;
|
253
|
+
else if (n_final >= 4229)
|
254
|
+
return 2.9925;
|
255
|
+
else if (n_final >= 3040)
|
256
|
+
return 2.99;
|
257
|
+
else if (n_final >= 2724)
|
258
|
+
return 2.989;
|
259
|
+
else if (n_final >= 1902)
|
260
|
+
return 2.985;
|
261
|
+
else if (n_final >= 1360)
|
262
|
+
return 2.98;
|
263
|
+
|
264
|
+
/* Note on the chosen precision: when calling it on smaller sample sizes,
|
265
|
+
the standard error of the separation depth will be larger, thus it's less
|
266
|
+
critical to get it right down to the smallest possible precision, while for
|
267
|
+
larger samples the standard error of the separation depth will be smaller */
|
268
|
+
}
|
269
|
+
|
270
|
+
for (size_t i = n_curr + 1; i <= n_final; i++)
|
271
|
+
curr += (-curr * (double)i + 3. * (double)i - 4.) / ((double)i * ((double)(i-1)));
|
272
|
+
return curr;
|
273
|
+
}
|
274
|
+
|
275
|
+
/* linear interpolation */
|
276
|
+
template <class ldouble_safe>
|
277
|
+
double expected_separation_depth(ldouble_safe n)
|
278
|
+
{
|
279
|
+
if (n >= THRESHOLD_EXACT_S)
|
280
|
+
return 3;
|
281
|
+
double s_l = expected_separation_depth((size_t) std::floor(n));
|
282
|
+
ldouble_safe u = std::ceil(n);
|
283
|
+
double s_u = s_l + (-s_l * u + 3. * u - 4.) / (u * (u - 1.));
|
284
|
+
double diff = n - std::floor(n);
|
285
|
+
return s_l + diff * s_u;
|
286
|
+
}
|
287
|
+
|
288
|
+
void increase_comb_counter(size_t ix_arr[], size_t st, size_t end, size_t n, double counter[], double exp_remainder)
|
289
|
+
{
|
290
|
+
size_t i, j;
|
291
|
+
size_t ncomb = calc_ncomb(n);
|
292
|
+
if (exp_remainder <= 1)
|
293
|
+
for (size_t el1 = st; el1 < end; el1++)
|
294
|
+
{
|
295
|
+
for (size_t el2 = el1 + 1; el2 <= end; el2++)
|
296
|
+
{
|
297
|
+
// counter[i * (n - (i+1)/2) + j - i - 1]++; /* beaware integer division */
|
298
|
+
i = ix_arr[el1]; j = ix_arr[el2];
|
299
|
+
counter[ix_comb(i, j, n, ncomb)]++;
|
300
|
+
}
|
301
|
+
}
|
302
|
+
else
|
303
|
+
for (size_t el1 = st; el1 < end; el1++)
|
304
|
+
{
|
305
|
+
for (size_t el2 = el1 + 1; el2 <= end; el2++)
|
306
|
+
{
|
307
|
+
i = ix_arr[el1]; j = ix_arr[el2];
|
308
|
+
counter[ix_comb(i, j, n, ncomb)] += exp_remainder;
|
309
|
+
}
|
310
|
+
}
|
311
|
+
}
|
312
|
+
|
313
|
+
void increase_comb_counter(size_t ix_arr[], size_t st, size_t end, size_t n,
|
314
|
+
double *restrict counter, double *restrict weights, double exp_remainder)
|
315
|
+
{
|
316
|
+
size_t i, j;
|
317
|
+
size_t ncomb = calc_ncomb(n);
|
318
|
+
if (exp_remainder <= 1)
|
319
|
+
for (size_t el1 = st; el1 < end; el1++)
|
320
|
+
{
|
321
|
+
for (size_t el2 = el1 + 1; el2 <= end; el2++)
|
322
|
+
{
|
323
|
+
i = ix_arr[el1]; j = ix_arr[el2];
|
324
|
+
counter[ix_comb(i, j, n, ncomb)] += weights[i] * weights[j];
|
325
|
+
}
|
326
|
+
}
|
327
|
+
else
|
328
|
+
for (size_t el1 = st; el1 < end; el1++)
|
329
|
+
{
|
330
|
+
for (size_t el2 = el1 + 1; el2 <= end; el2++)
|
331
|
+
{
|
332
|
+
i = ix_arr[el1]; j = ix_arr[el2];
|
333
|
+
counter[ix_comb(i, j, n, ncomb)] += weights[i] * weights[j] * exp_remainder;
|
334
|
+
}
|
335
|
+
}
|
336
|
+
}
|
337
|
+
|
338
|
+
/* Note to self: don't try merge this into a template with the one above, as the other one has 'restrict' qualifier */
|
339
|
+
void increase_comb_counter(size_t ix_arr[], size_t st, size_t end, size_t n,
|
340
|
+
double counter[], hashed_map<size_t, double> &weights, double exp_remainder)
|
341
|
+
{
|
342
|
+
size_t i, j;
|
343
|
+
size_t ncomb = calc_ncomb(n);
|
344
|
+
if (exp_remainder <= 1)
|
345
|
+
for (size_t el1 = st; el1 < end; el1++)
|
346
|
+
{
|
347
|
+
for (size_t el2 = el1 + 1; el2 <= end; el2++)
|
348
|
+
{
|
349
|
+
i = ix_arr[el1]; j = ix_arr[el2];
|
350
|
+
counter[ix_comb(i, j, n, ncomb)] += weights[i] * weights[j];
|
351
|
+
}
|
352
|
+
}
|
353
|
+
else
|
354
|
+
for (size_t el1 = st; el1 < end; el1++)
|
355
|
+
{
|
356
|
+
for (size_t el2 = el1 + 1; el2 <= end; el2++)
|
357
|
+
{
|
358
|
+
i = ix_arr[el1]; j = ix_arr[el2];
|
359
|
+
counter[ix_comb(i, j, n, ncomb)] += weights[i] * weights[j] * exp_remainder;
|
360
|
+
}
|
361
|
+
}
|
362
|
+
}
|
363
|
+
|
364
|
+
void increase_comb_counter_in_groups(size_t ix_arr[], size_t st, size_t end, size_t split_ix, size_t n,
|
365
|
+
double counter[], double exp_remainder)
|
366
|
+
{
|
367
|
+
size_t *ptr_split_ix = std::lower_bound(ix_arr + st, ix_arr + end + 1, split_ix);
|
368
|
+
size_t n_group = std::distance(ix_arr + st, ptr_split_ix);
|
369
|
+
n = n - split_ix;
|
370
|
+
|
371
|
+
if (exp_remainder <= 1)
|
372
|
+
for (size_t ix1 = st; ix1 < st + n_group; ix1++)
|
373
|
+
for (size_t ix2 = st + n_group; ix2 <= end; ix2++)
|
374
|
+
counter[ix_arr[ix1] * n + ix_arr[ix2] - split_ix]++;
|
375
|
+
else
|
376
|
+
for (size_t ix1 = st; ix1 < st + n_group; ix1++)
|
377
|
+
for (size_t ix2 = st + n_group; ix2 <= end; ix2++)
|
378
|
+
counter[ix_arr[ix1] * n + ix_arr[ix2] - split_ix] += exp_remainder;
|
379
|
+
}
|
380
|
+
|
381
|
+
void increase_comb_counter_in_groups(size_t ix_arr[], size_t st, size_t end, size_t split_ix, size_t n,
|
382
|
+
double *restrict counter, double *restrict weights, double exp_remainder)
|
383
|
+
{
|
384
|
+
size_t *ptr_split_ix = std::lower_bound(ix_arr + st, ix_arr + end + 1, split_ix);
|
385
|
+
size_t n_group = std::distance(ix_arr + st, ptr_split_ix);
|
386
|
+
n = n - split_ix;
|
387
|
+
|
388
|
+
if (exp_remainder <= 1)
|
389
|
+
for (size_t ix1 = st; ix1 < st + n_group; ix1++)
|
390
|
+
for (size_t ix2 = st + n_group; ix2 <= end; ix2++)
|
391
|
+
counter[ix_arr[ix1] * n + ix_arr[ix2] - split_ix]
|
392
|
+
+=
|
393
|
+
weights[ix_arr[ix1]] * weights[ix_arr[ix2]];
|
394
|
+
else
|
395
|
+
for (size_t ix1 = st; ix1 < st + n_group; ix1++)
|
396
|
+
for (size_t ix2 = st + n_group; ix2 <= end; ix2++)
|
397
|
+
counter[ix_arr[ix1] * n + ix_arr[ix2] - split_ix]
|
398
|
+
+=
|
399
|
+
weights[ix_arr[ix1]] * weights[ix_arr[ix2]] * exp_remainder;
|
400
|
+
}
|
401
|
+
|
402
|
+
void tmat_to_dense(double *restrict tmat, double *restrict dmat, size_t n, double fill_diag)
|
403
|
+
{
|
404
|
+
size_t ncomb = calc_ncomb(n);
|
405
|
+
for (size_t i = 0; i < (n-1); i++)
|
406
|
+
{
|
407
|
+
for (size_t j = i + 1; j < n; j++)
|
408
|
+
{
|
409
|
+
// dmat[i + j * n] = dmat[j + i * n] = tmat[i * (n - (i+1)/2) + j - i - 1];
|
410
|
+
dmat[i + j * n] = dmat[j + i * n] = tmat[ix_comb(i, j, n, ncomb)];
|
411
|
+
}
|
412
|
+
}
|
413
|
+
for (size_t i = 0; i < n; i++)
|
414
|
+
dmat[i + i * n] = fill_diag;
|
415
|
+
}
|
416
|
+
|
417
|
+
template <class real_t>
|
418
|
+
void build_btree_sampler(std::vector<double> &btree_weights, real_t *restrict sample_weights,
|
419
|
+
size_t nrows, size_t &restrict log2_n, size_t &restrict btree_offset)
|
420
|
+
{
|
421
|
+
/* build a perfectly-balanced binary search tree in which each node will
|
422
|
+
hold the sum of the weights of its children */
|
423
|
+
log2_n = log2ceil(nrows);
|
424
|
+
if (btree_weights.empty())
|
425
|
+
btree_weights.resize(pow2(log2_n + 1), 0);
|
426
|
+
else
|
427
|
+
btree_weights.assign(btree_weights.size(), 0);
|
428
|
+
btree_offset = pow2(log2_n) - 1;
|
429
|
+
|
430
|
+
for (size_t ix = 0; ix < nrows; ix++)
|
431
|
+
btree_weights[ix + btree_offset] = std::fmax(0., sample_weights[ix]);
|
432
|
+
for (size_t ix = btree_weights.size() - 1; ix > 0; ix--)
|
433
|
+
btree_weights[ix_parent(ix)] += btree_weights[ix];
|
434
|
+
|
435
|
+
if (std::isnan(btree_weights[0]) || btree_weights[0] <= 0)
|
436
|
+
{
|
437
|
+
fprintf(stderr, "Numeric precision error with sample weights, will not use them.\n");
|
438
|
+
log2_n = 0;
|
439
|
+
btree_weights.clear();
|
440
|
+
btree_weights.shrink_to_fit();
|
441
|
+
}
|
442
|
+
}
|
443
|
+
|
444
|
+
template <class real_t, class ldouble_safe>
|
445
|
+
void sample_random_rows(std::vector<size_t> &restrict ix_arr, size_t nrows, bool with_replacement,
|
446
|
+
RNG_engine &rnd_generator, std::vector<size_t> &restrict ix_all,
|
447
|
+
real_t *restrict sample_weights, std::vector<double> &restrict btree_weights,
|
448
|
+
size_t log2_n, size_t btree_offset, std::vector<bool> &is_repeated)
|
449
|
+
{
|
450
|
+
size_t ntake = ix_arr.size();
|
451
|
+
|
452
|
+
/* if with replacement, just generate random uniform numbers */
|
453
|
+
if (with_replacement)
|
454
|
+
{
|
455
|
+
if (sample_weights == NULL)
|
456
|
+
{
|
457
|
+
std::uniform_int_distribution<size_t> runif(0, nrows - 1);
|
458
|
+
for (size_t &ix : ix_arr)
|
459
|
+
ix = runif(rnd_generator);
|
460
|
+
}
|
461
|
+
|
462
|
+
else
|
463
|
+
{
|
464
|
+
std::discrete_distribution<size_t> runif(sample_weights, sample_weights + nrows);
|
465
|
+
for (size_t &ix : ix_arr)
|
466
|
+
ix = runif(rnd_generator);
|
467
|
+
}
|
468
|
+
}
|
469
|
+
|
470
|
+
/* if all the elements are needed, don't bother with any sampling */
|
471
|
+
else if (ntake == nrows)
|
472
|
+
{
|
473
|
+
std::iota(ix_arr.begin(), ix_arr.end(), (size_t)0);
|
474
|
+
}
|
475
|
+
|
476
|
+
|
477
|
+
/* if there are sample weights, use binary trees to keep track and update weight
|
478
|
+
https://stackoverflow.com/questions/57599509/c-random-non-repeated-integers-with-weights */
|
479
|
+
else if (sample_weights != NULL)
|
480
|
+
{
|
481
|
+
/* TODO: here could instead generate only 1 random number from zero to the full weight,
|
482
|
+
and then subtract from it as it goes down every level. Would have less precision
|
483
|
+
but should still work fine. */
|
484
|
+
|
485
|
+
double rnd_subrange, w_left;
|
486
|
+
double curr_subrange;
|
487
|
+
size_t curr_ix;
|
488
|
+
for (size_t &ix : ix_arr)
|
489
|
+
{
|
490
|
+
/* go down the tree by drawing a random number and
|
491
|
+
checking if it falls in the left or right ranges */
|
492
|
+
curr_ix = 0;
|
493
|
+
curr_subrange = btree_weights[0];
|
494
|
+
for (size_t lev = 0; lev < log2_n; lev++)
|
495
|
+
{
|
496
|
+
rnd_subrange = std::uniform_real_distribution<double>(0., curr_subrange)(rnd_generator);
|
497
|
+
w_left = btree_weights[ix_child(curr_ix)];
|
498
|
+
curr_ix = ix_child(curr_ix) + (rnd_subrange >= w_left);
|
499
|
+
curr_subrange = btree_weights[curr_ix];
|
500
|
+
}
|
501
|
+
|
502
|
+
/* finally, determine element to choose in this iteration */
|
503
|
+
ix = curr_ix - btree_offset;
|
504
|
+
|
505
|
+
/* now remove the weight of the chosen element */
|
506
|
+
btree_weights[curr_ix] = 0;
|
507
|
+
for (size_t lev = 0; lev < log2_n; lev++)
|
508
|
+
{
|
509
|
+
curr_ix = ix_parent(curr_ix);
|
510
|
+
btree_weights[curr_ix] = btree_weights[ix_child(curr_ix)]
|
511
|
+
+ btree_weights[ix_child(curr_ix) + 1];
|
512
|
+
}
|
513
|
+
}
|
514
|
+
}
|
515
|
+
|
516
|
+
/* if no sample weights and not with replacement (most common case expected),
|
517
|
+
then use different algorithms depending on the sampled fraction */
|
518
|
+
else
|
519
|
+
{
|
520
|
+
|
521
|
+
/* if sampling a larger fraction, fill an array enumerating the rows, shuffle, and take first N */
|
522
|
+
if (ntake >= (nrows / 2))
|
523
|
+
{
|
524
|
+
|
525
|
+
if (ix_all.empty())
|
526
|
+
ix_all.resize(nrows);
|
527
|
+
|
528
|
+
/* in order for random seeds to always be reproducible, don't re-use previous shuffles */
|
529
|
+
std::iota(ix_all.begin(), ix_all.end(), (size_t)0);
|
530
|
+
|
531
|
+
/* If the number of sampled elements is large, do a full shuffle, enjoy simd-instructs when copying over */
|
532
|
+
if (ntake >= ((nrows * 3)/4))
|
533
|
+
{
|
534
|
+
std::shuffle(ix_all.begin(), ix_all.end(), rnd_generator);
|
535
|
+
ix_arr.assign(ix_all.begin(), ix_all.begin() + ntake);
|
536
|
+
}
|
537
|
+
|
538
|
+
/* otherwise, do only a partial shuffle (use Yates algorithm) and copy elements along the way */
|
539
|
+
else
|
540
|
+
{
|
541
|
+
size_t chosen;
|
542
|
+
for (size_t i = nrows - 1; i >= nrows - ntake; i--)
|
543
|
+
{
|
544
|
+
chosen = std::uniform_int_distribution<size_t>(0, i)(rnd_generator);
|
545
|
+
ix_arr[nrows - i - 1] = ix_all[chosen];
|
546
|
+
ix_all[chosen] = ix_all[i];
|
547
|
+
}
|
548
|
+
}
|
549
|
+
|
550
|
+
}
|
551
|
+
|
552
|
+
/* If the sample size is small, use Floyd's random sampling algorithm
|
553
|
+
https://stackoverflow.com/questions/2394246/algorithm-to-select-a-single-random-combination-of-values */
|
554
|
+
else
|
555
|
+
{
|
556
|
+
|
557
|
+
size_t candidate;
|
558
|
+
|
559
|
+
/* if the sample size is relatively large, use a temporary boolean vector */
|
560
|
+
if (((ldouble_safe)ntake / (ldouble_safe)nrows) > (1. / 50.))
|
561
|
+
{
|
562
|
+
|
563
|
+
if (is_repeated.empty())
|
564
|
+
is_repeated.resize(nrows, false);
|
565
|
+
else
|
566
|
+
is_repeated.assign(is_repeated.size(), false);
|
567
|
+
|
568
|
+
for (size_t rnd_ix = nrows - ntake; rnd_ix < nrows; rnd_ix++)
|
569
|
+
{
|
570
|
+
candidate = std::uniform_int_distribution<size_t>(0, rnd_ix)(rnd_generator);
|
571
|
+
if (is_repeated[candidate])
|
572
|
+
{
|
573
|
+
ix_arr[ntake - (nrows - rnd_ix)] = rnd_ix;
|
574
|
+
is_repeated[rnd_ix] = true;
|
575
|
+
}
|
576
|
+
|
577
|
+
else
|
578
|
+
{
|
579
|
+
ix_arr[ntake - (nrows - rnd_ix)] = candidate;
|
580
|
+
is_repeated[candidate] = true;
|
581
|
+
}
|
582
|
+
}
|
583
|
+
|
584
|
+
}
|
585
|
+
|
586
|
+
/* if the sample size is very small, use an unordered set */
|
587
|
+
else
|
588
|
+
{
|
589
|
+
|
590
|
+
hashed_set<size_t> repeated_set;
|
591
|
+
repeated_set.reserve(ntake);
|
592
|
+
for (size_t rnd_ix = nrows - ntake; rnd_ix < nrows; rnd_ix++)
|
593
|
+
{
|
594
|
+
candidate = std::uniform_int_distribution<size_t>(0, rnd_ix)(rnd_generator);
|
595
|
+
if (repeated_set.find(candidate) == repeated_set.end()) /* TODO: switch to C++20 'contains' */
|
596
|
+
{
|
597
|
+
ix_arr[ntake - (nrows - rnd_ix)] = candidate;
|
598
|
+
repeated_set.insert(candidate);
|
599
|
+
}
|
600
|
+
|
601
|
+
else
|
602
|
+
{
|
603
|
+
ix_arr[ntake - (nrows - rnd_ix)] = rnd_ix;
|
604
|
+
repeated_set.insert(rnd_ix);
|
605
|
+
}
|
606
|
+
}
|
607
|
+
|
608
|
+
}
|
609
|
+
|
610
|
+
}
|
611
|
+
|
612
|
+
}
|
613
|
+
}
|
614
|
+
|
615
|
+
/* https://stackoverflow.com/questions/57599509/c-random-non-repeated-integers-with-weights */
|
616
|
+
template <class real_t>
|
617
|
+
void weighted_shuffle(size_t *restrict outp, size_t n, real_t *restrict weights, double *restrict buffer_arr, RNG_engine &rnd_generator)
|
618
|
+
{
|
619
|
+
/* determine smallest power of two that is larger than N */
|
620
|
+
size_t tree_levels = log2ceil(n);
|
621
|
+
|
622
|
+
/* initialize vector with place-holders for perfectly-balanced tree */
|
623
|
+
std::fill(buffer_arr, buffer_arr + pow2(tree_levels + 1), (double)0);
|
624
|
+
|
625
|
+
/* compute sums for the tree leaves at each node */
|
626
|
+
size_t offset = pow2(tree_levels) - 1;
|
627
|
+
for (size_t ix = 0; ix < n; ix++) {
|
628
|
+
buffer_arr[ix + offset] = std::fmax(0., weights[ix]);
|
629
|
+
}
|
630
|
+
for (size_t ix = pow2(tree_levels+1) - 1; ix > 0; ix--) {
|
631
|
+
buffer_arr[ix_parent(ix)] += buffer_arr[ix];
|
632
|
+
}
|
633
|
+
|
634
|
+
/* if the weights are invalid, produce an unweighted shuffle */
|
635
|
+
if (std::isnan(buffer_arr[0]) || buffer_arr[0] <= 0)
|
636
|
+
{
|
637
|
+
std::iota(outp, outp + n, (size_t)0);
|
638
|
+
std::shuffle(outp, outp + n, rnd_generator);
|
639
|
+
return;
|
640
|
+
}
|
641
|
+
|
642
|
+
/* sample according to uniform distribution */
|
643
|
+
double rnd_subrange, w_left;
|
644
|
+
double curr_subrange;
|
645
|
+
size_t curr_ix;
|
646
|
+
|
647
|
+
for (size_t el = 0; el < n; el++)
|
648
|
+
{
|
649
|
+
/* go down the tree by drawing a random number and
|
650
|
+
checking if it falls in the left or right sub-ranges */
|
651
|
+
curr_ix = 0;
|
652
|
+
curr_subrange = buffer_arr[0];
|
653
|
+
for (size_t lev = 0; lev < tree_levels; lev++)
|
654
|
+
{
|
655
|
+
rnd_subrange = std::uniform_real_distribution<double>(0., curr_subrange)(rnd_generator);
|
656
|
+
w_left = buffer_arr[ix_child(curr_ix)];
|
657
|
+
curr_ix = ix_child(curr_ix) + (rnd_subrange >= w_left);
|
658
|
+
curr_subrange = buffer_arr[curr_ix];
|
659
|
+
}
|
660
|
+
|
661
|
+
/* finally, add element from this iteration */
|
662
|
+
outp[el] = curr_ix - offset;
|
663
|
+
|
664
|
+
/* now remove the weight of the chosen element */
|
665
|
+
buffer_arr[curr_ix] = 0;
|
666
|
+
for (size_t lev = 0; lev < tree_levels; lev++)
|
667
|
+
{
|
668
|
+
curr_ix = ix_parent(curr_ix);
|
669
|
+
buffer_arr[curr_ix] = buffer_arr[ix_child(curr_ix)]
|
670
|
+
+ buffer_arr[ix_child(curr_ix) + 1];
|
671
|
+
}
|
672
|
+
}
|
673
|
+
}
|
674
|
+
|
675
|
+
double sample_random_uniform(double xmin, double xmax, RNG_engine &rng) noexcept
|
676
|
+
{
|
677
|
+
double out;
|
678
|
+
std::uniform_real_distribution<double> runif(xmin, xmax);
|
679
|
+
for (int attempt = 0; attempt < 100; attempt++)
|
680
|
+
{
|
681
|
+
out = runif(rng);
|
682
|
+
if (likely(out < xmax)) return out;
|
683
|
+
}
|
684
|
+
return xmin;
|
685
|
+
}
|
686
|
+
|
687
|
+
template <class ldouble_safe>
|
688
|
+
template <class other_t>
|
689
|
+
ColumnSampler<ldouble_safe>& ColumnSampler<ldouble_safe>::operator=(const ColumnSampler<other_t> &other)
|
690
|
+
{
|
691
|
+
this->col_indices = other.col_indices;
|
692
|
+
this->tree_weights = other.tree_weights;
|
693
|
+
this->curr_pos = other.curr_pos;
|
694
|
+
this->curr_col = other.curr_col;
|
695
|
+
this->last_given = other.last_given;
|
696
|
+
this->n_cols = other.n_cols;
|
697
|
+
this->tree_levels = other.tree_levels;
|
698
|
+
this->offset = other.offset;
|
699
|
+
this->n_dropped = other.n_dropped;
|
700
|
+
return *this;
|
701
|
+
}
|
702
|
+
|
703
|
+
/* This one samples with replacement. When using weights, the algorithm is the
|
704
|
+
same as for the row sampler, but keeping the weights after taking each iteration. */
|
705
|
+
/* TODO: this column sampler could use coroutines from C++20 once compilers implement them. */
|
706
|
+
template <class ldouble_safe>
|
707
|
+
template <class real_t>
|
708
|
+
void ColumnSampler<ldouble_safe>::initialize(real_t weights[], size_t n_cols)
|
709
|
+
{
|
710
|
+
this->n_cols = n_cols;
|
711
|
+
this->tree_levels = log2ceil(n_cols);
|
712
|
+
if (this->tree_weights.empty())
|
713
|
+
this->tree_weights.resize(pow2(this->tree_levels + 1), 0);
|
714
|
+
else {
|
715
|
+
if (this->tree_weights.size() != pow2(this->tree_levels + 1))
|
716
|
+
this->tree_weights.resize(this->tree_levels);
|
717
|
+
std::fill(this->tree_weights.begin(), this->tree_weights.end(), 0.);
|
718
|
+
}
|
719
|
+
|
720
|
+
/* compute sums for the tree leaves at each node */
|
721
|
+
this->offset = pow2(this->tree_levels) - 1;
|
722
|
+
for (size_t ix = 0; ix < this->n_cols; ix++)
|
723
|
+
this->tree_weights[ix + this->offset] = std::fmax(0., weights[ix]);
|
724
|
+
for (size_t ix = this->tree_weights.size() - 1; ix > 0; ix--)
|
725
|
+
this->tree_weights[ix_parent(ix)] += this->tree_weights[ix];
|
726
|
+
|
727
|
+
/* if the weights are invalid, make it an unweighted sampler */
|
728
|
+
if (unlikely(std::isnan(this->tree_weights[0]) || this->tree_weights[0] <= 0))
|
729
|
+
{
|
730
|
+
this->drop_weights();
|
731
|
+
}
|
732
|
+
|
733
|
+
this->n_dropped = 0;
|
734
|
+
}
|
735
|
+
|
736
|
+
template <class ldouble_safe>
|
737
|
+
void ColumnSampler<ldouble_safe>::drop_weights()
|
738
|
+
{
|
739
|
+
this->tree_weights.clear();
|
740
|
+
this->tree_weights.shrink_to_fit();
|
741
|
+
this->initialize(n_cols);
|
742
|
+
this->n_dropped = 0;
|
743
|
+
}
|
744
|
+
|
745
|
+
template <class ldouble_safe>
|
746
|
+
bool ColumnSampler<ldouble_safe>::has_weights()
|
747
|
+
{
|
748
|
+
return !this->tree_weights.empty();
|
749
|
+
}
|
750
|
+
|
751
|
+
template <class ldouble_safe>
|
752
|
+
void ColumnSampler<ldouble_safe>::initialize(size_t n_cols)
|
753
|
+
{
|
754
|
+
if (!this->has_weights())
|
755
|
+
{
|
756
|
+
this->n_cols = n_cols;
|
757
|
+
this->curr_pos = n_cols;
|
758
|
+
this->col_indices.resize(n_cols);
|
759
|
+
std::iota(this->col_indices.begin(), this->col_indices.end(), (size_t)0);
|
760
|
+
}
|
761
|
+
}
|
762
|
+
|
763
|
+
/* TODO: this one should instead call the same function for sampling rows,
|
764
|
+
and should be done at the time of initialization so as to avoid allocating
|
765
|
+
and filling the whole array. That way it'd be faster and use less memory. */
|
766
|
+
template <class ldouble_safe>
|
767
|
+
void ColumnSampler<ldouble_safe>::leave_m_cols(size_t m, RNG_engine &rnd_generator)
|
768
|
+
{
|
769
|
+
if (m == 0 || m >= this->n_cols)
|
770
|
+
return;
|
771
|
+
|
772
|
+
if (!this->has_weights())
|
773
|
+
{
|
774
|
+
size_t chosen;
|
775
|
+
if (m <= this->n_cols / 4)
|
776
|
+
{
|
777
|
+
for (this->curr_pos = 0; this->curr_pos < m; this->curr_pos++)
|
778
|
+
{
|
779
|
+
chosen = std::uniform_int_distribution<size_t>(0, this->n_cols - this->curr_pos - 1)(rnd_generator);
|
780
|
+
std::swap(this->col_indices[this->curr_pos + chosen], this->col_indices[this->curr_pos]);
|
781
|
+
}
|
782
|
+
}
|
783
|
+
|
784
|
+
else if ((ldouble_safe)m >= (ldouble_safe)(3./4.) * (ldouble_safe)this->n_cols)
|
785
|
+
{
|
786
|
+
for (this->curr_pos = this->n_cols-1; this->curr_pos > this->n_cols - m; this->curr_pos--)
|
787
|
+
{
|
788
|
+
chosen = std::uniform_int_distribution<size_t>(0, this->curr_pos)(rnd_generator);
|
789
|
+
std::swap(this->col_indices[chosen], this->col_indices[this->curr_pos]);
|
790
|
+
}
|
791
|
+
this->curr_pos = m;
|
792
|
+
}
|
793
|
+
|
794
|
+
else
|
795
|
+
{
|
796
|
+
std::shuffle(this->col_indices.begin(), this->col_indices.end(), rnd_generator);
|
797
|
+
this->curr_pos = m;
|
798
|
+
}
|
799
|
+
}
|
800
|
+
|
801
|
+
else
|
802
|
+
{
|
803
|
+
std::vector<double> curr_weights = this->tree_weights;
|
804
|
+
std::fill(this->tree_weights.begin(), this->tree_weights.end(), 0.);
|
805
|
+
double rnd_subrange, w_left;
|
806
|
+
double curr_subrange;
|
807
|
+
size_t curr_ix;
|
808
|
+
|
809
|
+
for (size_t col = 0; col < m; col++)
|
810
|
+
{
|
811
|
+
curr_ix = 0;
|
812
|
+
curr_subrange = curr_weights[0];
|
813
|
+
if (curr_subrange <= 0)
|
814
|
+
{
|
815
|
+
if (col == 0)
|
816
|
+
{
|
817
|
+
this->drop_weights();
|
818
|
+
return;
|
819
|
+
}
|
820
|
+
|
821
|
+
else
|
822
|
+
{
|
823
|
+
m = col;
|
824
|
+
goto rebuild_tree;
|
825
|
+
}
|
826
|
+
}
|
827
|
+
|
828
|
+
for (size_t lev = 0; lev < this->tree_levels; lev++)
|
829
|
+
{
|
830
|
+
rnd_subrange = std::uniform_real_distribution<double>(0., curr_subrange)(rnd_generator);
|
831
|
+
w_left = curr_weights[ix_child(curr_ix)];
|
832
|
+
curr_ix = ix_child(curr_ix) + (rnd_subrange >= w_left);
|
833
|
+
curr_subrange = curr_weights[curr_ix];
|
834
|
+
}
|
835
|
+
|
836
|
+
this->tree_weights[curr_ix] = curr_weights[curr_ix];
|
837
|
+
|
838
|
+
/* now remove the weight of the chosen element */
|
839
|
+
curr_weights[curr_ix] = 0;
|
840
|
+
for (size_t lev = 0; lev < this->tree_levels; lev++)
|
841
|
+
{
|
842
|
+
curr_ix = ix_parent(curr_ix);
|
843
|
+
curr_weights[curr_ix] = curr_weights[ix_child(curr_ix)]
|
844
|
+
+ curr_weights[ix_child(curr_ix) + 1];
|
845
|
+
}
|
846
|
+
}
|
847
|
+
|
848
|
+
/* rebuild the tree after getting new weights */
|
849
|
+
rebuild_tree:
|
850
|
+
for (size_t ix = this->tree_weights.size() - 1; ix > 0; ix--)
|
851
|
+
this->tree_weights[ix_parent(ix)] += this->tree_weights[ix];
|
852
|
+
|
853
|
+
this->n_dropped = this->n_cols - m;
|
854
|
+
}
|
855
|
+
}
|
856
|
+
|
857
|
+
template <class ldouble_safe>
|
858
|
+
void ColumnSampler<ldouble_safe>::drop_col(size_t col, size_t nobs_left)
|
859
|
+
{
|
860
|
+
if (!this->has_weights())
|
861
|
+
{
|
862
|
+
if (this->col_indices[this->last_given] == col)
|
863
|
+
{
|
864
|
+
std::swap(this->col_indices[this->last_given], this->col_indices[--this->curr_pos]);
|
865
|
+
}
|
866
|
+
|
867
|
+
else if (this->curr_pos > 4*nobs_left)
|
868
|
+
{
|
869
|
+
return;
|
870
|
+
}
|
871
|
+
|
872
|
+
else
|
873
|
+
{
|
874
|
+
for (size_t ix = 0; ix < this->curr_pos; ix++)
|
875
|
+
{
|
876
|
+
if (this->col_indices[ix] == col)
|
877
|
+
{
|
878
|
+
std::swap(this->col_indices[ix], this->col_indices[--this->curr_pos]);
|
879
|
+
break;
|
880
|
+
}
|
881
|
+
}
|
882
|
+
}
|
883
|
+
|
884
|
+
if (this->curr_col) this->curr_col--;
|
885
|
+
}
|
886
|
+
|
887
|
+
else
|
888
|
+
{
|
889
|
+
this->n_dropped++;
|
890
|
+
size_t curr_ix = col + this->offset;
|
891
|
+
this->tree_weights[curr_ix] = 0.;
|
892
|
+
for (size_t lev = 0; lev < this->tree_levels; lev++)
|
893
|
+
{
|
894
|
+
curr_ix = ix_parent(curr_ix);
|
895
|
+
this->tree_weights[curr_ix] = this->tree_weights[ix_child(curr_ix)]
|
896
|
+
+ this->tree_weights[ix_child(curr_ix) + 1];
|
897
|
+
}
|
898
|
+
}
|
899
|
+
}
|
900
|
+
|
901
|
+
template <class ldouble_safe>
|
902
|
+
void ColumnSampler<ldouble_safe>::drop_col(size_t col)
|
903
|
+
{
|
904
|
+
this->drop_col(col, SIZE_MAX);
|
905
|
+
}
|
906
|
+
|
907
|
+
/* to be used exclusively when initializing the density calculator,
|
908
|
+
and only when 'col_indices' is a straight range with no dropped columns */
|
909
|
+
template <class ldouble_safe>
|
910
|
+
void ColumnSampler<ldouble_safe>::drop_from_tail(size_t col)
|
911
|
+
{
|
912
|
+
std::swap(this->col_indices[col], this->col_indices[--this->curr_pos]);
|
913
|
+
}
|
914
|
+
|
915
|
+
template <class ldouble_safe>
|
916
|
+
void ColumnSampler<ldouble_safe>::prepare_full_pass()
|
917
|
+
{
|
918
|
+
this->curr_col = 0;
|
919
|
+
|
920
|
+
if (this->has_weights())
|
921
|
+
{
|
922
|
+
if (this->col_indices.size() < this->n_cols)
|
923
|
+
this->col_indices.resize(this->n_cols);
|
924
|
+
this->curr_pos = 0;
|
925
|
+
for (size_t col = 0; col < this->n_cols; col++)
|
926
|
+
{
|
927
|
+
if (this->tree_weights[col + this->offset] > 0)
|
928
|
+
this->col_indices[this->curr_pos++] = col;
|
929
|
+
}
|
930
|
+
}
|
931
|
+
}
|
932
|
+
|
933
|
+
template <class ldouble_safe>
|
934
|
+
bool ColumnSampler<ldouble_safe>::sample_col(size_t &col, RNG_engine &rnd_generator)
|
935
|
+
{
|
936
|
+
if (!this->has_weights())
|
937
|
+
{
|
938
|
+
switch(this->curr_pos)
|
939
|
+
{
|
940
|
+
case 0: return false;
|
941
|
+
case 1:
|
942
|
+
{
|
943
|
+
this->last_given = 0;
|
944
|
+
col = this->col_indices[0];
|
945
|
+
return true;
|
946
|
+
}
|
947
|
+
default:
|
948
|
+
{
|
949
|
+
this->last_given = std::uniform_int_distribution<size_t>(0, this->curr_pos-1)(rnd_generator);
|
950
|
+
col = this->col_indices[this->last_given];
|
951
|
+
return true;
|
952
|
+
}
|
953
|
+
}
|
954
|
+
}
|
955
|
+
|
956
|
+
else
|
957
|
+
{
|
958
|
+
/* TODO: here could instead generate only 1 random number from zero to the full weight,
|
959
|
+
and then subtract from it as it goes down every level. Would have less precision
|
960
|
+
but should still work fine. */
|
961
|
+
size_t curr_ix = 0;
|
962
|
+
double rnd_subrange, w_left;
|
963
|
+
double curr_subrange = this->tree_weights[0];
|
964
|
+
if (curr_subrange <= 0)
|
965
|
+
return false;
|
966
|
+
|
967
|
+
for (size_t lev = 0; lev < tree_levels; lev++)
|
968
|
+
{
|
969
|
+
rnd_subrange = std::uniform_real_distribution<double>(0., curr_subrange)(rnd_generator);
|
970
|
+
w_left = this->tree_weights[ix_child(curr_ix)];
|
971
|
+
curr_ix = ix_child(curr_ix) + (rnd_subrange >= w_left);
|
972
|
+
curr_subrange = this->tree_weights[curr_ix];
|
973
|
+
}
|
974
|
+
|
975
|
+
col = curr_ix - this->offset;
|
976
|
+
return true;
|
977
|
+
}
|
978
|
+
}
|
979
|
+
|
980
|
+
template <class ldouble_safe>
|
981
|
+
bool ColumnSampler<ldouble_safe>::sample_col(size_t &col)
|
982
|
+
{
|
983
|
+
if (this->curr_pos == this->curr_col || this->curr_pos == 0)
|
984
|
+
return false;
|
985
|
+
this->last_given = this->curr_col;
|
986
|
+
col = this->col_indices[this->curr_col++];
|
987
|
+
return true;
|
988
|
+
}
|
989
|
+
|
990
|
+
template <class ldouble_safe>
|
991
|
+
void ColumnSampler<ldouble_safe>::shuffle_remainder(RNG_engine &rnd_generator)
|
992
|
+
{
|
993
|
+
if (!this->has_weights())
|
994
|
+
{
|
995
|
+
this->prepare_full_pass();
|
996
|
+
std::shuffle(this->col_indices.begin(),
|
997
|
+
this->col_indices.begin() + this->curr_pos,
|
998
|
+
rnd_generator);
|
999
|
+
}
|
1000
|
+
|
1001
|
+
else
|
1002
|
+
{
|
1003
|
+
if (this->tree_weights[0] <= 0)
|
1004
|
+
return;
|
1005
|
+
std::vector<double> curr_weights = this->tree_weights;
|
1006
|
+
this->curr_pos = 0;
|
1007
|
+
this->curr_col = 0;
|
1008
|
+
|
1009
|
+
if (this->col_indices.size() < this->n_cols)
|
1010
|
+
this->col_indices.resize(this->n_cols);
|
1011
|
+
|
1012
|
+
double rnd_subrange, w_left;
|
1013
|
+
double curr_subrange;
|
1014
|
+
size_t curr_ix;
|
1015
|
+
|
1016
|
+
for (this->curr_pos = 0; this->curr_pos < this->n_cols; this->curr_pos++)
|
1017
|
+
{
|
1018
|
+
curr_ix = 0;
|
1019
|
+
curr_subrange = curr_weights[0];
|
1020
|
+
if (curr_subrange <= 0)
|
1021
|
+
return;
|
1022
|
+
|
1023
|
+
for (size_t lev = 0; lev < this->tree_levels; lev++)
|
1024
|
+
{
|
1025
|
+
rnd_subrange = std::uniform_real_distribution<double>(0., curr_subrange)(rnd_generator);
|
1026
|
+
w_left = curr_weights[ix_child(curr_ix)];
|
1027
|
+
curr_ix = ix_child(curr_ix) + (rnd_subrange >= w_left);
|
1028
|
+
curr_subrange = curr_weights[curr_ix];
|
1029
|
+
}
|
1030
|
+
|
1031
|
+
/* finally, add element from this iteration */
|
1032
|
+
this->col_indices[this->curr_pos] = curr_ix - this->offset;
|
1033
|
+
|
1034
|
+
/* now remove the weight of the chosen element */
|
1035
|
+
curr_weights[curr_ix] = 0;
|
1036
|
+
for (size_t lev = 0; lev < this->tree_levels; lev++)
|
1037
|
+
{
|
1038
|
+
curr_ix = ix_parent(curr_ix);
|
1039
|
+
curr_weights[curr_ix] = curr_weights[ix_child(curr_ix)]
|
1040
|
+
+ curr_weights[ix_child(curr_ix) + 1];
|
1041
|
+
}
|
1042
|
+
}
|
1043
|
+
}
|
1044
|
+
}
|
1045
|
+
|
1046
|
+
template <class ldouble_safe>
|
1047
|
+
size_t ColumnSampler<ldouble_safe>::get_remaining_cols()
|
1048
|
+
{
|
1049
|
+
if (!this->has_weights())
|
1050
|
+
return this->curr_pos;
|
1051
|
+
else
|
1052
|
+
return this->n_cols - this->n_dropped;
|
1053
|
+
}
|
1054
|
+
|
1055
|
+
template <class ldouble_safe>
|
1056
|
+
void ColumnSampler<ldouble_safe>::get_array_remaining_cols(std::vector<size_t> &restrict cols)
|
1057
|
+
{
|
1058
|
+
if (!this->has_weights())
|
1059
|
+
{
|
1060
|
+
cols.assign(this->col_indices.begin(), this->col_indices.begin() + this->curr_pos);
|
1061
|
+
std::sort(cols.begin(), cols.begin() + this->curr_pos);
|
1062
|
+
}
|
1063
|
+
|
1064
|
+
else
|
1065
|
+
{
|
1066
|
+
size_t n_rem = 0;
|
1067
|
+
for (size_t col = 0; col < this->n_cols; col++)
|
1068
|
+
{
|
1069
|
+
if (this->tree_weights[col + this->offset] > 0)
|
1070
|
+
{
|
1071
|
+
cols[n_rem++] = col;
|
1072
|
+
}
|
1073
|
+
}
|
1074
|
+
}
|
1075
|
+
}
|
1076
|
+
|
1077
|
+
template<class ldouble_safe, class real_t>
|
1078
|
+
bool SingleNodeColumnSampler<ldouble_safe, real_t>::initialize
|
1079
|
+
(
|
1080
|
+
double *restrict weights,
|
1081
|
+
std::vector<size_t> *col_indices,
|
1082
|
+
size_t curr_pos,
|
1083
|
+
size_t n_sample,
|
1084
|
+
bool backup_weights
|
1085
|
+
)
|
1086
|
+
{
|
1087
|
+
if (!curr_pos) return false;
|
1088
|
+
|
1089
|
+
this->col_indices = col_indices->data();
|
1090
|
+
this->curr_pos = curr_pos;
|
1091
|
+
this->n_left = this->curr_pos;
|
1092
|
+
this->weights_orig = weights;
|
1093
|
+
|
1094
|
+
if (n_sample > std::max(log2ceil(this->curr_pos), (size_t)3))
|
1095
|
+
{
|
1096
|
+
this->using_tree = true;
|
1097
|
+
this->backup_weights = false;
|
1098
|
+
|
1099
|
+
if (this->used_weights.empty()) {
|
1100
|
+
this->used_weights.reserve(col_indices->size());
|
1101
|
+
this->mapped_indices.reserve(col_indices->size());
|
1102
|
+
this->tree_weights.reserve(2 * col_indices->size());
|
1103
|
+
}
|
1104
|
+
|
1105
|
+
this->used_weights.resize(this->curr_pos);
|
1106
|
+
this->mapped_indices.resize(this->curr_pos);
|
1107
|
+
|
1108
|
+
for (size_t col = 0; col < this->curr_pos; col++) {
|
1109
|
+
this->mapped_indices[col] = this->col_indices[col];
|
1110
|
+
this->used_weights[col] = weights[this->col_indices[col]];
|
1111
|
+
if (!weights[this->col_indices[col]]) this->n_left--;
|
1112
|
+
}
|
1113
|
+
|
1114
|
+
this->tree_weights.resize(0);
|
1115
|
+
build_btree_sampler(this->tree_weights, this->used_weights.data(),
|
1116
|
+
this->curr_pos, this->tree_levels, this->offset);
|
1117
|
+
|
1118
|
+
this->n_inf = 0;
|
1119
|
+
if (std::isinf(this->tree_weights[0]))
|
1120
|
+
{
|
1121
|
+
if (this->mapped_inf_indices.empty())
|
1122
|
+
this->mapped_inf_indices.resize(this->curr_pos);
|
1123
|
+
|
1124
|
+
for (size_t col = 0; col < this->curr_pos; col++)
|
1125
|
+
{
|
1126
|
+
if (std::isinf(weights[this->col_indices[col]]))
|
1127
|
+
{
|
1128
|
+
this->mapped_inf_indices[this->n_inf++] = this->col_indices[col];
|
1129
|
+
weights[this->col_indices[col]] = 0;
|
1130
|
+
}
|
1131
|
+
|
1132
|
+
else
|
1133
|
+
{
|
1134
|
+
this->mapped_indices[col - this->n_inf] = this->col_indices[col];
|
1135
|
+
this->used_weights[col - this->n_inf] = weights[this->col_indices[col]];
|
1136
|
+
}
|
1137
|
+
}
|
1138
|
+
|
1139
|
+
this->tree_weights.resize(0);
|
1140
|
+
build_btree_sampler(this->tree_weights, this->used_weights.data(),
|
1141
|
+
this->curr_pos - this->n_inf, this->tree_levels, this->offset);
|
1142
|
+
}
|
1143
|
+
|
1144
|
+
this->used_weights.resize(0);
|
1145
|
+
|
1146
|
+
if (this->tree_weights[0] <= 0 && !this->n_inf)
|
1147
|
+
return false;
|
1148
|
+
}
|
1149
|
+
|
1150
|
+
else
|
1151
|
+
{
|
1152
|
+
this->using_tree = false;
|
1153
|
+
this->backup_weights = backup_weights;
|
1154
|
+
|
1155
|
+
if (this->backup_weights)
|
1156
|
+
{
|
1157
|
+
if (this->weights_own.empty())
|
1158
|
+
this->weights_own.resize(col_indices->size());
|
1159
|
+
this->weights_own.assign(weights, weights + this->curr_pos);
|
1160
|
+
}
|
1161
|
+
|
1162
|
+
this->cumw = 0;
|
1163
|
+
for (size_t col = 0; col < this->curr_pos; col++) {
|
1164
|
+
this->cumw += weights[this->col_indices[col]];
|
1165
|
+
if (!weights[this->col_indices[col]]) this->n_left--;
|
1166
|
+
}
|
1167
|
+
|
1168
|
+
if (std::isnan(this->cumw))
|
1169
|
+
throw std::runtime_error("NAs encountered. Try using a different value for 'missing_action'.\n");
|
1170
|
+
|
1171
|
+
/* if it's infinite, will choose among columns with infinite weight first */
|
1172
|
+
this->n_inf = 0;
|
1173
|
+
if (std::isinf(this->cumw))
|
1174
|
+
{
|
1175
|
+
if (this->inifinite_weights.empty())
|
1176
|
+
this->inifinite_weights.resize(col_indices->size());
|
1177
|
+
else
|
1178
|
+
this->inifinite_weights.assign(col_indices->size(), false);
|
1179
|
+
|
1180
|
+
this->cumw = 0;
|
1181
|
+
for (size_t col = 0; col < this->curr_pos; col++)
|
1182
|
+
{
|
1183
|
+
if (std::isinf(weights[this->col_indices[col]])) {
|
1184
|
+
this->n_inf++;
|
1185
|
+
this->inifinite_weights[this->col_indices[col]] = true;
|
1186
|
+
weights[this->col_indices[col]] = 0;
|
1187
|
+
}
|
1188
|
+
|
1189
|
+
else {
|
1190
|
+
this->cumw += weights[this->col_indices[col]];
|
1191
|
+
}
|
1192
|
+
}
|
1193
|
+
}
|
1194
|
+
|
1195
|
+
if (!this->cumw && !this->n_inf) return false;
|
1196
|
+
}
|
1197
|
+
|
1198
|
+
return true;
|
1199
|
+
}
|
1200
|
+
|
1201
|
+
template <class ldouble_safe, class real_t>
|
1202
|
+
bool SingleNodeColumnSampler<ldouble_safe, real_t>::sample_col(size_t &col_chosen, RNG_engine &rnd_generator)
|
1203
|
+
{
|
1204
|
+
if (!this->using_tree)
|
1205
|
+
{
|
1206
|
+
if (this->backup_weights)
|
1207
|
+
this->weights_orig = this->weights_own.data();
|
1208
|
+
|
1209
|
+
/* if there's infinites, choose uniformly at random from them */
|
1210
|
+
if (this->n_inf)
|
1211
|
+
{
|
1212
|
+
size_t chosen = std::uniform_int_distribution<size_t>(0, this->n_inf-1)(rnd_generator);
|
1213
|
+
size_t curr = 0;
|
1214
|
+
for (size_t col = 0; col < this->curr_pos; col++)
|
1215
|
+
{
|
1216
|
+
curr += inifinite_weights[this->col_indices[col]];
|
1217
|
+
if (curr == chosen)
|
1218
|
+
{
|
1219
|
+
col_chosen = this->col_indices[col];
|
1220
|
+
this->n_inf--;
|
1221
|
+
this->inifinite_weights[col_chosen] = false;
|
1222
|
+
this->n_left--;
|
1223
|
+
return true;
|
1224
|
+
}
|
1225
|
+
}
|
1226
|
+
assert(0);
|
1227
|
+
}
|
1228
|
+
|
1229
|
+
if (!this->n_left) return false;
|
1230
|
+
|
1231
|
+
/* due to the way this is calculated, there can be large roundoff errors and even negatives */
|
1232
|
+
if (this->cumw <= 0)
|
1233
|
+
{
|
1234
|
+
this->cumw = 0;
|
1235
|
+
for (size_t col = 0; col < this->curr_pos; col++)
|
1236
|
+
this->cumw += this->weights_orig[this->col_indices[col]];
|
1237
|
+
if (unlikely(this->cumw <= 0))
|
1238
|
+
unexpected_error();
|
1239
|
+
}
|
1240
|
+
|
1241
|
+
/* if there are no infinites, choose a column according to weight */
|
1242
|
+
ldouble_safe chosen = std::uniform_real_distribution<ldouble_safe>((ldouble_safe)0, this->cumw)(rnd_generator);
|
1243
|
+
ldouble_safe cumw_ = 0;
|
1244
|
+
for (size_t col = 0; col < this->curr_pos; col++)
|
1245
|
+
{
|
1246
|
+
cumw_ += this->weights_orig[this->col_indices[col]];
|
1247
|
+
if (cumw_ >= chosen)
|
1248
|
+
{
|
1249
|
+
col_chosen = this->col_indices[col];
|
1250
|
+
this->cumw -= this->weights_orig[col_chosen];
|
1251
|
+
this->weights_orig[col_chosen] = 0;
|
1252
|
+
this->n_left--;
|
1253
|
+
return true;
|
1254
|
+
}
|
1255
|
+
}
|
1256
|
+
col_chosen = this->col_indices[this->curr_pos-1];
|
1257
|
+
this->cumw -= this->weights_orig[col_chosen];
|
1258
|
+
this->weights_orig[col_chosen] = 0;
|
1259
|
+
this->n_left--;
|
1260
|
+
return true;
|
1261
|
+
}
|
1262
|
+
|
1263
|
+
else
|
1264
|
+
{
|
1265
|
+
/* if there's infinites, choose uniformly at random from them */
|
1266
|
+
if (this->n_inf)
|
1267
|
+
{
|
1268
|
+
size_t chosen = std::uniform_int_distribution<size_t>(0, this->n_inf-1)(rnd_generator);
|
1269
|
+
col_chosen = this->mapped_inf_indices[chosen];
|
1270
|
+
std::swap(this->mapped_inf_indices[chosen], this->mapped_inf_indices[--this->n_inf]);
|
1271
|
+
this->n_left--;
|
1272
|
+
return true;
|
1273
|
+
}
|
1274
|
+
|
1275
|
+
else
|
1276
|
+
{
|
1277
|
+
/* TODO: should standardize all these tree traversals into one.
|
1278
|
+
This one in particular could do with sampling only a single
|
1279
|
+
random number as it will not typically require exhausting all
|
1280
|
+
options like the usual column sampler. */
|
1281
|
+
if (!this->n_left) return false;
|
1282
|
+
size_t curr_ix = 0;
|
1283
|
+
double rnd_subrange, w_left;
|
1284
|
+
double curr_subrange = this->tree_weights[0];
|
1285
|
+
if (curr_subrange <= 0)
|
1286
|
+
return false;
|
1287
|
+
|
1288
|
+
for (size_t lev = 0; lev < tree_levels; lev++)
|
1289
|
+
{
|
1290
|
+
rnd_subrange = std::uniform_real_distribution<double>(0., curr_subrange)(rnd_generator);
|
1291
|
+
w_left = this->tree_weights[ix_child(curr_ix)];
|
1292
|
+
curr_ix = ix_child(curr_ix) + (rnd_subrange >= w_left);
|
1293
|
+
curr_subrange = this->tree_weights[curr_ix];
|
1294
|
+
}
|
1295
|
+
|
1296
|
+
col_chosen = this->mapped_indices[curr_ix - this->offset];
|
1297
|
+
|
1298
|
+
this->tree_weights[curr_ix] = 0.;
|
1299
|
+
for (size_t lev = 0; lev < this->tree_levels; lev++)
|
1300
|
+
{
|
1301
|
+
curr_ix = ix_parent(curr_ix);
|
1302
|
+
this->tree_weights[curr_ix] = this->tree_weights[ix_child(curr_ix)]
|
1303
|
+
+ this->tree_weights[ix_child(curr_ix) + 1];
|
1304
|
+
}
|
1305
|
+
|
1306
|
+
this->n_left--;
|
1307
|
+
return true;
|
1308
|
+
}
|
1309
|
+
}
|
1310
|
+
}
|
1311
|
+
|
1312
|
+
template <class ldouble_safe, class real_t>
|
1313
|
+
void SingleNodeColumnSampler<ldouble_safe, real_t>::backup(SingleNodeColumnSampler &other, size_t ncols_tot)
|
1314
|
+
{
|
1315
|
+
other.n_inf = this->n_inf;
|
1316
|
+
other.n_left = this->n_left;
|
1317
|
+
other.using_tree = this->using_tree;
|
1318
|
+
|
1319
|
+
if (this->using_tree)
|
1320
|
+
{
|
1321
|
+
if (other.tree_weights.empty())
|
1322
|
+
{
|
1323
|
+
other.tree_weights.reserve(ncols_tot);
|
1324
|
+
other.mapped_inf_indices.reserve(ncols_tot);
|
1325
|
+
}
|
1326
|
+
other.tree_weights.assign(this->tree_weights.begin(), this->tree_weights.end());
|
1327
|
+
other.mapped_inf_indices.assign(this->mapped_inf_indices.begin(), this->mapped_inf_indices.end());
|
1328
|
+
}
|
1329
|
+
|
1330
|
+
else
|
1331
|
+
{
|
1332
|
+
other.cumw = this->cumw;
|
1333
|
+
if (this->backup_weights)
|
1334
|
+
{
|
1335
|
+
if (other.weights_own.empty())
|
1336
|
+
other.weights_own.reserve(ncols_tot);
|
1337
|
+
|
1338
|
+
other.weights_own.resize(this->n_left);
|
1339
|
+
for (size_t col = 0; col < this->n_left; col++)
|
1340
|
+
other.weights_own[col] = this->weights_own[this->col_indices[col]];
|
1341
|
+
}
|
1342
|
+
|
1343
|
+
if (this->inifinite_weights.size())
|
1344
|
+
{
|
1345
|
+
if (other.inifinite_weights.empty())
|
1346
|
+
other.inifinite_weights.reserve(ncols_tot);
|
1347
|
+
|
1348
|
+
other.inifinite_weights.resize(this->n_left);
|
1349
|
+
for (size_t col = 0; col < this->n_left; col++)
|
1350
|
+
other.inifinite_weights[col] = this->inifinite_weights[this->col_indices[col]];
|
1351
|
+
}
|
1352
|
+
}
|
1353
|
+
}
|
1354
|
+
|
1355
|
+
template <class ldouble_safe, class real_t>
|
1356
|
+
void SingleNodeColumnSampler<ldouble_safe, real_t>::restore(const SingleNodeColumnSampler &other)
|
1357
|
+
{
|
1358
|
+
this->n_inf = other.n_inf;
|
1359
|
+
this->n_left = other.n_left;
|
1360
|
+
this->using_tree = other.using_tree;
|
1361
|
+
|
1362
|
+
if (this->using_tree)
|
1363
|
+
{
|
1364
|
+
this->tree_weights.assign(other.tree_weights.begin(), other.tree_weights.end());
|
1365
|
+
this->mapped_inf_indices.assign(other.mapped_inf_indices.begin(), other.mapped_inf_indices.end());
|
1366
|
+
}
|
1367
|
+
|
1368
|
+
else
|
1369
|
+
{
|
1370
|
+
this->cumw = other.cumw;
|
1371
|
+
if (this->backup_weights)
|
1372
|
+
{
|
1373
|
+
for (size_t col = 0; col < this->n_left; col++)
|
1374
|
+
this->weights_own[this->col_indices[col]] = other.weights_own[col];
|
1375
|
+
}
|
1376
|
+
|
1377
|
+
if (this->inifinite_weights.size())
|
1378
|
+
{
|
1379
|
+
for (size_t col = 0; col < this->n_left; col++)
|
1380
|
+
this->inifinite_weights[this->col_indices[col]] = other.inifinite_weights[col];
|
1381
|
+
}
|
1382
|
+
}
|
1383
|
+
}
|
1384
|
+
|
1385
|
+
template <class ldouble_safe, class real_t>
|
1386
|
+
void DensityCalculator<ldouble_safe, real_t>::initialize(size_t max_depth, int max_categ, bool reserve_counts, ScoringMetric scoring_metric)
|
1387
|
+
{
|
1388
|
+
this->multipliers.reserve(max_depth+3);
|
1389
|
+
this->multipliers.clear();
|
1390
|
+
if (scoring_metric != AdjDensity)
|
1391
|
+
this->multipliers.push_back(0);
|
1392
|
+
else
|
1393
|
+
this->multipliers.push_back(1);
|
1394
|
+
|
1395
|
+
if (reserve_counts)
|
1396
|
+
{
|
1397
|
+
this->counts.resize(max_categ);
|
1398
|
+
}
|
1399
|
+
}
|
1400
|
+
|
1401
|
+
template <class ldouble_safe, class real_t>
|
1402
|
+
template <class InputData>
|
1403
|
+
void DensityCalculator<ldouble_safe, real_t>::initialize_bdens(const InputData &input_data,
|
1404
|
+
const ModelParams &model_params,
|
1405
|
+
std::vector<size_t> &ix_arr,
|
1406
|
+
ColumnSampler<ldouble_safe> &col_sampler)
|
1407
|
+
{
|
1408
|
+
this->fast_bratio = model_params.fast_bratio;
|
1409
|
+
if (this->fast_bratio)
|
1410
|
+
{
|
1411
|
+
this->multipliers.reserve(model_params.max_depth + 3);
|
1412
|
+
this->multipliers.push_back(0);
|
1413
|
+
}
|
1414
|
+
|
1415
|
+
if (input_data.range_low != NULL || input_data.ncat_ != NULL)
|
1416
|
+
{
|
1417
|
+
if (input_data.ncols_numeric)
|
1418
|
+
{
|
1419
|
+
this->queue_box.reserve(model_params.max_depth+3);
|
1420
|
+
this->box_low.assign(input_data.range_low, input_data.range_low + input_data.ncols_numeric);
|
1421
|
+
this->box_high.assign(input_data.range_high, input_data.range_high + input_data.ncols_numeric);
|
1422
|
+
}
|
1423
|
+
|
1424
|
+
if (input_data.ncols_categ)
|
1425
|
+
{
|
1426
|
+
this->queue_ncat.reserve(model_params.max_depth+2);
|
1427
|
+
this->ncat.assign(input_data.ncat_, input_data.ncat_ + input_data.ncols_categ);
|
1428
|
+
}
|
1429
|
+
|
1430
|
+
if (!this->fast_bratio)
|
1431
|
+
{
|
1432
|
+
if (input_data.ncols_numeric)
|
1433
|
+
{
|
1434
|
+
this->ranges.resize(input_data.ncols_numeric);
|
1435
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
1436
|
+
this->ranges[col] = this->box_high[col] - this->box_low[col];
|
1437
|
+
}
|
1438
|
+
|
1439
|
+
if (input_data.ncols_categ)
|
1440
|
+
{
|
1441
|
+
this->ncat_orig = this->ncat;
|
1442
|
+
}
|
1443
|
+
}
|
1444
|
+
|
1445
|
+
return;
|
1446
|
+
}
|
1447
|
+
|
1448
|
+
if (input_data.ncols_numeric)
|
1449
|
+
{
|
1450
|
+
this->queue_box.reserve(model_params.max_depth+3);
|
1451
|
+
this->box_low.resize(input_data.ncols_numeric);
|
1452
|
+
this->box_high.resize(input_data.ncols_numeric);
|
1453
|
+
if (!this->fast_bratio)
|
1454
|
+
this->ranges.resize(input_data.ncols_numeric);
|
1455
|
+
}
|
1456
|
+
if (input_data.ncols_categ)
|
1457
|
+
{
|
1458
|
+
this->queue_ncat.reserve(model_params.max_depth+2);
|
1459
|
+
}
|
1460
|
+
bool unsplittable = false;
|
1461
|
+
|
1462
|
+
size_t npresent = 0;
|
1463
|
+
std::vector<signed char> categ_present;
|
1464
|
+
if (input_data.ncols_categ)
|
1465
|
+
{
|
1466
|
+
categ_present.resize(input_data.max_categ);
|
1467
|
+
}
|
1468
|
+
|
1469
|
+
|
1470
|
+
col_sampler.prepare_full_pass();
|
1471
|
+
size_t col;
|
1472
|
+
while (col_sampler.sample_col(col))
|
1473
|
+
{
|
1474
|
+
if (col < input_data.ncols_numeric)
|
1475
|
+
{
|
1476
|
+
if (input_data.Xc_indptr != NULL)
|
1477
|
+
{
|
1478
|
+
get_range((size_t*)ix_arr.data(), (size_t)0, ix_arr.size()-(size_t)1, col,
|
1479
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
1480
|
+
model_params.missing_action, this->box_low[col], this->box_high[col], unsplittable);
|
1481
|
+
}
|
1482
|
+
|
1483
|
+
else
|
1484
|
+
{
|
1485
|
+
get_range((size_t*)ix_arr.data(), input_data.numeric_data + input_data.nrows * col, (size_t)0, ix_arr.size()-(size_t)1,
|
1486
|
+
model_params.missing_action, this->box_low[col], this->box_high[col], unsplittable);
|
1487
|
+
}
|
1488
|
+
|
1489
|
+
|
1490
|
+
if (unsplittable)
|
1491
|
+
{
|
1492
|
+
this->box_low[col] = 0;
|
1493
|
+
this->box_high[col] = 0;
|
1494
|
+
if (!this->fast_bratio)
|
1495
|
+
this->ranges[col] = 0;
|
1496
|
+
col_sampler.drop_col(col);
|
1497
|
+
}
|
1498
|
+
|
1499
|
+
if (!this->fast_bratio)
|
1500
|
+
{
|
1501
|
+
this->ranges[col] = (ldouble_safe)this->box_high[col] - (ldouble_safe)this->box_low[col];
|
1502
|
+
this->ranges[col] = std::fmax(this->ranges[col], (ldouble_safe)0);
|
1503
|
+
}
|
1504
|
+
}
|
1505
|
+
|
1506
|
+
else
|
1507
|
+
{
|
1508
|
+
get_categs((size_t*)ix_arr.data(),
|
1509
|
+
input_data.categ_data + input_data.nrows * (col - input_data.ncols_numeric),
|
1510
|
+
(size_t)0, ix_arr.size()-(size_t)1, input_data.ncat[col],
|
1511
|
+
model_params.missing_action, categ_present.data(), npresent, unsplittable);
|
1512
|
+
|
1513
|
+
if (unsplittable)
|
1514
|
+
{
|
1515
|
+
this->ncat[col - input_data.ncols_numeric] = 1;
|
1516
|
+
col_sampler.drop_col(col);
|
1517
|
+
}
|
1518
|
+
|
1519
|
+
else
|
1520
|
+
{
|
1521
|
+
this->ncat[col - input_data.ncols_numeric] = npresent;
|
1522
|
+
}
|
1523
|
+
}
|
1524
|
+
}
|
1525
|
+
|
1526
|
+
if (!this->fast_bratio)
|
1527
|
+
this->ncat_orig = this->ncat;
|
1528
|
+
}
|
1529
|
+
|
1530
|
+
template<class ldouble_safe, class real_t>
|
1531
|
+
template <class InputData>
|
1532
|
+
void DensityCalculator<ldouble_safe, real_t>::initialize_bdens_ext(const InputData &input_data,
|
1533
|
+
const ModelParams &model_params,
|
1534
|
+
std::vector<size_t> &ix_arr,
|
1535
|
+
ColumnSampler<ldouble_safe> &col_sampler,
|
1536
|
+
bool col_sampler_is_fresh)
|
1537
|
+
{
|
1538
|
+
this->vals_ext_box.reserve(model_params.max_depth + 3);
|
1539
|
+
this->queue_ext_box.reserve(model_params.max_depth + 3);
|
1540
|
+
this->vals_ext_box.push_back(0);
|
1541
|
+
|
1542
|
+
if (input_data.range_low != NULL)
|
1543
|
+
{
|
1544
|
+
this->box_low.assign(input_data.range_low, input_data.range_low + input_data.ncols_numeric);
|
1545
|
+
this->box_high.assign(input_data.range_high, input_data.range_high + input_data.ncols_numeric);
|
1546
|
+
return;
|
1547
|
+
}
|
1548
|
+
|
1549
|
+
this->box_low.resize(input_data.ncols_numeric);
|
1550
|
+
this->box_high.resize(input_data.ncols_numeric);
|
1551
|
+
bool unsplittable = false;
|
1552
|
+
|
1553
|
+
/* TODO: find out if there's an optimal point for choosing one or the other loop
|
1554
|
+
when using 'leave_m_cols' and when using 'prob_pick_col_by_range', then fill in the
|
1555
|
+
lines that are commented out. */
|
1556
|
+
// if (!input_data.ncols_categ || model_params.ncols_per_tree < input_data.ncols_numeric)
|
1557
|
+
if (input_data.ncols_numeric)
|
1558
|
+
{
|
1559
|
+
col_sampler.prepare_full_pass();
|
1560
|
+
size_t col;
|
1561
|
+
while (col_sampler.sample_col(col))
|
1562
|
+
{
|
1563
|
+
if (col >= input_data.ncols_numeric)
|
1564
|
+
continue;
|
1565
|
+
if (input_data.Xc_indptr != NULL)
|
1566
|
+
{
|
1567
|
+
get_range((size_t*)ix_arr.data(), (size_t)0, ix_arr.size()-(size_t)1, col,
|
1568
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
1569
|
+
model_params.missing_action, this->box_low[col], this->box_high[col], unsplittable);
|
1570
|
+
}
|
1571
|
+
|
1572
|
+
else
|
1573
|
+
{
|
1574
|
+
get_range((size_t*)ix_arr.data(), input_data.numeric_data + input_data.nrows * col, (size_t)0, ix_arr.size()-(size_t)1,
|
1575
|
+
model_params.missing_action, this->box_low[col], this->box_high[col], unsplittable);
|
1576
|
+
}
|
1577
|
+
|
1578
|
+
if (unsplittable)
|
1579
|
+
{
|
1580
|
+
this->box_low[col] = 0;
|
1581
|
+
this->box_high[col] = 0;
|
1582
|
+
col_sampler.drop_col(col);
|
1583
|
+
}
|
1584
|
+
}
|
1585
|
+
}
|
1586
|
+
|
1587
|
+
// else if (input_data.ncols_numeric)
|
1588
|
+
// {
|
1589
|
+
// size_t n_unsplittable = 0;
|
1590
|
+
// std::vector<size_t> unsplittable_cols;
|
1591
|
+
// if (col_sampler_is_fresh && !col_sampler.has_weights())
|
1592
|
+
// unsplittable_cols.reserve(input_data.ncols_numeric);
|
1593
|
+
|
1594
|
+
// /* TODO: this will do unnecessary calculations when using 'leave_m_cols' */
|
1595
|
+
// for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
1596
|
+
// {
|
1597
|
+
// if (input_data.Xc_indptr != NULL)
|
1598
|
+
// {
|
1599
|
+
// get_range((size_t*)ix_arr.data(), (size_t)0, ix_arr.size()-(size_t)1, col,
|
1600
|
+
// input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
1601
|
+
// model_params.missing_action, this->box_low[col], this->box_high[col], unsplittable);
|
1602
|
+
// }
|
1603
|
+
|
1604
|
+
// else
|
1605
|
+
// {
|
1606
|
+
// get_range((size_t*)ix_arr.data(), input_data.numeric_data + input_data.nrows * col, (size_t)0, ix_arr.size()-(size_t)1,
|
1607
|
+
// model_params.missing_action, this->box_low[col], this->box_high[col], unsplittable);
|
1608
|
+
// }
|
1609
|
+
|
1610
|
+
// if (unsplittable)
|
1611
|
+
// {
|
1612
|
+
// this->box_low[col] = 0;
|
1613
|
+
// this->box_high[col] = 0;
|
1614
|
+
// n_unsplittable++;
|
1615
|
+
// if (col_sampler.has_weights())
|
1616
|
+
// col_sampler.drop_col(col);
|
1617
|
+
// else if (col_sampler_is_fresh)
|
1618
|
+
// unsplittable_cols.push_back(col);
|
1619
|
+
// }
|
1620
|
+
// }
|
1621
|
+
|
1622
|
+
// if (n_unsplittable && col_sampler_is_fresh && !col_sampler.has_weights())
|
1623
|
+
// {
|
1624
|
+
// #if (__cplusplus >= 202002L)
|
1625
|
+
// for (auto col : unsplittable_cols | std::views::reverse)
|
1626
|
+
// col_sampler.drop_from_tail(col);
|
1627
|
+
// #else
|
1628
|
+
// for (size_t inv_col = 0; inv_col < unsplittable_cols.size(); inv_col++)
|
1629
|
+
// {
|
1630
|
+
// size_t col = unsplittable_cols.size() - inv_col - 1;
|
1631
|
+
// col_sampler.drop_from_tail(unsplittable_cols[col]);
|
1632
|
+
// }
|
1633
|
+
// #endif
|
1634
|
+
// }
|
1635
|
+
|
1636
|
+
// else if (n_unsplittable > model_params.sample_size / 16 && !col_sampler_is_fresh && !col_sampler.has_weights())
|
1637
|
+
// {
|
1638
|
+
// /* TODO */
|
1639
|
+
// }
|
1640
|
+
// }
|
1641
|
+
}
|
1642
|
+
|
1643
|
+
template<class ldouble_safe, class real_t>
|
1644
|
+
void DensityCalculator<ldouble_safe, real_t>::push_density(double xmin, double xmax, double split_point)
|
1645
|
+
{
|
1646
|
+
if (std::isinf(xmax) || std::isinf(xmin) || std::isnan(xmin) || std::isnan(xmax) || std::isnan(split_point))
|
1647
|
+
{
|
1648
|
+
this->multipliers.push_back(0);
|
1649
|
+
return;
|
1650
|
+
}
|
1651
|
+
|
1652
|
+
double range = std::fmax(xmax - xmin, std::numeric_limits<double>::min());
|
1653
|
+
double dleft = std::fmax(split_point - xmin, std::numeric_limits<double>::min());
|
1654
|
+
double dright = std::fmax(xmax - split_point, std::numeric_limits<double>::min());
|
1655
|
+
double mult_left = std::log(dleft / range);
|
1656
|
+
double mult_right = std::log(dright / range);
|
1657
|
+
while (std::isinf(mult_left))
|
1658
|
+
{
|
1659
|
+
dleft = std::nextafter(dleft, (mult_left < 0)? HUGE_VAL : (-HUGE_VAL));
|
1660
|
+
mult_left = std::log(dleft / range);
|
1661
|
+
}
|
1662
|
+
while (std::isinf(mult_right))
|
1663
|
+
{
|
1664
|
+
dright = std::nextafter(dright, (mult_right < 0)? HUGE_VAL : (-HUGE_VAL));
|
1665
|
+
mult_right = std::log(dright / range);
|
1666
|
+
}
|
1667
|
+
|
1668
|
+
mult_left = std::isnan(mult_left)? 0 : mult_left;
|
1669
|
+
mult_right = std::isnan(mult_right)? 0 : mult_right;
|
1670
|
+
|
1671
|
+
ldouble_safe curr = this->multipliers.back();
|
1672
|
+
this->multipliers.push_back(curr + mult_right);
|
1673
|
+
this->multipliers.push_back(curr + mult_left);
|
1674
|
+
}
|
1675
|
+
|
1676
|
+
template<class ldouble_safe, class real_t>
|
1677
|
+
void DensityCalculator<ldouble_safe, real_t>::push_density(int n_left, int n_present)
|
1678
|
+
{
|
1679
|
+
this->push_density(0., (double)n_present, (double)n_left);
|
1680
|
+
}
|
1681
|
+
|
1682
|
+
/* For single category splits */
|
1683
|
+
template<class ldouble_safe, class real_t>
|
1684
|
+
void DensityCalculator<ldouble_safe, real_t>::push_density(size_t counts[], int ncat)
|
1685
|
+
{
|
1686
|
+
/* this one assumes 'categ_present' has entries 0/1 for missing/present */
|
1687
|
+
int n_present = 0;
|
1688
|
+
for (int cat = 0; cat < ncat; cat++)
|
1689
|
+
n_present += counts[cat] > 0;
|
1690
|
+
this->push_density(0., (double)n_present, 1.);
|
1691
|
+
}
|
1692
|
+
|
1693
|
+
/* For single category splits */
|
1694
|
+
template<class ldouble_safe, class real_t>
|
1695
|
+
void DensityCalculator<ldouble_safe, real_t>::push_density(int n_present)
|
1696
|
+
{
|
1697
|
+
this->push_density(0., (double)n_present, 1.);
|
1698
|
+
}
|
1699
|
+
|
1700
|
+
/* For binary categorical splits */
|
1701
|
+
template<class ldouble_safe, class real_t>
|
1702
|
+
void DensityCalculator<ldouble_safe, real_t>::push_density()
|
1703
|
+
{
|
1704
|
+
this->multipliers.push_back(0);
|
1705
|
+
this->multipliers.push_back(0);
|
1706
|
+
}
|
1707
|
+
|
1708
|
+
template<class ldouble_safe, class real_t>
|
1709
|
+
void DensityCalculator<ldouble_safe, real_t>::push_adj(double xmin, double xmax, double split_point, double pct_tree_left, ScoringMetric scoring_metric)
|
1710
|
+
{
|
1711
|
+
double range = std::fmax(xmax - xmin, std::numeric_limits<double>::min());
|
1712
|
+
double dleft = std::fmax(split_point - xmin, std::numeric_limits<double>::min());
|
1713
|
+
double dright = std::fmax(xmax - split_point, std::numeric_limits<double>::min());
|
1714
|
+
double chunk_left = dleft / range;
|
1715
|
+
double chunk_right = dright / range;
|
1716
|
+
if (std::isinf(xmax) || std::isinf(xmin) || std::isnan(xmin) || std::isnan(xmax) || std::isnan(split_point))
|
1717
|
+
{
|
1718
|
+
chunk_left = pct_tree_left;
|
1719
|
+
chunk_right = 1. - pct_tree_left;
|
1720
|
+
goto add_chunks;
|
1721
|
+
}
|
1722
|
+
|
1723
|
+
if (std::isnan(chunk_left) || std::isnan(chunk_right))
|
1724
|
+
{
|
1725
|
+
chunk_left = 0.5;
|
1726
|
+
chunk_right = 0.5;
|
1727
|
+
}
|
1728
|
+
|
1729
|
+
chunk_left = pct_tree_left / chunk_left;
|
1730
|
+
chunk_right = (1. - pct_tree_left) / chunk_right;
|
1731
|
+
|
1732
|
+
add_chunks:
|
1733
|
+
chunk_left = 2. / (1. + .5/chunk_left);
|
1734
|
+
chunk_right = 2. / (1. + .5/chunk_right);
|
1735
|
+
// chunk_left = 2. / (1. + 1./chunk_left);
|
1736
|
+
// chunk_right = 2. / (1. + 1./chunk_right);
|
1737
|
+
// chunk_left = 2. - std::exp2(1. - chunk_left);
|
1738
|
+
// chunk_right = 2. - std::exp2(1. - chunk_right);
|
1739
|
+
|
1740
|
+
ldouble_safe curr = this->multipliers.back();
|
1741
|
+
if (scoring_metric == AdjDepth)
|
1742
|
+
{
|
1743
|
+
this->multipliers.push_back(curr + chunk_right);
|
1744
|
+
this->multipliers.push_back(curr + chunk_left);
|
1745
|
+
}
|
1746
|
+
|
1747
|
+
else
|
1748
|
+
{
|
1749
|
+
this->multipliers.push_back(std::fmax(curr * chunk_right, (ldouble_safe)std::numeric_limits<double>::epsilon()));
|
1750
|
+
this->multipliers.push_back(std::fmax(curr * chunk_left, (ldouble_safe)std::numeric_limits<double>::epsilon()));
|
1751
|
+
}
|
1752
|
+
}
|
1753
|
+
|
1754
|
+
template<class ldouble_safe, class real_t>
|
1755
|
+
void DensityCalculator<ldouble_safe, real_t>::push_adj(signed char *restrict categ_present, size_t *restrict counts, int ncat, ScoringMetric scoring_metric)
|
1756
|
+
{
|
1757
|
+
/* this one assumes 'categ_present' has entries -1/0/1 for missing/right/left */
|
1758
|
+
int cnt_cat_left = 0;
|
1759
|
+
int cnt_cat = 0;
|
1760
|
+
size_t cnt = 0;
|
1761
|
+
size_t cnt_left = 0;
|
1762
|
+
for (int cat = 0; cat < ncat; cat++)
|
1763
|
+
{
|
1764
|
+
if (counts[cat] > 0)
|
1765
|
+
{
|
1766
|
+
cnt += counts[cat];
|
1767
|
+
cnt_cat_left += categ_present[cat];
|
1768
|
+
cnt_left += categ_present[cat]? counts[cat] : 0;
|
1769
|
+
cnt_cat++;
|
1770
|
+
}
|
1771
|
+
}
|
1772
|
+
|
1773
|
+
double pct_tree_left = (ldouble_safe)cnt_left / (ldouble_safe)cnt;
|
1774
|
+
this->push_adj(0., (double)cnt_cat, (double)cnt_cat_left, pct_tree_left, scoring_metric);
|
1775
|
+
}
|
1776
|
+
|
1777
|
+
/* For single category splits */
|
1778
|
+
template<class ldouble_safe, class real_t>
|
1779
|
+
void DensityCalculator<ldouble_safe, real_t>::push_adj(size_t *restrict counts, int ncat, int chosen_cat, ScoringMetric scoring_metric)
|
1780
|
+
{
|
1781
|
+
/* this one assumes 'categ_present' has entries 0/1 for missing/present */
|
1782
|
+
int cnt_cat = 0;
|
1783
|
+
size_t cnt = 0;
|
1784
|
+
for (int cat = 0; cat < ncat; cat++)
|
1785
|
+
{
|
1786
|
+
cnt += counts[cat];
|
1787
|
+
cnt_cat += counts[cat] > 0;
|
1788
|
+
}
|
1789
|
+
|
1790
|
+
double pct_tree_left = (ldouble_safe)counts[chosen_cat] / (ldouble_safe)cnt;
|
1791
|
+
this->push_adj(0., (double)cnt_cat, 1., pct_tree_left, scoring_metric);
|
1792
|
+
}
|
1793
|
+
|
1794
|
+
/* For binary categorical splits */
|
1795
|
+
template<class ldouble_safe, class real_t>
|
1796
|
+
void DensityCalculator<ldouble_safe, real_t>::push_adj(double pct_tree_left, ScoringMetric scoring_metric)
|
1797
|
+
{
|
1798
|
+
this->push_adj(0., 1., 0.5, pct_tree_left, scoring_metric);
|
1799
|
+
}
|
1800
|
+
|
1801
|
+
template<class ldouble_safe, class real_t>
|
1802
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens(double split_point, size_t col)
|
1803
|
+
{
|
1804
|
+
if (this->fast_bratio)
|
1805
|
+
this->push_bdens_fast_route(split_point, col);
|
1806
|
+
else
|
1807
|
+
this->push_bdens_internal(split_point, col);
|
1808
|
+
}
|
1809
|
+
|
1810
|
+
template<class ldouble_safe, class real_t>
|
1811
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens_internal(double split_point, size_t col)
|
1812
|
+
{
|
1813
|
+
this->queue_box.push_back(this->box_high[col]);
|
1814
|
+
this->box_high[col] = split_point;
|
1815
|
+
}
|
1816
|
+
|
1817
|
+
template<class ldouble_safe, class real_t>
|
1818
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens_fast_route(double split_point, size_t col)
|
1819
|
+
{
|
1820
|
+
ldouble_safe curr_range = (ldouble_safe)this->box_high[col] - (ldouble_safe)this->box_low[col];
|
1821
|
+
ldouble_safe fraction_left = ((ldouble_safe)split_point - (ldouble_safe)this->box_low[col]) / curr_range;
|
1822
|
+
ldouble_safe fraction_right = ((ldouble_safe)this->box_high[col] - (ldouble_safe)split_point) / curr_range;
|
1823
|
+
fraction_left = std::fmax(fraction_left, (ldouble_safe)std::numeric_limits<double>::min());
|
1824
|
+
fraction_left = std::fmin(fraction_left, (ldouble_safe)(1. - std::numeric_limits<double>::epsilon()));
|
1825
|
+
fraction_left = std::log(fraction_left);
|
1826
|
+
fraction_left += this->multipliers.back();
|
1827
|
+
fraction_right = std::fmax(fraction_right, (ldouble_safe)std::numeric_limits<double>::min());
|
1828
|
+
fraction_right = std::fmin(fraction_right, (ldouble_safe)(1. - std::numeric_limits<double>::epsilon()));
|
1829
|
+
fraction_right = std::log(fraction_right);
|
1830
|
+
fraction_right += this->multipliers.back();
|
1831
|
+
this->multipliers.push_back(fraction_right);
|
1832
|
+
this->multipliers.push_back(fraction_left);
|
1833
|
+
|
1834
|
+
this->push_bdens_internal(split_point, col);
|
1835
|
+
}
|
1836
|
+
|
1837
|
+
template<class ldouble_safe, class real_t>
|
1838
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens(int ncat_branch_left, size_t col)
|
1839
|
+
{
|
1840
|
+
if (this->fast_bratio)
|
1841
|
+
this->push_bdens_fast_route(ncat_branch_left, col);
|
1842
|
+
else
|
1843
|
+
this->push_bdens_internal(ncat_branch_left, col);
|
1844
|
+
}
|
1845
|
+
|
1846
|
+
template<class ldouble_safe, class real_t>
|
1847
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens_internal(int ncat_branch_left, size_t col)
|
1848
|
+
{
|
1849
|
+
this->queue_ncat.push_back(this->ncat[col]);
|
1850
|
+
this->ncat[col] = ncat_branch_left;
|
1851
|
+
}
|
1852
|
+
|
1853
|
+
template<class ldouble_safe, class real_t>
|
1854
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens_fast_route(int ncat_branch_left, size_t col)
|
1855
|
+
{
|
1856
|
+
double fraction_left = std::log((double)ncat_branch_left / this->ncat[col]);
|
1857
|
+
double fraction_right = std::log((double)(this->ncat[col] - ncat_branch_left) / this->ncat[col]);
|
1858
|
+
ldouble_safe curr = this->multipliers.back();
|
1859
|
+
this->multipliers.push_back(curr + fraction_right);
|
1860
|
+
this->multipliers.push_back(curr + fraction_left);
|
1861
|
+
|
1862
|
+
this->push_bdens_internal(ncat_branch_left, col);
|
1863
|
+
}
|
1864
|
+
|
1865
|
+
template<class ldouble_safe, class real_t>
|
1866
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens(const std::vector<signed char> &cat_split, size_t col)
|
1867
|
+
{
|
1868
|
+
if (this->fast_bratio)
|
1869
|
+
this->push_bdens_fast_route(cat_split, col);
|
1870
|
+
else
|
1871
|
+
this->push_bdens_internal(cat_split, col);
|
1872
|
+
}
|
1873
|
+
|
1874
|
+
template<class ldouble_safe, class real_t>
|
1875
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens_internal(const std::vector<signed char> &cat_split, size_t col)
|
1876
|
+
{
|
1877
|
+
int ncat_branch_left = 0;
|
1878
|
+
for (auto el : cat_split)
|
1879
|
+
ncat_branch_left += el == 1;
|
1880
|
+
this->push_bdens_internal(ncat_branch_left, col);
|
1881
|
+
}
|
1882
|
+
|
1883
|
+
template<class ldouble_safe, class real_t>
|
1884
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens_fast_route(const std::vector<signed char> &cat_split, size_t col)
|
1885
|
+
{
|
1886
|
+
int ncat_branch_left = 0;
|
1887
|
+
for (auto el : cat_split)
|
1888
|
+
ncat_branch_left += el == 1;
|
1889
|
+
this->push_bdens_fast_route(ncat_branch_left, col);
|
1890
|
+
}
|
1891
|
+
|
1892
|
+
template<class ldouble_safe, class real_t>
|
1893
|
+
void DensityCalculator<ldouble_safe, real_t>::push_bdens_ext(const IsoHPlane &hplane, const ModelParams &model_params)
|
1894
|
+
{
|
1895
|
+
double x1, x2;
|
1896
|
+
double xlow = 0, xhigh = 0;
|
1897
|
+
size_t col;
|
1898
|
+
size_t col_num = 0;
|
1899
|
+
size_t col_cat = 0;
|
1900
|
+
|
1901
|
+
for (size_t col_outer = 0; col_outer < hplane.col_num.size(); col_outer++)
|
1902
|
+
{
|
1903
|
+
switch (hplane.col_type[col_outer])
|
1904
|
+
{
|
1905
|
+
case Numeric:
|
1906
|
+
{
|
1907
|
+
col = hplane.col_num[col_outer];
|
1908
|
+
x1 = hplane.coef[col_num] * (this->box_low[col] - hplane.mean[col_num]);
|
1909
|
+
x2 = hplane.coef[col_num] * (this->box_high[col] - hplane.mean[col_num]);
|
1910
|
+
xlow += std::fmin(x1, x2);
|
1911
|
+
xhigh += std::fmax(x1, x2);
|
1912
|
+
break;
|
1913
|
+
}
|
1914
|
+
|
1915
|
+
case Categorical:
|
1916
|
+
{
|
1917
|
+
switch (model_params.cat_split_type)
|
1918
|
+
{
|
1919
|
+
case SingleCateg:
|
1920
|
+
{
|
1921
|
+
xlow += std::fmin(hplane.fill_new[col_cat], 0.);
|
1922
|
+
xhigh += std::fmax(hplane.fill_new[col_cat], 0.);
|
1923
|
+
break;
|
1924
|
+
}
|
1925
|
+
|
1926
|
+
case SubSet:
|
1927
|
+
{
|
1928
|
+
xlow += *std::min_element(hplane.cat_coef[col_cat].begin(), hplane.cat_coef[col_cat].end());
|
1929
|
+
xhigh += *std::max_element(hplane.cat_coef[col_cat].begin(), hplane.cat_coef[col_cat].end());
|
1930
|
+
break;
|
1931
|
+
}
|
1932
|
+
}
|
1933
|
+
break;
|
1934
|
+
}
|
1935
|
+
|
1936
|
+
default:
|
1937
|
+
{
|
1938
|
+
assert(0);
|
1939
|
+
}
|
1940
|
+
}
|
1941
|
+
}
|
1942
|
+
|
1943
|
+
double chunk_left;
|
1944
|
+
double chunk_right;
|
1945
|
+
double xdiff = xhigh - xlow;
|
1946
|
+
|
1947
|
+
if (model_params.scoring_metric != BoxedDensity)
|
1948
|
+
{
|
1949
|
+
chunk_left = (hplane.split_point - xlow) / xdiff;
|
1950
|
+
chunk_right = (xhigh - hplane.split_point) / xdiff;
|
1951
|
+
chunk_left = std::fmin(chunk_left, std::numeric_limits<double>::min());
|
1952
|
+
chunk_left = std::fmax(chunk_left, 1.-std::numeric_limits<double>::epsilon());
|
1953
|
+
chunk_right = std::fmin(chunk_right, std::numeric_limits<double>::min());
|
1954
|
+
chunk_right = std::fmax(chunk_right, 1.-std::numeric_limits<double>::epsilon());
|
1955
|
+
}
|
1956
|
+
|
1957
|
+
else
|
1958
|
+
{
|
1959
|
+
chunk_left = xdiff / (hplane.split_point - xlow);
|
1960
|
+
chunk_right = xdiff / (xhigh - hplane.split_point);
|
1961
|
+
chunk_left = std::fmin(chunk_left, 1.);
|
1962
|
+
chunk_right = std::fmin(chunk_right, 1.);
|
1963
|
+
}
|
1964
|
+
|
1965
|
+
this->queue_ext_box.push_back(std::log(chunk_right) + this->vals_ext_box.back());
|
1966
|
+
this->vals_ext_box.push_back(std::log(chunk_left) + this->vals_ext_box.back());
|
1967
|
+
}
|
1968
|
+
|
1969
|
+
template<class ldouble_safe, class real_t>
|
1970
|
+
void DensityCalculator<ldouble_safe, real_t>::pop()
|
1971
|
+
{
|
1972
|
+
this->multipliers.pop_back();
|
1973
|
+
}
|
1974
|
+
|
1975
|
+
template<class ldouble_safe, class real_t>
|
1976
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_right()
|
1977
|
+
{
|
1978
|
+
this->multipliers.pop_back();
|
1979
|
+
}
|
1980
|
+
|
1981
|
+
template<class ldouble_safe, class real_t>
|
1982
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens(size_t col)
|
1983
|
+
{
|
1984
|
+
if (this->fast_bratio)
|
1985
|
+
this->pop_bdens_fast_route(col);
|
1986
|
+
else
|
1987
|
+
this->pop_bdens_internal(col);
|
1988
|
+
}
|
1989
|
+
|
1990
|
+
template<class ldouble_safe, class real_t>
|
1991
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_internal(size_t col)
|
1992
|
+
{
|
1993
|
+
double old_high = this->queue_box.back();
|
1994
|
+
this->queue_box.pop_back();
|
1995
|
+
this->queue_box.push_back(this->box_low[col]);
|
1996
|
+
this->box_low[col] = this->box_high[col];
|
1997
|
+
this->box_high[col] = old_high;
|
1998
|
+
}
|
1999
|
+
|
2000
|
+
template<class ldouble_safe, class real_t>
|
2001
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_fast_route(size_t col)
|
2002
|
+
{
|
2003
|
+
this->multipliers.pop_back();
|
2004
|
+
this->pop_bdens_internal(col);
|
2005
|
+
}
|
2006
|
+
|
2007
|
+
template<class ldouble_safe, class real_t>
|
2008
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_right(size_t col)
|
2009
|
+
{
|
2010
|
+
if (this->fast_bratio)
|
2011
|
+
this->pop_bdens_right_fast_route(col);
|
2012
|
+
else
|
2013
|
+
this->pop_bdens_right_internal(col);
|
2014
|
+
}
|
2015
|
+
|
2016
|
+
template<class ldouble_safe, class real_t>
|
2017
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_right_internal(size_t col)
|
2018
|
+
{
|
2019
|
+
double old_low = this->queue_box.back();
|
2020
|
+
this->queue_box.pop_back();
|
2021
|
+
this->box_low[col] = old_low;
|
2022
|
+
}
|
2023
|
+
|
2024
|
+
template<class ldouble_safe, class real_t>
|
2025
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_right_fast_route(size_t col)
|
2026
|
+
{
|
2027
|
+
this->multipliers.pop_back();
|
2028
|
+
this->pop_bdens_right_internal(col);
|
2029
|
+
}
|
2030
|
+
|
2031
|
+
template<class ldouble_safe, class real_t>
|
2032
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_cat(size_t col)
|
2033
|
+
{
|
2034
|
+
if (this->fast_bratio)
|
2035
|
+
this->pop_bdens_cat_fast_route(col);
|
2036
|
+
else
|
2037
|
+
this->pop_bdens_cat_internal(col);
|
2038
|
+
}
|
2039
|
+
|
2040
|
+
template<class ldouble_safe, class real_t>
|
2041
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_cat_internal(size_t col)
|
2042
|
+
{
|
2043
|
+
int old_ncat = this->queue_ncat.back();
|
2044
|
+
this->ncat[col] = old_ncat - this->ncat[col];
|
2045
|
+
}
|
2046
|
+
|
2047
|
+
template<class ldouble_safe, class real_t>
|
2048
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_cat_fast_route(size_t col)
|
2049
|
+
{
|
2050
|
+
this->multipliers.pop_back();
|
2051
|
+
this->pop_bdens_cat_internal(col);
|
2052
|
+
}
|
2053
|
+
|
2054
|
+
template<class ldouble_safe, class real_t>
|
2055
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_cat_right(size_t col)
|
2056
|
+
{
|
2057
|
+
if (this->fast_bratio)
|
2058
|
+
this->pop_bdens_cat_right_fast_route(col);
|
2059
|
+
else
|
2060
|
+
this->pop_bdens_cat_right_internal(col);
|
2061
|
+
}
|
2062
|
+
|
2063
|
+
template<class ldouble_safe, class real_t>
|
2064
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_cat_right_internal(size_t col)
|
2065
|
+
{
|
2066
|
+
int old_ncat = this->queue_ncat.back();
|
2067
|
+
this->queue_ncat.pop_back();
|
2068
|
+
this->ncat[col] = old_ncat;
|
2069
|
+
}
|
2070
|
+
|
2071
|
+
template<class ldouble_safe, class real_t>
|
2072
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_cat_right_fast_route(size_t col)
|
2073
|
+
{
|
2074
|
+
this->multipliers.pop_back();
|
2075
|
+
this->pop_bdens_cat_right_internal(col);
|
2076
|
+
}
|
2077
|
+
|
2078
|
+
template<class ldouble_safe, class real_t>
|
2079
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_ext()
|
2080
|
+
{
|
2081
|
+
this->vals_ext_box.pop_back();
|
2082
|
+
this->vals_ext_box.push_back(this->queue_ext_box.back());
|
2083
|
+
this->queue_ext_box.pop_back();
|
2084
|
+
}
|
2085
|
+
|
2086
|
+
template<class ldouble_safe, class real_t>
|
2087
|
+
void DensityCalculator<ldouble_safe, real_t>::pop_bdens_ext_right()
|
2088
|
+
{
|
2089
|
+
this->vals_ext_box.pop_back();
|
2090
|
+
}
|
2091
|
+
|
2092
|
+
/* this outputs the logarithm of the density */
|
2093
|
+
template<class ldouble_safe, class real_t>
|
2094
|
+
double DensityCalculator<ldouble_safe, real_t>::calc_density(ldouble_safe remainder, size_t sample_size)
|
2095
|
+
{
|
2096
|
+
return std::log(remainder) - std::log((ldouble_safe)sample_size) - this->multipliers.back();
|
2097
|
+
}
|
2098
|
+
|
2099
|
+
template<class ldouble_safe, class real_t>
|
2100
|
+
ldouble_safe DensityCalculator<ldouble_safe, real_t>::calc_adj_depth()
|
2101
|
+
{
|
2102
|
+
ldouble_safe out = this->multipliers.back();
|
2103
|
+
return std::fmax(out, (ldouble_safe)std::numeric_limits<double>::min());
|
2104
|
+
}
|
2105
|
+
|
2106
|
+
template<class ldouble_safe, class real_t>
|
2107
|
+
double DensityCalculator<ldouble_safe, real_t>::calc_adj_density()
|
2108
|
+
{
|
2109
|
+
return this->multipliers.back();
|
2110
|
+
}
|
2111
|
+
|
2112
|
+
/* this outputs the logarithm of the density */
|
2113
|
+
template<class ldouble_safe, class real_t>
|
2114
|
+
ldouble_safe DensityCalculator<ldouble_safe, real_t>::calc_bratio_inv_log()
|
2115
|
+
{
|
2116
|
+
if (!this->multipliers.empty())
|
2117
|
+
return -this->multipliers.back();
|
2118
|
+
|
2119
|
+
ldouble_safe sum_log_switdh = 0;
|
2120
|
+
ldouble_safe ratio_col;
|
2121
|
+
for (size_t col = 0; col < this->ranges.size(); col++)
|
2122
|
+
{
|
2123
|
+
if (!this->ranges[col]) continue;
|
2124
|
+
ratio_col = this->ranges[col] / ((ldouble_safe)this->box_high[col] - (ldouble_safe)this->box_low[col]);
|
2125
|
+
ratio_col = std::fmax(ratio_col, (ldouble_safe)1);
|
2126
|
+
sum_log_switdh += std::log(ratio_col);
|
2127
|
+
}
|
2128
|
+
|
2129
|
+
for (size_t col = 0; col < this->ncat.size(); col++)
|
2130
|
+
{
|
2131
|
+
if (this->ncat_orig[col] <= 1) continue;
|
2132
|
+
sum_log_switdh += std::log((double)this->ncat_orig[col] / (double)this->ncat[col]);
|
2133
|
+
}
|
2134
|
+
|
2135
|
+
return sum_log_switdh;
|
2136
|
+
}
|
2137
|
+
|
2138
|
+
template<class ldouble_safe, class real_t>
|
2139
|
+
ldouble_safe DensityCalculator<ldouble_safe, real_t>::calc_bratio_log()
|
2140
|
+
{
|
2141
|
+
if (!this->multipliers.empty())
|
2142
|
+
return this->multipliers.back();
|
2143
|
+
|
2144
|
+
ldouble_safe sum_log_switdh = 0;
|
2145
|
+
ldouble_safe ratio_col;
|
2146
|
+
for (size_t col = 0; col < this->ranges.size(); col++)
|
2147
|
+
{
|
2148
|
+
if (!this->ranges[col]) continue;
|
2149
|
+
ratio_col = ((ldouble_safe)this->box_high[col] - (ldouble_safe)this->box_low[col]) / this->ranges[col];
|
2150
|
+
ratio_col = std::fmax(ratio_col, (ldouble_safe)std::numeric_limits<double>::min());
|
2151
|
+
ratio_col = std::fmin(ratio_col, (ldouble_safe)(1. - std::numeric_limits<double>::epsilon()));
|
2152
|
+
sum_log_switdh += std::log(ratio_col);
|
2153
|
+
}
|
2154
|
+
|
2155
|
+
for (size_t col = 0; col < this->ncat.size(); col++)
|
2156
|
+
{
|
2157
|
+
if (this->ncat_orig[col] <= 1) continue;
|
2158
|
+
sum_log_switdh += std::log((double)this->ncat[col] / (double)this->ncat_orig[col]);
|
2159
|
+
}
|
2160
|
+
|
2161
|
+
return sum_log_switdh;
|
2162
|
+
}
|
2163
|
+
|
2164
|
+
/* this does NOT output the logarithm of the density */
|
2165
|
+
template<class ldouble_safe, class real_t>
|
2166
|
+
double DensityCalculator<ldouble_safe, real_t>::calc_bratio()
|
2167
|
+
{
|
2168
|
+
return std::exp(this->calc_bratio_log());
|
2169
|
+
}
|
2170
|
+
|
2171
|
+
const double MIN_DENS = std::log(std::numeric_limits<double>::min());
|
2172
|
+
|
2173
|
+
/* this outputs the logarithm of the density */
|
2174
|
+
template<class ldouble_safe, class real_t>
|
2175
|
+
double DensityCalculator<ldouble_safe, real_t>::calc_bdens(ldouble_safe remainder, size_t sample_size)
|
2176
|
+
{
|
2177
|
+
double out = std::log(remainder) - std::log((ldouble_safe)sample_size) - this->calc_bratio_inv_log();
|
2178
|
+
return std::fmax(out, MIN_DENS);
|
2179
|
+
}
|
2180
|
+
|
2181
|
+
/* this outputs the logarithm of the density */
|
2182
|
+
template<class ldouble_safe, class real_t>
|
2183
|
+
double DensityCalculator<ldouble_safe, real_t>::calc_bdens2(ldouble_safe remainder, size_t sample_size)
|
2184
|
+
{
|
2185
|
+
double out = std::log(remainder) - std::log((ldouble_safe)sample_size) - this->calc_bratio_log();
|
2186
|
+
return std::fmax(out, MIN_DENS);
|
2187
|
+
}
|
2188
|
+
|
2189
|
+
/* this outputs the logarithm of the density */
|
2190
|
+
template<class ldouble_safe, class real_t>
|
2191
|
+
ldouble_safe DensityCalculator<ldouble_safe, real_t>::calc_bratio_log_ext()
|
2192
|
+
{
|
2193
|
+
return this->vals_ext_box.back();
|
2194
|
+
}
|
2195
|
+
|
2196
|
+
template<class ldouble_safe, class real_t>
|
2197
|
+
double DensityCalculator<ldouble_safe, real_t>::calc_bratio_ext()
|
2198
|
+
{
|
2199
|
+
double out = std::exp(this->calc_bratio_log_ext());
|
2200
|
+
return std::fmax(out, std::numeric_limits<double>::min());
|
2201
|
+
}
|
2202
|
+
|
2203
|
+
/* this outputs the logarithm of the density */
|
2204
|
+
template<class ldouble_safe, class real_t>
|
2205
|
+
double DensityCalculator<ldouble_safe, real_t>::calc_bdens_ext(ldouble_safe remainder, size_t sample_size)
|
2206
|
+
{
|
2207
|
+
double out = std::log(remainder) - std::log((ldouble_safe)sample_size) - this->calc_bratio_log_ext();
|
2208
|
+
return std::fmax(out, MIN_DENS);
|
2209
|
+
}
|
2210
|
+
|
2211
|
+
template<class ldouble_safe, class real_t>
|
2212
|
+
void DensityCalculator<ldouble_safe, real_t>::save_range(double xmin, double xmax)
|
2213
|
+
{
|
2214
|
+
this->xmin = xmin;
|
2215
|
+
this->xmax = xmax;
|
2216
|
+
}
|
2217
|
+
|
2218
|
+
template<class ldouble_safe, class real_t>
|
2219
|
+
void DensityCalculator<ldouble_safe, real_t>::restore_range(double &restrict xmin, double &restrict xmax)
|
2220
|
+
{
|
2221
|
+
xmin = this->xmin;
|
2222
|
+
xmax = this->xmax;
|
2223
|
+
}
|
2224
|
+
|
2225
|
+
template<class ldouble_safe, class real_t>
|
2226
|
+
void DensityCalculator<ldouble_safe, real_t>::save_counts(size_t *restrict cat_counts, int ncat)
|
2227
|
+
{
|
2228
|
+
this->counts.assign(cat_counts, cat_counts + ncat);
|
2229
|
+
}
|
2230
|
+
|
2231
|
+
template<class ldouble_safe, class real_t>
|
2232
|
+
void DensityCalculator<ldouble_safe, real_t>::save_n_present_and_left(signed char *restrict split_left, int ncat)
|
2233
|
+
{
|
2234
|
+
this->n_present = 0;
|
2235
|
+
this->n_left = 0;
|
2236
|
+
for (int cat = 0; cat < ncat; cat++)
|
2237
|
+
{
|
2238
|
+
this->n_present += split_left[cat] >= 0;
|
2239
|
+
this->n_left += split_left[cat] == 1;
|
2240
|
+
}
|
2241
|
+
}
|
2242
|
+
|
2243
|
+
template<class ldouble_safe, class real_t>
|
2244
|
+
void DensityCalculator<ldouble_safe, real_t>::save_n_present(size_t *restrict cat_counts, int ncat)
|
2245
|
+
{
|
2246
|
+
this->n_present = 0;
|
2247
|
+
for (int cat = 0; cat < ncat; cat++)
|
2248
|
+
this->n_present += cat_counts[cat] > 0;
|
2249
|
+
}
|
2250
|
+
|
2251
|
+
/* For hyperplane intersections */
|
2252
|
+
size_t divide_subset_split(size_t ix_arr[], double x[], size_t st, size_t end, double split_point) noexcept
|
2253
|
+
{
|
2254
|
+
size_t temp;
|
2255
|
+
size_t st_orig = st;
|
2256
|
+
for (size_t row = st_orig; row <= end; row++)
|
2257
|
+
{
|
2258
|
+
if (x[row - st_orig] <= split_point)
|
2259
|
+
{
|
2260
|
+
temp = ix_arr[st];
|
2261
|
+
ix_arr[st] = ix_arr[row];
|
2262
|
+
ix_arr[row] = temp;
|
2263
|
+
st++;
|
2264
|
+
}
|
2265
|
+
}
|
2266
|
+
return st;
|
2267
|
+
}
|
2268
|
+
|
2269
|
+
/* For numerical columns */
|
2270
|
+
template <class real_t>
|
2271
|
+
void divide_subset_split(size_t *restrict ix_arr, real_t x[], size_t st, size_t end, double split_point,
|
2272
|
+
MissingAction missing_action, size_t &restrict st_NA, size_t &restrict end_NA, size_t &restrict split_ix) noexcept
|
2273
|
+
{
|
2274
|
+
size_t temp;
|
2275
|
+
|
2276
|
+
/* if NAs are not to be bothered with, just need to do a single pass */
|
2277
|
+
if (missing_action == Fail)
|
2278
|
+
{
|
2279
|
+
/* move to the left if it's l.e. split point */
|
2280
|
+
for (size_t row = st; row <= end; row++)
|
2281
|
+
{
|
2282
|
+
if (x[ix_arr[row]] <= split_point)
|
2283
|
+
{
|
2284
|
+
temp = ix_arr[st];
|
2285
|
+
ix_arr[st] = ix_arr[row];
|
2286
|
+
ix_arr[row] = temp;
|
2287
|
+
st++;
|
2288
|
+
}
|
2289
|
+
}
|
2290
|
+
split_ix = st;
|
2291
|
+
}
|
2292
|
+
|
2293
|
+
/* otherwise, first put to the left all l.e. and not NA, then all NAs to the end of the left */
|
2294
|
+
else
|
2295
|
+
{
|
2296
|
+
for (size_t row = st; row <= end; row++)
|
2297
|
+
{
|
2298
|
+
if (!std::isnan(x[ix_arr[row]]) && x[ix_arr[row]] <= split_point)
|
2299
|
+
{
|
2300
|
+
temp = ix_arr[st];
|
2301
|
+
ix_arr[st] = ix_arr[row];
|
2302
|
+
ix_arr[row] = temp;
|
2303
|
+
st++;
|
2304
|
+
}
|
2305
|
+
}
|
2306
|
+
st_NA = st;
|
2307
|
+
|
2308
|
+
for (size_t row = st; row <= end; row++)
|
2309
|
+
{
|
2310
|
+
if (unlikely(std::isnan(x[ix_arr[row]])))
|
2311
|
+
{
|
2312
|
+
temp = ix_arr[st];
|
2313
|
+
ix_arr[st] = ix_arr[row];
|
2314
|
+
ix_arr[row] = temp;
|
2315
|
+
st++;
|
2316
|
+
}
|
2317
|
+
}
|
2318
|
+
end_NA = st;
|
2319
|
+
}
|
2320
|
+
}
|
2321
|
+
|
2322
|
+
/* For sparse numeric columns */
|
2323
|
+
template <class real_t, class sparse_ix>
|
2324
|
+
void divide_subset_split(size_t *restrict ix_arr, size_t st, size_t end, size_t col_num,
|
2325
|
+
real_t Xc[], sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr, double split_point,
|
2326
|
+
MissingAction missing_action, size_t &restrict st_NA, size_t &restrict end_NA, size_t &restrict split_ix) noexcept
|
2327
|
+
{
|
2328
|
+
/* TODO: this is a mess, needs refactoring */
|
2329
|
+
/* TODO: when moving zeros, would be better to instead move by '>' (opposite as in here) */
|
2330
|
+
/* TODO: should create an extra version to go along with 'predict' that would
|
2331
|
+
add the range penalty right here to spare operations. */
|
2332
|
+
if (Xc_indptr[col_num] == Xc_indptr[col_num + 1])
|
2333
|
+
{
|
2334
|
+
if (missing_action == Fail)
|
2335
|
+
{
|
2336
|
+
split_ix = (0 <= split_point)? (end+1) : st;
|
2337
|
+
}
|
2338
|
+
|
2339
|
+
else
|
2340
|
+
{
|
2341
|
+
st_NA = (0 <= split_point)? (end+1) : st;
|
2342
|
+
end_NA = (0 <= split_point)? (end+1) : st;
|
2343
|
+
}
|
2344
|
+
|
2345
|
+
}
|
2346
|
+
|
2347
|
+
size_t st_col = Xc_indptr[col_num];
|
2348
|
+
size_t end_col = Xc_indptr[col_num + 1] - 1;
|
2349
|
+
size_t curr_pos = st_col;
|
2350
|
+
size_t ind_end_col = Xc_ind[end_col];
|
2351
|
+
size_t temp;
|
2352
|
+
bool move_zeros = 0 <= split_point;
|
2353
|
+
size_t *ptr_st = std::lower_bound(ix_arr + st, ix_arr + end + 1, Xc_ind[st_col]);
|
2354
|
+
|
2355
|
+
if (move_zeros && ptr_st > ix_arr + st)
|
2356
|
+
st = ptr_st - ix_arr;
|
2357
|
+
|
2358
|
+
if (missing_action == Fail)
|
2359
|
+
{
|
2360
|
+
if (move_zeros)
|
2361
|
+
{
|
2362
|
+
for (size_t *row = ptr_st;
|
2363
|
+
row != ix_arr + end + 1;
|
2364
|
+
)
|
2365
|
+
{
|
2366
|
+
if (curr_pos >= end_col + 1)
|
2367
|
+
{
|
2368
|
+
for (size_t *r = row; r <= ix_arr + end; r++)
|
2369
|
+
{
|
2370
|
+
temp = ix_arr[st];
|
2371
|
+
ix_arr[st] = *r;
|
2372
|
+
*r = temp;
|
2373
|
+
st++;
|
2374
|
+
}
|
2375
|
+
break;
|
2376
|
+
}
|
2377
|
+
|
2378
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
2379
|
+
{
|
2380
|
+
if (Xc[curr_pos] <= split_point)
|
2381
|
+
{
|
2382
|
+
temp = ix_arr[st];
|
2383
|
+
ix_arr[st] = *row;
|
2384
|
+
*row = temp;
|
2385
|
+
st++;
|
2386
|
+
}
|
2387
|
+
if (curr_pos == end_col && row < ix_arr + end)
|
2388
|
+
{
|
2389
|
+
for (size_t *r = row + 1; r <= ix_arr + end; r++)
|
2390
|
+
{
|
2391
|
+
temp = ix_arr[st];
|
2392
|
+
ix_arr[st] = *r;
|
2393
|
+
*r = temp;
|
2394
|
+
st++;
|
2395
|
+
}
|
2396
|
+
}
|
2397
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
2398
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
2399
|
+
}
|
2400
|
+
|
2401
|
+
else
|
2402
|
+
{
|
2403
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
2404
|
+
{
|
2405
|
+
while (row <= ix_arr + end && Xc_ind[curr_pos] > (sparse_ix)(*row))
|
2406
|
+
{
|
2407
|
+
temp = ix_arr[st];
|
2408
|
+
ix_arr[st] = *row;
|
2409
|
+
*row = temp;
|
2410
|
+
st++; row++;
|
2411
|
+
}
|
2412
|
+
}
|
2413
|
+
|
2414
|
+
else
|
2415
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
2416
|
+
}
|
2417
|
+
}
|
2418
|
+
}
|
2419
|
+
|
2420
|
+
else /* don't move zeros */
|
2421
|
+
{
|
2422
|
+
for (size_t *row = ptr_st;
|
2423
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
2424
|
+
)
|
2425
|
+
{
|
2426
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
2427
|
+
{
|
2428
|
+
if (Xc[curr_pos] <= split_point)
|
2429
|
+
{
|
2430
|
+
temp = ix_arr[st];
|
2431
|
+
ix_arr[st] = *row;
|
2432
|
+
*row = temp;
|
2433
|
+
st++;
|
2434
|
+
}
|
2435
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
2436
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
2437
|
+
}
|
2438
|
+
|
2439
|
+
else
|
2440
|
+
{
|
2441
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
2442
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
2443
|
+
else
|
2444
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
2445
|
+
}
|
2446
|
+
}
|
2447
|
+
}
|
2448
|
+
|
2449
|
+
split_ix = st;
|
2450
|
+
}
|
2451
|
+
|
2452
|
+
else /* can have NAs */
|
2453
|
+
{
|
2454
|
+
|
2455
|
+
bool has_NAs = false;
|
2456
|
+
if (move_zeros)
|
2457
|
+
{
|
2458
|
+
for (size_t *row = ptr_st;
|
2459
|
+
row != ix_arr + end + 1;
|
2460
|
+
)
|
2461
|
+
{
|
2462
|
+
if (curr_pos >= end_col + 1)
|
2463
|
+
{
|
2464
|
+
for (size_t *r = row; r <= ix_arr + end; r++)
|
2465
|
+
{
|
2466
|
+
temp = ix_arr[st];
|
2467
|
+
ix_arr[st] = *r;
|
2468
|
+
*r = temp;
|
2469
|
+
st++;
|
2470
|
+
}
|
2471
|
+
break;
|
2472
|
+
}
|
2473
|
+
|
2474
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
2475
|
+
{
|
2476
|
+
if (unlikely(std::isnan(Xc[curr_pos])))
|
2477
|
+
has_NAs = true;
|
2478
|
+
else if (Xc[curr_pos] <= split_point)
|
2479
|
+
{
|
2480
|
+
temp = ix_arr[st];
|
2481
|
+
ix_arr[st] = *row;
|
2482
|
+
*row = temp;
|
2483
|
+
st++;
|
2484
|
+
}
|
2485
|
+
if (curr_pos == end_col && row < ix_arr + end)
|
2486
|
+
for (size_t *r = row + 1; r <= ix_arr + end; r++)
|
2487
|
+
{
|
2488
|
+
temp = ix_arr[st];
|
2489
|
+
ix_arr[st] = *r;
|
2490
|
+
*r = temp;
|
2491
|
+
st++;
|
2492
|
+
}
|
2493
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
2494
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
2495
|
+
}
|
2496
|
+
|
2497
|
+
else
|
2498
|
+
{
|
2499
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
2500
|
+
{
|
2501
|
+
while (row <= ix_arr + end && Xc_ind[curr_pos] > (sparse_ix)(*row))
|
2502
|
+
{
|
2503
|
+
temp = ix_arr[st];
|
2504
|
+
ix_arr[st] = *row;
|
2505
|
+
*row = temp;
|
2506
|
+
st++; row++;
|
2507
|
+
}
|
2508
|
+
}
|
2509
|
+
|
2510
|
+
else
|
2511
|
+
{
|
2512
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
2513
|
+
}
|
2514
|
+
}
|
2515
|
+
}
|
2516
|
+
}
|
2517
|
+
|
2518
|
+
else /* don't move zeros */
|
2519
|
+
{
|
2520
|
+
for (size_t *row = ptr_st;
|
2521
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
2522
|
+
)
|
2523
|
+
{
|
2524
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
2525
|
+
{
|
2526
|
+
if (unlikely(std::isnan(Xc[curr_pos]))) has_NAs = true;
|
2527
|
+
if (!std::isnan(Xc[curr_pos]) && Xc[curr_pos] <= split_point)
|
2528
|
+
{
|
2529
|
+
temp = ix_arr[st];
|
2530
|
+
ix_arr[st] = *row;
|
2531
|
+
*row = temp;
|
2532
|
+
st++;
|
2533
|
+
}
|
2534
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
2535
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
2536
|
+
}
|
2537
|
+
|
2538
|
+
else
|
2539
|
+
{
|
2540
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
2541
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
2542
|
+
else
|
2543
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
2544
|
+
}
|
2545
|
+
}
|
2546
|
+
}
|
2547
|
+
|
2548
|
+
|
2549
|
+
st_NA = st;
|
2550
|
+
if (has_NAs)
|
2551
|
+
{
|
2552
|
+
curr_pos = st_col;
|
2553
|
+
std::sort(ix_arr + st, ix_arr + end + 1);
|
2554
|
+
for (size_t *row = ix_arr + st;
|
2555
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
2556
|
+
)
|
2557
|
+
{
|
2558
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
2559
|
+
{
|
2560
|
+
if (unlikely(std::isnan(Xc[curr_pos])))
|
2561
|
+
{
|
2562
|
+
temp = ix_arr[st];
|
2563
|
+
ix_arr[st] = *row;
|
2564
|
+
*row = temp;
|
2565
|
+
st++;
|
2566
|
+
}
|
2567
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
2568
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
2569
|
+
}
|
2570
|
+
|
2571
|
+
else
|
2572
|
+
{
|
2573
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
2574
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
2575
|
+
else
|
2576
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
2577
|
+
}
|
2578
|
+
}
|
2579
|
+
}
|
2580
|
+
end_NA = st;
|
2581
|
+
|
2582
|
+
}
|
2583
|
+
|
2584
|
+
}
|
2585
|
+
|
2586
|
+
/* For categorical columns split by subset */
|
2587
|
+
void divide_subset_split(size_t *restrict ix_arr, int x[], size_t st, size_t end, signed char split_categ[],
|
2588
|
+
MissingAction missing_action, size_t &restrict st_NA, size_t &restrict end_NA, size_t &restrict split_ix) noexcept
|
2589
|
+
{
|
2590
|
+
size_t temp;
|
2591
|
+
|
2592
|
+
/* if NAs are not to be bothered with, just need to do a single pass */
|
2593
|
+
if (missing_action == Fail)
|
2594
|
+
{
|
2595
|
+
/* move to the left if it's l.e. than the split point */
|
2596
|
+
for (size_t row = st; row <= end; row++)
|
2597
|
+
{
|
2598
|
+
if (split_categ[ x[ix_arr[row]] ] == 1)
|
2599
|
+
{
|
2600
|
+
temp = ix_arr[st];
|
2601
|
+
ix_arr[st] = ix_arr[row];
|
2602
|
+
ix_arr[row] = temp;
|
2603
|
+
st++;
|
2604
|
+
}
|
2605
|
+
}
|
2606
|
+
split_ix = st;
|
2607
|
+
}
|
2608
|
+
|
2609
|
+
/* otherwise, first put to the left all l.e. and not NA, then all NAs to the end of the left */
|
2610
|
+
else
|
2611
|
+
{
|
2612
|
+
for (size_t row = st; row <= end; row++)
|
2613
|
+
{
|
2614
|
+
if (x[ix_arr[row]] >= 0 && split_categ[ x[ix_arr[row]] ] == 1)
|
2615
|
+
{
|
2616
|
+
temp = ix_arr[st];
|
2617
|
+
ix_arr[st] = ix_arr[row];
|
2618
|
+
ix_arr[row] = temp;
|
2619
|
+
st++;
|
2620
|
+
}
|
2621
|
+
}
|
2622
|
+
st_NA = st;
|
2623
|
+
|
2624
|
+
for (size_t row = st; row <= end; row++)
|
2625
|
+
{
|
2626
|
+
if (x[ix_arr[row]] < 0)
|
2627
|
+
{
|
2628
|
+
temp = ix_arr[st];
|
2629
|
+
ix_arr[st] = ix_arr[row];
|
2630
|
+
ix_arr[row] = temp;
|
2631
|
+
st++;
|
2632
|
+
}
|
2633
|
+
}
|
2634
|
+
end_NA = st;
|
2635
|
+
}
|
2636
|
+
}
|
2637
|
+
|
2638
|
+
/* For categorical columns split by subset, used at prediction time (with similarity) */
|
2639
|
+
void divide_subset_split(size_t *restrict ix_arr, int x[], size_t st, size_t end, signed char split_categ[],
|
2640
|
+
int ncat, MissingAction missing_action, NewCategAction new_cat_action,
|
2641
|
+
bool move_new_to_left, size_t &restrict st_NA, size_t &restrict end_NA, size_t &restrict split_ix) noexcept
|
2642
|
+
{
|
2643
|
+
size_t temp;
|
2644
|
+
int cval;
|
2645
|
+
|
2646
|
+
/* if NAs are not to be bothered with, just need to do a single pass */
|
2647
|
+
if (missing_action == Fail && new_cat_action != Weighted)
|
2648
|
+
{
|
2649
|
+
/* in this case, will need to fill 'split_ix', otherwise need to fill 'st_NA' and 'end_NA' */
|
2650
|
+
if (new_cat_action == Smallest && move_new_to_left)
|
2651
|
+
{
|
2652
|
+
for (size_t row = st; row <= end; row++)
|
2653
|
+
{
|
2654
|
+
cval = x[ix_arr[row]];
|
2655
|
+
if (cval >= ncat || split_categ[cval] == 1 || split_categ[cval] == (-1))
|
2656
|
+
{
|
2657
|
+
temp = ix_arr[st];
|
2658
|
+
ix_arr[st] = ix_arr[row];
|
2659
|
+
ix_arr[row] = temp;
|
2660
|
+
st++;
|
2661
|
+
}
|
2662
|
+
}
|
2663
|
+
}
|
2664
|
+
|
2665
|
+
else if (new_cat_action == Random)
|
2666
|
+
{
|
2667
|
+
for (size_t row = st; row <= end; row++)
|
2668
|
+
{
|
2669
|
+
cval = x[ix_arr[row]];
|
2670
|
+
cval = (cval >= ncat)? (cval % ncat) : cval;
|
2671
|
+
if (split_categ[cval] == 1)
|
2672
|
+
{
|
2673
|
+
temp = ix_arr[st];
|
2674
|
+
ix_arr[st] = ix_arr[row];
|
2675
|
+
ix_arr[row] = temp;
|
2676
|
+
st++;
|
2677
|
+
}
|
2678
|
+
}
|
2679
|
+
}
|
2680
|
+
|
2681
|
+
else
|
2682
|
+
{
|
2683
|
+
for (size_t row = st; row <= end; row++)
|
2684
|
+
{
|
2685
|
+
cval = x[ix_arr[row]];
|
2686
|
+
if (cval < ncat && split_categ[cval] == 1)
|
2687
|
+
{
|
2688
|
+
temp = ix_arr[st];
|
2689
|
+
ix_arr[st] = ix_arr[row];
|
2690
|
+
ix_arr[row] = temp;
|
2691
|
+
st++;
|
2692
|
+
}
|
2693
|
+
}
|
2694
|
+
}
|
2695
|
+
|
2696
|
+
split_ix = st;
|
2697
|
+
}
|
2698
|
+
|
2699
|
+
/* if there are new categories, and their direction was decided at random,
|
2700
|
+
can just reuse what was randomly decided for previous columns by taking
|
2701
|
+
a remainder w.r.t. the number of previous columns. Note however that this
|
2702
|
+
will not be an unbiased decision if the model used a gain criterion. */
|
2703
|
+
else if (new_cat_action == Random)
|
2704
|
+
{
|
2705
|
+
if (missing_action == Impute && !move_new_to_left)
|
2706
|
+
{
|
2707
|
+
for (size_t row = st; row <= end; row++)
|
2708
|
+
{
|
2709
|
+
cval = x[ix_arr[row]];
|
2710
|
+
cval = (cval >= ncat)? (cval % ncat) : cval;
|
2711
|
+
if (cval < 0 || split_categ[cval] == 1)
|
2712
|
+
{
|
2713
|
+
temp = ix_arr[st];
|
2714
|
+
ix_arr[st] = ix_arr[row];
|
2715
|
+
ix_arr[row] = temp;
|
2716
|
+
st++;
|
2717
|
+
}
|
2718
|
+
}
|
2719
|
+
}
|
2720
|
+
|
2721
|
+
else
|
2722
|
+
{
|
2723
|
+
for (size_t row = st; row <= end; row++)
|
2724
|
+
{
|
2725
|
+
cval = x[ix_arr[row]];
|
2726
|
+
cval = (cval >= ncat)? (cval % ncat) : cval;
|
2727
|
+
if (cval >= 0 && split_categ[cval] == 1)
|
2728
|
+
{
|
2729
|
+
temp = ix_arr[st];
|
2730
|
+
ix_arr[st] = ix_arr[row];
|
2731
|
+
ix_arr[row] = temp;
|
2732
|
+
st++;
|
2733
|
+
}
|
2734
|
+
}
|
2735
|
+
}
|
2736
|
+
st_NA = st;
|
2737
|
+
|
2738
|
+
if (!(missing_action == Impute && !move_new_to_left))
|
2739
|
+
{
|
2740
|
+
for (size_t row = st; row <= end; row++)
|
2741
|
+
{
|
2742
|
+
if (unlikely(x[ix_arr[row]] < 0))
|
2743
|
+
{
|
2744
|
+
temp = ix_arr[st];
|
2745
|
+
ix_arr[st] = ix_arr[row];
|
2746
|
+
ix_arr[row] = temp;
|
2747
|
+
st++;
|
2748
|
+
}
|
2749
|
+
}
|
2750
|
+
}
|
2751
|
+
end_NA = st;
|
2752
|
+
}
|
2753
|
+
|
2754
|
+
/* otherwise, first put to the left all l.e. and not NA, then all NAs to the end of the left */
|
2755
|
+
else
|
2756
|
+
{
|
2757
|
+
/* Note: if having 'new_cat_action'='Smallest' and 'missing_action'='Impute', missing values
|
2758
|
+
and new categories will necessarily go into different branches, thus it's possible to do
|
2759
|
+
all the movements in one pass if certain conditions match. */
|
2760
|
+
|
2761
|
+
if (new_cat_action == Smallest && move_new_to_left)
|
2762
|
+
{
|
2763
|
+
for (size_t row = st; row <= end; row++)
|
2764
|
+
{
|
2765
|
+
cval = x[ix_arr[row]];
|
2766
|
+
if (cval >= 0 && (cval >= ncat || split_categ[cval] == 1 || split_categ[cval] == (-1)))
|
2767
|
+
{
|
2768
|
+
temp = ix_arr[st];
|
2769
|
+
ix_arr[st] = ix_arr[row];
|
2770
|
+
ix_arr[row] = temp;
|
2771
|
+
st++;
|
2772
|
+
}
|
2773
|
+
}
|
2774
|
+
}
|
2775
|
+
|
2776
|
+
else if (missing_action == Impute && !move_new_to_left)
|
2777
|
+
{
|
2778
|
+
for (size_t row = st; row <= end; row++)
|
2779
|
+
{
|
2780
|
+
cval = x[ix_arr[row]];
|
2781
|
+
if (cval < ncat && (cval < 0 || split_categ[cval] == 1))
|
2782
|
+
{
|
2783
|
+
temp = ix_arr[st];
|
2784
|
+
ix_arr[st] = ix_arr[row];
|
2785
|
+
ix_arr[row] = temp;
|
2786
|
+
st++;
|
2787
|
+
}
|
2788
|
+
}
|
2789
|
+
}
|
2790
|
+
|
2791
|
+
else
|
2792
|
+
{
|
2793
|
+
for (size_t row = st; row <= end; row++)
|
2794
|
+
{
|
2795
|
+
cval = x[ix_arr[row]];
|
2796
|
+
if (cval >= 0 && cval < ncat && split_categ[cval] == 1)
|
2797
|
+
{
|
2798
|
+
temp = ix_arr[st];
|
2799
|
+
ix_arr[st] = ix_arr[row];
|
2800
|
+
ix_arr[row] = temp;
|
2801
|
+
st++;
|
2802
|
+
}
|
2803
|
+
}
|
2804
|
+
}
|
2805
|
+
|
2806
|
+
st_NA = st;
|
2807
|
+
|
2808
|
+
if (new_cat_action == Weighted && missing_action == Divide)
|
2809
|
+
{
|
2810
|
+
for (size_t row = st; row <= end; row++)
|
2811
|
+
{
|
2812
|
+
cval = x[ix_arr[row]];
|
2813
|
+
if (cval < 0 || cval >= ncat || split_categ[cval] == (-1))
|
2814
|
+
{
|
2815
|
+
temp = ix_arr[st];
|
2816
|
+
ix_arr[st] = ix_arr[row];
|
2817
|
+
ix_arr[row] = temp;
|
2818
|
+
st++;
|
2819
|
+
}
|
2820
|
+
}
|
2821
|
+
}
|
2822
|
+
|
2823
|
+
else if (new_cat_action == Weighted)
|
2824
|
+
{
|
2825
|
+
for (size_t row = st; row <= end; row++)
|
2826
|
+
{
|
2827
|
+
cval = x[ix_arr[row]];
|
2828
|
+
if (cval >= 0 && (cval >= ncat || split_categ[cval] == (-1)))
|
2829
|
+
{
|
2830
|
+
temp = ix_arr[st];
|
2831
|
+
ix_arr[st] = ix_arr[row];
|
2832
|
+
ix_arr[row] = temp;
|
2833
|
+
st++;
|
2834
|
+
}
|
2835
|
+
}
|
2836
|
+
}
|
2837
|
+
|
2838
|
+
else if (missing_action == Divide)
|
2839
|
+
{
|
2840
|
+
for (size_t row = st; row <= end; row++)
|
2841
|
+
{
|
2842
|
+
if (unlikely(x[ix_arr[row]] < 0))
|
2843
|
+
{
|
2844
|
+
temp = ix_arr[st];
|
2845
|
+
ix_arr[st] = ix_arr[row];
|
2846
|
+
ix_arr[row] = temp;
|
2847
|
+
st++;
|
2848
|
+
}
|
2849
|
+
}
|
2850
|
+
}
|
2851
|
+
|
2852
|
+
end_NA = st;
|
2853
|
+
}
|
2854
|
+
}
|
2855
|
+
|
2856
|
+
/* For categoricals split on a single category */
|
2857
|
+
void divide_subset_split(size_t *restrict ix_arr, int x[], size_t st, size_t end, int split_categ,
|
2858
|
+
MissingAction missing_action, size_t &restrict st_NA, size_t &restrict end_NA, size_t &restrict split_ix) noexcept
|
2859
|
+
{
|
2860
|
+
size_t temp;
|
2861
|
+
|
2862
|
+
/* if NAs are not to be bothered with, just need to do a single pass */
|
2863
|
+
if (missing_action == Fail)
|
2864
|
+
{
|
2865
|
+
/* move to the left if it's equal to the chosen category */
|
2866
|
+
for (size_t row = st; row <= end; row++)
|
2867
|
+
{
|
2868
|
+
if (x[ix_arr[row]] == split_categ)
|
2869
|
+
{
|
2870
|
+
temp = ix_arr[st];
|
2871
|
+
ix_arr[st] = ix_arr[row];
|
2872
|
+
ix_arr[row] = temp;
|
2873
|
+
st++;
|
2874
|
+
}
|
2875
|
+
}
|
2876
|
+
split_ix = st;
|
2877
|
+
}
|
2878
|
+
|
2879
|
+
/* otherwise, first put to the left all equal to chosen and not NA, then all NAs to the end of the left */
|
2880
|
+
else
|
2881
|
+
{
|
2882
|
+
for (size_t row = st; row <= end; row++)
|
2883
|
+
{
|
2884
|
+
if (x[ix_arr[row]] == split_categ)
|
2885
|
+
{
|
2886
|
+
temp = ix_arr[st];
|
2887
|
+
ix_arr[st] = ix_arr[row];
|
2888
|
+
ix_arr[row] = temp;
|
2889
|
+
st++;
|
2890
|
+
}
|
2891
|
+
}
|
2892
|
+
st_NA = st;
|
2893
|
+
|
2894
|
+
for (size_t row = st; row <= end; row++)
|
2895
|
+
{
|
2896
|
+
if (unlikely(x[ix_arr[row]] < 0))
|
2897
|
+
{
|
2898
|
+
temp = ix_arr[st];
|
2899
|
+
ix_arr[st] = ix_arr[row];
|
2900
|
+
ix_arr[row] = temp;
|
2901
|
+
st++;
|
2902
|
+
}
|
2903
|
+
}
|
2904
|
+
end_NA = st;
|
2905
|
+
}
|
2906
|
+
}
|
2907
|
+
|
2908
|
+
/* For categoricals split on sub-set that turned out to have 2 categories only (prediction-time) */
|
2909
|
+
void divide_subset_split(size_t *restrict ix_arr, int x[], size_t st, size_t end,
|
2910
|
+
MissingAction missing_action, NewCategAction new_cat_action,
|
2911
|
+
bool move_new_to_left, size_t &restrict st_NA, size_t &restrict end_NA, size_t &restrict split_ix) noexcept
|
2912
|
+
{
|
2913
|
+
size_t temp;
|
2914
|
+
|
2915
|
+
/* if NAs are not to be bothered with, just need to do a single pass */
|
2916
|
+
if (missing_action == Fail)
|
2917
|
+
{
|
2918
|
+
/* move to the left if it's l.e. than the split point */
|
2919
|
+
if (new_cat_action == Smallest && move_new_to_left)
|
2920
|
+
{
|
2921
|
+
for (size_t row = st; row <= end; row++)
|
2922
|
+
{
|
2923
|
+
if (x[ix_arr[row]] == 0 || x[ix_arr[row]] > 1)
|
2924
|
+
{
|
2925
|
+
temp = ix_arr[st];
|
2926
|
+
ix_arr[st] = ix_arr[row];
|
2927
|
+
ix_arr[row] = temp;
|
2928
|
+
st++;
|
2929
|
+
}
|
2930
|
+
}
|
2931
|
+
}
|
2932
|
+
|
2933
|
+
else
|
2934
|
+
{
|
2935
|
+
for (size_t row = st; row <= end; row++)
|
2936
|
+
{
|
2937
|
+
if (x[ix_arr[row]] == 0)
|
2938
|
+
{
|
2939
|
+
temp = ix_arr[st];
|
2940
|
+
ix_arr[st] = ix_arr[row];
|
2941
|
+
ix_arr[row] = temp;
|
2942
|
+
st++;
|
2943
|
+
}
|
2944
|
+
}
|
2945
|
+
}
|
2946
|
+
split_ix = st;
|
2947
|
+
}
|
2948
|
+
|
2949
|
+
/* otherwise, first put to the left all l.e. and not NA, then all NAs to the end of the left */
|
2950
|
+
else
|
2951
|
+
{
|
2952
|
+
if (new_cat_action == Smallest && move_new_to_left)
|
2953
|
+
{
|
2954
|
+
for (size_t row = st; row <= end; row++)
|
2955
|
+
{
|
2956
|
+
if (x[ix_arr[row]] == 0 || x[ix_arr[row]] > 1)
|
2957
|
+
{
|
2958
|
+
temp = ix_arr[st];
|
2959
|
+
ix_arr[st] = ix_arr[row];
|
2960
|
+
ix_arr[row] = temp;
|
2961
|
+
st++;
|
2962
|
+
}
|
2963
|
+
}
|
2964
|
+
st_NA = st;
|
2965
|
+
|
2966
|
+
for (size_t row = st; row <= end; row++)
|
2967
|
+
{
|
2968
|
+
if (unlikely(x[ix_arr[row]] < 0))
|
2969
|
+
{
|
2970
|
+
temp = ix_arr[st];
|
2971
|
+
ix_arr[st] = ix_arr[row];
|
2972
|
+
ix_arr[row] = temp;
|
2973
|
+
st++;
|
2974
|
+
}
|
2975
|
+
}
|
2976
|
+
end_NA = st;
|
2977
|
+
}
|
2978
|
+
|
2979
|
+
else
|
2980
|
+
{
|
2981
|
+
for (size_t row = st; row <= end; row++)
|
2982
|
+
{
|
2983
|
+
if (x[ix_arr[row]] == 0)
|
2984
|
+
{
|
2985
|
+
temp = ix_arr[st];
|
2986
|
+
ix_arr[st] = ix_arr[row];
|
2987
|
+
ix_arr[row] = temp;
|
2988
|
+
st++;
|
2989
|
+
}
|
2990
|
+
}
|
2991
|
+
st_NA = st;
|
2992
|
+
|
2993
|
+
for (size_t row = st; row <= end; row++)
|
2994
|
+
{
|
2995
|
+
if (unlikely(x[ix_arr[row]] < 0))
|
2996
|
+
{
|
2997
|
+
temp = ix_arr[st];
|
2998
|
+
ix_arr[st] = ix_arr[row];
|
2999
|
+
ix_arr[row] = temp;
|
3000
|
+
st++;
|
3001
|
+
}
|
3002
|
+
}
|
3003
|
+
end_NA = st;
|
3004
|
+
}
|
3005
|
+
}
|
3006
|
+
}
|
3007
|
+
|
3008
|
+
/* for regular numeric columns */
|
3009
|
+
template <class real_t>
|
3010
|
+
void get_range(size_t ix_arr[], real_t *restrict x, size_t st, size_t end,
|
3011
|
+
MissingAction missing_action, double &restrict xmin, double &restrict xmax, bool &unsplittable) noexcept
|
3012
|
+
{
|
3013
|
+
xmin = HUGE_VAL;
|
3014
|
+
xmax = -HUGE_VAL;
|
3015
|
+
double xval;
|
3016
|
+
|
3017
|
+
if (missing_action == Fail)
|
3018
|
+
{
|
3019
|
+
for (size_t row = st; row <= end; row++)
|
3020
|
+
{
|
3021
|
+
xval = x[ix_arr[row]];
|
3022
|
+
xmin = (xval < xmin)? xval : xmin;
|
3023
|
+
xmax = (xval > xmax)? xval : xmax;
|
3024
|
+
}
|
3025
|
+
}
|
3026
|
+
|
3027
|
+
|
3028
|
+
else
|
3029
|
+
{
|
3030
|
+
for (size_t row = st; row <= end; row++)
|
3031
|
+
{
|
3032
|
+
xval = x[ix_arr[row]];
|
3033
|
+
xmin = std::fmin(xmin, xval);
|
3034
|
+
xmax = std::fmax(xmax, xval);
|
3035
|
+
}
|
3036
|
+
}
|
3037
|
+
|
3038
|
+
unsplittable = (xmin == xmax) || (xmin == HUGE_VAL && xmax == -HUGE_VAL) || std::isnan(xmin) || std::isnan(xmax);
|
3039
|
+
}
|
3040
|
+
|
3041
|
+
template <class real_t>
|
3042
|
+
void get_range(real_t *restrict x, size_t n,
|
3043
|
+
MissingAction missing_action, double &restrict xmin, double &restrict xmax, bool &unsplittable) noexcept
|
3044
|
+
{
|
3045
|
+
xmin = HUGE_VAL;
|
3046
|
+
xmax = -HUGE_VAL;
|
3047
|
+
|
3048
|
+
if (missing_action == Fail)
|
3049
|
+
{
|
3050
|
+
for (size_t row = 0; row < n; row++)
|
3051
|
+
{
|
3052
|
+
xmin = (x[row] < xmin)? x[row] : xmin;
|
3053
|
+
xmax = (x[row] > xmax)? x[row] : xmax;
|
3054
|
+
}
|
3055
|
+
}
|
3056
|
+
|
3057
|
+
|
3058
|
+
else
|
3059
|
+
{
|
3060
|
+
for (size_t row = 0; row < n; row++)
|
3061
|
+
{
|
3062
|
+
xmin = std::fmin(xmin, x[row]);
|
3063
|
+
xmax = std::fmax(xmax, x[row]);
|
3064
|
+
}
|
3065
|
+
}
|
3066
|
+
|
3067
|
+
unsplittable = (xmin == xmax) || (xmin == HUGE_VAL && xmax == -HUGE_VAL) || std::isnan(xmin) || std::isnan(xmax);
|
3068
|
+
}
|
3069
|
+
|
3070
|
+
/* for sparse inputs */
|
3071
|
+
template <class real_t, class sparse_ix>
|
3072
|
+
void get_range(size_t *restrict ix_arr, size_t st, size_t end, size_t col_num,
|
3073
|
+
real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
3074
|
+
MissingAction missing_action, double &restrict xmin, double &restrict xmax, bool &unsplittable) noexcept
|
3075
|
+
{
|
3076
|
+
/* ix_arr must already be sorted beforehand */
|
3077
|
+
xmin = HUGE_VAL;
|
3078
|
+
xmax = -HUGE_VAL;
|
3079
|
+
|
3080
|
+
size_t st_col = Xc_indptr[col_num];
|
3081
|
+
size_t end_col = Xc_indptr[col_num + 1];
|
3082
|
+
size_t nnz_col = end_col - st_col;
|
3083
|
+
end_col--;
|
3084
|
+
size_t curr_pos = st_col;
|
3085
|
+
|
3086
|
+
if (!nnz_col ||
|
3087
|
+
Xc_ind[st_col] > (sparse_ix)ix_arr[end] ||
|
3088
|
+
(sparse_ix)ix_arr[st] > Xc_ind[end_col]
|
3089
|
+
)
|
3090
|
+
{
|
3091
|
+
unsplittable = true;
|
3092
|
+
return;
|
3093
|
+
}
|
3094
|
+
|
3095
|
+
if (nnz_col < end - st + 1 ||
|
3096
|
+
Xc_ind[st_col] > (sparse_ix)ix_arr[st] ||
|
3097
|
+
Xc_ind[end_col] < (sparse_ix)ix_arr[end]
|
3098
|
+
)
|
3099
|
+
{
|
3100
|
+
xmin = 0;
|
3101
|
+
xmax = 0;
|
3102
|
+
}
|
3103
|
+
|
3104
|
+
size_t ind_end_col = Xc_ind[end_col];
|
3105
|
+
size_t nmatches = 0;
|
3106
|
+
|
3107
|
+
if (missing_action == Fail)
|
3108
|
+
{
|
3109
|
+
for (size_t *row = std::lower_bound(ix_arr + st, ix_arr + end + 1, Xc_ind[st_col]);
|
3110
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
3111
|
+
)
|
3112
|
+
{
|
3113
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
3114
|
+
{
|
3115
|
+
nmatches++;
|
3116
|
+
xmin = (Xc[curr_pos] < xmin)? Xc[curr_pos] : xmin;
|
3117
|
+
xmax = (Xc[curr_pos] > xmax)? Xc[curr_pos] : xmax;
|
3118
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
3119
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
3120
|
+
}
|
3121
|
+
|
3122
|
+
else
|
3123
|
+
{
|
3124
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
3125
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
3126
|
+
else
|
3127
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
3128
|
+
}
|
3129
|
+
}
|
3130
|
+
}
|
3131
|
+
|
3132
|
+
else /* can have NAs */
|
3133
|
+
{
|
3134
|
+
for (size_t *row = std::lower_bound(ix_arr + st, ix_arr + end + 1, Xc_ind[st_col]);
|
3135
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
3136
|
+
)
|
3137
|
+
{
|
3138
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
3139
|
+
{
|
3140
|
+
nmatches++;
|
3141
|
+
xmin = std::fmin(xmin, Xc[curr_pos]);
|
3142
|
+
xmax = std::fmax(xmax, Xc[curr_pos]);
|
3143
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
3144
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
3145
|
+
}
|
3146
|
+
|
3147
|
+
else
|
3148
|
+
{
|
3149
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
3150
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
3151
|
+
else
|
3152
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
3153
|
+
}
|
3154
|
+
}
|
3155
|
+
|
3156
|
+
}
|
3157
|
+
|
3158
|
+
if (nmatches < (end - st + 1))
|
3159
|
+
{
|
3160
|
+
xmin = std::fmin(xmin, 0);
|
3161
|
+
xmax = std::fmax(xmax, 0);
|
3162
|
+
}
|
3163
|
+
|
3164
|
+
unsplittable = (xmin == xmax) || (xmin == HUGE_VAL && xmax == -HUGE_VAL) || std::isnan(xmin) || std::isnan(xmax);
|
3165
|
+
}
|
3166
|
+
|
3167
|
+
template <class real_t, class sparse_ix>
|
3168
|
+
void get_range(size_t col_num, size_t nrows,
|
3169
|
+
real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
3170
|
+
MissingAction missing_action, double &restrict xmin, double &restrict xmax, bool &unsplittable) noexcept
|
3171
|
+
{
|
3172
|
+
xmin = HUGE_VAL;
|
3173
|
+
xmax = -HUGE_VAL;
|
3174
|
+
|
3175
|
+
if ((size_t)(Xc_indptr[col_num+1] - Xc_indptr[col_num]) < nrows)
|
3176
|
+
{
|
3177
|
+
xmin = 0;
|
3178
|
+
xmax = 0;
|
3179
|
+
}
|
3180
|
+
|
3181
|
+
if (missing_action == Fail)
|
3182
|
+
{
|
3183
|
+
for (auto ix = Xc_indptr[col_num]; ix < Xc_indptr[col_num+1]; ix++)
|
3184
|
+
{
|
3185
|
+
xmin = (Xc[ix] < xmin)? Xc[ix] : xmin;
|
3186
|
+
xmax = (Xc[ix] > xmax)? Xc[ix] : xmax;
|
3187
|
+
}
|
3188
|
+
}
|
3189
|
+
|
3190
|
+
else
|
3191
|
+
{
|
3192
|
+
for (auto ix = Xc_indptr[col_num]; ix < Xc_indptr[col_num+1]; ix++)
|
3193
|
+
{
|
3194
|
+
if (unlikely(std::isinf(Xc[ix]))) continue;
|
3195
|
+
xmin = std::fmin(xmin, Xc[ix]);
|
3196
|
+
xmax = std::fmax(xmax, Xc[ix]);
|
3197
|
+
}
|
3198
|
+
}
|
3199
|
+
|
3200
|
+
unsplittable = (xmin == xmax) || (xmin == HUGE_VAL && xmax == -HUGE_VAL) || std::isnan(xmin) || std::isnan(xmax);
|
3201
|
+
}
|
3202
|
+
|
3203
|
+
|
3204
|
+
void get_categs(size_t *restrict ix_arr, int x[], size_t st, size_t end, int ncat,
|
3205
|
+
MissingAction missing_action, signed char categs[], size_t &restrict npresent, bool &unsplittable) noexcept
|
3206
|
+
{
|
3207
|
+
std::fill(categs, categs + ncat, -1);
|
3208
|
+
npresent = 0;
|
3209
|
+
for (size_t row = st; row <= end; row++)
|
3210
|
+
if (likely(x[ix_arr[row]] >= 0))
|
3211
|
+
categs[x[ix_arr[row]]] = 1;
|
3212
|
+
|
3213
|
+
npresent = std::accumulate(categs,
|
3214
|
+
categs + ncat,
|
3215
|
+
(size_t)0,
|
3216
|
+
[](const size_t a, const signed char b){return a + (b > 0);}
|
3217
|
+
);
|
3218
|
+
|
3219
|
+
unsplittable = npresent < 2;
|
3220
|
+
}
|
3221
|
+
|
3222
|
+
template <class real_t>
|
3223
|
+
bool check_more_than_two_unique_values(size_t ix_arr[], size_t st, size_t end, real_t x[], MissingAction missing_action)
|
3224
|
+
{
|
3225
|
+
if (end - st <= 1) return false;
|
3226
|
+
|
3227
|
+
if (missing_action == Fail)
|
3228
|
+
{
|
3229
|
+
real_t x0 = x[ix_arr[st]];
|
3230
|
+
for (size_t ix = st+1; ix <= end; ix++)
|
3231
|
+
{
|
3232
|
+
if (x[ix_arr[ix]] != x0) return true;
|
3233
|
+
}
|
3234
|
+
}
|
3235
|
+
|
3236
|
+
else
|
3237
|
+
{
|
3238
|
+
real_t x0;
|
3239
|
+
size_t ix;
|
3240
|
+
for (ix = st; ix <= end; ix++)
|
3241
|
+
{
|
3242
|
+
if (likely(!is_na_or_inf(x[ix_arr[ix]])))
|
3243
|
+
{
|
3244
|
+
x0 = x[ix_arr[ix]];
|
3245
|
+
ix++;
|
3246
|
+
break;
|
3247
|
+
}
|
3248
|
+
}
|
3249
|
+
|
3250
|
+
for (; ix <= end; ix++)
|
3251
|
+
{
|
3252
|
+
if (!is_na_or_inf(x[ix_arr[ix]]) && x[ix_arr[ix]] != x0)
|
3253
|
+
return true;
|
3254
|
+
}
|
3255
|
+
}
|
3256
|
+
|
3257
|
+
return false;
|
3258
|
+
}
|
3259
|
+
|
3260
|
+
bool check_more_than_two_unique_values(size_t ix_arr[], size_t st, size_t end, int x[], MissingAction missing_action)
|
3261
|
+
{
|
3262
|
+
if (end - st <= 1) return false;
|
3263
|
+
|
3264
|
+
if (missing_action == Fail)
|
3265
|
+
{
|
3266
|
+
int x0 = x[ix_arr[st]];
|
3267
|
+
for (size_t ix = st+1; ix <= end; ix++)
|
3268
|
+
{
|
3269
|
+
if (x[ix_arr[ix]] != x0) return true;
|
3270
|
+
}
|
3271
|
+
}
|
3272
|
+
|
3273
|
+
else
|
3274
|
+
{
|
3275
|
+
int x0;
|
3276
|
+
size_t ix;
|
3277
|
+
for (ix = st; ix <= end; ix++)
|
3278
|
+
{
|
3279
|
+
if (x[ix_arr[ix]] >= 0)
|
3280
|
+
{
|
3281
|
+
x0 = x[ix_arr[ix]];
|
3282
|
+
ix++;
|
3283
|
+
break;
|
3284
|
+
}
|
3285
|
+
}
|
3286
|
+
|
3287
|
+
for (; ix <= end; ix++)
|
3288
|
+
{
|
3289
|
+
if (x[ix_arr[ix]] >= 0 && x[ix_arr[ix]] != x0)
|
3290
|
+
return true;
|
3291
|
+
}
|
3292
|
+
}
|
3293
|
+
|
3294
|
+
return false;
|
3295
|
+
}
|
3296
|
+
|
3297
|
+
template <class real_t, class sparse_ix>
|
3298
|
+
bool check_more_than_two_unique_values(size_t *restrict ix_arr, size_t st, size_t end, size_t col,
|
3299
|
+
sparse_ix *restrict Xc_indptr, sparse_ix *restrict Xc_ind, real_t *restrict Xc,
|
3300
|
+
MissingAction missing_action)
|
3301
|
+
{
|
3302
|
+
if (end - st <= 1) return false;
|
3303
|
+
if (Xc_indptr[col+1] == Xc_indptr[col]) return false;
|
3304
|
+
bool has_zeros = (end - st + 1) > (size_t)(Xc_indptr[col+1] - Xc_indptr[col]);
|
3305
|
+
if (has_zeros && !is_na_or_inf(Xc[Xc_indptr[col]]) && Xc[Xc_indptr[col]] != 0) return true;
|
3306
|
+
|
3307
|
+
size_t st_col = Xc_indptr[col];
|
3308
|
+
size_t end_col = Xc_indptr[col + 1] - 1;
|
3309
|
+
size_t curr_pos = st_col;
|
3310
|
+
size_t ind_end_col = Xc_ind[end_col];
|
3311
|
+
|
3312
|
+
/* 'ix_arr' should be sorted beforehand */
|
3313
|
+
/* TODO: refactor this */
|
3314
|
+
real_t x0 = 0;
|
3315
|
+
size_t *row;
|
3316
|
+
for (row = std::lower_bound(ix_arr + st, ix_arr + end + 1, Xc_ind[st_col]);
|
3317
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
3318
|
+
)
|
3319
|
+
{
|
3320
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
3321
|
+
{
|
3322
|
+
if (is_na_or_inf(Xc[curr_pos]) || (has_zeros && Xc[curr_pos] == 0))
|
3323
|
+
{
|
3324
|
+
if (row == ix_arr + end || curr_pos == end_col) return false;
|
3325
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
3326
|
+
}
|
3327
|
+
|
3328
|
+
x0 = Xc[curr_pos];
|
3329
|
+
if (has_zeros) return true;
|
3330
|
+
else if (x0 == 0) has_zeros = true;
|
3331
|
+
if (row == ix_arr + end || curr_pos == end_col) return false;
|
3332
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
3333
|
+
break;
|
3334
|
+
}
|
3335
|
+
|
3336
|
+
else
|
3337
|
+
{
|
3338
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
3339
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
3340
|
+
else
|
3341
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
3342
|
+
}
|
3343
|
+
}
|
3344
|
+
|
3345
|
+
for (;
|
3346
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
3347
|
+
)
|
3348
|
+
{
|
3349
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
3350
|
+
{
|
3351
|
+
if (is_na_or_inf(Xc[curr_pos]) || (has_zeros && Xc[curr_pos] == 0))
|
3352
|
+
{
|
3353
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
3354
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
3355
|
+
}
|
3356
|
+
|
3357
|
+
else if (Xc[curr_pos] != x0)
|
3358
|
+
{
|
3359
|
+
return true;
|
3360
|
+
}
|
3361
|
+
|
3362
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
3363
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
3364
|
+
}
|
3365
|
+
|
3366
|
+
else
|
3367
|
+
{
|
3368
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
3369
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
3370
|
+
else
|
3371
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
3372
|
+
}
|
3373
|
+
}
|
3374
|
+
|
3375
|
+
return false;
|
3376
|
+
}
|
3377
|
+
|
3378
|
+
template <class real_t, class sparse_ix>
|
3379
|
+
bool check_more_than_two_unique_values(size_t nrows, size_t col,
|
3380
|
+
sparse_ix *restrict Xc_indptr, sparse_ix *restrict Xc_ind, real_t *restrict Xc,
|
3381
|
+
MissingAction missing_action)
|
3382
|
+
{
|
3383
|
+
if (nrows <= 1) return false;
|
3384
|
+
if (Xc_indptr[col+1] == Xc_indptr[col]) return false;
|
3385
|
+
bool has_zeros = nrows > (size_t)(Xc_indptr[col+1] - Xc_indptr[col]);
|
3386
|
+
if (has_zeros && !is_na_or_inf(Xc[Xc_indptr[col]]) && Xc[Xc_indptr[col]] != 0) return true;
|
3387
|
+
|
3388
|
+
real_t x0 = 0;
|
3389
|
+
sparse_ix ix;
|
3390
|
+
for (ix = Xc_indptr[col]; ix < Xc_indptr[col+1]; ix++)
|
3391
|
+
{
|
3392
|
+
if (!is_na_or_inf(Xc[ix]))
|
3393
|
+
{
|
3394
|
+
if (has_zeros && Xc[ix] == 0) continue;
|
3395
|
+
if (has_zeros) return true;
|
3396
|
+
else if (Xc[ix] == 0) has_zeros = true;
|
3397
|
+
x0 = Xc[ix];
|
3398
|
+
ix++;
|
3399
|
+
break;
|
3400
|
+
}
|
3401
|
+
}
|
3402
|
+
|
3403
|
+
for (ix = Xc_indptr[col]; ix < Xc_indptr[col+1]; ix++)
|
3404
|
+
{
|
3405
|
+
if (!is_na_or_inf(Xc[ix]))
|
3406
|
+
{
|
3407
|
+
if (has_zeros && Xc[ix] == 0) continue;
|
3408
|
+
if (Xc[ix] != x0) return true;
|
3409
|
+
}
|
3410
|
+
}
|
3411
|
+
|
3412
|
+
return false;
|
3413
|
+
}
|
3414
|
+
|
3415
|
+
void count_categs(size_t *restrict ix_arr, size_t st, size_t end, int x[], int ncat, size_t *restrict counts)
|
3416
|
+
{
|
3417
|
+
std::fill(counts, counts + ncat, (size_t)0);
|
3418
|
+
for (size_t row = st; row <= end; row++)
|
3419
|
+
if (likely(x[ix_arr[row]] >= 0))
|
3420
|
+
counts[x[ix_arr[row]]]++;
|
3421
|
+
}
|
3422
|
+
|
3423
|
+
int count_ncateg_in_col(const int x[], const size_t n, const int ncat, unsigned char buffer[])
|
3424
|
+
{
|
3425
|
+
memset(buffer, 0, ncat*sizeof(char));
|
3426
|
+
for (size_t ix = 0; ix < n; ix++)
|
3427
|
+
{
|
3428
|
+
if (likely(x[ix] >= 0)) buffer[x[ix]] = true;
|
3429
|
+
}
|
3430
|
+
|
3431
|
+
int ncat_present = 0;
|
3432
|
+
for (int cat = 0; cat < ncat; cat++)
|
3433
|
+
ncat_present += buffer[cat];
|
3434
|
+
return ncat_present;
|
3435
|
+
}
|
3436
|
+
|
3437
|
+
template <class ldouble_safe>
|
3438
|
+
ldouble_safe calculate_sum_weights(std::vector<size_t> &ix_arr, size_t st, size_t end, size_t curr_depth,
|
3439
|
+
std::vector<double> &weights_arr, hashed_map<size_t, double> &weights_map)
|
3440
|
+
{
|
3441
|
+
if (curr_depth > 0 && !weights_arr.empty())
|
3442
|
+
return std::accumulate(ix_arr.begin() + st,
|
3443
|
+
ix_arr.begin() + end + 1,
|
3444
|
+
(ldouble_safe)0,
|
3445
|
+
[&weights_arr](const ldouble_safe a, const size_t ix){return a + weights_arr[ix];});
|
3446
|
+
else if (curr_depth > 0 && !weights_map.empty())
|
3447
|
+
return std::accumulate(ix_arr.begin() + st,
|
3448
|
+
ix_arr.begin() + end + 1,
|
3449
|
+
(ldouble_safe)0,
|
3450
|
+
[&weights_map](const ldouble_safe a, const size_t ix){return a + weights_map[ix];});
|
3451
|
+
else
|
3452
|
+
return -HUGE_VAL;
|
3453
|
+
}
|
3454
|
+
|
3455
|
+
template <class real_t>
|
3456
|
+
size_t move_NAs_to_front(size_t ix_arr[], size_t st, size_t end, real_t x[])
|
3457
|
+
{
|
3458
|
+
size_t st_non_na = st;
|
3459
|
+
size_t temp;
|
3460
|
+
|
3461
|
+
for (size_t row = st; row <= end; row++)
|
3462
|
+
{
|
3463
|
+
if (unlikely(is_na_or_inf(x[ix_arr[row]])))
|
3464
|
+
{
|
3465
|
+
temp = ix_arr[st_non_na];
|
3466
|
+
ix_arr[st_non_na] = ix_arr[row];
|
3467
|
+
ix_arr[row] = temp;
|
3468
|
+
st_non_na++;
|
3469
|
+
}
|
3470
|
+
}
|
3471
|
+
|
3472
|
+
return st_non_na;
|
3473
|
+
}
|
3474
|
+
|
3475
|
+
template <class real_t, class sparse_ix>
|
3476
|
+
size_t move_NAs_to_front(size_t *restrict ix_arr, size_t st, size_t end, size_t col_num, real_t Xc[], sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr)
|
3477
|
+
{
|
3478
|
+
size_t st_non_na = st;
|
3479
|
+
size_t temp;
|
3480
|
+
|
3481
|
+
size_t st_col = Xc_indptr[col_num];
|
3482
|
+
size_t end_col = Xc_indptr[col_num + 1] - 1;
|
3483
|
+
size_t curr_pos = st_col;
|
3484
|
+
size_t ind_end_col = Xc_ind[end_col];
|
3485
|
+
std::sort(ix_arr + st, ix_arr + end + 1);
|
3486
|
+
size_t *ptr_st = std::lower_bound(ix_arr + st, ix_arr + end + 1, Xc_ind[st_col]);
|
3487
|
+
|
3488
|
+
for (size_t *row = ptr_st;
|
3489
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
3490
|
+
)
|
3491
|
+
{
|
3492
|
+
if (Xc_ind[curr_pos] == *row)
|
3493
|
+
{
|
3494
|
+
if (unlikely(is_na_or_inf(Xc[curr_pos])))
|
3495
|
+
{
|
3496
|
+
temp = ix_arr[st_non_na];
|
3497
|
+
ix_arr[st_non_na] = *row;
|
3498
|
+
*row = temp;
|
3499
|
+
st_non_na++;
|
3500
|
+
}
|
3501
|
+
|
3502
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
3503
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
3504
|
+
}
|
3505
|
+
|
3506
|
+
else
|
3507
|
+
{
|
3508
|
+
if (Xc_ind[curr_pos] > *row)
|
3509
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
3510
|
+
else
|
3511
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
3512
|
+
}
|
3513
|
+
}
|
3514
|
+
|
3515
|
+
return st_non_na;
|
3516
|
+
}
|
3517
|
+
|
3518
|
+
size_t move_NAs_to_front(size_t ix_arr[], size_t st, size_t end, int x[])
|
3519
|
+
{
|
3520
|
+
size_t st_non_na = st;
|
3521
|
+
size_t temp;
|
3522
|
+
|
3523
|
+
for (size_t row = st; row <= end; row++)
|
3524
|
+
{
|
3525
|
+
if (unlikely(x[ix_arr[row]] < 0))
|
3526
|
+
{
|
3527
|
+
temp = ix_arr[st_non_na];
|
3528
|
+
ix_arr[st_non_na] = ix_arr[row];
|
3529
|
+
ix_arr[row] = temp;
|
3530
|
+
st_non_na++;
|
3531
|
+
}
|
3532
|
+
}
|
3533
|
+
|
3534
|
+
return st_non_na;
|
3535
|
+
}
|
3536
|
+
|
3537
|
+
size_t center_NAs(size_t ix_arr[], size_t st_left, size_t st, size_t curr_pos)
|
3538
|
+
{
|
3539
|
+
size_t temp;
|
3540
|
+
for (size_t row = st_left; row < st; row++)
|
3541
|
+
{
|
3542
|
+
temp = ix_arr[--curr_pos];
|
3543
|
+
ix_arr[curr_pos] = ix_arr[row];
|
3544
|
+
ix_arr[row] = temp;
|
3545
|
+
}
|
3546
|
+
|
3547
|
+
return curr_pos;
|
3548
|
+
}
|
3549
|
+
|
3550
|
+
/* FIXME / TODO: this calculation would not take weight into account */
|
3551
|
+
/* Here:
|
3552
|
+
- 'ix_arr' should be partitioned putting the NAs and Infs at the beginning: [st_orig, st)
|
3553
|
+
- the rest of the range [st, end] should be sorted in ascending order
|
3554
|
+
The output should have a filled-in 'x' with median values, plus a re-sorted 'ix_arr'
|
3555
|
+
taking into account that now the median values are in the middle. */
|
3556
|
+
template <class real_t>
|
3557
|
+
void fill_NAs_with_median(size_t *restrict ix_arr, size_t st_orig, size_t st, size_t end, real_t *restrict x,
|
3558
|
+
double *restrict buffer_imputed_x, double *restrict xmedian)
|
3559
|
+
{
|
3560
|
+
size_t tot = end - st + 1;
|
3561
|
+
size_t idx_half = st + div2(tot);
|
3562
|
+
bool is_odd = (tot % 2) != 0;
|
3563
|
+
|
3564
|
+
if (is_odd)
|
3565
|
+
{
|
3566
|
+
*xmedian = x[ix_arr[idx_half]];
|
3567
|
+
idx_half--;
|
3568
|
+
}
|
3569
|
+
|
3570
|
+
else
|
3571
|
+
{
|
3572
|
+
idx_half--;
|
3573
|
+
double xlow = x[ix_arr[idx_half]];
|
3574
|
+
double xhigh = x[ix_arr[idx_half+(size_t)1]];
|
3575
|
+
*xmedian = xlow + (xhigh-xlow)/2.;
|
3576
|
+
}
|
3577
|
+
|
3578
|
+
for (size_t ix = st_orig; ix < st; ix++)
|
3579
|
+
buffer_imputed_x[ix_arr[ix]] = (*xmedian);
|
3580
|
+
for (size_t ix = st; ix <= end; ix++)
|
3581
|
+
buffer_imputed_x[ix_arr[ix]] = x[ix_arr[ix]];
|
3582
|
+
|
3583
|
+
/* 'ix_arr' can be resorted in-place, but the logic is a bit complex */
|
3584
|
+
/* step 1: move all NAs to their place by swapping them with the lower-half
|
3585
|
+
in ascending order (after this, the lower half will be unordered).
|
3586
|
+
along the way, copy the indices that claim the places where earlier
|
3587
|
+
there were missing values. these copied indices will be sorted in
|
3588
|
+
descending order at the end, as they were inserted in reverse order. */
|
3589
|
+
size_t end_pointer = idx_half;
|
3590
|
+
size_t n_move = std::min(st-st_orig, idx_half-st+1);
|
3591
|
+
size_t temp;
|
3592
|
+
for (size_t ix = st_orig; ix < st_orig + n_move; ix++)
|
3593
|
+
{
|
3594
|
+
temp = ix_arr[end_pointer];
|
3595
|
+
ix_arr[end_pointer] = ix_arr[ix];
|
3596
|
+
ix_arr[ix] = temp;
|
3597
|
+
end_pointer--;
|
3598
|
+
}
|
3599
|
+
|
3600
|
+
/* step 2: reverse the indices that were moved to the beginning so
|
3601
|
+
as to maintain the sorting order */
|
3602
|
+
std::reverse(ix_arr + st_orig, ix_arr + st_orig + n_move);
|
3603
|
+
/* step 3: rotate the total number of elements by the number of moved elements */
|
3604
|
+
size_t n_unmoved = (idx_half - st + 1) - n_move;
|
3605
|
+
std::rotate(ix_arr + st_orig,
|
3606
|
+
ix_arr + st_orig + n_move,
|
3607
|
+
ix_arr + st_orig + n_move + n_unmoved);
|
3608
|
+
}
|
3609
|
+
|
3610
|
+
template <class real_t, class sparse_ix>
|
3611
|
+
void todense(size_t *restrict ix_arr, size_t st, size_t end,
|
3612
|
+
size_t col_num, real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
3613
|
+
double *restrict buffer_arr)
|
3614
|
+
{
|
3615
|
+
std::fill(buffer_arr, buffer_arr + (end - st + 1), (double)0);
|
3616
|
+
|
3617
|
+
size_t st_col = Xc_indptr[col_num];
|
3618
|
+
size_t end_col = Xc_indptr[col_num + 1] - 1;
|
3619
|
+
size_t curr_pos = st_col;
|
3620
|
+
size_t ind_end_col = Xc_ind[end_col];
|
3621
|
+
size_t *ptr_st = std::lower_bound(ix_arr + st, ix_arr + end + 1, Xc_ind[st_col]);
|
3622
|
+
|
3623
|
+
for (size_t *row = ptr_st;
|
3624
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
3625
|
+
)
|
3626
|
+
{
|
3627
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
3628
|
+
{
|
3629
|
+
buffer_arr[row - (ix_arr + st)] = Xc[curr_pos];
|
3630
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
3631
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
3632
|
+
}
|
3633
|
+
|
3634
|
+
else
|
3635
|
+
{
|
3636
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
3637
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
3638
|
+
else
|
3639
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
3640
|
+
}
|
3641
|
+
}
|
3642
|
+
}
|
3643
|
+
|
3644
|
+
|
3645
|
+
template <class real_t>
|
3646
|
+
void colmajor_to_rowmajor(real_t *restrict X, size_t nrows, size_t ncols, std::vector<double> &X_row_major)
|
3647
|
+
{
|
3648
|
+
X_row_major.resize(nrows * ncols);
|
3649
|
+
for (size_t row = 0; row < nrows; row++)
|
3650
|
+
for (size_t col = 0; col < ncols; col++)
|
3651
|
+
X_row_major[row + col*nrows] = X[col + row*ncols];
|
3652
|
+
}
|
3653
|
+
|
3654
|
+
template <class real_t, class sparse_ix>
|
3655
|
+
void colmajor_to_rowmajor(real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
3656
|
+
size_t nrows, size_t ncols,
|
3657
|
+
std::vector<double> &Xr, std::vector<size_t> &Xr_ind, std::vector<size_t> &Xr_indptr)
|
3658
|
+
{
|
3659
|
+
/* First convert to COO */
|
3660
|
+
size_t nnz = Xc_indptr[ncols];
|
3661
|
+
std::vector<size_t> row_indices(nnz);
|
3662
|
+
for (size_t col = 0; col < ncols; col++)
|
3663
|
+
{
|
3664
|
+
for (sparse_ix ix = Xc_indptr[col]; ix < Xc_indptr[col+1]; ix++)
|
3665
|
+
{
|
3666
|
+
row_indices[ix] = Xc_ind[ix];
|
3667
|
+
}
|
3668
|
+
}
|
3669
|
+
|
3670
|
+
/* Then copy the data argsorted by rows */
|
3671
|
+
std::vector<size_t> argsorted_indices(nnz);
|
3672
|
+
std::iota(argsorted_indices.begin(), argsorted_indices.end(), (size_t)0);
|
3673
|
+
std::stable_sort(argsorted_indices.begin(), argsorted_indices.end(),
|
3674
|
+
[&row_indices](const size_t a, const size_t b)
|
3675
|
+
{return row_indices[a] < row_indices[b];});
|
3676
|
+
Xr.resize(nnz);
|
3677
|
+
Xr_ind.resize(nnz);
|
3678
|
+
for (size_t ix = 0; ix < nnz; ix++)
|
3679
|
+
{
|
3680
|
+
Xr[ix] = Xc[argsorted_indices[ix]];
|
3681
|
+
Xr_ind[ix] = Xc_ind[argsorted_indices[ix]];
|
3682
|
+
}
|
3683
|
+
|
3684
|
+
/* Now build the index pointer */
|
3685
|
+
Xr_indptr.resize(nrows+1);
|
3686
|
+
size_t curr_row = 0;
|
3687
|
+
size_t curr_n = 0;
|
3688
|
+
for (size_t ix = 0; ix < nnz; ix++)
|
3689
|
+
{
|
3690
|
+
if (row_indices[argsorted_indices[ix]] != curr_row)
|
3691
|
+
{
|
3692
|
+
Xr_indptr[curr_row+1] = curr_n;
|
3693
|
+
curr_n = 0;
|
3694
|
+
curr_row = row_indices[argsorted_indices[ix]];
|
3695
|
+
}
|
3696
|
+
|
3697
|
+
else
|
3698
|
+
{
|
3699
|
+
curr_n++;
|
3700
|
+
}
|
3701
|
+
}
|
3702
|
+
for (size_t row = 1; row < nrows; row++)
|
3703
|
+
Xr_indptr[row+1] += Xr_indptr[row];
|
3704
|
+
}
|
3705
|
+
|
3706
|
+
|
3707
|
+
bool interrupt_switch = false;
|
3708
|
+
bool handle_is_locked = false;
|
3709
|
+
|
3710
|
+
/* Function to handle interrupt signals */
|
3711
|
+
void set_interrup_global_variable(int s)
|
3712
|
+
{
|
3713
|
+
#pragma omp critical
|
3714
|
+
{
|
3715
|
+
interrupt_switch = true;
|
3716
|
+
}
|
3717
|
+
}
|
3718
|
+
|
3719
|
+
void check_interrupt_switch(SignalSwitcher &ss)
|
3720
|
+
{
|
3721
|
+
if (interrupt_switch)
|
3722
|
+
{
|
3723
|
+
ss.restore_handle();
|
3724
|
+
fprintf(stderr, "Error: procedure was interrupted\n");
|
3725
|
+
raise(SIGINT);
|
3726
|
+
#ifdef _FOR_R
|
3727
|
+
Rcpp::checkUserInterrupt();
|
3728
|
+
#elif !defined(DONT_THROW_ON_INTERRUPT)
|
3729
|
+
throw std::runtime_error("Error: procedure was interrupted.\n");
|
3730
|
+
#endif
|
3731
|
+
}
|
3732
|
+
}
|
3733
|
+
|
3734
|
+
#ifdef _FOR_PYTHON
|
3735
|
+
bool cy_check_interrupt_switch()
|
3736
|
+
{
|
3737
|
+
return interrupt_switch;
|
3738
|
+
}
|
3739
|
+
|
3740
|
+
void cy_tick_off_interrupt_switch()
|
3741
|
+
{
|
3742
|
+
interrupt_switch = false;
|
3743
|
+
}
|
3744
|
+
#endif
|
3745
|
+
|
3746
|
+
SignalSwitcher::SignalSwitcher()
|
3747
|
+
{
|
3748
|
+
#pragma omp critical
|
3749
|
+
{
|
3750
|
+
if (!handle_is_locked)
|
3751
|
+
{
|
3752
|
+
handle_is_locked = true;
|
3753
|
+
interrupt_switch = false;
|
3754
|
+
this->old_sig = signal(SIGINT, set_interrup_global_variable);
|
3755
|
+
this->is_active = true;
|
3756
|
+
}
|
3757
|
+
|
3758
|
+
else {
|
3759
|
+
this->is_active = false;
|
3760
|
+
}
|
3761
|
+
}
|
3762
|
+
}
|
3763
|
+
|
3764
|
+
SignalSwitcher::~SignalSwitcher()
|
3765
|
+
{
|
3766
|
+
#ifndef _FOR_PYTHON
|
3767
|
+
#pragma omp critical
|
3768
|
+
{
|
3769
|
+
if (this->is_active && handle_is_locked)
|
3770
|
+
interrupt_switch = false;
|
3771
|
+
}
|
3772
|
+
#endif
|
3773
|
+
this->restore_handle();
|
3774
|
+
}
|
3775
|
+
|
3776
|
+
void SignalSwitcher::restore_handle()
|
3777
|
+
{
|
3778
|
+
#pragma omp critical
|
3779
|
+
{
|
3780
|
+
if (this->is_active && handle_is_locked)
|
3781
|
+
{
|
3782
|
+
signal(SIGINT, this->old_sig);
|
3783
|
+
this->is_active = false;
|
3784
|
+
handle_is_locked = false;
|
3785
|
+
}
|
3786
|
+
}
|
3787
|
+
}
|
3788
|
+
|
3789
|
+
bool has_long_double()
|
3790
|
+
{
|
3791
|
+
#ifndef NO_LONG_DOUBLE
|
3792
|
+
return sizeof(long double) > sizeof(double);
|
3793
|
+
#else
|
3794
|
+
return false;
|
3795
|
+
#endif
|
3796
|
+
}
|
3797
|
+
|
3798
|
+
/* Return the #def'd constants from standard header. This is in order to determine if the return
|
3799
|
+
value from the 'fit_model' function is a success or failure within Cython, which does not
|
3800
|
+
allow importing #def'd macro values. */
|
3801
|
+
int return_EXIT_SUCCESS()
|
3802
|
+
{
|
3803
|
+
return EXIT_SUCCESS;
|
3804
|
+
}
|
3805
|
+
int return_EXIT_FAILURE()
|
3806
|
+
{
|
3807
|
+
return EXIT_FAILURE;
|
3808
|
+
}
|