isotree 0.2.2 → 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 +8 -1
- data/LICENSE.txt +2 -2
- data/README.md +32 -14
- data/ext/isotree/ext.cpp +144 -31
- data/ext/isotree/extconf.rb +7 -7
- data/lib/isotree/isolation_forest.rb +110 -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 +38 -104
- 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 +1,1886 @@
|
|
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
|
+
|
66
|
+
/* Calculate distance or similarity or kernel/proximity between data points
|
67
|
+
*
|
68
|
+
* Parameters
|
69
|
+
* ==========
|
70
|
+
* - numeric_data[nrows * ncols_numeric]
|
71
|
+
* Pointer to numeric data for which to make calculations. If not using 'indexer', must be
|
72
|
+
* ordered by columns like Fortran, not ordered by rows like C (i.e. entries 1..n contain
|
73
|
+
* column 0, n+1..2n column 1, etc.), while if using 'indexer', may be passed in either
|
74
|
+
* row-major or column-major format (with row-major being faster).
|
75
|
+
* If categorical data is passed, must be in the same storage order (row-major / column-major)
|
76
|
+
* as numerical data (whether dense or sparse).
|
77
|
+
* The column order must be the same as in the data that was used to fit the model.
|
78
|
+
* If making calculations between two sets of observations/rows (see documentation for 'rmat'),
|
79
|
+
* the first group is assumed to be the earlier rows here.
|
80
|
+
* Pass NULL if there are no dense numeric columns.
|
81
|
+
* Can only pass one of 'numeric_data' or 'Xc' + 'Xc_ind' + 'Xc_indptr'.
|
82
|
+
* - categ_data[nrows * ncols_categ]
|
83
|
+
* Pointer to categorical data for which to make calculations. If not using 'indexer', must be
|
84
|
+
* ordered by columns like Fortran, not ordered by rows like C (i.e. entries 1..n contain
|
85
|
+
* column 0, n+1..2n column 1, etc.), while if using 'indexer', may be passed in either
|
86
|
+
* row-major or column-major format (with row-major being faster).
|
87
|
+
* If numerical data is passed, must be in the same storage order (row-major / column-major)
|
88
|
+
* as categorical data (whether the numerical data is dense or sparse).
|
89
|
+
* Each category should be represented as an integer, and these integers must start at zero and
|
90
|
+
* be in consecutive order - i.e. if category '3' is present, category '2' must have also been
|
91
|
+
* present when the model was fit (note that they are not treated as being ordinal, this is just
|
92
|
+
* an encoding). Missing values should be encoded as negative numbers such as (-1). The encoding
|
93
|
+
* must be the same as was used in the data to which the model was fit.
|
94
|
+
* Pass NULL if there are no categorical columns.
|
95
|
+
* If making calculations between two sets of observations/rows (see documentation for 'rmat'),
|
96
|
+
* the first group is assumed to be the earlier rows here.
|
97
|
+
* - Xc[nnz]
|
98
|
+
* Pointer to numeric data in sparse numeric matrix in CSC format (column-compressed),
|
99
|
+
* or optionally in CSR format (row-compressed) if using 'indexer' and passing 'is_col_major=false'
|
100
|
+
* (not recommended as the calculations will be slower if sparse data is passed as CSR).
|
101
|
+
* If categorical data is passed, must be in the same storage order (row-major or CSR / column-major or CSC)
|
102
|
+
* as numerical data (whether dense or sparse).
|
103
|
+
* Pass NULL if there are no sparse numeric columns.
|
104
|
+
* Can only pass one of 'numeric_data' or 'Xc' + 'Xc_ind' + 'Xc_indptr'.
|
105
|
+
* - Xc_ind[nnz]
|
106
|
+
* Pointer to row indices to which each non-zero entry in 'Xc' corresponds
|
107
|
+
* (column indices if 'Xc' is in CSR format).
|
108
|
+
* Must be in sorted order, otherwise results will be incorrect.
|
109
|
+
* Pass NULL if there are no sparse numeric columns in CSC or CSR format.
|
110
|
+
* - Xc_indptr[ncols_categ + 1]
|
111
|
+
* Pointer to column index pointers that tell at entry [col] where does column 'col'
|
112
|
+
* start and at entry [col + 1] where does column 'col' end
|
113
|
+
* (row index pointers if 'Xc' is passed in CSR format).
|
114
|
+
* Pass NULL if there are no sparse numeric columns in CSC or CSR format.
|
115
|
+
* If making calculations between two sets of observations/rows (see documentation for 'rmat'),
|
116
|
+
* the first group is assumed to be the earlier rows here.
|
117
|
+
* - nrows
|
118
|
+
* Number of rows in 'numeric_data', 'Xc', 'categ_data'.
|
119
|
+
* - use_long_double
|
120
|
+
* Whether to use 'long double' (extended precision) type for the calculations. This makes them
|
121
|
+
* more accurate (provided that the compiler used has wider long doubles than doubles), but
|
122
|
+
* slower - especially in platforms in which 'long double' is a software-emulated type (e.g.
|
123
|
+
* Power8 platforms).
|
124
|
+
* - nthreads
|
125
|
+
* Number of parallel threads to use. Note that, the more threads, the more memory will be
|
126
|
+
* allocated, even if the thread does not end up being used (with one exception being kernel calculations
|
127
|
+
* with respect to reference points in an idexer). Ignored when not building with OpenMP support.
|
128
|
+
* - assume_full_distr
|
129
|
+
* Whether to assume that the fitted model represents a full population distribution (will use a
|
130
|
+
* standardizing criterion assuming infinite sample, and the results of the similarity between two points
|
131
|
+
* at prediction time will not depend on the prescence of any third point that is similar to them, but will
|
132
|
+
* differ more compared to the pairwise distances between points from which the model was fit). If passing
|
133
|
+
* 'false', will calculate pairwise distances as if the new observations at prediction time were added to
|
134
|
+
* the sample to which each tree was fit, which will make the distances between two points potentially vary
|
135
|
+
* according to other newly introduced points.
|
136
|
+
* This was added for experimentation purposes only and it's not recommended to pass 'false'.
|
137
|
+
* Note that when calculating distances using 'indexer', there
|
138
|
+
* might be slight discrepancies between the numbers produced with or without the indexer due to what
|
139
|
+
* are considered "additional" observations in this calculation.
|
140
|
+
* This is ignored when passing 'as_kernel=true'.
|
141
|
+
* - standardize_dist
|
142
|
+
* Whether to standardize the resulting average separation depths between rows according
|
143
|
+
* to the expected average separation depth in a similar way as when predicting outlierness,
|
144
|
+
* in order to obtain a standardized distance. If passing 'false', will output the average
|
145
|
+
* separation depth instead.
|
146
|
+
* If passing 'as_kernel=true', this indicates whether to output a fraction (if 'true') or
|
147
|
+
* the raw number of matching trees (if 'false').
|
148
|
+
* - as_kernel
|
149
|
+
* Whether to calculate the "similarities" as isolation kernel or proximity matrix, which counts
|
150
|
+
* the proportion of trees in which two observations end up in the same terminal node. This is
|
151
|
+
* typically much faster than separation-based distance, but is typically not as good quality.
|
152
|
+
* Note that, for kernel calculations, the indexer is only used if it has reference points stored on it.
|
153
|
+
* - model_outputs
|
154
|
+
* Pointer to fitted single-variable model object from function 'fit_iforest'. Pass NULL
|
155
|
+
* if the calculations are to be made from an extended model. Can only pass one of
|
156
|
+
* 'model_outputs' and 'model_outputs_ext'.
|
157
|
+
* - model_outputs_ext
|
158
|
+
* Pointer to fitted extended model object from function 'fit_iforest'. Pass NULL
|
159
|
+
* if the calculations are to be made from a single-variable model. Can only pass one of
|
160
|
+
* 'model_outputs' and 'model_outputs_ext'.
|
161
|
+
* - tmat[nrows * (nrows - 1) / 2] (out)
|
162
|
+
* Pointer to array where the resulting pairwise distances or average separation depths or kernels will
|
163
|
+
* be written into. As the output is a symmetric matrix, this function will only fill in the
|
164
|
+
* upper-triangular part, in which entry 0 <= i < j < n will be located at position
|
165
|
+
* p(i,j) = (i * (n - (i+1)/2) + j - i - 1).
|
166
|
+
* Can be converted to a dense square matrix through function 'tmat_to_dense'.
|
167
|
+
* The array must already be initialized to zeros.
|
168
|
+
* If calculating distance/separation from a group of points to another group of points,
|
169
|
+
* pass NULL here and use 'rmat' instead.
|
170
|
+
* - rmat[nrows1 * nrows2] (out)
|
171
|
+
* Pointer to array where to write the distances or separation depths or kernels between each row in
|
172
|
+
* one set of observations and each row in a different set of observations. If doing these
|
173
|
+
* calculations for all pairs of observations/rows, pass 'tmat' instead.
|
174
|
+
* Will take the first group of observations as the rows in this matrix, and the second
|
175
|
+
* group as the columns. The groups are assumed to be in the same data arrays, with the
|
176
|
+
* first group corresponding to the earlier rows there.
|
177
|
+
* This matrix will be used in row-major order (i.e. entries 1..nrows2 contain the first row from nrows1).
|
178
|
+
* Must be already initialized to zeros.
|
179
|
+
* If passing 'use_indexed_references=true' plus an indexer object with reference points, this
|
180
|
+
* array should have dimension [nrows, n_references].
|
181
|
+
* Ignored when 'tmat' is passed.
|
182
|
+
* - n_from
|
183
|
+
* When calculating distances between two groups of points, this indicates the number of
|
184
|
+
* observations/rows belonging to the first group (the rows in 'rmat'), which will be
|
185
|
+
* assumed to be the first 'n_from' rows.
|
186
|
+
* Ignored when 'tmat' is passed or when 'use_indexed_references=true' plus an indexer with
|
187
|
+
* references are passed.
|
188
|
+
* - use_indexed_references
|
189
|
+
* Whether to calculate distances with respect to reference points stored in the indexer
|
190
|
+
* object, if it has any. This is only supported with 'assume_full_distr=true' or with 'as_kernel=true'.
|
191
|
+
* If passing 'use_indexed_references=true', then 'tmat' must be NULL, and 'rmat' must
|
192
|
+
* be of dimension [nrows, n_references].
|
193
|
+
* - indexer
|
194
|
+
* Pointer to associated tree indexer for the model being used, if it was constructed,
|
195
|
+
* which can be used to speed up distance calculations, assuming that it was built with
|
196
|
+
* option 'with_distances=true'. If it does not contain node distances, it will not be used.
|
197
|
+
* Pass NULL if the indexer has not been constructed or was constructed with 'with_distances=false'.
|
198
|
+
* If it contains reference points and passing 'use_indexed_references=true', distances will be
|
199
|
+
* calculated between between the input data passed here and the reference points stored in this object.
|
200
|
+
* If passing 'as_kernel=true', the indexer can only be used for calculating kernels with respect to
|
201
|
+
* reference points in the indexer, otherwise it will not be used (which also means that the data must be
|
202
|
+
* passed in column-major order for all kernel calculations that are not with respect to reference points
|
203
|
+
* from an indexer).
|
204
|
+
* - is_col_major
|
205
|
+
* Whether the data comes in column-major order. If using 'indexer', predictions are also possible
|
206
|
+
* (and are even faster for the case of dense-only data) if passing the data in row-major format.
|
207
|
+
* Without 'indexer' (and with 'as_kernel=true' but without reference points in the idnexer), data
|
208
|
+
* may only be passed in column-major format.
|
209
|
+
* If there is sparse numeric data, it is highly suggested to pass it in CSC/column-major format.
|
210
|
+
* - ld_numeric
|
211
|
+
* If passing 'is_col_major=false', this indicates the leading dimension of the array 'numeric_data'.
|
212
|
+
* Typically, this corresponds to the number of columns, but may be larger (the array will
|
213
|
+
* be accessed assuming that row 'n' starts at 'numeric_data + n*ld_numeric'). If passing
|
214
|
+
* 'numeric_data' in column-major order, this is ignored and will be assumed that the
|
215
|
+
* leading dimension corresponds to the number of rows. This is ignored when passing numeric
|
216
|
+
* data in sparse format.
|
217
|
+
* Note that data in row-major order is only accepted when using 'indexer'.
|
218
|
+
* - ld_categ
|
219
|
+
* If passing 'is_col_major=false', this indicates the leading dimension of the array 'categ_data'.
|
220
|
+
* Typically, this corresponds to the number of columns, but may be larger (the array will
|
221
|
+
* be accessed assuming that row 'n' starts at 'categ_data + n*ld_categ'). If passing
|
222
|
+
* 'categ_data' in column-major order, this is ignored and will be assumed that the
|
223
|
+
* leading dimension corresponds to the number of rows.
|
224
|
+
* Note that data in row-major order is only accepted when using 'indexer'.
|
225
|
+
*/
|
226
|
+
template <class real_t, class sparse_ix>
|
227
|
+
void calc_similarity(real_t numeric_data[], int categ_data[],
|
228
|
+
real_t Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
229
|
+
size_t nrows, bool use_long_double, int nthreads,
|
230
|
+
bool assume_full_distr, bool standardize_dist, bool as_kernel,
|
231
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
232
|
+
double tmat[], double rmat[], size_t n_from, bool use_indexed_references,
|
233
|
+
TreesIndexer *indexer, bool is_col_major, size_t ld_numeric, size_t ld_categ)
|
234
|
+
{
|
235
|
+
if (use_long_double && !has_long_double()) {
|
236
|
+
use_long_double = false;
|
237
|
+
fprintf(stderr, "Passed 'use_long_double=true', but library was compiled without long double support.\n");
|
238
|
+
}
|
239
|
+
#ifndef NO_LONG_DOUBLE
|
240
|
+
if (likely(!use_long_double))
|
241
|
+
#endif
|
242
|
+
calc_similarity_internal<real_t, sparse_ix, double>(
|
243
|
+
numeric_data, categ_data,
|
244
|
+
Xc, Xc_ind, Xc_indptr,
|
245
|
+
nrows, nthreads,
|
246
|
+
assume_full_distr, standardize_dist, as_kernel,
|
247
|
+
model_outputs, model_outputs_ext,
|
248
|
+
tmat, rmat, n_from, use_indexed_references,
|
249
|
+
indexer, is_col_major, ld_numeric, ld_categ
|
250
|
+
);
|
251
|
+
#ifndef NO_LONG_DOUBLE
|
252
|
+
else
|
253
|
+
calc_similarity_internal<real_t, sparse_ix, long double>(
|
254
|
+
numeric_data, categ_data,
|
255
|
+
Xc, Xc_ind, Xc_indptr,
|
256
|
+
nrows, nthreads,
|
257
|
+
assume_full_distr, standardize_dist, as_kernel,
|
258
|
+
model_outputs, model_outputs_ext,
|
259
|
+
tmat, rmat, n_from, use_indexed_references,
|
260
|
+
indexer, is_col_major, ld_numeric, ld_categ
|
261
|
+
);
|
262
|
+
#endif
|
263
|
+
}
|
264
|
+
|
265
|
+
template <class real_t, class sparse_ix, class ldouble_safe>
|
266
|
+
void calc_similarity_internal(
|
267
|
+
real_t numeric_data[], int categ_data[],
|
268
|
+
real_t Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
269
|
+
size_t nrows, int nthreads,
|
270
|
+
bool assume_full_distr, bool standardize_dist, bool as_kernel,
|
271
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
272
|
+
double tmat[], double rmat[], size_t n_from, bool use_indexed_references,
|
273
|
+
TreesIndexer *indexer, bool is_col_major, size_t ld_numeric, size_t ld_categ)
|
274
|
+
{
|
275
|
+
if (nrows < 2 && (!use_indexed_references || indexer == NULL || indexer->indices.empty() || indexer->indices.front().reference_points.empty()))
|
276
|
+
throw std::runtime_error("Cannot calculate distances from less than 2 rows.\n");
|
277
|
+
if (as_kernel && (tmat != NULL || !use_indexed_references || (indexer != NULL && !indexer->indices.empty() && indexer->indices.front().reference_points.empty())))
|
278
|
+
indexer = NULL;
|
279
|
+
|
280
|
+
if (indexer != NULL && model_outputs != NULL)
|
281
|
+
{
|
282
|
+
if (model_outputs->missing_action == Divide) {
|
283
|
+
indexer = NULL;
|
284
|
+
if (use_indexed_references) throw std::runtime_error("Invalid indexer - cannot use references from it.\n");
|
285
|
+
}
|
286
|
+
if (model_outputs->new_cat_action == Weighted && model_outputs->cat_split_type == SubSet && categ_data != NULL) {
|
287
|
+
indexer = NULL;
|
288
|
+
if (use_indexed_references) throw std::runtime_error("Invalid indexer - cannot use references from it.\n");
|
289
|
+
}
|
290
|
+
}
|
291
|
+
if (
|
292
|
+
!as_kernel &&
|
293
|
+
indexer != NULL &&
|
294
|
+
(indexer->indices.empty() || indexer->indices.front().node_distances.empty())
|
295
|
+
) {
|
296
|
+
if (use_indexed_references && !indexer->indices.empty() && !indexer->indices.front().reference_points.empty())
|
297
|
+
throw std::runtime_error("Indexer was built without distances. Cannot use references from it.\n");
|
298
|
+
else {
|
299
|
+
indexer = NULL;
|
300
|
+
fprintf(stderr, "Indexer has no pre-computed distances, will not be used for distance calculations.\n");
|
301
|
+
}
|
302
|
+
}
|
303
|
+
if (
|
304
|
+
!is_col_major &&
|
305
|
+
indexer == NULL &&
|
306
|
+
(
|
307
|
+
Xc_indptr != NULL
|
308
|
+
||
|
309
|
+
(nrows != 1 &&
|
310
|
+
((numeric_data != NULL && ld_numeric > 1) || (categ_data != NULL && ld_categ > 1)))
|
311
|
+
)
|
312
|
+
)
|
313
|
+
throw std::runtime_error("Cannot calculate distances with row-major data without indexer.\n");
|
314
|
+
if (indexer != NULL)
|
315
|
+
{
|
316
|
+
if (use_indexed_references && tmat == NULL && !indexer->indices.empty() && !indexer->indices.front().reference_points.empty())
|
317
|
+
{
|
318
|
+
if (unlikely(!assume_full_distr))
|
319
|
+
throw std::runtime_error("Cannot calculate distances to reference points in indexer with 'assume_full_distr=false'.\n");
|
320
|
+
|
321
|
+
if (!as_kernel)
|
322
|
+
{
|
323
|
+
calc_similarity_from_indexer_with_references(
|
324
|
+
numeric_data, categ_data,
|
325
|
+
Xc, Xc_ind, Xc_indptr,
|
326
|
+
nrows, nthreads, standardize_dist,
|
327
|
+
model_outputs, model_outputs_ext,
|
328
|
+
rmat,
|
329
|
+
indexer, is_col_major, ld_numeric, ld_categ
|
330
|
+
);
|
331
|
+
}
|
332
|
+
|
333
|
+
else
|
334
|
+
{
|
335
|
+
kernel_to_references(*indexer,
|
336
|
+
model_outputs, model_outputs_ext,
|
337
|
+
numeric_data, categ_data,
|
338
|
+
Xc, Xc_ind, Xc_indptr,
|
339
|
+
is_col_major, ld_numeric, ld_categ,
|
340
|
+
nrows, nthreads,
|
341
|
+
rmat,
|
342
|
+
standardize_dist);
|
343
|
+
}
|
344
|
+
}
|
345
|
+
|
346
|
+
else
|
347
|
+
{
|
348
|
+
if (as_kernel) goto skip_indexer_if_kernel;
|
349
|
+
calc_similarity_from_indexer(
|
350
|
+
numeric_data, categ_data,
|
351
|
+
Xc, Xc_ind, Xc_indptr,
|
352
|
+
nrows, nthreads, assume_full_distr, standardize_dist,
|
353
|
+
model_outputs, model_outputs_ext,
|
354
|
+
tmat, rmat, n_from,
|
355
|
+
indexer, is_col_major, ld_numeric, ld_categ
|
356
|
+
);
|
357
|
+
}
|
358
|
+
|
359
|
+
return;
|
360
|
+
}
|
361
|
+
skip_indexer_if_kernel:
|
362
|
+
|
363
|
+
PredictionData<real_t, sparse_ix>
|
364
|
+
prediction_data = {numeric_data, categ_data, nrows,
|
365
|
+
false, 0, 0,
|
366
|
+
Xc, Xc_ind, Xc_indptr,
|
367
|
+
NULL, NULL, NULL};
|
368
|
+
|
369
|
+
size_t ntrees = (model_outputs != NULL)? model_outputs->trees.size() : model_outputs_ext->hplanes.size();
|
370
|
+
|
371
|
+
if (tmat != NULL) n_from = 0;
|
372
|
+
|
373
|
+
if (n_from == 0) {
|
374
|
+
#if SIZE_MAX == UINT32_MAX
|
375
|
+
size_t lim_rows = (size_t)UINT16_MAX - (size_t)1;
|
376
|
+
#elif SIZE_MAX == UINT64_MAX
|
377
|
+
size_t lim_rows = (size_t)UINT32_MAX - (size_t)1;
|
378
|
+
#else
|
379
|
+
size_t lim_rows = (size_t)std::ceil(std::sqrt((ldouble_safe)SIZE_MAX));
|
380
|
+
#endif
|
381
|
+
if (nrows > lim_rows)
|
382
|
+
throw std::runtime_error("Number of rows implies too large distance matrix (integer overflow).");
|
383
|
+
}
|
384
|
+
|
385
|
+
if ((size_t)nthreads > ntrees)
|
386
|
+
nthreads = (int)ntrees;
|
387
|
+
#ifdef _OPENMP
|
388
|
+
std::vector<WorkerForSimilarity> worker_memory(nthreads);
|
389
|
+
#else
|
390
|
+
std::vector<WorkerForSimilarity> worker_memory(1);
|
391
|
+
nthreads = 1;
|
392
|
+
#endif
|
393
|
+
|
394
|
+
/* Global variable that determines if the procedure receives a stop signal */
|
395
|
+
SignalSwitcher ss = SignalSwitcher();
|
396
|
+
check_interrupt_switch(ss);
|
397
|
+
#if defined(DONT_THROW_ON_INTERRUPT)
|
398
|
+
if (interrupt_switch) return;
|
399
|
+
#endif
|
400
|
+
/* For handling exceptions */
|
401
|
+
bool threw_exception = false;
|
402
|
+
std::exception_ptr ex = NULL;
|
403
|
+
|
404
|
+
if (
|
405
|
+
tmat == NULL &&
|
406
|
+
use_indexed_references &&
|
407
|
+
indexer != NULL &&
|
408
|
+
!indexer->indices.empty() &&
|
409
|
+
!indexer->indices.front().reference_points.empty() &&
|
410
|
+
(as_kernel || !indexer->indices.front().node_distances.empty())
|
411
|
+
) {
|
412
|
+
n_from = indexer->indices.front().reference_points.size();
|
413
|
+
}
|
414
|
+
|
415
|
+
if (model_outputs != NULL)
|
416
|
+
{
|
417
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) \
|
418
|
+
shared(ntrees, worker_memory, prediction_data, model_outputs, ex, threw_exception, n_from)
|
419
|
+
for (size_t_for tree = 0; tree < (decltype(tree))ntrees; tree++)
|
420
|
+
{
|
421
|
+
if (threw_exception || interrupt_switch) continue;
|
422
|
+
try
|
423
|
+
{
|
424
|
+
initialize_worker_for_sim(worker_memory[omp_get_thread_num()], prediction_data,
|
425
|
+
model_outputs, NULL, n_from, assume_full_distr);
|
426
|
+
traverse_tree_sim<PredictionData<real_t, sparse_ix>, ldouble_safe>(
|
427
|
+
worker_memory[omp_get_thread_num()],
|
428
|
+
prediction_data,
|
429
|
+
*model_outputs,
|
430
|
+
model_outputs->trees[tree],
|
431
|
+
(size_t)0,
|
432
|
+
as_kernel);
|
433
|
+
}
|
434
|
+
|
435
|
+
catch (...)
|
436
|
+
{
|
437
|
+
#pragma omp critical
|
438
|
+
{
|
439
|
+
if (!threw_exception)
|
440
|
+
{
|
441
|
+
threw_exception = true;
|
442
|
+
ex = std::current_exception();
|
443
|
+
}
|
444
|
+
}
|
445
|
+
}
|
446
|
+
}
|
447
|
+
}
|
448
|
+
|
449
|
+
else
|
450
|
+
{
|
451
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) \
|
452
|
+
shared(ntrees, worker_memory, prediction_data, model_outputs_ext, ex, threw_exception, n_from)
|
453
|
+
for (size_t_for hplane = 0; hplane < (decltype(hplane))ntrees; hplane++)
|
454
|
+
{
|
455
|
+
if (threw_exception || interrupt_switch) continue;
|
456
|
+
try
|
457
|
+
{
|
458
|
+
initialize_worker_for_sim(worker_memory[omp_get_thread_num()], prediction_data,
|
459
|
+
NULL, model_outputs_ext, n_from, assume_full_distr);
|
460
|
+
traverse_hplane_sim<PredictionData<real_t, sparse_ix>, ldouble_safe>(
|
461
|
+
worker_memory[omp_get_thread_num()],
|
462
|
+
prediction_data,
|
463
|
+
*model_outputs_ext,
|
464
|
+
model_outputs_ext->hplanes[hplane],
|
465
|
+
(size_t)0,
|
466
|
+
as_kernel);
|
467
|
+
}
|
468
|
+
|
469
|
+
catch (...)
|
470
|
+
{
|
471
|
+
#pragma omp critical
|
472
|
+
{
|
473
|
+
if (!threw_exception)
|
474
|
+
{
|
475
|
+
threw_exception = true;
|
476
|
+
ex = std::current_exception();
|
477
|
+
}
|
478
|
+
}
|
479
|
+
}
|
480
|
+
}
|
481
|
+
}
|
482
|
+
|
483
|
+
check_interrupt_switch(ss);
|
484
|
+
#if defined(DONT_THROW_ON_INTERRUPT)
|
485
|
+
if (interrupt_switch) return;
|
486
|
+
#endif
|
487
|
+
|
488
|
+
if (threw_exception)
|
489
|
+
std::rethrow_exception(ex);
|
490
|
+
|
491
|
+
/* gather and transform the results */
|
492
|
+
gather_sim_result< PredictionData<real_t, sparse_ix>,
|
493
|
+
InputData<real_t, sparse_ix>,
|
494
|
+
WorkerMemory<ImputedData<sparse_ix, ldouble_safe>, ldouble_safe, real_t> >
|
495
|
+
(&worker_memory, NULL,
|
496
|
+
&prediction_data, NULL,
|
497
|
+
model_outputs, model_outputs_ext,
|
498
|
+
tmat, rmat, n_from,
|
499
|
+
ntrees, assume_full_distr,
|
500
|
+
standardize_dist, as_kernel, nthreads);
|
501
|
+
|
502
|
+
check_interrupt_switch(ss);
|
503
|
+
#if defined(DONT_THROW_ON_INTERRUPT)
|
504
|
+
if (interrupt_switch) return;
|
505
|
+
#endif
|
506
|
+
}
|
507
|
+
|
508
|
+
template <class PredictionData, class ldouble_safe>
|
509
|
+
void traverse_tree_sim(WorkerForSimilarity &workspace,
|
510
|
+
PredictionData &prediction_data,
|
511
|
+
IsoForest &model_outputs,
|
512
|
+
std::vector<IsoTree> &trees,
|
513
|
+
size_t curr_tree,
|
514
|
+
const bool as_kernel)
|
515
|
+
{
|
516
|
+
if (interrupt_switch)
|
517
|
+
return;
|
518
|
+
|
519
|
+
if (workspace.st == workspace.end)
|
520
|
+
return;
|
521
|
+
|
522
|
+
if (workspace.tmat_sep.empty())
|
523
|
+
{
|
524
|
+
std::sort(workspace.ix_arr.begin() + workspace.st, workspace.ix_arr.begin() + workspace.end + 1);
|
525
|
+
if (workspace.ix_arr[workspace.st] >= workspace.n_from)
|
526
|
+
return;
|
527
|
+
if (workspace.ix_arr[workspace.end] < workspace.n_from)
|
528
|
+
return;
|
529
|
+
}
|
530
|
+
|
531
|
+
/* Note: the first separation step will not be added here, as it simply consists of adding +1
|
532
|
+
to every combination regardless. It has to be added at the end in 'gather_sim_result' to
|
533
|
+
obtain the average separation depth. */
|
534
|
+
if (trees[curr_tree].tree_left == 0)
|
535
|
+
{
|
536
|
+
ldouble_safe rem = (ldouble_safe) trees[curr_tree].remainder;
|
537
|
+
if (workspace.weights_arr.empty())
|
538
|
+
{
|
539
|
+
if (!as_kernel)
|
540
|
+
{
|
541
|
+
rem += (ldouble_safe)(workspace.end - workspace.st + 1);
|
542
|
+
if (!workspace.tmat_sep.empty())
|
543
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
544
|
+
prediction_data.nrows, workspace.tmat_sep.data(),
|
545
|
+
workspace.assume_full_distr? 3. : expected_separation_depth(rem));
|
546
|
+
else if (!workspace.rmat.empty())
|
547
|
+
increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end,
|
548
|
+
workspace.n_from, prediction_data.nrows, workspace.rmat.data(),
|
549
|
+
workspace.assume_full_distr? 3. : expected_separation_depth(rem));
|
550
|
+
}
|
551
|
+
|
552
|
+
else
|
553
|
+
{
|
554
|
+
if (!workspace.tmat_sep.empty())
|
555
|
+
{
|
556
|
+
size_t i_, j_;
|
557
|
+
for (size_t i = workspace.st; i < workspace.end; i++)
|
558
|
+
{
|
559
|
+
i_ = workspace.ix_arr[i];
|
560
|
+
for (size_t j = i + 1; j <= workspace.end; j++)
|
561
|
+
{
|
562
|
+
j_ = workspace.ix_arr[j];
|
563
|
+
workspace.tmat_sep[ix_comb(i_, j_, prediction_data.nrows, workspace.tmat_sep.size())]++;
|
564
|
+
}
|
565
|
+
}
|
566
|
+
}
|
567
|
+
|
568
|
+
else if (!workspace.rmat.empty())
|
569
|
+
{
|
570
|
+
size_t n_group = std::distance(workspace.ix_arr.begin() + workspace.st,
|
571
|
+
std::lower_bound(workspace.ix_arr.begin() + workspace.st,
|
572
|
+
workspace.ix_arr.begin() + workspace.end + 1,
|
573
|
+
workspace.n_from));
|
574
|
+
double *restrict rmat_this;
|
575
|
+
for (size_t i = workspace.st; i < workspace.st + n_group; i++)
|
576
|
+
{
|
577
|
+
rmat_this = workspace.rmat.data() + workspace.ix_arr[i]*workspace.n_from;
|
578
|
+
for (size_t j = workspace.st + n_group; j <= workspace.end; j++)
|
579
|
+
{
|
580
|
+
rmat_this[workspace.ix_arr[j] - workspace.n_from]++;
|
581
|
+
}
|
582
|
+
}
|
583
|
+
}
|
584
|
+
}
|
585
|
+
}
|
586
|
+
|
587
|
+
else
|
588
|
+
{
|
589
|
+
if (!as_kernel)
|
590
|
+
{
|
591
|
+
if (!workspace.assume_full_distr)
|
592
|
+
{
|
593
|
+
rem += std::accumulate(workspace.ix_arr.begin() + workspace.st,
|
594
|
+
workspace.ix_arr.begin() + workspace.end,
|
595
|
+
(ldouble_safe) 0.,
|
596
|
+
[&workspace](ldouble_safe curr, size_t ix)
|
597
|
+
{return curr + (ldouble_safe)workspace.weights_arr[ix];}
|
598
|
+
);
|
599
|
+
}
|
600
|
+
|
601
|
+
if (!workspace.tmat_sep.empty())
|
602
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
603
|
+
prediction_data.nrows, workspace.tmat_sep.data(),
|
604
|
+
workspace.weights_arr.data(),
|
605
|
+
workspace.assume_full_distr? 3. : expected_separation_depth(rem));
|
606
|
+
else if (!workspace.rmat.empty())
|
607
|
+
increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end,
|
608
|
+
workspace.n_from, prediction_data.nrows,
|
609
|
+
workspace.rmat.data(), workspace.weights_arr.data(),
|
610
|
+
workspace.assume_full_distr? 3. : expected_separation_depth(rem));
|
611
|
+
}
|
612
|
+
|
613
|
+
else
|
614
|
+
{
|
615
|
+
if (!workspace.tmat_sep.empty())
|
616
|
+
{
|
617
|
+
size_t i_, j_;
|
618
|
+
double w_this;
|
619
|
+
for (size_t i = workspace.st; i < workspace.end; i++)
|
620
|
+
{
|
621
|
+
i_ = workspace.ix_arr[i];
|
622
|
+
w_this = workspace.weights_arr[i_];
|
623
|
+
for (size_t j = i + 1; j <= workspace.end; j++)
|
624
|
+
{
|
625
|
+
j_ = workspace.ix_arr[j];
|
626
|
+
workspace.tmat_sep[ix_comb(i_, j_, prediction_data.nrows, workspace.tmat_sep.size())]
|
627
|
+
+=
|
628
|
+
w_this * workspace.weights_arr[j_];
|
629
|
+
}
|
630
|
+
}
|
631
|
+
}
|
632
|
+
|
633
|
+
else if (!workspace.rmat.empty())
|
634
|
+
{
|
635
|
+
size_t n_group = std::distance(workspace.ix_arr.begin() + workspace.st,
|
636
|
+
std::lower_bound(workspace.ix_arr.begin() + workspace.st,
|
637
|
+
workspace.ix_arr.begin() + workspace.end + 1,
|
638
|
+
workspace.n_from));
|
639
|
+
double *restrict rmat_this;
|
640
|
+
double w_this;
|
641
|
+
size_t i_, j_;
|
642
|
+
for (size_t i = workspace.st; i < workspace.st + n_group; i++)
|
643
|
+
{
|
644
|
+
i_ = workspace.ix_arr[i];
|
645
|
+
rmat_this = workspace.rmat.data() + i_*workspace.n_from;
|
646
|
+
w_this = workspace.weights_arr[i_];
|
647
|
+
for (size_t j = workspace.st + n_group; j <= workspace.end; j++)
|
648
|
+
{
|
649
|
+
j_ = workspace.ix_arr[j];
|
650
|
+
rmat_this[j_ - workspace.n_from]
|
651
|
+
+=
|
652
|
+
w_this * workspace.weights_arr[j_];
|
653
|
+
}
|
654
|
+
}
|
655
|
+
}
|
656
|
+
}
|
657
|
+
}
|
658
|
+
return;
|
659
|
+
}
|
660
|
+
|
661
|
+
else if (curr_tree > 0 && !as_kernel)
|
662
|
+
{
|
663
|
+
if (!workspace.tmat_sep.empty())
|
664
|
+
{
|
665
|
+
if (workspace.weights_arr.empty())
|
666
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
667
|
+
prediction_data.nrows, workspace.tmat_sep.data(), -1.);
|
668
|
+
else
|
669
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
670
|
+
prediction_data.nrows, workspace.tmat_sep.data(),
|
671
|
+
workspace.weights_arr.data(), -1.);
|
672
|
+
}
|
673
|
+
else if (!workspace.rmat.empty())
|
674
|
+
{
|
675
|
+
if (workspace.weights_arr.empty())
|
676
|
+
increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end,
|
677
|
+
workspace.n_from, prediction_data.nrows, workspace.rmat.data(), -1.);
|
678
|
+
else
|
679
|
+
increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end,
|
680
|
+
workspace.n_from, prediction_data.nrows,
|
681
|
+
workspace.rmat.data(), workspace.weights_arr.data(), -1.);
|
682
|
+
}
|
683
|
+
}
|
684
|
+
|
685
|
+
|
686
|
+
/* divide according to tree */
|
687
|
+
if (prediction_data.Xc_indptr != NULL && !workspace.tmat_sep.empty())
|
688
|
+
std::sort(workspace.ix_arr.begin() + workspace.st, workspace.ix_arr.begin() + workspace.end + 1);
|
689
|
+
size_t st_NA, end_NA, split_ix;
|
690
|
+
switch (trees[curr_tree].col_type)
|
691
|
+
{
|
692
|
+
case Numeric:
|
693
|
+
{
|
694
|
+
if (prediction_data.Xc_indptr == NULL)
|
695
|
+
divide_subset_split(workspace.ix_arr.data(),
|
696
|
+
prediction_data.numeric_data + prediction_data.nrows * trees[curr_tree].col_num,
|
697
|
+
workspace.st, workspace.end, trees[curr_tree].num_split,
|
698
|
+
model_outputs.missing_action, st_NA, end_NA, split_ix);
|
699
|
+
else
|
700
|
+
divide_subset_split(workspace.ix_arr.data(), workspace.st, workspace.end, trees[curr_tree].col_num,
|
701
|
+
prediction_data.Xc, prediction_data.Xc_ind, prediction_data.Xc_indptr,
|
702
|
+
trees[curr_tree].num_split, model_outputs.missing_action,
|
703
|
+
st_NA, end_NA, split_ix);
|
704
|
+
break;
|
705
|
+
}
|
706
|
+
|
707
|
+
case Categorical:
|
708
|
+
{
|
709
|
+
switch(model_outputs.cat_split_type)
|
710
|
+
{
|
711
|
+
case SingleCateg:
|
712
|
+
{
|
713
|
+
divide_subset_split(workspace.ix_arr.data(),
|
714
|
+
prediction_data.categ_data + prediction_data.nrows * trees[curr_tree].col_num,
|
715
|
+
workspace.st, workspace.end, trees[curr_tree].chosen_cat,
|
716
|
+
model_outputs.missing_action, st_NA, end_NA, split_ix);
|
717
|
+
break;
|
718
|
+
}
|
719
|
+
|
720
|
+
case SubSet:
|
721
|
+
{
|
722
|
+
if (!trees[curr_tree].cat_split.size())
|
723
|
+
divide_subset_split(workspace.ix_arr.data(),
|
724
|
+
prediction_data.categ_data + prediction_data.nrows * trees[curr_tree].col_num,
|
725
|
+
workspace.st, workspace.end,
|
726
|
+
model_outputs.missing_action, model_outputs.new_cat_action,
|
727
|
+
trees[curr_tree].pct_tree_left < .5, st_NA, end_NA, split_ix);
|
728
|
+
else
|
729
|
+
divide_subset_split(workspace.ix_arr.data(),
|
730
|
+
prediction_data.categ_data + prediction_data.nrows * trees[curr_tree].col_num,
|
731
|
+
workspace.st, workspace.end, trees[curr_tree].cat_split.data(),
|
732
|
+
(int) trees[curr_tree].cat_split.size(),
|
733
|
+
model_outputs.missing_action, model_outputs.new_cat_action,
|
734
|
+
(bool)(trees[curr_tree].pct_tree_left < .5), st_NA, end_NA, split_ix);
|
735
|
+
break;
|
736
|
+
}
|
737
|
+
}
|
738
|
+
break;
|
739
|
+
}
|
740
|
+
|
741
|
+
default:
|
742
|
+
{
|
743
|
+
assert(0);
|
744
|
+
break;
|
745
|
+
}
|
746
|
+
}
|
747
|
+
|
748
|
+
|
749
|
+
/* continue splitting recursively */
|
750
|
+
size_t orig_end = workspace.end;
|
751
|
+
if (model_outputs.new_cat_action == Weighted && model_outputs.cat_split_type == SubSet && prediction_data.categ_data != NULL) {
|
752
|
+
if (model_outputs.missing_action == Fail && trees[curr_tree].col_type == Numeric) {
|
753
|
+
st_NA = split_ix;
|
754
|
+
end_NA = split_ix;
|
755
|
+
}
|
756
|
+
goto missing_action_divide;
|
757
|
+
}
|
758
|
+
switch (model_outputs.missing_action)
|
759
|
+
{
|
760
|
+
case Impute:
|
761
|
+
{
|
762
|
+
split_ix = (trees[curr_tree].pct_tree_left >= .5)? end_NA : st_NA;
|
763
|
+
}
|
764
|
+
|
765
|
+
case Fail:
|
766
|
+
{
|
767
|
+
if (split_ix > workspace.st)
|
768
|
+
{
|
769
|
+
workspace.end = split_ix - 1;
|
770
|
+
traverse_tree_sim<PredictionData, ldouble_safe>(
|
771
|
+
workspace,
|
772
|
+
prediction_data,
|
773
|
+
model_outputs,
|
774
|
+
trees,
|
775
|
+
trees[curr_tree].tree_left,
|
776
|
+
as_kernel);
|
777
|
+
}
|
778
|
+
|
779
|
+
|
780
|
+
if (split_ix <= orig_end)
|
781
|
+
{
|
782
|
+
workspace.st = split_ix;
|
783
|
+
workspace.end = orig_end;
|
784
|
+
traverse_tree_sim<PredictionData, ldouble_safe>(
|
785
|
+
workspace,
|
786
|
+
prediction_data,
|
787
|
+
model_outputs,
|
788
|
+
trees,
|
789
|
+
trees[curr_tree].tree_right,
|
790
|
+
as_kernel);
|
791
|
+
}
|
792
|
+
break;
|
793
|
+
}
|
794
|
+
|
795
|
+
case Divide: /* new_cat_action = 'Weighted' will also fall here */
|
796
|
+
{
|
797
|
+
/* TODO: this one should also have a parameter 'changed_weoghts' like during fitting */
|
798
|
+
missing_action_divide:
|
799
|
+
/* TODO: maybe here it shouldn't copy the whole ix_arr,
|
800
|
+
but then it'd need to re-generate it from outside too */
|
801
|
+
std::vector<double> weights_arr;
|
802
|
+
std::vector<size_t> ix_arr;
|
803
|
+
if (end_NA > workspace.st)
|
804
|
+
{
|
805
|
+
weights_arr.assign(workspace.weights_arr.begin(),
|
806
|
+
workspace.weights_arr.begin() + end_NA);
|
807
|
+
ix_arr.assign(workspace.ix_arr.begin(),
|
808
|
+
workspace.ix_arr.begin() + end_NA);
|
809
|
+
}
|
810
|
+
|
811
|
+
if (end_NA > workspace.st)
|
812
|
+
{
|
813
|
+
workspace.end = end_NA - 1;
|
814
|
+
for (size_t row = st_NA; row < end_NA; row++)
|
815
|
+
workspace.weights_arr[workspace.ix_arr[row]] *= trees[curr_tree].pct_tree_left;
|
816
|
+
traverse_tree_sim<PredictionData, ldouble_safe>(
|
817
|
+
workspace,
|
818
|
+
prediction_data,
|
819
|
+
model_outputs,
|
820
|
+
trees,
|
821
|
+
trees[curr_tree].tree_left,
|
822
|
+
as_kernel);
|
823
|
+
}
|
824
|
+
|
825
|
+
if (st_NA <= orig_end)
|
826
|
+
{
|
827
|
+
workspace.st = st_NA;
|
828
|
+
workspace.end = orig_end;
|
829
|
+
if (!weights_arr.empty())
|
830
|
+
{
|
831
|
+
std::copy(weights_arr.begin(),
|
832
|
+
weights_arr.end(),
|
833
|
+
workspace.weights_arr.begin());
|
834
|
+
std::copy(ix_arr.begin(),
|
835
|
+
ix_arr.end(),
|
836
|
+
workspace.ix_arr.begin());
|
837
|
+
weights_arr.clear();
|
838
|
+
weights_arr.shrink_to_fit();
|
839
|
+
ix_arr.clear();
|
840
|
+
ix_arr.shrink_to_fit();
|
841
|
+
}
|
842
|
+
|
843
|
+
for (size_t row = st_NA; row < end_NA; row++)
|
844
|
+
workspace.weights_arr[workspace.ix_arr[row]] *= (1. - trees[curr_tree].pct_tree_left);
|
845
|
+
traverse_tree_sim<PredictionData, ldouble_safe>(
|
846
|
+
workspace,
|
847
|
+
prediction_data,
|
848
|
+
model_outputs,
|
849
|
+
trees,
|
850
|
+
trees[curr_tree].tree_right,
|
851
|
+
as_kernel);
|
852
|
+
}
|
853
|
+
break;
|
854
|
+
}
|
855
|
+
}
|
856
|
+
}
|
857
|
+
|
858
|
+
template <class PredictionData, class ldouble_safe>
|
859
|
+
void traverse_hplane_sim(WorkerForSimilarity &workspace,
|
860
|
+
PredictionData &prediction_data,
|
861
|
+
ExtIsoForest &model_outputs,
|
862
|
+
std::vector<IsoHPlane> &hplanes,
|
863
|
+
size_t curr_tree,
|
864
|
+
const bool as_kernel)
|
865
|
+
{
|
866
|
+
if (interrupt_switch)
|
867
|
+
return;
|
868
|
+
|
869
|
+
if (workspace.st == workspace.end)
|
870
|
+
return;
|
871
|
+
|
872
|
+
if (workspace.tmat_sep.empty())
|
873
|
+
{
|
874
|
+
std::sort(workspace.ix_arr.begin() + workspace.st, workspace.ix_arr.begin() + workspace.end + 1);
|
875
|
+
if (workspace.ix_arr[workspace.st] >= workspace.n_from)
|
876
|
+
return;
|
877
|
+
if (workspace.ix_arr[workspace.end] < workspace.n_from)
|
878
|
+
return;
|
879
|
+
}
|
880
|
+
|
881
|
+
/* Note: the first separation step will not be added here, as it simply consists of adding +1
|
882
|
+
to every combination regardless. It has to be added at the end in 'gather_sim_result' to
|
883
|
+
obtain the average separation depth. */
|
884
|
+
if (hplanes[curr_tree].hplane_left == 0)
|
885
|
+
{
|
886
|
+
if (!as_kernel)
|
887
|
+
{
|
888
|
+
if (!workspace.tmat_sep.empty())
|
889
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
890
|
+
prediction_data.nrows, workspace.tmat_sep.data(),
|
891
|
+
workspace.assume_full_distr? 3. :
|
892
|
+
expected_separation_depth((ldouble_safe) hplanes[curr_tree].remainder
|
893
|
+
+ (ldouble_safe)(workspace.end - workspace.st + 1))
|
894
|
+
);
|
895
|
+
else if (!workspace.rmat.empty())
|
896
|
+
increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.n_from,
|
897
|
+
prediction_data.nrows, workspace.rmat.data(),
|
898
|
+
workspace.assume_full_distr? 3. :
|
899
|
+
expected_separation_depth((ldouble_safe) hplanes[curr_tree].remainder
|
900
|
+
+ (ldouble_safe)(workspace.end - workspace.st + 1))
|
901
|
+
);
|
902
|
+
}
|
903
|
+
|
904
|
+
else
|
905
|
+
{
|
906
|
+
if (!workspace.tmat_sep.empty())
|
907
|
+
{
|
908
|
+
size_t i_, j_;
|
909
|
+
for (size_t i = workspace.st; i < workspace.end; i++)
|
910
|
+
{
|
911
|
+
i_ = workspace.ix_arr[i];
|
912
|
+
for (size_t j = i + 1; j <= workspace.end; j++)
|
913
|
+
{
|
914
|
+
j_ = workspace.ix_arr[j];
|
915
|
+
workspace.tmat_sep[ix_comb(i_, j_, prediction_data.nrows, workspace.tmat_sep.size())]++;
|
916
|
+
}
|
917
|
+
}
|
918
|
+
}
|
919
|
+
|
920
|
+
else if (!workspace.rmat.empty())
|
921
|
+
{
|
922
|
+
size_t n_group = std::distance(workspace.ix_arr.begin() + workspace.st,
|
923
|
+
std::lower_bound(workspace.ix_arr.begin() + workspace.st,
|
924
|
+
workspace.ix_arr.begin() + workspace.end + 1,
|
925
|
+
workspace.n_from));
|
926
|
+
double *restrict rmat_this;
|
927
|
+
for (size_t i = workspace.st; i < workspace.st + n_group; i++)
|
928
|
+
{
|
929
|
+
rmat_this = workspace.rmat.data() + workspace.ix_arr[i]*workspace.n_from;
|
930
|
+
for (size_t j = workspace.st + n_group; j <= workspace.end; j++)
|
931
|
+
{
|
932
|
+
rmat_this[workspace.ix_arr[j] - workspace.n_from]++;
|
933
|
+
}
|
934
|
+
}
|
935
|
+
}
|
936
|
+
}
|
937
|
+
return;
|
938
|
+
}
|
939
|
+
|
940
|
+
else if (curr_tree > 0 && !as_kernel)
|
941
|
+
{
|
942
|
+
if (!workspace.tmat_sep.empty())
|
943
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
944
|
+
prediction_data.nrows, workspace.tmat_sep.data(), -1.);
|
945
|
+
else if (!workspace.rmat.empty())
|
946
|
+
increase_comb_counter_in_groups(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.n_from,
|
947
|
+
prediction_data.nrows, workspace.rmat.data(), -1.);
|
948
|
+
}
|
949
|
+
|
950
|
+
if (prediction_data.Xc_indptr != NULL && workspace.tmat_sep.size())
|
951
|
+
std::sort(workspace.ix_arr.begin() + workspace.st, workspace.ix_arr.begin() + workspace.end + 1);
|
952
|
+
|
953
|
+
/* reconstruct linear combination */
|
954
|
+
size_t ncols_numeric = 0;
|
955
|
+
size_t ncols_categ = 0;
|
956
|
+
std::fill(workspace.comb_val.begin(), workspace.comb_val.begin() + (workspace.end - workspace.st + 1), 0);
|
957
|
+
double unused;
|
958
|
+
if (prediction_data.categ_data != NULL || prediction_data.Xc_indptr != NULL)
|
959
|
+
{
|
960
|
+
for (size_t col = 0; col < hplanes[curr_tree].col_num.size(); col++)
|
961
|
+
{
|
962
|
+
switch(hplanes[curr_tree].col_type[col])
|
963
|
+
{
|
964
|
+
case Numeric:
|
965
|
+
{
|
966
|
+
if (prediction_data.Xc_indptr == NULL)
|
967
|
+
add_linear_comb(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.comb_val.data(),
|
968
|
+
prediction_data.numeric_data + prediction_data.nrows * hplanes[curr_tree].col_num[col],
|
969
|
+
hplanes[curr_tree].coef[ncols_numeric], (double)0, hplanes[curr_tree].mean[ncols_numeric],
|
970
|
+
(model_outputs.missing_action == Fail)? unused : hplanes[curr_tree].fill_val[col],
|
971
|
+
model_outputs.missing_action, NULL, NULL, false);
|
972
|
+
else
|
973
|
+
add_linear_comb(workspace.ix_arr.data(), workspace.st, workspace.end,
|
974
|
+
hplanes[curr_tree].col_num[col], workspace.comb_val.data(),
|
975
|
+
prediction_data.Xc, prediction_data.Xc_ind, prediction_data.Xc_indptr,
|
976
|
+
hplanes[curr_tree].coef[ncols_numeric], (double)0, hplanes[curr_tree].mean[ncols_numeric],
|
977
|
+
(model_outputs.missing_action == Fail)? unused : hplanes[curr_tree].fill_val[col],
|
978
|
+
model_outputs.missing_action, NULL, NULL, false);
|
979
|
+
ncols_numeric++;
|
980
|
+
break;
|
981
|
+
}
|
982
|
+
|
983
|
+
case Categorical:
|
984
|
+
{
|
985
|
+
switch(model_outputs.cat_split_type)
|
986
|
+
{
|
987
|
+
case SingleCateg:
|
988
|
+
{
|
989
|
+
add_linear_comb<ldouble_safe>(
|
990
|
+
workspace.ix_arr.data(), workspace.st, workspace.end, workspace.comb_val.data(),
|
991
|
+
prediction_data.categ_data + prediction_data.nrows * hplanes[curr_tree].col_num[col],
|
992
|
+
(int)0, NULL, hplanes[curr_tree].fill_new[ncols_categ],
|
993
|
+
hplanes[curr_tree].chosen_cat[ncols_categ],
|
994
|
+
(model_outputs.missing_action == Fail)? unused : hplanes[curr_tree].fill_val[col],
|
995
|
+
workspace.comb_val[0], NULL, NULL, model_outputs.new_cat_action,
|
996
|
+
model_outputs.missing_action, SingleCateg, false);
|
997
|
+
break;
|
998
|
+
}
|
999
|
+
|
1000
|
+
case SubSet:
|
1001
|
+
{
|
1002
|
+
add_linear_comb<ldouble_safe>(
|
1003
|
+
workspace.ix_arr.data(), workspace.st, workspace.end, workspace.comb_val.data(),
|
1004
|
+
prediction_data.categ_data + prediction_data.nrows * hplanes[curr_tree].col_num[col],
|
1005
|
+
(int) hplanes[curr_tree].cat_coef[ncols_categ].size(),
|
1006
|
+
hplanes[curr_tree].cat_coef[ncols_categ].data(), (double) 0, (int) 0,
|
1007
|
+
(model_outputs.missing_action == Fail)? unused : hplanes[curr_tree].fill_val[col],
|
1008
|
+
hplanes[curr_tree].fill_new[ncols_categ], NULL, NULL,
|
1009
|
+
model_outputs.new_cat_action, model_outputs.missing_action, SubSet, false);
|
1010
|
+
break;
|
1011
|
+
}
|
1012
|
+
}
|
1013
|
+
ncols_categ++;
|
1014
|
+
break;
|
1015
|
+
}
|
1016
|
+
|
1017
|
+
default:
|
1018
|
+
{
|
1019
|
+
assert(0);
|
1020
|
+
break;
|
1021
|
+
}
|
1022
|
+
}
|
1023
|
+
}
|
1024
|
+
}
|
1025
|
+
|
1026
|
+
|
1027
|
+
else /* faster version for numerical-only */
|
1028
|
+
{
|
1029
|
+
for (size_t col = 0; col < hplanes[curr_tree].col_num.size(); col++)
|
1030
|
+
add_linear_comb(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.comb_val.data(),
|
1031
|
+
prediction_data.numeric_data + prediction_data.nrows * hplanes[curr_tree].col_num[col],
|
1032
|
+
hplanes[curr_tree].coef[col], (double)0, hplanes[curr_tree].mean[col],
|
1033
|
+
(model_outputs.missing_action == Fail)? unused : hplanes[curr_tree].fill_val[col],
|
1034
|
+
model_outputs.missing_action, NULL, NULL, false);
|
1035
|
+
}
|
1036
|
+
|
1037
|
+
/* divide data */
|
1038
|
+
size_t split_ix = divide_subset_split(workspace.ix_arr.data(), workspace.comb_val.data(),
|
1039
|
+
workspace.st, workspace.end, hplanes[curr_tree].split_point);
|
1040
|
+
|
1041
|
+
/* continue splitting recursively */
|
1042
|
+
size_t orig_end = workspace.end;
|
1043
|
+
if (split_ix > workspace.st)
|
1044
|
+
{
|
1045
|
+
workspace.end = split_ix - 1;
|
1046
|
+
traverse_hplane_sim<PredictionData, ldouble_safe>(
|
1047
|
+
workspace,
|
1048
|
+
prediction_data,
|
1049
|
+
model_outputs,
|
1050
|
+
hplanes,
|
1051
|
+
hplanes[curr_tree].hplane_left,
|
1052
|
+
as_kernel);
|
1053
|
+
}
|
1054
|
+
|
1055
|
+
if (split_ix <= orig_end)
|
1056
|
+
{
|
1057
|
+
workspace.st = split_ix;
|
1058
|
+
workspace.end = orig_end;
|
1059
|
+
traverse_hplane_sim<PredictionData, ldouble_safe>(
|
1060
|
+
workspace,
|
1061
|
+
prediction_data,
|
1062
|
+
model_outputs,
|
1063
|
+
hplanes,
|
1064
|
+
hplanes[curr_tree].hplane_right,
|
1065
|
+
as_kernel);
|
1066
|
+
}
|
1067
|
+
|
1068
|
+
}
|
1069
|
+
|
1070
|
+
template <class PredictionData, class InputData, class WorkerMemory>
|
1071
|
+
void gather_sim_result(std::vector<WorkerForSimilarity> *worker_memory,
|
1072
|
+
std::vector<WorkerMemory> *worker_memory_m,
|
1073
|
+
PredictionData *prediction_data, InputData *input_data,
|
1074
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
1075
|
+
double *restrict tmat, double *restrict rmat, size_t n_from,
|
1076
|
+
size_t ntrees, bool assume_full_distr,
|
1077
|
+
bool standardize_dist, bool as_kernel, int nthreads)
|
1078
|
+
{
|
1079
|
+
if (interrupt_switch)
|
1080
|
+
return;
|
1081
|
+
|
1082
|
+
size_t nrows = (prediction_data != NULL)? prediction_data->nrows : input_data->nrows;
|
1083
|
+
size_t ncomb = calc_ncomb(nrows);
|
1084
|
+
size_t n_to = (prediction_data != NULL)? (prediction_data->nrows - n_from) : 0;
|
1085
|
+
|
1086
|
+
#ifdef _OPENMP
|
1087
|
+
if (nthreads > 1)
|
1088
|
+
{
|
1089
|
+
if (worker_memory != NULL)
|
1090
|
+
{
|
1091
|
+
for (WorkerForSimilarity &w : *worker_memory)
|
1092
|
+
{
|
1093
|
+
if (!w.tmat_sep.empty())
|
1094
|
+
{
|
1095
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(ncomb, tmat, w, worker_memory)
|
1096
|
+
for (size_t_for ix = 0; ix < (decltype(ix))ncomb; ix++)
|
1097
|
+
tmat[ix] += w.tmat_sep[ix];
|
1098
|
+
}
|
1099
|
+
else if (!w.rmat.empty())
|
1100
|
+
{
|
1101
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(rmat, w, worker_memory)
|
1102
|
+
for (size_t_for ix = 0; ix < (decltype(ix))w.rmat.size(); ix++)
|
1103
|
+
rmat[ix] += w.rmat[ix];
|
1104
|
+
}
|
1105
|
+
}
|
1106
|
+
}
|
1107
|
+
|
1108
|
+
else
|
1109
|
+
{
|
1110
|
+
for (WorkerMemory &w : *worker_memory_m)
|
1111
|
+
{
|
1112
|
+
if (!w.tmat_sep.empty())
|
1113
|
+
{
|
1114
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(ncomb, tmat, w, worker_memory_m)
|
1115
|
+
for (size_t_for ix = 0; ix < (decltype(ix))ncomb; ix++)
|
1116
|
+
tmat[ix] += w.tmat_sep[ix];
|
1117
|
+
}
|
1118
|
+
}
|
1119
|
+
}
|
1120
|
+
}
|
1121
|
+
|
1122
|
+
else
|
1123
|
+
#endif
|
1124
|
+
{
|
1125
|
+
if (worker_memory != NULL)
|
1126
|
+
{
|
1127
|
+
if (!(*worker_memory)[0].tmat_sep.empty())
|
1128
|
+
std::copy((*worker_memory)[0].tmat_sep.begin(), (*worker_memory)[0].tmat_sep.end(), tmat);
|
1129
|
+
else
|
1130
|
+
std::copy((*worker_memory)[0].rmat.begin(), (*worker_memory)[0].rmat.end(), rmat);
|
1131
|
+
}
|
1132
|
+
|
1133
|
+
else
|
1134
|
+
{
|
1135
|
+
std::copy((*worker_memory_m)[0].tmat_sep.begin(), (*worker_memory_m)[0].tmat_sep.end(), tmat);
|
1136
|
+
}
|
1137
|
+
}
|
1138
|
+
|
1139
|
+
double ntrees_dbl = (double) ntrees;
|
1140
|
+
if (standardize_dist)
|
1141
|
+
{
|
1142
|
+
if (as_kernel)
|
1143
|
+
{
|
1144
|
+
if (tmat != NULL)
|
1145
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1146
|
+
tmat[ix] /= ntrees_dbl;
|
1147
|
+
else
|
1148
|
+
for (size_t ix = 0; ix < (n_from * n_to); ix++)
|
1149
|
+
rmat[ix] /= ntrees_dbl;
|
1150
|
+
return;
|
1151
|
+
}
|
1152
|
+
|
1153
|
+
|
1154
|
+
/* Note: the separation distances up this point are missing the first hop, which is always
|
1155
|
+
a +1 to every combination. Thus, it needs to be added back for the average separation depth.
|
1156
|
+
For the standardized metric, it takes the expected divisor as 2(=3-1) instead of 3, given
|
1157
|
+
that every combination will always get a +1 at the beginning. Since what's obtained here
|
1158
|
+
is a sum across all trees, adding this +1 means adding the number of trees. */
|
1159
|
+
double div_trees = ntrees_dbl;
|
1160
|
+
if (assume_full_distr)
|
1161
|
+
{
|
1162
|
+
div_trees *= 2;
|
1163
|
+
}
|
1164
|
+
|
1165
|
+
else if (input_data != NULL)
|
1166
|
+
{
|
1167
|
+
div_trees *= (expected_separation_depth(input_data->nrows) - 1);
|
1168
|
+
}
|
1169
|
+
|
1170
|
+
else
|
1171
|
+
{
|
1172
|
+
div_trees *= ((
|
1173
|
+
(model_outputs != NULL)?
|
1174
|
+
expected_separation_depth_hotstart(model_outputs->exp_avg_sep,
|
1175
|
+
model_outputs->orig_sample_size,
|
1176
|
+
model_outputs->orig_sample_size + prediction_data->nrows)
|
1177
|
+
:
|
1178
|
+
expected_separation_depth_hotstart(model_outputs_ext->exp_avg_sep,
|
1179
|
+
model_outputs_ext->orig_sample_size,
|
1180
|
+
model_outputs_ext->orig_sample_size + prediction_data->nrows)
|
1181
|
+
) - 1);
|
1182
|
+
}
|
1183
|
+
|
1184
|
+
|
1185
|
+
if (tmat != NULL)
|
1186
|
+
#ifndef _WIN32
|
1187
|
+
#pragma omp simd
|
1188
|
+
#endif
|
1189
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1190
|
+
tmat[ix] = std::exp2( - tmat[ix] / div_trees);
|
1191
|
+
else
|
1192
|
+
#ifndef _WIN32
|
1193
|
+
#pragma omp simd
|
1194
|
+
#endif
|
1195
|
+
for (size_t ix = 0; ix < (n_from * n_to); ix++)
|
1196
|
+
rmat[ix] = std::exp2( - rmat[ix] / div_trees);
|
1197
|
+
}
|
1198
|
+
|
1199
|
+
else
|
1200
|
+
{
|
1201
|
+
if (as_kernel) return;
|
1202
|
+
|
1203
|
+
if (tmat != NULL)
|
1204
|
+
#ifndef _WIN32
|
1205
|
+
#pragma omp simd
|
1206
|
+
#endif
|
1207
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1208
|
+
tmat[ix] = (tmat[ix] + ntrees) / ntrees_dbl;
|
1209
|
+
else
|
1210
|
+
#ifndef _WIN32
|
1211
|
+
#pragma omp simd
|
1212
|
+
#endif
|
1213
|
+
for (size_t ix = 0; ix < (n_from * n_to); ix++)
|
1214
|
+
rmat[ix] = (rmat[ix] + ntrees) / ntrees_dbl;
|
1215
|
+
}
|
1216
|
+
}
|
1217
|
+
|
1218
|
+
template <class PredictionData>
|
1219
|
+
void initialize_worker_for_sim(WorkerForSimilarity &workspace,
|
1220
|
+
PredictionData &prediction_data,
|
1221
|
+
IsoForest *model_outputs,
|
1222
|
+
ExtIsoForest *model_outputs_ext,
|
1223
|
+
size_t n_from,
|
1224
|
+
bool assume_full_distr)
|
1225
|
+
{
|
1226
|
+
workspace.st = 0;
|
1227
|
+
workspace.end = prediction_data.nrows - 1;
|
1228
|
+
workspace.n_from = n_from;
|
1229
|
+
workspace.assume_full_distr = assume_full_distr; /* doesn't need to have one copy per worker */
|
1230
|
+
|
1231
|
+
if (workspace.ix_arr.empty())
|
1232
|
+
{
|
1233
|
+
workspace.ix_arr.resize(prediction_data.nrows);
|
1234
|
+
std::iota(workspace.ix_arr.begin(), workspace.ix_arr.end(), (size_t)0);
|
1235
|
+
if (!n_from)
|
1236
|
+
workspace.tmat_sep.resize(calc_ncomb(prediction_data.nrows), 0);
|
1237
|
+
else
|
1238
|
+
workspace.rmat.resize((prediction_data.nrows - n_from) * n_from, 0);
|
1239
|
+
}
|
1240
|
+
|
1241
|
+
if (model_outputs != NULL &&
|
1242
|
+
(model_outputs->missing_action == Divide ||
|
1243
|
+
(model_outputs->new_cat_action == Weighted && model_outputs->cat_split_type == SubSet && prediction_data.categ_data != NULL)))
|
1244
|
+
{
|
1245
|
+
if (workspace.weights_arr.empty())
|
1246
|
+
workspace.weights_arr.resize(prediction_data.nrows, 1.);
|
1247
|
+
else
|
1248
|
+
std::fill(workspace.weights_arr.begin(), workspace.weights_arr.end(), 1.);
|
1249
|
+
}
|
1250
|
+
|
1251
|
+
if (model_outputs_ext != NULL)
|
1252
|
+
{
|
1253
|
+
if (workspace.comb_val.empty())
|
1254
|
+
workspace.comb_val.resize(prediction_data.nrows, 0);
|
1255
|
+
else
|
1256
|
+
std::fill(workspace.comb_val.begin(), workspace.comb_val.end(), 0);
|
1257
|
+
}
|
1258
|
+
}
|
1259
|
+
|
1260
|
+
template <class real_t, class sparse_ix>
|
1261
|
+
void calc_similarity_from_indexer
|
1262
|
+
(
|
1263
|
+
real_t *restrict numeric_data, int *restrict categ_data,
|
1264
|
+
real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
1265
|
+
size_t nrows, int nthreads, bool assume_full_distr, bool standardize_dist,
|
1266
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
1267
|
+
double *restrict tmat, double *restrict rmat, size_t n_from,
|
1268
|
+
TreesIndexer *indexer, bool is_col_major, size_t ld_numeric, size_t ld_categ
|
1269
|
+
)
|
1270
|
+
{
|
1271
|
+
SignalSwitcher ss;
|
1272
|
+
size_t ntrees = (model_outputs != NULL)? model_outputs->trees.size() : model_outputs_ext->hplanes.size();
|
1273
|
+
std::vector<sparse_ix> terminal_indices(nrows * ntrees);
|
1274
|
+
std::unique_ptr<double[]> ignored(new double[nrows]);
|
1275
|
+
predict_iforest(numeric_data, categ_data,
|
1276
|
+
is_col_major, ld_numeric, ld_categ,
|
1277
|
+
is_col_major? Xc : nullptr, is_col_major? Xc_ind : nullptr, is_col_major? Xc_indptr : nullptr,
|
1278
|
+
is_col_major? (real_t*)nullptr : Xc, is_col_major? (sparse_ix*)nullptr : Xc_ind, is_col_major? (sparse_ix*)nullptr : Xc_indptr,
|
1279
|
+
nrows, nthreads, false,
|
1280
|
+
model_outputs, model_outputs_ext,
|
1281
|
+
ignored.get(), terminal_indices.data(),
|
1282
|
+
(double*)NULL,
|
1283
|
+
indexer);
|
1284
|
+
ignored.reset();
|
1285
|
+
|
1286
|
+
#ifndef _OPENMP
|
1287
|
+
nthreads = 1;
|
1288
|
+
#endif
|
1289
|
+
|
1290
|
+
check_interrupt_switch(ss);
|
1291
|
+
|
1292
|
+
if (n_from == 0)
|
1293
|
+
{
|
1294
|
+
size_t ncomb = calc_ncomb(nrows);
|
1295
|
+
std::fill_n(tmat, ncomb, 0.);
|
1296
|
+
|
1297
|
+
std::vector<std::vector<double>> sum_separations(nthreads);
|
1298
|
+
if (nthreads != 1) {
|
1299
|
+
for (auto &v : sum_separations) v.resize(ncomb);
|
1300
|
+
}
|
1301
|
+
|
1302
|
+
std::vector<std::vector<size_t>> thread_argsorted_nodes(nthreads);
|
1303
|
+
for (auto &v : thread_argsorted_nodes) v.resize(nrows);
|
1304
|
+
|
1305
|
+
std::vector<std::vector<size_t>> thread_sorted_nodes(nthreads);
|
1306
|
+
for (auto &v : thread_sorted_nodes) v.reserve(nrows); /* <- could shrink to max number of terminal nodes */
|
1307
|
+
|
1308
|
+
|
1309
|
+
bool threw_exception = false;
|
1310
|
+
std::exception_ptr ex = NULL;
|
1311
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) \
|
1312
|
+
shared(model_outputs, model_outputs_ext, nthreads, indexer, nrows, ncomb, terminal_indices, \
|
1313
|
+
sum_separations, thread_argsorted_nodes, thread_sorted_nodes, tmat, \
|
1314
|
+
threw_exception, ex)
|
1315
|
+
for (size_t_for tree = 0; tree < (decltype(tree))ntrees; tree++)
|
1316
|
+
{
|
1317
|
+
if (interrupt_switch || threw_exception) continue;
|
1318
|
+
|
1319
|
+
if (unlikely(indexer->indices[tree].n_terminal <= 1))
|
1320
|
+
{
|
1321
|
+
for (auto &el : sum_separations[omp_get_thread_num()]) el += 1.;
|
1322
|
+
continue;
|
1323
|
+
}
|
1324
|
+
|
1325
|
+
double *restrict ptr_this_sep = sum_separations[omp_get_thread_num()].data();
|
1326
|
+
if (nthreads == 1) ptr_this_sep = tmat;
|
1327
|
+
double *restrict node_dist_this = indexer->indices[tree].node_distances.data();
|
1328
|
+
double *restrict node_depths_this = indexer->indices[tree].node_depths.data();
|
1329
|
+
size_t n_terminal_this = indexer->indices[tree].n_terminal;
|
1330
|
+
size_t ncomb_this = calc_ncomb(n_terminal_this);
|
1331
|
+
std::vector<IsoTree> *tree_this = (model_outputs != NULL)? &model_outputs->trees[tree] : nullptr;
|
1332
|
+
std::vector<IsoHPlane> *hplane_this = (model_outputs_ext != NULL)? &model_outputs_ext->hplanes[tree] : nullptr;
|
1333
|
+
sparse_ix *restrict terminal_indices_this = terminal_indices.data() + nrows * tree;
|
1334
|
+
size_t i, j;
|
1335
|
+
double add_round;
|
1336
|
+
|
1337
|
+
if (assume_full_distr)
|
1338
|
+
{
|
1339
|
+
for (size_t el1 = 0; el1 < nrows-1; el1++)
|
1340
|
+
{
|
1341
|
+
i = terminal_indices_this[el1];
|
1342
|
+
for (size_t el2 = el1+1; el2 < nrows; el2++)
|
1343
|
+
{
|
1344
|
+
j = terminal_indices_this[el2];
|
1345
|
+
if (unlikely(i == j))
|
1346
|
+
add_round = node_depths_this[i] + 3.;
|
1347
|
+
else
|
1348
|
+
add_round = node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
|
1349
|
+
ptr_this_sep[ix_comb(el1, el2, nrows, ncomb)] += add_round;
|
1350
|
+
}
|
1351
|
+
}
|
1352
|
+
}
|
1353
|
+
|
1354
|
+
else
|
1355
|
+
{
|
1356
|
+
hashed_set<size_t> nodes_w_repeated;
|
1357
|
+
try
|
1358
|
+
{
|
1359
|
+
nodes_w_repeated.reserve(n_terminal_this);
|
1360
|
+
for (size_t el1 = 0; el1 < nrows-1; el1++)
|
1361
|
+
{
|
1362
|
+
i = terminal_indices_this[el1];
|
1363
|
+
for (size_t el2 = el1+1; el2 < nrows; el2++)
|
1364
|
+
{
|
1365
|
+
j = terminal_indices_this[el2];
|
1366
|
+
if (unlikely(i == j))
|
1367
|
+
nodes_w_repeated.insert(i);
|
1368
|
+
else
|
1369
|
+
ptr_this_sep[ix_comb(el1, el2, nrows, ncomb)]
|
1370
|
+
+=
|
1371
|
+
node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
|
1372
|
+
}
|
1373
|
+
}
|
1374
|
+
}
|
1375
|
+
|
1376
|
+
catch (...)
|
1377
|
+
{
|
1378
|
+
#pragma omp critical
|
1379
|
+
{
|
1380
|
+
if (!threw_exception)
|
1381
|
+
{
|
1382
|
+
threw_exception = true;
|
1383
|
+
ex = std::current_exception();
|
1384
|
+
}
|
1385
|
+
}
|
1386
|
+
}
|
1387
|
+
|
1388
|
+
if (likely(!nodes_w_repeated.empty()))
|
1389
|
+
{
|
1390
|
+
std::vector<size_t> *restrict argsorted_nodes = &thread_argsorted_nodes[omp_get_thread_num()];
|
1391
|
+
std::iota(argsorted_nodes->begin(), argsorted_nodes->end(), (size_t)0);
|
1392
|
+
std::sort(argsorted_nodes->begin(), argsorted_nodes->end(),
|
1393
|
+
[&terminal_indices_this](const size_t a, const size_t b)
|
1394
|
+
{return terminal_indices_this[a] < terminal_indices_this[b];});
|
1395
|
+
std::vector<size_t>::iterator curr_begin = argsorted_nodes->begin();
|
1396
|
+
std::vector<size_t>::iterator new_begin;
|
1397
|
+
|
1398
|
+
std::vector<size_t> *restrict sorted_nodes = &thread_sorted_nodes[omp_get_thread_num()];
|
1399
|
+
sorted_nodes->assign(nodes_w_repeated.begin(), nodes_w_repeated.end());
|
1400
|
+
std::sort(sorted_nodes->begin(), sorted_nodes->end());
|
1401
|
+
for (size_t node_ix : *sorted_nodes)
|
1402
|
+
{
|
1403
|
+
curr_begin = std::lower_bound(curr_begin, argsorted_nodes->end(),
|
1404
|
+
node_ix,
|
1405
|
+
[&terminal_indices_this](const size_t &a, const size_t &b)
|
1406
|
+
{return (size_t)terminal_indices_this[a] < b;});
|
1407
|
+
new_begin = std::upper_bound(curr_begin, argsorted_nodes->end(),
|
1408
|
+
node_ix,
|
1409
|
+
[&terminal_indices_this](const size_t &a, const size_t &b)
|
1410
|
+
{return a < (size_t)terminal_indices_this[b];});
|
1411
|
+
size_t n_this = std::distance(curr_begin, new_begin);
|
1412
|
+
double sep_this
|
1413
|
+
=
|
1414
|
+
n_this
|
1415
|
+
+
|
1416
|
+
((tree_this != NULL)?
|
1417
|
+
(*tree_this)[node_ix].remainder
|
1418
|
+
:
|
1419
|
+
(*hplane_this)[node_ix].remainder);
|
1420
|
+
double sep_this_ = expected_separation_depth(sep_this) + node_depths_this[node_ix];
|
1421
|
+
|
1422
|
+
size_t i, j;
|
1423
|
+
for (size_t el1 = 0; el1 < n_this-1; el1++)
|
1424
|
+
{
|
1425
|
+
i = *(curr_begin + el1);
|
1426
|
+
for (size_t el2 = el1+1; el2 < n_this; el2++)
|
1427
|
+
{
|
1428
|
+
j = *(curr_begin + el2);
|
1429
|
+
ptr_this_sep[ix_comb(i, j, nrows, ncomb)] += sep_this_;
|
1430
|
+
}
|
1431
|
+
}
|
1432
|
+
|
1433
|
+
curr_begin = new_begin;
|
1434
|
+
}
|
1435
|
+
}
|
1436
|
+
|
1437
|
+
}
|
1438
|
+
}
|
1439
|
+
|
1440
|
+
check_interrupt_switch(ss);
|
1441
|
+
|
1442
|
+
if (threw_exception)
|
1443
|
+
std::rethrow_exception(ex);
|
1444
|
+
|
1445
|
+
if (nthreads == 1)
|
1446
|
+
{
|
1447
|
+
/* Here 'tmat' already contains the sum of separations */
|
1448
|
+
}
|
1449
|
+
|
1450
|
+
else
|
1451
|
+
{
|
1452
|
+
for (int tid = 0; tid < nthreads; tid++)
|
1453
|
+
{
|
1454
|
+
double *restrict seps_thread = sum_separations[tid].data();
|
1455
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1456
|
+
tmat[ix] += seps_thread[ix];
|
1457
|
+
}
|
1458
|
+
}
|
1459
|
+
|
1460
|
+
check_interrupt_switch(ss);
|
1461
|
+
|
1462
|
+
if (standardize_dist)
|
1463
|
+
{
|
1464
|
+
double divisor;
|
1465
|
+
if (assume_full_distr)
|
1466
|
+
divisor = (double)(ntrees * 2);
|
1467
|
+
else
|
1468
|
+
divisor = (double)ntrees * ((model_outputs != NULL)? model_outputs->exp_avg_sep : model_outputs_ext->exp_avg_sep);
|
1469
|
+
|
1470
|
+
if (assume_full_distr)
|
1471
|
+
{
|
1472
|
+
double ntrees_dbl = (double)ntrees;
|
1473
|
+
#ifndef _WIN32
|
1474
|
+
#pragma omp simd
|
1475
|
+
#endif
|
1476
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1477
|
+
tmat[ix] = std::exp2( - (tmat[ix] - ntrees_dbl) / divisor);
|
1478
|
+
}
|
1479
|
+
|
1480
|
+
else
|
1481
|
+
{
|
1482
|
+
#ifndef _WIN32
|
1483
|
+
#pragma omp simd
|
1484
|
+
#endif
|
1485
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1486
|
+
tmat[ix] = std::exp2( - tmat[ix] / divisor);
|
1487
|
+
}
|
1488
|
+
}
|
1489
|
+
|
1490
|
+
else
|
1491
|
+
{
|
1492
|
+
double divisor = (double)ntrees;
|
1493
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1494
|
+
tmat[ix] /= divisor;
|
1495
|
+
}
|
1496
|
+
|
1497
|
+
check_interrupt_switch(ss);
|
1498
|
+
}
|
1499
|
+
|
1500
|
+
/* TODO: merge this with the block above, can simplify lots of things by a couple if-elses */
|
1501
|
+
else /* has 'rmat' / 'nfrom>0' */
|
1502
|
+
{
|
1503
|
+
size_t n_to = nrows - n_from;
|
1504
|
+
size_t ncomb = n_from * n_to;
|
1505
|
+
std::fill_n(rmat, ncomb, 0.);
|
1506
|
+
|
1507
|
+
std::vector<std::vector<double>> sum_separations(nthreads);
|
1508
|
+
if (nthreads != 1) {
|
1509
|
+
for (auto &v : sum_separations) v.resize(ncomb);
|
1510
|
+
}
|
1511
|
+
|
1512
|
+
std::vector<std::vector<size_t>> thread_argsorted_nodes(nthreads);
|
1513
|
+
for (auto &v : thread_argsorted_nodes) v.resize(nrows);
|
1514
|
+
|
1515
|
+
std::vector<std::vector<size_t>> thread_doubly_argsorted(nthreads);
|
1516
|
+
for (auto &v : thread_doubly_argsorted) v.reserve(nrows);
|
1517
|
+
|
1518
|
+
std::vector<std::vector<size_t>> thread_sorted_nodes(nthreads);
|
1519
|
+
for (auto &v : thread_sorted_nodes) v.reserve(nrows); /* <- could shrink to max number of terminal nodes */
|
1520
|
+
|
1521
|
+
bool threw_exception = false;
|
1522
|
+
std::exception_ptr ex = NULL;
|
1523
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) \
|
1524
|
+
shared(model_outputs, model_outputs_ext, nthreads, indexer, nrows, ncomb, terminal_indices, \
|
1525
|
+
sum_separations, thread_argsorted_nodes, thread_sorted_nodes, thread_doubly_argsorted, rmat, n_to, n_from, \
|
1526
|
+
threw_exception, ex)
|
1527
|
+
for (size_t_for tree = 0; tree < (decltype(tree))ntrees; tree++)
|
1528
|
+
{
|
1529
|
+
if (interrupt_switch || threw_exception) continue;
|
1530
|
+
|
1531
|
+
if (unlikely(indexer->indices[tree].n_terminal <= 1))
|
1532
|
+
{
|
1533
|
+
for (auto &el : sum_separations[omp_get_thread_num()]) el += 1.;
|
1534
|
+
continue;
|
1535
|
+
}
|
1536
|
+
|
1537
|
+
double *restrict ptr_this_sep = sum_separations[omp_get_thread_num()].data();
|
1538
|
+
if (nthreads == 1) ptr_this_sep = rmat;
|
1539
|
+
double *restrict node_dist_this = indexer->indices[tree].node_distances.data();
|
1540
|
+
double *restrict node_depths_this = indexer->indices[tree].node_depths.data();
|
1541
|
+
size_t n_terminal_this = indexer->indices[tree].n_terminal;
|
1542
|
+
size_t ncomb_this = calc_ncomb(n_terminal_this);
|
1543
|
+
std::vector<IsoTree> *tree_this = (model_outputs != NULL)? &model_outputs->trees[tree] : nullptr;
|
1544
|
+
std::vector<IsoHPlane> *hplane_this = (model_outputs_ext != NULL)? &model_outputs_ext->hplanes[tree] : nullptr;
|
1545
|
+
sparse_ix *restrict terminal_indices_this = terminal_indices.data() + nrows * tree;
|
1546
|
+
size_t i, j;
|
1547
|
+
double add_round;
|
1548
|
+
|
1549
|
+
if (assume_full_distr)
|
1550
|
+
{
|
1551
|
+
for (size_t el1 = 0; el1 < n_from; el1++)
|
1552
|
+
{
|
1553
|
+
i = terminal_indices_this[el1];
|
1554
|
+
double *ptr_this_sep_ = ptr_this_sep + el1*n_to;
|
1555
|
+
for (size_t el2 = n_from; el2 < nrows; el2++)
|
1556
|
+
{
|
1557
|
+
j = terminal_indices_this[el2];
|
1558
|
+
if (unlikely(i == j))
|
1559
|
+
add_round = node_depths_this[i] + 3.;
|
1560
|
+
else
|
1561
|
+
add_round = node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
|
1562
|
+
ptr_this_sep_[el2-n_from] += add_round;
|
1563
|
+
}
|
1564
|
+
}
|
1565
|
+
}
|
1566
|
+
|
1567
|
+
else
|
1568
|
+
{
|
1569
|
+
hashed_set<size_t> nodes_w_repeated;
|
1570
|
+
try
|
1571
|
+
{
|
1572
|
+
nodes_w_repeated.reserve(n_terminal_this);
|
1573
|
+
for (size_t el1 = 0; el1 < n_from; el1++)
|
1574
|
+
{
|
1575
|
+
i = terminal_indices_this[el1];
|
1576
|
+
double *ptr_this_sep_ = ptr_this_sep + el1*n_to;
|
1577
|
+
for (size_t el2 = n_from; el2 < nrows; el2++)
|
1578
|
+
{
|
1579
|
+
j = terminal_indices_this[el2];
|
1580
|
+
if (unlikely(i == j))
|
1581
|
+
nodes_w_repeated.insert(i);
|
1582
|
+
else
|
1583
|
+
ptr_this_sep_[el2-n_from]
|
1584
|
+
+=
|
1585
|
+
node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
|
1586
|
+
}
|
1587
|
+
}
|
1588
|
+
|
1589
|
+
if (likely(!nodes_w_repeated.empty()))
|
1590
|
+
{
|
1591
|
+
std::vector<size_t> *restrict argsorted_nodes = &thread_argsorted_nodes[omp_get_thread_num()];
|
1592
|
+
std::iota(argsorted_nodes->begin(), argsorted_nodes->end(), (size_t)0);
|
1593
|
+
std::sort(argsorted_nodes->begin(), argsorted_nodes->end(),
|
1594
|
+
[&terminal_indices_this](const size_t a, const size_t b)
|
1595
|
+
{return terminal_indices_this[a] < terminal_indices_this[b];});
|
1596
|
+
std::vector<size_t>::iterator curr_begin = argsorted_nodes->begin();
|
1597
|
+
std::vector<size_t>::iterator new_begin;
|
1598
|
+
|
1599
|
+
std::vector<size_t> *restrict sorted_nodes = &thread_sorted_nodes[omp_get_thread_num()];
|
1600
|
+
sorted_nodes->assign(nodes_w_repeated.begin(), nodes_w_repeated.end());
|
1601
|
+
std::sort(sorted_nodes->begin(), sorted_nodes->end());
|
1602
|
+
for (size_t node_ix : *sorted_nodes)
|
1603
|
+
{
|
1604
|
+
curr_begin = std::lower_bound(curr_begin, argsorted_nodes->end(),
|
1605
|
+
node_ix,
|
1606
|
+
[&terminal_indices_this](const size_t &a, const size_t &b)
|
1607
|
+
{return (size_t)terminal_indices_this[a] < b;});
|
1608
|
+
new_begin = std::upper_bound(curr_begin, argsorted_nodes->end(),
|
1609
|
+
node_ix,
|
1610
|
+
[&terminal_indices_this](const size_t &a, const size_t &b)
|
1611
|
+
{return a < (size_t)terminal_indices_this[b];});
|
1612
|
+
size_t n_this = std::distance(curr_begin, new_begin);
|
1613
|
+
if (unlikely(!n_this)) unexpected_error();
|
1614
|
+
double sep_this
|
1615
|
+
=
|
1616
|
+
n_this
|
1617
|
+
+
|
1618
|
+
((tree_this != NULL)?
|
1619
|
+
(*tree_this)[node_ix].remainder
|
1620
|
+
:
|
1621
|
+
(*hplane_this)[node_ix].remainder);
|
1622
|
+
double sep_this_ = expected_separation_depth(sep_this) + node_depths_this[node_ix];
|
1623
|
+
|
1624
|
+
std::vector<size_t> *restrict doubly_argsorted = &thread_doubly_argsorted[omp_get_thread_num()];
|
1625
|
+
doubly_argsorted->assign(curr_begin, curr_begin + n_this);
|
1626
|
+
std::sort(doubly_argsorted->begin(), doubly_argsorted->end());
|
1627
|
+
std::vector<size_t>::iterator pos_n_from = std::lower_bound(doubly_argsorted->begin(),
|
1628
|
+
doubly_argsorted->end(),
|
1629
|
+
n_from);
|
1630
|
+
if (pos_n_from == doubly_argsorted->end()) unexpected_error();
|
1631
|
+
size_t n1 = std::distance(doubly_argsorted->begin(), pos_n_from);
|
1632
|
+
size_t i, j;
|
1633
|
+
double *ptr_this_sep__;
|
1634
|
+
for (size_t el1 = 0; el1 < n1; el1++)
|
1635
|
+
{
|
1636
|
+
i = (*doubly_argsorted)[el1];
|
1637
|
+
ptr_this_sep__ = ptr_this_sep + i*n_to;
|
1638
|
+
for (size_t el2 = n1; el2 < n_this; el2++)
|
1639
|
+
{
|
1640
|
+
j = (*doubly_argsorted)[el2];
|
1641
|
+
ptr_this_sep__[j-n_from] += sep_this_;
|
1642
|
+
}
|
1643
|
+
}
|
1644
|
+
|
1645
|
+
curr_begin = new_begin;
|
1646
|
+
}
|
1647
|
+
}
|
1648
|
+
}
|
1649
|
+
|
1650
|
+
catch (...)
|
1651
|
+
{
|
1652
|
+
#pragma omp critical
|
1653
|
+
{
|
1654
|
+
if (!threw_exception)
|
1655
|
+
{
|
1656
|
+
threw_exception = true;
|
1657
|
+
ex = std::current_exception();
|
1658
|
+
}
|
1659
|
+
}
|
1660
|
+
}
|
1661
|
+
}
|
1662
|
+
}
|
1663
|
+
|
1664
|
+
check_interrupt_switch(ss);
|
1665
|
+
|
1666
|
+
if (threw_exception)
|
1667
|
+
std::rethrow_exception(ex);
|
1668
|
+
|
1669
|
+
if (nthreads == 1)
|
1670
|
+
{
|
1671
|
+
/* Here 'rmat' already contains the sum of separations */
|
1672
|
+
}
|
1673
|
+
|
1674
|
+
else
|
1675
|
+
{
|
1676
|
+
for (int tid = 0; tid < nthreads; tid++)
|
1677
|
+
{
|
1678
|
+
double *restrict seps_thread = sum_separations[tid].data();
|
1679
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1680
|
+
rmat[ix] += seps_thread[ix];
|
1681
|
+
}
|
1682
|
+
}
|
1683
|
+
|
1684
|
+
check_interrupt_switch(ss);
|
1685
|
+
|
1686
|
+
if (standardize_dist)
|
1687
|
+
{
|
1688
|
+
double divisor;
|
1689
|
+
if (assume_full_distr)
|
1690
|
+
divisor = (double)(ntrees * 2);
|
1691
|
+
else
|
1692
|
+
divisor = (double)ntrees * ((model_outputs != NULL)? model_outputs->exp_avg_sep : model_outputs_ext->exp_avg_sep);
|
1693
|
+
|
1694
|
+
if (assume_full_distr)
|
1695
|
+
{
|
1696
|
+
double ntrees_dbl = (double)ntrees;
|
1697
|
+
#ifndef _WIN32
|
1698
|
+
#pragma omp simd
|
1699
|
+
#endif
|
1700
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1701
|
+
rmat[ix] = std::exp2( - (rmat[ix] - ntrees_dbl) / divisor);
|
1702
|
+
}
|
1703
|
+
|
1704
|
+
else
|
1705
|
+
{
|
1706
|
+
#ifndef _WIN32
|
1707
|
+
#pragma omp simd
|
1708
|
+
#endif
|
1709
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1710
|
+
rmat[ix] = std::exp2( - rmat[ix] / divisor);
|
1711
|
+
}
|
1712
|
+
}
|
1713
|
+
|
1714
|
+
else
|
1715
|
+
{
|
1716
|
+
double divisor = (double)ntrees;
|
1717
|
+
for (size_t ix = 0; ix < ncomb; ix++)
|
1718
|
+
rmat[ix] /= divisor;
|
1719
|
+
}
|
1720
|
+
|
1721
|
+
check_interrupt_switch(ss);
|
1722
|
+
}
|
1723
|
+
}
|
1724
|
+
|
1725
|
+
template <class real_t, class sparse_ix>
|
1726
|
+
void calc_similarity_from_indexer_with_references
|
1727
|
+
(
|
1728
|
+
real_t *restrict numeric_data, int *restrict categ_data,
|
1729
|
+
real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
1730
|
+
size_t nrows, int nthreads, bool standardize_dist,
|
1731
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
1732
|
+
double *restrict rmat,
|
1733
|
+
TreesIndexer *indexer, bool is_col_major, size_t ld_numeric, size_t ld_categ
|
1734
|
+
)
|
1735
|
+
{
|
1736
|
+
size_t n_ref = get_number_of_reference_points(*indexer);
|
1737
|
+
if (unlikely(!n_ref)) unexpected_error();
|
1738
|
+
|
1739
|
+
SignalSwitcher ss;
|
1740
|
+
|
1741
|
+
size_t ntrees = (model_outputs != NULL)? model_outputs->trees.size() : model_outputs_ext->hplanes.size();
|
1742
|
+
std::vector<sparse_ix> terminal_indices(nrows * ntrees);
|
1743
|
+
std::unique_ptr<double[]> ignored(new double[nrows]);
|
1744
|
+
predict_iforest(numeric_data, categ_data,
|
1745
|
+
is_col_major, ld_numeric, ld_categ,
|
1746
|
+
is_col_major? Xc : nullptr, is_col_major? Xc_ind : nullptr, is_col_major? Xc_indptr : nullptr,
|
1747
|
+
is_col_major? (real_t*)nullptr : Xc, is_col_major? (sparse_ix*)nullptr : Xc_ind, is_col_major? (sparse_ix*)nullptr : Xc_indptr,
|
1748
|
+
nrows, nthreads, false,
|
1749
|
+
model_outputs, model_outputs_ext,
|
1750
|
+
ignored.get(), terminal_indices.data(),
|
1751
|
+
(double*)NULL,
|
1752
|
+
indexer);
|
1753
|
+
ignored.reset();
|
1754
|
+
|
1755
|
+
#ifndef _OPENMP
|
1756
|
+
nthreads = 1;
|
1757
|
+
#endif
|
1758
|
+
|
1759
|
+
check_interrupt_switch(ss);
|
1760
|
+
|
1761
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) \
|
1762
|
+
shared(rmat, terminal_indices, nrows, n_ref, indexer, ntrees)
|
1763
|
+
for (size_t_for row = 0; row < (decltype(row))nrows; row++)
|
1764
|
+
{
|
1765
|
+
if (interrupt_switch) continue;
|
1766
|
+
|
1767
|
+
size_t i, j;
|
1768
|
+
size_t n_terminal_this;
|
1769
|
+
size_t ncomb_this;
|
1770
|
+
size_t *restrict ref_this;
|
1771
|
+
sparse_ix *restrict ind_this;
|
1772
|
+
double *restrict node_depths_this;
|
1773
|
+
double *restrict node_dist_this;
|
1774
|
+
double *rmat_this = rmat + row*n_ref;
|
1775
|
+
memset(rmat_this, 0, n_ref*sizeof(double));
|
1776
|
+
for (size_t tree = 0; tree < ntrees; tree++)
|
1777
|
+
{
|
1778
|
+
ref_this = indexer->indices[tree].reference_points.data();
|
1779
|
+
ind_this = terminal_indices.data() + tree*nrows;
|
1780
|
+
node_depths_this = indexer->indices[tree].node_depths.data();
|
1781
|
+
n_terminal_this = indexer->indices[tree].n_terminal;
|
1782
|
+
node_dist_this = indexer->indices[tree].node_distances.data();
|
1783
|
+
ncomb_this = calc_ncomb(n_terminal_this);
|
1784
|
+
for (size_t ref = 0; ref < n_ref; ref++)
|
1785
|
+
{
|
1786
|
+
i = ind_this[row];
|
1787
|
+
j = ref_this[ref];
|
1788
|
+
|
1789
|
+
if (unlikely(i == j))
|
1790
|
+
rmat_this[ref] += node_depths_this[i] + 3.;
|
1791
|
+
else
|
1792
|
+
rmat_this[ref] += node_dist_this[ix_comb(i, j, n_terminal_this, ncomb_this)];
|
1793
|
+
}
|
1794
|
+
}
|
1795
|
+
}
|
1796
|
+
|
1797
|
+
check_interrupt_switch(ss);
|
1798
|
+
|
1799
|
+
size_t size_rmat = nrows * n_ref;
|
1800
|
+
if (standardize_dist)
|
1801
|
+
{
|
1802
|
+
double ntrees_dbl = (double)ntrees;
|
1803
|
+
double div_trees = (double)(mult2(ntrees));
|
1804
|
+
#ifndef _WIN32
|
1805
|
+
#pragma omp simd
|
1806
|
+
#endif
|
1807
|
+
for (size_t ix = 0; ix < size_rmat; ix++)
|
1808
|
+
rmat[ix] = std::exp2( - (rmat[ix] - ntrees_dbl) / div_trees);
|
1809
|
+
}
|
1810
|
+
|
1811
|
+
else
|
1812
|
+
{
|
1813
|
+
double div_trees = (double)ntrees;
|
1814
|
+
for (size_t ix = 0; ix < size_rmat; ix++)
|
1815
|
+
rmat[ix] /= div_trees;
|
1816
|
+
}
|
1817
|
+
|
1818
|
+
check_interrupt_switch(ss);
|
1819
|
+
}
|
1820
|
+
|
1821
|
+
template <class real_t, class sparse_ix>
|
1822
|
+
void kernel_to_references(TreesIndexer &indexer,
|
1823
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
1824
|
+
real_t *restrict numeric_data, int *restrict categ_data,
|
1825
|
+
real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
1826
|
+
bool is_col_major, size_t ld_numeric, size_t ld_categ,
|
1827
|
+
size_t nrows, int nthreads,
|
1828
|
+
double *restrict rmat,
|
1829
|
+
bool standardize)
|
1830
|
+
{
|
1831
|
+
size_t ntrees = indexer.indices.size();
|
1832
|
+
size_t n_ref = indexer.indices.front().reference_points.size();
|
1833
|
+
|
1834
|
+
SignalSwitcher ss;
|
1835
|
+
|
1836
|
+
std::unique_ptr<sparse_ix[]> terminal_indices(new sparse_ix[nrows*ntrees]);
|
1837
|
+
std::unique_ptr<double[]> ignored(new double[nrows]);
|
1838
|
+
predict_iforest(numeric_data, categ_data,
|
1839
|
+
is_col_major, ld_numeric, ld_categ,
|
1840
|
+
is_col_major? Xc : nullptr, is_col_major? Xc_ind : nullptr, is_col_major? Xc_indptr : nullptr,
|
1841
|
+
is_col_major? (real_t*)nullptr : Xc, is_col_major? (sparse_ix*)nullptr : Xc_ind, is_col_major? (sparse_ix*)nullptr : Xc_indptr,
|
1842
|
+
nrows, nthreads, false,
|
1843
|
+
model_outputs, model_outputs_ext,
|
1844
|
+
ignored.get(), terminal_indices.get(),
|
1845
|
+
(double*)NULL,
|
1846
|
+
&indexer);
|
1847
|
+
ignored.reset();
|
1848
|
+
|
1849
|
+
check_interrupt_switch(ss);
|
1850
|
+
|
1851
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) \
|
1852
|
+
shared(indexer, terminal_indices, nrows, ntrees, n_ref, rmat)
|
1853
|
+
for (size_t_for row = 0; row < (decltype(row))nrows; row++)
|
1854
|
+
{
|
1855
|
+
if (interrupt_switch) continue;
|
1856
|
+
|
1857
|
+
SingleTreeIndex *restrict index_node;
|
1858
|
+
size_t idx_this;
|
1859
|
+
sparse_ix *restrict terminal_indices_this = terminal_indices.get() + row;
|
1860
|
+
double *restrict rmat_this = rmat + row*n_ref;
|
1861
|
+
memset(rmat_this, 0, n_ref*sizeof(double));
|
1862
|
+
|
1863
|
+
for (size_t tree = 0; tree < ntrees; tree++)
|
1864
|
+
{
|
1865
|
+
idx_this = terminal_indices_this[tree*nrows];
|
1866
|
+
index_node = &indexer.indices[tree];
|
1867
|
+
for (size_t ind = index_node->reference_indptr[idx_this];
|
1868
|
+
ind < index_node->reference_indptr[idx_this + 1];
|
1869
|
+
ind++)
|
1870
|
+
{
|
1871
|
+
rmat_this[index_node->reference_mapping[ind]]++;
|
1872
|
+
}
|
1873
|
+
}
|
1874
|
+
}
|
1875
|
+
|
1876
|
+
check_interrupt_switch(ss);
|
1877
|
+
|
1878
|
+
if (standardize)
|
1879
|
+
{
|
1880
|
+
double ntrees_dbl = (double)ntrees;
|
1881
|
+
for (size_t ix = 0; ix < nrows*n_ref; ix++)
|
1882
|
+
rmat[ix] /= ntrees_dbl;
|
1883
|
+
}
|
1884
|
+
|
1885
|
+
check_interrupt_switch(ss);
|
1886
|
+
}
|