faiss 0.1.4 → 0.2.1
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +26 -1
- data/README.md +15 -3
- data/ext/faiss/ext.cpp +12 -308
- data/ext/faiss/extconf.rb +5 -2
- data/ext/faiss/index.cpp +189 -0
- data/ext/faiss/index_binary.cpp +75 -0
- data/ext/faiss/kmeans.cpp +40 -0
- data/ext/faiss/numo.hpp +867 -0
- data/ext/faiss/pca_matrix.cpp +33 -0
- data/ext/faiss/product_quantizer.cpp +53 -0
- data/ext/faiss/utils.cpp +13 -0
- data/ext/faiss/utils.h +5 -0
- data/lib/faiss.rb +0 -5
- data/lib/faiss/version.rb +1 -1
- data/vendor/faiss/faiss/AutoTune.cpp +292 -291
- data/vendor/faiss/faiss/AutoTune.h +55 -56
- data/vendor/faiss/faiss/Clustering.cpp +334 -195
- data/vendor/faiss/faiss/Clustering.h +88 -35
- data/vendor/faiss/faiss/IVFlib.cpp +171 -195
- data/vendor/faiss/faiss/IVFlib.h +48 -51
- data/vendor/faiss/faiss/Index.cpp +85 -103
- data/vendor/faiss/faiss/Index.h +54 -48
- data/vendor/faiss/faiss/Index2Layer.cpp +139 -164
- data/vendor/faiss/faiss/Index2Layer.h +22 -22
- data/vendor/faiss/faiss/IndexBinary.cpp +45 -37
- data/vendor/faiss/faiss/IndexBinary.h +140 -132
- data/vendor/faiss/faiss/IndexBinaryFlat.cpp +73 -53
- data/vendor/faiss/faiss/IndexBinaryFlat.h +29 -24
- data/vendor/faiss/faiss/IndexBinaryFromFloat.cpp +46 -43
- data/vendor/faiss/faiss/IndexBinaryFromFloat.h +16 -15
- data/vendor/faiss/faiss/IndexBinaryHNSW.cpp +215 -232
- data/vendor/faiss/faiss/IndexBinaryHNSW.h +25 -24
- data/vendor/faiss/faiss/IndexBinaryHash.cpp +182 -177
- data/vendor/faiss/faiss/IndexBinaryHash.h +41 -34
- data/vendor/faiss/faiss/IndexBinaryIVF.cpp +489 -461
- data/vendor/faiss/faiss/IndexBinaryIVF.h +97 -68
- data/vendor/faiss/faiss/IndexFlat.cpp +116 -147
- data/vendor/faiss/faiss/IndexFlat.h +35 -46
- data/vendor/faiss/faiss/IndexHNSW.cpp +372 -348
- data/vendor/faiss/faiss/IndexHNSW.h +57 -41
- data/vendor/faiss/faiss/IndexIVF.cpp +474 -454
- data/vendor/faiss/faiss/IndexIVF.h +146 -113
- data/vendor/faiss/faiss/IndexIVFFlat.cpp +248 -250
- data/vendor/faiss/faiss/IndexIVFFlat.h +48 -51
- data/vendor/faiss/faiss/IndexIVFPQ.cpp +457 -516
- data/vendor/faiss/faiss/IndexIVFPQ.h +74 -66
- data/vendor/faiss/faiss/IndexIVFPQFastScan.cpp +406 -372
- data/vendor/faiss/faiss/IndexIVFPQFastScan.h +82 -57
- data/vendor/faiss/faiss/IndexIVFPQR.cpp +104 -102
- data/vendor/faiss/faiss/IndexIVFPQR.h +33 -28
- data/vendor/faiss/faiss/IndexIVFSpectralHash.cpp +125 -133
- data/vendor/faiss/faiss/IndexIVFSpectralHash.h +19 -21
- data/vendor/faiss/faiss/IndexLSH.cpp +75 -96
- data/vendor/faiss/faiss/IndexLSH.h +21 -26
- data/vendor/faiss/faiss/IndexLattice.cpp +42 -56
- data/vendor/faiss/faiss/IndexLattice.h +11 -16
- data/vendor/faiss/faiss/IndexNNDescent.cpp +231 -0
- data/vendor/faiss/faiss/IndexNNDescent.h +72 -0
- data/vendor/faiss/faiss/IndexNSG.cpp +303 -0
- data/vendor/faiss/faiss/IndexNSG.h +85 -0
- data/vendor/faiss/faiss/IndexPQ.cpp +405 -464
- data/vendor/faiss/faiss/IndexPQ.h +64 -67
- data/vendor/faiss/faiss/IndexPQFastScan.cpp +143 -170
- data/vendor/faiss/faiss/IndexPQFastScan.h +46 -32
- data/vendor/faiss/faiss/IndexPreTransform.cpp +120 -150
- data/vendor/faiss/faiss/IndexPreTransform.h +33 -36
- data/vendor/faiss/faiss/IndexRefine.cpp +115 -131
- data/vendor/faiss/faiss/IndexRefine.h +22 -23
- data/vendor/faiss/faiss/IndexReplicas.cpp +147 -153
- data/vendor/faiss/faiss/IndexReplicas.h +62 -56
- data/vendor/faiss/faiss/IndexResidual.cpp +291 -0
- data/vendor/faiss/faiss/IndexResidual.h +152 -0
- data/vendor/faiss/faiss/IndexScalarQuantizer.cpp +120 -155
- data/vendor/faiss/faiss/IndexScalarQuantizer.h +41 -45
- data/vendor/faiss/faiss/IndexShards.cpp +256 -240
- data/vendor/faiss/faiss/IndexShards.h +85 -73
- data/vendor/faiss/faiss/MatrixStats.cpp +112 -97
- data/vendor/faiss/faiss/MatrixStats.h +7 -10
- data/vendor/faiss/faiss/MetaIndexes.cpp +135 -157
- data/vendor/faiss/faiss/MetaIndexes.h +40 -34
- data/vendor/faiss/faiss/MetricType.h +7 -7
- data/vendor/faiss/faiss/VectorTransform.cpp +652 -474
- data/vendor/faiss/faiss/VectorTransform.h +61 -89
- data/vendor/faiss/faiss/clone_index.cpp +77 -73
- data/vendor/faiss/faiss/clone_index.h +4 -9
- data/vendor/faiss/faiss/gpu/GpuAutoTune.cpp +33 -38
- data/vendor/faiss/faiss/gpu/GpuAutoTune.h +11 -9
- data/vendor/faiss/faiss/gpu/GpuCloner.cpp +197 -170
- data/vendor/faiss/faiss/gpu/GpuCloner.h +53 -35
- data/vendor/faiss/faiss/gpu/GpuClonerOptions.cpp +12 -14
- data/vendor/faiss/faiss/gpu/GpuClonerOptions.h +27 -25
- data/vendor/faiss/faiss/gpu/GpuDistance.h +116 -112
- data/vendor/faiss/faiss/gpu/GpuFaissAssert.h +1 -2
- data/vendor/faiss/faiss/gpu/GpuIndex.h +134 -137
- data/vendor/faiss/faiss/gpu/GpuIndexBinaryFlat.h +76 -73
- data/vendor/faiss/faiss/gpu/GpuIndexFlat.h +173 -162
- data/vendor/faiss/faiss/gpu/GpuIndexIVF.h +67 -64
- data/vendor/faiss/faiss/gpu/GpuIndexIVFFlat.h +89 -86
- data/vendor/faiss/faiss/gpu/GpuIndexIVFPQ.h +150 -141
- data/vendor/faiss/faiss/gpu/GpuIndexIVFScalarQuantizer.h +101 -103
- data/vendor/faiss/faiss/gpu/GpuIndicesOptions.h +17 -16
- data/vendor/faiss/faiss/gpu/GpuResources.cpp +116 -128
- data/vendor/faiss/faiss/gpu/GpuResources.h +182 -186
- data/vendor/faiss/faiss/gpu/StandardGpuResources.cpp +433 -422
- data/vendor/faiss/faiss/gpu/StandardGpuResources.h +131 -130
- data/vendor/faiss/faiss/gpu/impl/InterleavedCodes.cpp +468 -456
- data/vendor/faiss/faiss/gpu/impl/InterleavedCodes.h +25 -19
- data/vendor/faiss/faiss/gpu/impl/RemapIndices.cpp +22 -20
- data/vendor/faiss/faiss/gpu/impl/RemapIndices.h +9 -8
- data/vendor/faiss/faiss/gpu/perf/IndexWrapper-inl.h +39 -44
- data/vendor/faiss/faiss/gpu/perf/IndexWrapper.h +16 -14
- data/vendor/faiss/faiss/gpu/perf/PerfClustering.cpp +77 -71
- data/vendor/faiss/faiss/gpu/perf/PerfIVFPQAdd.cpp +109 -88
- data/vendor/faiss/faiss/gpu/perf/WriteIndex.cpp +75 -64
- data/vendor/faiss/faiss/gpu/test/TestCodePacking.cpp +230 -215
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexBinaryFlat.cpp +80 -86
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexFlat.cpp +284 -277
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFFlat.cpp +416 -416
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFPQ.cpp +611 -517
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFScalarQuantizer.cpp +166 -164
- data/vendor/faiss/faiss/gpu/test/TestGpuMemoryException.cpp +61 -53
- data/vendor/faiss/faiss/gpu/test/TestUtils.cpp +274 -238
- data/vendor/faiss/faiss/gpu/test/TestUtils.h +73 -57
- data/vendor/faiss/faiss/gpu/test/demo_ivfpq_indexing_gpu.cpp +47 -50
- data/vendor/faiss/faiss/gpu/utils/DeviceUtils.h +79 -72
- data/vendor/faiss/faiss/gpu/utils/StackDeviceMemory.cpp +140 -146
- data/vendor/faiss/faiss/gpu/utils/StackDeviceMemory.h +69 -71
- data/vendor/faiss/faiss/gpu/utils/StaticUtils.h +21 -16
- data/vendor/faiss/faiss/gpu/utils/Timer.cpp +25 -29
- data/vendor/faiss/faiss/gpu/utils/Timer.h +30 -29
- data/vendor/faiss/faiss/impl/AdditiveQuantizer.cpp +270 -0
- data/vendor/faiss/faiss/impl/AdditiveQuantizer.h +115 -0
- data/vendor/faiss/faiss/impl/AuxIndexStructures.cpp +90 -120
- data/vendor/faiss/faiss/impl/AuxIndexStructures.h +81 -65
- data/vendor/faiss/faiss/impl/FaissAssert.h +73 -58
- data/vendor/faiss/faiss/impl/FaissException.cpp +56 -48
- data/vendor/faiss/faiss/impl/FaissException.h +41 -29
- data/vendor/faiss/faiss/impl/HNSW.cpp +595 -611
- data/vendor/faiss/faiss/impl/HNSW.h +179 -200
- data/vendor/faiss/faiss/impl/LocalSearchQuantizer.cpp +672 -0
- data/vendor/faiss/faiss/impl/LocalSearchQuantizer.h +172 -0
- data/vendor/faiss/faiss/impl/NNDescent.cpp +487 -0
- data/vendor/faiss/faiss/impl/NNDescent.h +154 -0
- data/vendor/faiss/faiss/impl/NSG.cpp +682 -0
- data/vendor/faiss/faiss/impl/NSG.h +199 -0
- data/vendor/faiss/faiss/impl/PolysemousTraining.cpp +484 -454
- data/vendor/faiss/faiss/impl/PolysemousTraining.h +52 -55
- data/vendor/faiss/faiss/impl/ProductQuantizer-inl.h +26 -47
- data/vendor/faiss/faiss/impl/ProductQuantizer.cpp +469 -459
- data/vendor/faiss/faiss/impl/ProductQuantizer.h +76 -87
- data/vendor/faiss/faiss/impl/ResidualQuantizer.cpp +448 -0
- data/vendor/faiss/faiss/impl/ResidualQuantizer.h +130 -0
- data/vendor/faiss/faiss/impl/ResultHandler.h +96 -132
- data/vendor/faiss/faiss/impl/ScalarQuantizer.cpp +648 -701
- data/vendor/faiss/faiss/impl/ScalarQuantizer.h +48 -46
- data/vendor/faiss/faiss/impl/ThreadedIndex-inl.h +129 -131
- data/vendor/faiss/faiss/impl/ThreadedIndex.h +61 -55
- data/vendor/faiss/faiss/impl/index_read.cpp +547 -479
- data/vendor/faiss/faiss/impl/index_write.cpp +497 -407
- data/vendor/faiss/faiss/impl/io.cpp +75 -94
- data/vendor/faiss/faiss/impl/io.h +31 -41
- data/vendor/faiss/faiss/impl/io_macros.h +40 -29
- data/vendor/faiss/faiss/impl/lattice_Zn.cpp +137 -186
- data/vendor/faiss/faiss/impl/lattice_Zn.h +40 -51
- data/vendor/faiss/faiss/impl/platform_macros.h +29 -8
- data/vendor/faiss/faiss/impl/pq4_fast_scan.cpp +77 -124
- data/vendor/faiss/faiss/impl/pq4_fast_scan.h +39 -48
- data/vendor/faiss/faiss/impl/pq4_fast_scan_search_1.cpp +41 -52
- data/vendor/faiss/faiss/impl/pq4_fast_scan_search_qbs.cpp +80 -117
- data/vendor/faiss/faiss/impl/simd_result_handlers.h +109 -137
- data/vendor/faiss/faiss/index_factory.cpp +269 -218
- data/vendor/faiss/faiss/index_factory.h +6 -7
- data/vendor/faiss/faiss/index_io.h +23 -26
- data/vendor/faiss/faiss/invlists/BlockInvertedLists.cpp +67 -75
- data/vendor/faiss/faiss/invlists/BlockInvertedLists.h +22 -24
- data/vendor/faiss/faiss/invlists/DirectMap.cpp +96 -112
- data/vendor/faiss/faiss/invlists/DirectMap.h +29 -33
- data/vendor/faiss/faiss/invlists/InvertedLists.cpp +307 -364
- data/vendor/faiss/faiss/invlists/InvertedLists.h +151 -151
- data/vendor/faiss/faiss/invlists/InvertedListsIOHook.cpp +29 -34
- data/vendor/faiss/faiss/invlists/InvertedListsIOHook.h +17 -18
- data/vendor/faiss/faiss/invlists/OnDiskInvertedLists.cpp +257 -293
- data/vendor/faiss/faiss/invlists/OnDiskInvertedLists.h +50 -45
- data/vendor/faiss/faiss/python/python_callbacks.cpp +23 -26
- data/vendor/faiss/faiss/python/python_callbacks.h +9 -16
- data/vendor/faiss/faiss/utils/AlignedTable.h +79 -44
- data/vendor/faiss/faiss/utils/Heap.cpp +40 -48
- data/vendor/faiss/faiss/utils/Heap.h +186 -209
- data/vendor/faiss/faiss/utils/WorkerThread.cpp +67 -76
- data/vendor/faiss/faiss/utils/WorkerThread.h +32 -33
- data/vendor/faiss/faiss/utils/distances.cpp +301 -310
- data/vendor/faiss/faiss/utils/distances.h +133 -118
- data/vendor/faiss/faiss/utils/distances_simd.cpp +456 -516
- data/vendor/faiss/faiss/utils/extra_distances-inl.h +117 -0
- data/vendor/faiss/faiss/utils/extra_distances.cpp +113 -232
- data/vendor/faiss/faiss/utils/extra_distances.h +30 -29
- data/vendor/faiss/faiss/utils/hamming-inl.h +260 -209
- data/vendor/faiss/faiss/utils/hamming.cpp +375 -469
- data/vendor/faiss/faiss/utils/hamming.h +62 -85
- data/vendor/faiss/faiss/utils/ordered_key_value.h +16 -18
- data/vendor/faiss/faiss/utils/partitioning.cpp +393 -318
- data/vendor/faiss/faiss/utils/partitioning.h +26 -21
- data/vendor/faiss/faiss/utils/quantize_lut.cpp +78 -66
- data/vendor/faiss/faiss/utils/quantize_lut.h +22 -20
- data/vendor/faiss/faiss/utils/random.cpp +39 -63
- data/vendor/faiss/faiss/utils/random.h +13 -16
- data/vendor/faiss/faiss/utils/simdlib.h +4 -2
- data/vendor/faiss/faiss/utils/simdlib_avx2.h +88 -85
- data/vendor/faiss/faiss/utils/simdlib_emulated.h +226 -165
- data/vendor/faiss/faiss/utils/simdlib_neon.h +832 -0
- data/vendor/faiss/faiss/utils/utils.cpp +304 -287
- data/vendor/faiss/faiss/utils/utils.h +53 -48
- metadata +31 -10
- data/lib/faiss/index.rb +0 -20
- data/lib/faiss/index_binary.rb +0 -20
- data/lib/faiss/kmeans.rb +0 -15
- data/lib/faiss/pca_matrix.rb +0 -15
- data/lib/faiss/product_quantizer.rb +0 -22
@@ -10,38 +10,39 @@
|
|
10
10
|
#ifndef FAISS_AUTO_TUNE_H
|
11
11
|
#define FAISS_AUTO_TUNE_H
|
12
12
|
|
13
|
-
#include <vector>
|
14
|
-
#include <unordered_map>
|
15
13
|
#include <stdint.h>
|
14
|
+
#include <unordered_map>
|
15
|
+
#include <vector>
|
16
16
|
|
17
17
|
#include <faiss/Index.h>
|
18
18
|
#include <faiss/IndexBinary.h>
|
19
19
|
|
20
20
|
namespace faiss {
|
21
21
|
|
22
|
-
|
23
22
|
/**
|
24
23
|
* Evaluation criterion. Returns a performance measure in [0,1],
|
25
24
|
* higher is better.
|
26
25
|
*/
|
27
26
|
struct AutoTuneCriterion {
|
28
27
|
typedef Index::idx_t idx_t;
|
29
|
-
idx_t nq;
|
30
|
-
idx_t nnn;
|
31
|
-
idx_t gt_nnn; ///< nb of GT NNs required to evaluate
|
28
|
+
idx_t nq; ///< nb of queries this criterion is evaluated on
|
29
|
+
idx_t nnn; ///< nb of NNs that the query should request
|
30
|
+
idx_t gt_nnn; ///< nb of GT NNs required to evaluate criterion
|
32
31
|
|
33
|
-
std::vector<float> gt_D;
|
34
|
-
std::vector<idx_t> gt_I;
|
32
|
+
std::vector<float> gt_D; ///< Ground-truth distances (size nq * gt_nnn)
|
33
|
+
std::vector<idx_t> gt_I; ///< Ground-truth indexes (size nq * gt_nnn)
|
35
34
|
|
36
|
-
AutoTuneCriterion
|
35
|
+
AutoTuneCriterion(idx_t nq, idx_t nnn);
|
37
36
|
|
38
37
|
/** Intitializes the gt_D and gt_I vectors. Must be called before evaluating
|
39
38
|
*
|
40
39
|
* @param gt_D_in size nq * gt_nnn
|
41
40
|
* @param gt_I_in size nq * gt_nnn
|
42
41
|
*/
|
43
|
-
void set_groundtruth
|
44
|
-
|
42
|
+
void set_groundtruth(
|
43
|
+
int gt_nnn,
|
44
|
+
const float* gt_D_in,
|
45
|
+
const idx_t* gt_I_in);
|
45
46
|
|
46
47
|
/** Evaluate the criterion.
|
47
48
|
*
|
@@ -49,29 +50,25 @@ struct AutoTuneCriterion {
|
|
49
50
|
* @param I size nq * nnn
|
50
51
|
* @return the criterion, between 0 and 1. Larger is better.
|
51
52
|
*/
|
52
|
-
virtual double evaluate
|
53
|
-
|
54
|
-
virtual ~AutoTuneCriterion () {}
|
53
|
+
virtual double evaluate(const float* D, const idx_t* I) const = 0;
|
55
54
|
|
55
|
+
virtual ~AutoTuneCriterion() {}
|
56
56
|
};
|
57
57
|
|
58
|
-
struct OneRecallAtRCriterion: AutoTuneCriterion {
|
59
|
-
|
58
|
+
struct OneRecallAtRCriterion : AutoTuneCriterion {
|
60
59
|
idx_t R;
|
61
60
|
|
62
|
-
OneRecallAtRCriterion
|
61
|
+
OneRecallAtRCriterion(idx_t nq, idx_t R);
|
63
62
|
|
64
63
|
double evaluate(const float* D, const idx_t* I) const override;
|
65
64
|
|
66
65
|
~OneRecallAtRCriterion() override {}
|
67
66
|
};
|
68
67
|
|
69
|
-
|
70
|
-
struct IntersectionCriterion: AutoTuneCriterion {
|
71
|
-
|
68
|
+
struct IntersectionCriterion : AutoTuneCriterion {
|
72
69
|
idx_t R;
|
73
70
|
|
74
|
-
IntersectionCriterion
|
71
|
+
IntersectionCriterion(idx_t nq, idx_t R);
|
75
72
|
|
76
73
|
double evaluate(const float* D, const idx_t* I) const override;
|
77
74
|
|
@@ -91,7 +88,7 @@ struct OperatingPoint {
|
|
91
88
|
double perf; ///< performance measure (output of a Criterion)
|
92
89
|
double t; ///< corresponding execution time (ms)
|
93
90
|
std::string key; ///< key that identifies this op pt
|
94
|
-
int64_t cno;
|
91
|
+
int64_t cno; ///< integer identifer
|
95
92
|
};
|
96
93
|
|
97
94
|
struct OperatingPoints {
|
@@ -102,27 +99,27 @@ struct OperatingPoints {
|
|
102
99
|
std::vector<OperatingPoint> optimal_pts;
|
103
100
|
|
104
101
|
// begins with a single operating point: t=0, perf=0
|
105
|
-
OperatingPoints
|
102
|
+
OperatingPoints();
|
106
103
|
|
107
104
|
/// add operating points from other to this, with a prefix to the keys
|
108
|
-
int merge_with
|
109
|
-
|
105
|
+
int merge_with(
|
106
|
+
const OperatingPoints& other,
|
107
|
+
const std::string& prefix = "");
|
110
108
|
|
111
|
-
void clear
|
109
|
+
void clear();
|
112
110
|
|
113
111
|
/// add a performance measure. Return whether it is an optimal point
|
114
|
-
bool add
|
112
|
+
bool add(double perf, double t, const std::string& key, size_t cno = 0);
|
115
113
|
|
116
114
|
/// get time required to obtain a given performance measure
|
117
|
-
double t_for_perf
|
115
|
+
double t_for_perf(double perf) const;
|
118
116
|
|
119
117
|
/// easy-to-read output
|
120
|
-
void display
|
118
|
+
void display(bool only_optimal = true) const;
|
121
119
|
|
122
120
|
/// output to a format easy to digest by gnuplot
|
123
|
-
void all_to_gnuplot
|
124
|
-
void optimal_to_gnuplot
|
125
|
-
|
121
|
+
void all_to_gnuplot(const char* fname) const;
|
122
|
+
void optimal_to_gnuplot(const char* fname) const;
|
126
123
|
};
|
127
124
|
|
128
125
|
/// possible values of a parameter, sorted from least to most expensive/accurate
|
@@ -156,41 +153,45 @@ struct ParameterSpace {
|
|
156
153
|
/// duration (to avoid jittering in MT mode)
|
157
154
|
double min_test_duration;
|
158
155
|
|
159
|
-
ParameterSpace
|
156
|
+
ParameterSpace();
|
160
157
|
|
161
158
|
/// nb of combinations, = product of values sizes
|
162
|
-
size_t n_combinations
|
159
|
+
size_t n_combinations() const;
|
163
160
|
|
164
161
|
/// returns whether combinations c1 >= c2 in the tuple sense
|
165
|
-
bool combination_ge
|
162
|
+
bool combination_ge(size_t c1, size_t c2) const;
|
166
163
|
|
167
164
|
/// get string representation of the combination
|
168
|
-
std::string combination_name
|
165
|
+
std::string combination_name(size_t cno) const;
|
169
166
|
|
170
167
|
/// print a description on stdout
|
171
|
-
void display
|
168
|
+
void display() const;
|
172
169
|
|
173
170
|
/// add a new parameter (or return it if it exists)
|
174
|
-
ParameterRange
|
171
|
+
ParameterRange& add_range(const std::string& name);
|
175
172
|
|
176
173
|
/// initialize with reasonable parameters for the index
|
177
|
-
virtual void initialize
|
174
|
+
virtual void initialize(const Index* index);
|
178
175
|
|
179
176
|
/// set a combination of parameters on an index
|
180
|
-
void set_index_parameters
|
177
|
+
void set_index_parameters(Index* index, size_t cno) const;
|
181
178
|
|
182
179
|
/// set a combination of parameters described by a string
|
183
|
-
void set_index_parameters
|
180
|
+
void set_index_parameters(Index* index, const char* param_string) const;
|
184
181
|
|
185
182
|
/// set one of the parameters, returns whether setting was successful
|
186
|
-
virtual void set_index_parameter
|
187
|
-
|
183
|
+
virtual void set_index_parameter(
|
184
|
+
Index* index,
|
185
|
+
const std::string& name,
|
186
|
+
double val) const;
|
188
187
|
|
189
188
|
/** find an upper bound on the performance and a lower bound on t
|
190
189
|
* for configuration cno given another operating point op */
|
191
|
-
void update_bounds
|
192
|
-
|
193
|
-
|
190
|
+
void update_bounds(
|
191
|
+
size_t cno,
|
192
|
+
const OperatingPoint& op,
|
193
|
+
double* upper_bound_perf,
|
194
|
+
double* lower_bound_t) const;
|
194
195
|
|
195
196
|
/** explore operating points
|
196
197
|
* @param index index to run on
|
@@ -198,18 +199,16 @@ struct ParameterSpace {
|
|
198
199
|
* @param crit selection criterion
|
199
200
|
* @param ops resulting operating points
|
200
201
|
*/
|
201
|
-
void explore
|
202
|
-
|
203
|
-
|
204
|
-
|
205
|
-
|
206
|
-
|
202
|
+
void explore(
|
203
|
+
Index* index,
|
204
|
+
size_t nq,
|
205
|
+
const float* xq,
|
206
|
+
const AutoTuneCriterion& crit,
|
207
|
+
OperatingPoints* ops) const;
|
208
|
+
|
209
|
+
virtual ~ParameterSpace() {}
|
207
210
|
};
|
208
211
|
|
209
|
-
|
210
|
-
|
211
212
|
} // namespace faiss
|
212
213
|
|
213
|
-
|
214
|
-
|
215
214
|
#endif
|
@@ -8,6 +8,7 @@
|
|
8
8
|
// -*- c++ -*-
|
9
9
|
|
10
10
|
#include <faiss/Clustering.h>
|
11
|
+
#include <faiss/VectorTransform.h>
|
11
12
|
#include <faiss/impl/AuxIndexStructures.h>
|
12
13
|
|
13
14
|
#include <cinttypes>
|
@@ -17,100 +18,100 @@
|
|
17
18
|
|
18
19
|
#include <omp.h>
|
19
20
|
|
20
|
-
#include <faiss/utils/utils.h>
|
21
|
-
#include <faiss/utils/random.h>
|
22
|
-
#include <faiss/utils/distances.h>
|
23
|
-
#include <faiss/impl/FaissAssert.h>
|
24
21
|
#include <faiss/IndexFlat.h>
|
22
|
+
#include <faiss/impl/FaissAssert.h>
|
23
|
+
#include <faiss/utils/distances.h>
|
24
|
+
#include <faiss/utils/random.h>
|
25
|
+
#include <faiss/utils/utils.h>
|
25
26
|
|
26
27
|
namespace faiss {
|
27
28
|
|
28
|
-
ClusteringParameters::ClusteringParameters
|
29
|
-
|
30
|
-
|
31
|
-
|
32
|
-
|
33
|
-
|
34
|
-
|
35
|
-
|
36
|
-
|
37
|
-
|
38
|
-
|
39
|
-
|
40
|
-
{}
|
29
|
+
ClusteringParameters::ClusteringParameters()
|
30
|
+
: niter(25),
|
31
|
+
nredo(1),
|
32
|
+
verbose(false),
|
33
|
+
spherical(false),
|
34
|
+
int_centroids(false),
|
35
|
+
update_index(false),
|
36
|
+
frozen_centroids(false),
|
37
|
+
min_points_per_centroid(39),
|
38
|
+
max_points_per_centroid(256),
|
39
|
+
seed(1234),
|
40
|
+
decode_block_size(32768) {}
|
41
41
|
// 39 corresponds to 10000 / 256 -> to avoid warnings on PQ tests with randu10k
|
42
42
|
|
43
|
+
Clustering::Clustering(int d, int k) : d(d), k(k) {}
|
43
44
|
|
44
|
-
Clustering::Clustering
|
45
|
-
|
46
|
-
|
47
|
-
Clustering::Clustering (int d, int k, const ClusteringParameters &cp):
|
48
|
-
ClusteringParameters (cp), d(d), k(k) {}
|
45
|
+
Clustering::Clustering(int d, int k, const ClusteringParameters& cp)
|
46
|
+
: ClusteringParameters(cp), d(d), k(k) {}
|
49
47
|
|
50
|
-
|
51
|
-
|
52
|
-
static double imbalance_factor (int n, int k, int64_t *assign) {
|
48
|
+
static double imbalance_factor(int n, int k, int64_t* assign) {
|
53
49
|
std::vector<int> hist(k, 0);
|
54
50
|
for (int i = 0; i < n; i++)
|
55
51
|
hist[assign[i]]++;
|
56
52
|
|
57
53
|
double tot = 0, uf = 0;
|
58
54
|
|
59
|
-
for (int i = 0
|
55
|
+
for (int i = 0; i < k; i++) {
|
60
56
|
tot += hist[i];
|
61
|
-
uf += hist[i] * (double)
|
57
|
+
uf += hist[i] * (double)hist[i];
|
62
58
|
}
|
63
59
|
uf = uf * k / (tot * tot);
|
64
60
|
|
65
61
|
return uf;
|
66
62
|
}
|
67
63
|
|
68
|
-
void Clustering::post_process_centroids
|
69
|
-
{
|
70
|
-
|
64
|
+
void Clustering::post_process_centroids() {
|
71
65
|
if (spherical) {
|
72
|
-
fvec_renorm_L2
|
66
|
+
fvec_renorm_L2(d, k, centroids.data());
|
73
67
|
}
|
74
68
|
|
75
69
|
if (int_centroids) {
|
76
70
|
for (size_t i = 0; i < centroids.size(); i++)
|
77
|
-
centroids[i] = roundf
|
71
|
+
centroids[i] = roundf(centroids[i]);
|
78
72
|
}
|
79
73
|
}
|
80
74
|
|
81
|
-
|
82
|
-
|
83
|
-
|
84
|
-
|
85
|
-
|
75
|
+
void Clustering::train(
|
76
|
+
idx_t nx,
|
77
|
+
const float* x_in,
|
78
|
+
Index& index,
|
79
|
+
const float* weights) {
|
80
|
+
train_encoded(
|
81
|
+
nx,
|
82
|
+
reinterpret_cast<const uint8_t*>(x_in),
|
83
|
+
nullptr,
|
84
|
+
index,
|
85
|
+
weights);
|
86
86
|
}
|
87
87
|
|
88
|
-
|
89
88
|
namespace {
|
90
89
|
|
91
90
|
using idx_t = Clustering::idx_t;
|
92
91
|
|
93
92
|
idx_t subsample_training_set(
|
94
|
-
|
95
|
-
|
96
|
-
|
97
|
-
|
98
|
-
|
99
|
-
|
93
|
+
const Clustering& clus,
|
94
|
+
idx_t nx,
|
95
|
+
const uint8_t* x,
|
96
|
+
size_t line_size,
|
97
|
+
const float* weights,
|
98
|
+
uint8_t** x_out,
|
99
|
+
float** weights_out) {
|
100
100
|
if (clus.verbose) {
|
101
101
|
printf("Sampling a subset of %zd / %" PRId64 " for training\n",
|
102
|
-
clus.k * clus.max_points_per_centroid,
|
102
|
+
clus.k * clus.max_points_per_centroid,
|
103
|
+
nx);
|
103
104
|
}
|
104
|
-
std::vector<int> perm
|
105
|
-
rand_perm
|
105
|
+
std::vector<int> perm(nx);
|
106
|
+
rand_perm(perm.data(), nx, clus.seed);
|
106
107
|
nx = clus.k * clus.max_points_per_centroid;
|
107
|
-
uint8_t
|
108
|
+
uint8_t* x_new = new uint8_t[nx * line_size];
|
108
109
|
*x_out = x_new;
|
109
110
|
for (idx_t i = 0; i < nx; i++) {
|
110
|
-
memcpy
|
111
|
+
memcpy(x_new + i * line_size, x + perm[i] * line_size, line_size);
|
111
112
|
}
|
112
113
|
if (weights) {
|
113
|
-
float
|
114
|
+
float* weights_new = new float[nx];
|
114
115
|
for (idx_t i = 0; i < nx; i++) {
|
115
116
|
weights_new[i] = weights[perm[i]];
|
116
117
|
}
|
@@ -134,20 +135,23 @@ idx_t subsample_training_set(
|
|
134
135
|
*
|
135
136
|
*/
|
136
137
|
|
137
|
-
void compute_centroids
|
138
|
-
|
139
|
-
|
140
|
-
|
141
|
-
|
142
|
-
|
143
|
-
|
144
|
-
|
138
|
+
void compute_centroids(
|
139
|
+
size_t d,
|
140
|
+
size_t k,
|
141
|
+
size_t n,
|
142
|
+
size_t k_frozen,
|
143
|
+
const uint8_t* x,
|
144
|
+
const Index* codec,
|
145
|
+
const int64_t* assign,
|
146
|
+
const float* weights,
|
147
|
+
float* hassign,
|
148
|
+
float* centroids) {
|
145
149
|
k -= k_frozen;
|
146
150
|
centroids += k_frozen * d;
|
147
151
|
|
148
|
-
memset
|
152
|
+
memset(centroids, 0, sizeof(*centroids) * d * k);
|
149
153
|
|
150
|
-
size_t line_size = codec ? codec->sa_code_size() : d * sizeof
|
154
|
+
size_t line_size = codec ? codec->sa_code_size() : d * sizeof(float);
|
151
155
|
|
152
156
|
#pragma omp parallel
|
153
157
|
{
|
@@ -157,20 +161,20 @@ void compute_centroids (size_t d, size_t k, size_t n,
|
|
157
161
|
// this thread is taking care of centroids c0:c1
|
158
162
|
size_t c0 = (k * rank) / nt;
|
159
163
|
size_t c1 = (k * (rank + 1)) / nt;
|
160
|
-
std::vector<float> decode_buffer
|
164
|
+
std::vector<float> decode_buffer(d);
|
161
165
|
|
162
166
|
for (size_t i = 0; i < n; i++) {
|
163
167
|
int64_t ci = assign[i];
|
164
|
-
assert
|
168
|
+
assert(ci >= 0 && ci < k + k_frozen);
|
165
169
|
ci -= k_frozen;
|
166
|
-
if (ci >= c0 && ci < c1)
|
167
|
-
float
|
168
|
-
const float
|
170
|
+
if (ci >= c0 && ci < c1) {
|
171
|
+
float* c = centroids + ci * d;
|
172
|
+
const float* xi;
|
169
173
|
if (!codec) {
|
170
174
|
xi = reinterpret_cast<const float*>(x + i * line_size);
|
171
175
|
} else {
|
172
|
-
float
|
173
|
-
codec->sa_decode
|
176
|
+
float* xif = decode_buffer.data();
|
177
|
+
codec->sa_decode(1, x + i * line_size, xif);
|
174
178
|
xi = xif;
|
175
179
|
}
|
176
180
|
if (weights) {
|
@@ -187,7 +191,6 @@ void compute_centroids (size_t d, size_t k, size_t n,
|
|
187
191
|
}
|
188
192
|
}
|
189
193
|
}
|
190
|
-
|
191
194
|
}
|
192
195
|
|
193
196
|
#pragma omp parallel for
|
@@ -196,12 +199,11 @@ void compute_centroids (size_t d, size_t k, size_t n,
|
|
196
199
|
continue;
|
197
200
|
}
|
198
201
|
float norm = 1 / hassign[ci];
|
199
|
-
float
|
202
|
+
float* c = centroids + ci * d;
|
200
203
|
for (size_t j = 0; j < d; j++) {
|
201
204
|
c[j] *= norm;
|
202
205
|
}
|
203
206
|
}
|
204
|
-
|
205
207
|
}
|
206
208
|
|
207
209
|
// a bit above machine epsilon for float16
|
@@ -214,29 +216,33 @@ void compute_centroids (size_t d, size_t k, size_t n,
|
|
214
216
|
*
|
215
217
|
* @return nb of spliting operations (larger is worse)
|
216
218
|
*/
|
217
|
-
int split_clusters
|
218
|
-
|
219
|
-
|
220
|
-
|
221
|
-
|
219
|
+
int split_clusters(
|
220
|
+
size_t d,
|
221
|
+
size_t k,
|
222
|
+
size_t n,
|
223
|
+
size_t k_frozen,
|
224
|
+
float* hassign,
|
225
|
+
float* centroids) {
|
222
226
|
k -= k_frozen;
|
223
227
|
centroids += k_frozen * d;
|
224
228
|
|
225
229
|
/* Take care of void clusters */
|
226
230
|
size_t nsplit = 0;
|
227
|
-
RandomGenerator rng
|
231
|
+
RandomGenerator rng(1234);
|
228
232
|
for (size_t ci = 0; ci < k; ci++) {
|
229
233
|
if (hassign[ci] == 0) { /* need to redefine a centroid */
|
230
234
|
size_t cj;
|
231
235
|
for (cj = 0; 1; cj = (cj + 1) % k) {
|
232
236
|
/* probability to pick this cluster for split */
|
233
|
-
float p = (hassign[cj] - 1.0) / (float)
|
234
|
-
float r = rng.rand_float
|
237
|
+
float p = (hassign[cj] - 1.0) / (float)(n - k);
|
238
|
+
float r = rng.rand_float();
|
235
239
|
if (r < p) {
|
236
240
|
break; /* found our cluster to be split */
|
237
241
|
}
|
238
242
|
}
|
239
|
-
memcpy
|
243
|
+
memcpy(centroids + ci * d,
|
244
|
+
centroids + cj * d,
|
245
|
+
sizeof(*centroids) * d);
|
240
246
|
|
241
247
|
/* small symmetric pertubation */
|
242
248
|
for (size_t j = 0; j < d; j++) {
|
@@ -257,30 +263,35 @@ int split_clusters (size_t d, size_t k, size_t n,
|
|
257
263
|
}
|
258
264
|
|
259
265
|
return nsplit;
|
260
|
-
|
261
266
|
}
|
262
267
|
|
263
|
-
|
264
|
-
|
265
|
-
|
266
|
-
|
267
|
-
|
268
|
-
|
269
|
-
|
270
|
-
|
271
|
-
|
272
|
-
|
273
|
-
|
274
|
-
|
275
|
-
|
276
|
-
|
277
|
-
|
278
|
-
|
279
|
-
|
280
|
-
|
281
|
-
|
268
|
+
}; // namespace
|
269
|
+
|
270
|
+
void Clustering::train_encoded(
|
271
|
+
idx_t nx,
|
272
|
+
const uint8_t* x_in,
|
273
|
+
const Index* codec,
|
274
|
+
Index& index,
|
275
|
+
const float* weights) {
|
276
|
+
FAISS_THROW_IF_NOT_FMT(
|
277
|
+
nx >= k,
|
278
|
+
"Number of training points (%" PRId64
|
279
|
+
") should be at least "
|
280
|
+
"as large as number of clusters (%zd)",
|
281
|
+
nx,
|
282
|
+
k);
|
283
|
+
|
284
|
+
FAISS_THROW_IF_NOT_FMT(
|
285
|
+
(!codec || codec->d == d),
|
286
|
+
"Codec dimension %d not the same as data dimension %d",
|
287
|
+
int(codec->d),
|
288
|
+
int(d));
|
289
|
+
|
290
|
+
FAISS_THROW_IF_NOT_FMT(
|
291
|
+
index.d == d,
|
282
292
|
"Index dimension %d not the same as data dimension %d",
|
283
|
-
int(index.d),
|
293
|
+
int(index.d),
|
294
|
+
int(d));
|
284
295
|
|
285
296
|
double t0 = getmillisecs();
|
286
297
|
|
@@ -288,67 +299,78 @@ void Clustering::train_encoded (idx_t nx, const uint8_t *x_in,
|
|
288
299
|
// Check for NaNs in input data. Normally it is the user's
|
289
300
|
// responsibility, but it may spare us some hard-to-debug
|
290
301
|
// reports.
|
291
|
-
const float
|
302
|
+
const float* x = reinterpret_cast<const float*>(x_in);
|
292
303
|
for (size_t i = 0; i < nx * d; i++) {
|
293
|
-
FAISS_THROW_IF_NOT_MSG
|
294
|
-
|
304
|
+
FAISS_THROW_IF_NOT_MSG(
|
305
|
+
std::isfinite(x[i]), "input contains NaN's or Inf's");
|
295
306
|
}
|
296
307
|
}
|
297
308
|
|
298
|
-
const uint8_t
|
299
|
-
std::unique_ptr<uint8_t
|
300
|
-
std::unique_ptr<float
|
309
|
+
const uint8_t* x = x_in;
|
310
|
+
std::unique_ptr<uint8_t[]> del1;
|
311
|
+
std::unique_ptr<float[]> del3;
|
301
312
|
size_t line_size = codec ? codec->sa_code_size() : sizeof(float) * d;
|
302
313
|
|
303
314
|
if (nx > k * max_points_per_centroid) {
|
304
|
-
uint8_t
|
305
|
-
float
|
306
|
-
nx = subsample_training_set
|
307
|
-
|
308
|
-
del1.reset
|
309
|
-
|
315
|
+
uint8_t* x_new;
|
316
|
+
float* weights_new;
|
317
|
+
nx = subsample_training_set(
|
318
|
+
*this, nx, x, line_size, weights, &x_new, &weights_new);
|
319
|
+
del1.reset(x_new);
|
320
|
+
x = x_new;
|
321
|
+
del3.reset(weights_new);
|
322
|
+
weights = weights_new;
|
310
323
|
} else if (nx < k * min_points_per_centroid) {
|
311
|
-
fprintf
|
312
|
-
|
313
|
-
|
314
|
-
|
324
|
+
fprintf(stderr,
|
325
|
+
"WARNING clustering %" PRId64
|
326
|
+
" points to %zd centroids: "
|
327
|
+
"please provide at least %" PRId64 " training points\n",
|
328
|
+
nx,
|
329
|
+
k,
|
330
|
+
idx_t(k) * min_points_per_centroid);
|
315
331
|
}
|
316
332
|
|
317
333
|
if (nx == k) {
|
318
334
|
// this is a corner case, just copy training set to clusters
|
319
335
|
if (verbose) {
|
320
|
-
printf("Number of training points (%" PRId64
|
321
|
-
"
|
336
|
+
printf("Number of training points (%" PRId64
|
337
|
+
") same as number of "
|
338
|
+
"clusters, just copying\n",
|
339
|
+
nx);
|
322
340
|
}
|
323
|
-
centroids.resize
|
341
|
+
centroids.resize(d * k);
|
324
342
|
if (!codec) {
|
325
|
-
memcpy
|
343
|
+
memcpy(centroids.data(), x_in, sizeof(float) * d * k);
|
326
344
|
} else {
|
327
|
-
codec->sa_decode
|
345
|
+
codec->sa_decode(nx, x_in, centroids.data());
|
328
346
|
}
|
329
347
|
|
330
348
|
// one fake iteration...
|
331
|
-
ClusteringIterationStats stats = {
|
332
|
-
iteration_stats.push_back
|
349
|
+
ClusteringIterationStats stats = {0.0, 0.0, 0.0, 1.0, 0};
|
350
|
+
iteration_stats.push_back(stats);
|
333
351
|
|
334
352
|
index.reset();
|
335
353
|
index.add(k, centroids.data());
|
336
354
|
return;
|
337
355
|
}
|
338
356
|
|
339
|
-
|
340
357
|
if (verbose) {
|
341
|
-
printf("Clustering %" PRId64
|
358
|
+
printf("Clustering %" PRId64
|
359
|
+
" points in %zdD to %zd clusters, "
|
342
360
|
"redo %d times, %d iterations\n",
|
343
|
-
nx,
|
361
|
+
nx,
|
362
|
+
d,
|
363
|
+
k,
|
364
|
+
nredo,
|
365
|
+
niter);
|
344
366
|
if (codec) {
|
345
367
|
printf("Input data encoded in %zd bytes per vector\n",
|
346
|
-
codec->sa_code_size
|
368
|
+
codec->sa_code_size());
|
347
369
|
}
|
348
370
|
}
|
349
371
|
|
350
|
-
std::unique_ptr<idx_t
|
351
|
-
std::unique_ptr<float
|
372
|
+
std::unique_ptr<idx_t[]> assign(new idx_t[nx]);
|
373
|
+
std::unique_ptr<float[]> dis(new float[nx]);
|
352
374
|
|
353
375
|
// remember best iteration for redo
|
354
376
|
bool lower_is_better = index.metric_type != METRIC_INNER_PRODUCT;
|
@@ -358,52 +380,49 @@ void Clustering::train_encoded (idx_t nx, const uint8_t *x_in,
|
|
358
380
|
|
359
381
|
// support input centroids
|
360
382
|
|
361
|
-
FAISS_THROW_IF_NOT_MSG
|
362
|
-
|
363
|
-
|
364
|
-
);
|
383
|
+
FAISS_THROW_IF_NOT_MSG(
|
384
|
+
centroids.size() % d == 0,
|
385
|
+
"size of provided input centroids not a multiple of dimension");
|
365
386
|
|
366
387
|
size_t n_input_centroids = centroids.size() / d;
|
367
388
|
|
368
389
|
if (verbose && n_input_centroids > 0) {
|
369
|
-
printf
|
370
|
-
|
390
|
+
printf(" Using %zd centroids provided as input (%sfrozen)\n",
|
391
|
+
n_input_centroids,
|
392
|
+
frozen_centroids ? "" : "not ");
|
371
393
|
}
|
372
394
|
|
373
395
|
double t_search_tot = 0;
|
374
396
|
if (verbose) {
|
375
|
-
printf(" Preprocessing in %.2f s\n",
|
376
|
-
(getmillisecs() - t0) / 1000.);
|
397
|
+
printf(" Preprocessing in %.2f s\n", (getmillisecs() - t0) / 1000.);
|
377
398
|
}
|
378
399
|
t0 = getmillisecs();
|
379
400
|
|
380
401
|
// temporary buffer to decode vectors during the optimization
|
381
|
-
std::vector<float> decode_buffer
|
382
|
-
(codec ? d * decode_block_size : 0);
|
402
|
+
std::vector<float> decode_buffer(codec ? d * decode_block_size : 0);
|
383
403
|
|
384
404
|
for (int redo = 0; redo < nredo; redo++) {
|
385
|
-
|
386
405
|
if (verbose && nredo > 1) {
|
387
406
|
printf("Outer iteration %d / %d\n", redo, nredo);
|
388
407
|
}
|
389
408
|
|
390
409
|
// initialize (remaining) centroids with random points from the dataset
|
391
|
-
centroids.resize
|
392
|
-
std::vector<int> perm
|
410
|
+
centroids.resize(d * k);
|
411
|
+
std::vector<int> perm(nx);
|
393
412
|
|
394
|
-
rand_perm
|
413
|
+
rand_perm(perm.data(), nx, seed + 1 + redo * 15486557L);
|
395
414
|
|
396
415
|
if (!codec) {
|
397
|
-
for (int i = n_input_centroids; i < k
|
398
|
-
memcpy
|
416
|
+
for (int i = n_input_centroids; i < k; i++) {
|
417
|
+
memcpy(¢roids[i * d], x + perm[i] * line_size, line_size);
|
399
418
|
}
|
400
419
|
} else {
|
401
|
-
for (int i = n_input_centroids; i < k
|
402
|
-
codec->sa_decode
|
420
|
+
for (int i = n_input_centroids; i < k; i++) {
|
421
|
+
codec->sa_decode(1, x + perm[i] * line_size, ¢roids[i * d]);
|
403
422
|
}
|
404
423
|
}
|
405
424
|
|
406
|
-
post_process_centroids
|
425
|
+
post_process_centroids();
|
407
426
|
|
408
427
|
// prepare the index
|
409
428
|
|
@@ -412,10 +431,10 @@ void Clustering::train_encoded (idx_t nx, const uint8_t *x_in,
|
|
412
431
|
}
|
413
432
|
|
414
433
|
if (!index.is_trained) {
|
415
|
-
index.train
|
434
|
+
index.train(k, centroids.data());
|
416
435
|
}
|
417
436
|
|
418
|
-
index.add
|
437
|
+
index.add(k, centroids.data());
|
419
438
|
|
420
439
|
// k-means iterations
|
421
440
|
|
@@ -424,18 +443,28 @@ void Clustering::train_encoded (idx_t nx, const uint8_t *x_in,
|
|
424
443
|
double t0s = getmillisecs();
|
425
444
|
|
426
445
|
if (!codec) {
|
427
|
-
index.search
|
428
|
-
|
446
|
+
index.search(
|
447
|
+
nx,
|
448
|
+
reinterpret_cast<const float*>(x),
|
449
|
+
1,
|
450
|
+
dis.get(),
|
451
|
+
assign.get());
|
429
452
|
} else {
|
430
453
|
// search by blocks of decode_block_size vectors
|
431
|
-
size_t code_size = codec->sa_code_size
|
454
|
+
size_t code_size = codec->sa_code_size();
|
432
455
|
for (size_t i0 = 0; i0 < nx; i0 += decode_block_size) {
|
433
456
|
size_t i1 = i0 + decode_block_size;
|
434
|
-
if (i1 > nx) {
|
435
|
-
|
436
|
-
|
437
|
-
|
438
|
-
|
457
|
+
if (i1 > nx) {
|
458
|
+
i1 = nx;
|
459
|
+
}
|
460
|
+
codec->sa_decode(
|
461
|
+
i1 - i0, x + code_size * i0, decode_buffer.data());
|
462
|
+
index.search(
|
463
|
+
i1 - i0,
|
464
|
+
decode_buffer.data(),
|
465
|
+
1,
|
466
|
+
dis.get() + i0,
|
467
|
+
assign.get() + i0);
|
439
468
|
}
|
440
469
|
}
|
441
470
|
|
@@ -449,61 +478,71 @@ void Clustering::train_encoded (idx_t nx, const uint8_t *x_in,
|
|
449
478
|
}
|
450
479
|
|
451
480
|
// update the centroids
|
452
|
-
std::vector<float> hassign
|
481
|
+
std::vector<float> hassign(k);
|
453
482
|
|
454
483
|
size_t k_frozen = frozen_centroids ? n_input_centroids : 0;
|
455
|
-
compute_centroids
|
456
|
-
|
457
|
-
|
458
|
-
|
459
|
-
|
460
|
-
|
461
|
-
|
462
|
-
|
463
|
-
|
464
|
-
|
484
|
+
compute_centroids(
|
485
|
+
d,
|
486
|
+
k,
|
487
|
+
nx,
|
488
|
+
k_frozen,
|
489
|
+
x,
|
490
|
+
codec,
|
491
|
+
assign.get(),
|
492
|
+
weights,
|
493
|
+
hassign.data(),
|
494
|
+
centroids.data());
|
495
|
+
|
496
|
+
int nsplit = split_clusters(
|
497
|
+
d, k, nx, k_frozen, hassign.data(), centroids.data());
|
465
498
|
|
466
499
|
// collect statistics
|
467
|
-
ClusteringIterationStats stats =
|
468
|
-
|
469
|
-
|
470
|
-
|
471
|
-
|
500
|
+
ClusteringIterationStats stats = {
|
501
|
+
obj,
|
502
|
+
(getmillisecs() - t0) / 1000.0,
|
503
|
+
t_search_tot / 1000,
|
504
|
+
imbalance_factor(nx, k, assign.get()),
|
505
|
+
nsplit};
|
472
506
|
iteration_stats.push_back(stats);
|
473
507
|
|
474
508
|
if (verbose) {
|
475
|
-
printf
|
476
|
-
|
477
|
-
|
478
|
-
|
479
|
-
|
509
|
+
printf(" Iteration %d (%.2f s, search %.2f s): "
|
510
|
+
"objective=%g imbalance=%.3f nsplit=%d \r",
|
511
|
+
i,
|
512
|
+
stats.time,
|
513
|
+
stats.time_search,
|
514
|
+
stats.obj,
|
515
|
+
stats.imbalance_factor,
|
516
|
+
nsplit);
|
517
|
+
fflush(stdout);
|
480
518
|
}
|
481
519
|
|
482
|
-
post_process_centroids
|
520
|
+
post_process_centroids();
|
483
521
|
|
484
522
|
// add centroids to index for the next iteration (or for output)
|
485
523
|
|
486
|
-
index.reset
|
524
|
+
index.reset();
|
487
525
|
if (update_index) {
|
488
|
-
index.train
|
526
|
+
index.train(k, centroids.data());
|
489
527
|
}
|
490
528
|
|
491
|
-
index.add
|
492
|
-
InterruptCallback::check
|
529
|
+
index.add(k, centroids.data());
|
530
|
+
InterruptCallback::check();
|
493
531
|
}
|
494
532
|
|
495
|
-
if (verbose)
|
533
|
+
if (verbose)
|
534
|
+
printf("\n");
|
496
535
|
if (nredo > 1) {
|
497
536
|
if ((lower_is_better && obj < best_obj) ||
|
498
537
|
(!lower_is_better && obj > best_obj)) {
|
499
538
|
if (verbose) {
|
500
|
-
printf
|
539
|
+
printf("Objective improved: keep new clusters\n");
|
501
540
|
}
|
502
541
|
best_centroids = centroids;
|
503
542
|
best_iteration_stats = iteration_stats;
|
504
543
|
best_obj = obj;
|
505
544
|
}
|
506
|
-
index.reset
|
545
|
+
index.reset();
|
507
546
|
}
|
508
547
|
}
|
509
548
|
if (nredo > 1) {
|
@@ -512,20 +551,120 @@ void Clustering::train_encoded (idx_t nx, const uint8_t *x_in,
|
|
512
551
|
index.reset();
|
513
552
|
index.add(k, best_centroids.data());
|
514
553
|
}
|
515
|
-
|
516
554
|
}
|
517
555
|
|
518
|
-
float kmeans_clustering
|
519
|
-
|
520
|
-
|
521
|
-
|
522
|
-
|
556
|
+
float kmeans_clustering(
|
557
|
+
size_t d,
|
558
|
+
size_t n,
|
559
|
+
size_t k,
|
560
|
+
const float* x,
|
561
|
+
float* centroids) {
|
562
|
+
Clustering clus(d, k);
|
523
563
|
clus.verbose = d * n * k > (1L << 30);
|
524
564
|
// display logs if > 1Gflop per iteration
|
525
|
-
IndexFlatL2 index
|
526
|
-
clus.train
|
565
|
+
IndexFlatL2 index(d);
|
566
|
+
clus.train(n, x, index);
|
527
567
|
memcpy(centroids, clus.centroids.data(), sizeof(*centroids) * d * k);
|
528
568
|
return clus.iteration_stats.back().obj;
|
529
569
|
}
|
530
570
|
|
571
|
+
/******************************************************************************
|
572
|
+
* ProgressiveDimClustering implementation
|
573
|
+
******************************************************************************/
|
574
|
+
|
575
|
+
ProgressiveDimClusteringParameters::ProgressiveDimClusteringParameters() {
|
576
|
+
progressive_dim_steps = 10;
|
577
|
+
apply_pca = true; // seems a good idea to do this by default
|
578
|
+
niter = 10; // reduce nb of iterations per step
|
579
|
+
}
|
580
|
+
|
581
|
+
Index* ProgressiveDimIndexFactory::operator()(int dim) {
|
582
|
+
return new IndexFlatL2(dim);
|
583
|
+
}
|
584
|
+
|
585
|
+
ProgressiveDimClustering::ProgressiveDimClustering(int d, int k) : d(d), k(k) {}
|
586
|
+
|
587
|
+
ProgressiveDimClustering::ProgressiveDimClustering(
|
588
|
+
int d,
|
589
|
+
int k,
|
590
|
+
const ProgressiveDimClusteringParameters& cp)
|
591
|
+
: ProgressiveDimClusteringParameters(cp), d(d), k(k) {}
|
592
|
+
|
593
|
+
namespace {
|
594
|
+
|
595
|
+
using idx_t = Index::idx_t;
|
596
|
+
|
597
|
+
void copy_columns(idx_t n, idx_t d1, const float* src, idx_t d2, float* dest) {
|
598
|
+
idx_t d = std::min(d1, d2);
|
599
|
+
for (idx_t i = 0; i < n; i++) {
|
600
|
+
memcpy(dest, src, sizeof(float) * d);
|
601
|
+
src += d1;
|
602
|
+
dest += d2;
|
603
|
+
}
|
604
|
+
}
|
605
|
+
|
606
|
+
}; // namespace
|
607
|
+
|
608
|
+
void ProgressiveDimClustering::train(
|
609
|
+
idx_t n,
|
610
|
+
const float* x,
|
611
|
+
ProgressiveDimIndexFactory& factory) {
|
612
|
+
int d_prev = 0;
|
613
|
+
|
614
|
+
PCAMatrix pca(d, d);
|
615
|
+
|
616
|
+
std::vector<float> xbuf;
|
617
|
+
if (apply_pca) {
|
618
|
+
if (verbose) {
|
619
|
+
printf("Training PCA transform\n");
|
620
|
+
}
|
621
|
+
pca.train(n, x);
|
622
|
+
if (verbose) {
|
623
|
+
printf("Apply PCA\n");
|
624
|
+
}
|
625
|
+
xbuf.resize(n * d);
|
626
|
+
pca.apply_noalloc(n, x, xbuf.data());
|
627
|
+
x = xbuf.data();
|
628
|
+
}
|
629
|
+
|
630
|
+
for (int iter = 0; iter < progressive_dim_steps; iter++) {
|
631
|
+
int di = int(pow(d, (1. + iter) / progressive_dim_steps));
|
632
|
+
if (verbose) {
|
633
|
+
printf("Progressive dim step %d: cluster in dimension %d\n",
|
634
|
+
iter,
|
635
|
+
di);
|
636
|
+
}
|
637
|
+
std::unique_ptr<Index> clustering_index(factory(di));
|
638
|
+
|
639
|
+
Clustering clus(di, k, *this);
|
640
|
+
if (d_prev > 0) {
|
641
|
+
// copy warm-start centroids (padded with 0s)
|
642
|
+
clus.centroids.resize(k * di);
|
643
|
+
copy_columns(
|
644
|
+
k, d_prev, centroids.data(), di, clus.centroids.data());
|
645
|
+
}
|
646
|
+
std::vector<float> xsub(n * di);
|
647
|
+
copy_columns(n, d, x, di, xsub.data());
|
648
|
+
|
649
|
+
clus.train(n, xsub.data(), *clustering_index.get());
|
650
|
+
|
651
|
+
centroids = clus.centroids;
|
652
|
+
iteration_stats.insert(
|
653
|
+
iteration_stats.end(),
|
654
|
+
clus.iteration_stats.begin(),
|
655
|
+
clus.iteration_stats.end());
|
656
|
+
|
657
|
+
d_prev = di;
|
658
|
+
}
|
659
|
+
|
660
|
+
if (apply_pca) {
|
661
|
+
if (verbose) {
|
662
|
+
printf("Revert PCA transform on centroids\n");
|
663
|
+
}
|
664
|
+
std::vector<float> cent_transformed(d * k);
|
665
|
+
pca.reverse_transform(k, centroids.data(), cent_transformed.data());
|
666
|
+
cent_transformed.swap(centroids);
|
667
|
+
}
|
668
|
+
}
|
669
|
+
|
531
670
|
} // namespace faiss
|