1 | /*********************************************************************** |
---|
2 | * Software License Agreement (BSD License) |
---|
3 | * |
---|
4 | * Copyright 2008-2009 Marius Muja (mariusm@cs.ubc.ca). All rights reserved. |
---|
5 | * Copyright 2008-2009 David G. Lowe (lowe@cs.ubc.ca). All rights reserved. |
---|
6 | * |
---|
7 | * THE BSD LICENSE |
---|
8 | * |
---|
9 | * Redistribution and use in source and binary forms, with or without |
---|
10 | * modification, are permitted provided that the following conditions |
---|
11 | * are met: |
---|
12 | * |
---|
13 | * 1. Redistributions of source code must retain the above copyright |
---|
14 | * notice, this list of conditions and the following disclaimer. |
---|
15 | * 2. Redistributions in binary form must reproduce the above copyright |
---|
16 | * notice, this list of conditions and the following disclaimer in the |
---|
17 | * documentation and/or other materials provided with the distribution. |
---|
18 | * |
---|
19 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR |
---|
20 | * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES |
---|
21 | * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. |
---|
22 | * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, |
---|
23 | * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT |
---|
24 | * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, |
---|
25 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY |
---|
26 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
---|
27 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF |
---|
28 | * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
---|
29 | *************************************************************************/ |
---|
30 | |
---|
31 | #ifndef FLANN_INDEX_TESTING_H_ |
---|
32 | #define FLANN_INDEX_TESTING_H_ |
---|
33 | |
---|
34 | #include <cstring> |
---|
35 | #include <cassert> |
---|
36 | #include <cmath> |
---|
37 | |
---|
38 | #include "flann/util/matrix.h" |
---|
39 | #include "flann/algorithms/nn_index.h" |
---|
40 | #include "flann/util/result_set.h" |
---|
41 | #include "flann/util/logger.h" |
---|
42 | #include "flann/util/timer.h" |
---|
43 | |
---|
44 | |
---|
45 | namespace flann |
---|
46 | { |
---|
47 | |
---|
48 | inline int countCorrectMatches(int* neighbors, int* groundTruth, int n) |
---|
49 | { |
---|
50 | int count = 0; |
---|
51 | for (int i=0; i<n; ++i) { |
---|
52 | for (int k=0; k<n; ++k) { |
---|
53 | if (neighbors[i]==groundTruth[k]) { |
---|
54 | count++; |
---|
55 | break; |
---|
56 | } |
---|
57 | } |
---|
58 | } |
---|
59 | return count; |
---|
60 | } |
---|
61 | |
---|
62 | |
---|
63 | template <typename Distance> |
---|
64 | typename Distance::ResultType computeDistanceRaport(const Matrix<typename Distance::ElementType>& inputData, typename Distance::ElementType* target, |
---|
65 | int* neighbors, int* groundTruth, int veclen, int n, const Distance& distance) |
---|
66 | { |
---|
67 | typedef typename Distance::ResultType DistanceType; |
---|
68 | |
---|
69 | DistanceType ret = 0; |
---|
70 | for (int i=0; i<n; ++i) { |
---|
71 | DistanceType den = distance(inputData[groundTruth[i]], target, veclen); |
---|
72 | DistanceType num = distance(inputData[neighbors[i]], target, veclen); |
---|
73 | |
---|
74 | if ((den==0)&&(num==0)) { |
---|
75 | ret += 1; |
---|
76 | } |
---|
77 | else { |
---|
78 | ret += num/den; |
---|
79 | } |
---|
80 | } |
---|
81 | |
---|
82 | return ret; |
---|
83 | } |
---|
84 | |
---|
85 | template <typename Distance> |
---|
86 | float search_with_ground_truth(NNIndex<Distance>& index, const Matrix<typename Distance::ElementType>& inputData, |
---|
87 | const Matrix<typename Distance::ElementType>& testData, const Matrix<int>& matches, int nn, int checks, |
---|
88 | float& time, typename Distance::ResultType& dist, const Distance& distance, int skipMatches) |
---|
89 | { |
---|
90 | typedef typename Distance::ResultType DistanceType; |
---|
91 | |
---|
92 | if (matches.cols<size_t(nn)) { |
---|
93 | Logger::info("matches.cols=%d, nn=%d\n",matches.cols,nn); |
---|
94 | |
---|
95 | throw FLANNException("Ground truth is not computed for as many neighbors as requested"); |
---|
96 | } |
---|
97 | |
---|
98 | KNNResultSet<DistanceType> resultSet(nn+skipMatches); |
---|
99 | SearchParams searchParams(checks); |
---|
100 | |
---|
101 | int* indices = new int[nn+skipMatches]; |
---|
102 | DistanceType* dists = new DistanceType[nn+skipMatches]; |
---|
103 | int* neighbors = indices + skipMatches; |
---|
104 | |
---|
105 | int correct = 0; |
---|
106 | DistanceType distR = 0; |
---|
107 | StartStopTimer t; |
---|
108 | int repeats = 0; |
---|
109 | while (t.value<0.2) { |
---|
110 | repeats++; |
---|
111 | t.start(); |
---|
112 | correct = 0; |
---|
113 | distR = 0; |
---|
114 | for (size_t i = 0; i < testData.rows; i++) { |
---|
115 | resultSet.clear(); |
---|
116 | index.findNeighbors(resultSet, testData[i], searchParams); |
---|
117 | resultSet.copy(indices,dists,nn+skipMatches); |
---|
118 | |
---|
119 | correct += countCorrectMatches(neighbors,matches[i], nn); |
---|
120 | distR += computeDistanceRaport<Distance>(inputData, testData[i], neighbors, matches[i], testData.cols, nn, distance); |
---|
121 | } |
---|
122 | t.stop(); |
---|
123 | } |
---|
124 | time = float(t.value/repeats); |
---|
125 | |
---|
126 | delete[] indices; |
---|
127 | delete[] dists; |
---|
128 | |
---|
129 | float precicion = (float)correct/(nn*testData.rows); |
---|
130 | |
---|
131 | dist = distR/(testData.rows*nn); |
---|
132 | |
---|
133 | Logger::info("%8d %10.4g %10.5g %10.5g %10.5g\n", |
---|
134 | checks, precicion, time, 1000.0 * time / testData.rows, dist); |
---|
135 | |
---|
136 | return precicion; |
---|
137 | } |
---|
138 | |
---|
139 | |
---|
140 | template <typename Distance> |
---|
141 | float test_index_checks(NNIndex<Distance>& index, const Matrix<typename Distance::ElementType>& inputData, |
---|
142 | const Matrix<typename Distance::ElementType>& testData, const Matrix<int>& matches, |
---|
143 | int checks, float& precision, const Distance& distance, int nn = 1, int skipMatches = 0) |
---|
144 | { |
---|
145 | typedef typename Distance::ResultType DistanceType; |
---|
146 | |
---|
147 | Logger::info(" Nodes Precision(%) Time(s) Time/vec(ms) Mean dist\n"); |
---|
148 | Logger::info("---------------------------------------------------------\n"); |
---|
149 | |
---|
150 | float time = 0; |
---|
151 | DistanceType dist = 0; |
---|
152 | precision = search_with_ground_truth(index, inputData, testData, matches, nn, checks, time, dist, distance, skipMatches); |
---|
153 | |
---|
154 | return time; |
---|
155 | } |
---|
156 | |
---|
157 | template <typename Distance> |
---|
158 | float test_index_precision(NNIndex<Distance>& index, const Matrix<typename Distance::ElementType>& inputData, |
---|
159 | const Matrix<typename Distance::ElementType>& testData, const Matrix<int>& matches, |
---|
160 | float precision, int& checks, const Distance& distance, int nn = 1, int skipMatches = 0) |
---|
161 | { |
---|
162 | typedef typename Distance::ResultType DistanceType; |
---|
163 | const float SEARCH_EPS = 0.001f; |
---|
164 | |
---|
165 | Logger::info(" Nodes Precision(%) Time(s) Time/vec(ms) Mean dist\n"); |
---|
166 | Logger::info("---------------------------------------------------------\n"); |
---|
167 | |
---|
168 | int c2 = 1; |
---|
169 | float p2; |
---|
170 | int c1 = 1; |
---|
171 | // float p1; |
---|
172 | float time; |
---|
173 | DistanceType dist; |
---|
174 | |
---|
175 | p2 = search_with_ground_truth(index, inputData, testData, matches, nn, c2, time, dist, distance, skipMatches); |
---|
176 | |
---|
177 | if (p2>precision) { |
---|
178 | Logger::info("Got as close as I can\n"); |
---|
179 | checks = c2; |
---|
180 | return time; |
---|
181 | } |
---|
182 | |
---|
183 | while (p2<precision) { |
---|
184 | c1 = c2; |
---|
185 | // p1 = p2; |
---|
186 | c2 *=2; |
---|
187 | p2 = search_with_ground_truth(index, inputData, testData, matches, nn, c2, time, dist, distance, skipMatches); |
---|
188 | } |
---|
189 | |
---|
190 | int cx; |
---|
191 | float realPrecision; |
---|
192 | if (fabs(p2-precision)>SEARCH_EPS) { |
---|
193 | Logger::info("Start linear estimation\n"); |
---|
194 | // after we got to values in the vecinity of the desired precision |
---|
195 | // use linear approximation get a better estimation |
---|
196 | |
---|
197 | cx = (c1+c2)/2; |
---|
198 | realPrecision = search_with_ground_truth(index, inputData, testData, matches, nn, cx, time, dist, distance, skipMatches); |
---|
199 | while (fabs(realPrecision-precision)>SEARCH_EPS) { |
---|
200 | |
---|
201 | if (realPrecision<precision) { |
---|
202 | c1 = cx; |
---|
203 | } |
---|
204 | else { |
---|
205 | c2 = cx; |
---|
206 | } |
---|
207 | cx = (c1+c2)/2; |
---|
208 | if (cx==c1) { |
---|
209 | Logger::info("Got as close as I can\n"); |
---|
210 | break; |
---|
211 | } |
---|
212 | realPrecision = search_with_ground_truth(index, inputData, testData, matches, nn, cx, time, dist, distance, skipMatches); |
---|
213 | } |
---|
214 | |
---|
215 | c2 = cx; |
---|
216 | p2 = realPrecision; |
---|
217 | |
---|
218 | } |
---|
219 | else { |
---|
220 | Logger::info("No need for linear estimation\n"); |
---|
221 | cx = c2; |
---|
222 | realPrecision = p2; |
---|
223 | } |
---|
224 | |
---|
225 | checks = cx; |
---|
226 | return time; |
---|
227 | } |
---|
228 | |
---|
229 | |
---|
230 | template <typename Distance> |
---|
231 | void test_index_precisions(NNIndex<Distance>& index, const Matrix<typename Distance::ElementType>& inputData, |
---|
232 | const Matrix<typename Distance::ElementType>& testData, const Matrix<int>& matches, |
---|
233 | float* precisions, int precisions_length, const Distance& distance, int nn = 1, int skipMatches = 0, float maxTime = 0) |
---|
234 | { |
---|
235 | typedef typename Distance::ResultType DistanceType; |
---|
236 | |
---|
237 | const float SEARCH_EPS = 0.001; |
---|
238 | |
---|
239 | // make sure precisions array is sorted |
---|
240 | std::sort(precisions, precisions+precisions_length); |
---|
241 | |
---|
242 | int pindex = 0; |
---|
243 | float precision = precisions[pindex]; |
---|
244 | |
---|
245 | Logger::info(" Nodes Precision(%) Time(s) Time/vec(ms) Mean dist\n"); |
---|
246 | Logger::info("---------------------------------------------------------\n"); |
---|
247 | |
---|
248 | int c2 = 1; |
---|
249 | float p2; |
---|
250 | |
---|
251 | int c1 = 1; |
---|
252 | float p1; |
---|
253 | |
---|
254 | float time; |
---|
255 | DistanceType dist; |
---|
256 | |
---|
257 | p2 = search_with_ground_truth(index, inputData, testData, matches, nn, c2, time, dist, distance, skipMatches); |
---|
258 | |
---|
259 | // if precision for 1 run down the tree is already |
---|
260 | // better then some of the requested precisions, then |
---|
261 | // skip those |
---|
262 | while (precisions[pindex]<p2 && pindex<precisions_length) { |
---|
263 | pindex++; |
---|
264 | } |
---|
265 | |
---|
266 | if (pindex==precisions_length) { |
---|
267 | Logger::info("Got as close as I can\n"); |
---|
268 | return; |
---|
269 | } |
---|
270 | |
---|
271 | for (int i=pindex; i<precisions_length; ++i) { |
---|
272 | |
---|
273 | precision = precisions[i]; |
---|
274 | while (p2<precision) { |
---|
275 | c1 = c2; |
---|
276 | p1 = p2; |
---|
277 | c2 *=2; |
---|
278 | p2 = search_with_ground_truth(index, inputData, testData, matches, nn, c2, time, dist, distance, skipMatches); |
---|
279 | if ((maxTime> 0)&&(time > maxTime)&&(p2<precision)) return; |
---|
280 | } |
---|
281 | |
---|
282 | int cx; |
---|
283 | float realPrecision; |
---|
284 | if (fabs(p2-precision)>SEARCH_EPS) { |
---|
285 | Logger::info("Start linear estimation\n"); |
---|
286 | // after we got to values in the vecinity of the desired precision |
---|
287 | // use linear approximation get a better estimation |
---|
288 | |
---|
289 | cx = (c1+c2)/2; |
---|
290 | realPrecision = search_with_ground_truth(index, inputData, testData, matches, nn, cx, time, dist, distance, skipMatches); |
---|
291 | while (fabs(realPrecision-precision)>SEARCH_EPS) { |
---|
292 | |
---|
293 | if (realPrecision<precision) { |
---|
294 | c1 = cx; |
---|
295 | } |
---|
296 | else { |
---|
297 | c2 = cx; |
---|
298 | } |
---|
299 | cx = (c1+c2)/2; |
---|
300 | if (cx==c1) { |
---|
301 | Logger::info("Got as close as I can\n"); |
---|
302 | break; |
---|
303 | } |
---|
304 | realPrecision = search_with_ground_truth(index, inputData, testData, matches, nn, cx, time, dist, distance, skipMatches); |
---|
305 | } |
---|
306 | |
---|
307 | c2 = cx; |
---|
308 | p2 = realPrecision; |
---|
309 | |
---|
310 | } |
---|
311 | else { |
---|
312 | Logger::info("No need for linear estimation\n"); |
---|
313 | cx = c2; |
---|
314 | realPrecision = p2; |
---|
315 | } |
---|
316 | |
---|
317 | } |
---|
318 | } |
---|
319 | |
---|
320 | } |
---|
321 | |
---|
322 | #endif //FLANN_INDEX_TESTING_H_ |
---|