faiss 0.1.7 → 0.2.3
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +18 -0
- data/README.md +7 -7
- data/ext/faiss/ext.cpp +1 -1
- data/ext/faiss/extconf.rb +8 -2
- data/ext/faiss/index.cpp +102 -69
- data/ext/faiss/index_binary.cpp +24 -30
- data/ext/faiss/kmeans.cpp +20 -16
- data/ext/faiss/numo.hpp +867 -0
- data/ext/faiss/pca_matrix.cpp +13 -14
- data/ext/faiss/product_quantizer.cpp +23 -24
- data/ext/faiss/utils.cpp +10 -37
- data/ext/faiss/utils.h +2 -13
- data/lib/faiss/version.rb +1 -1
- data/lib/faiss.rb +0 -5
- 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 +26 -12
- 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
@@ -0,0 +1,172 @@
|
|
1
|
+
/**
|
2
|
+
* Copyright (c) Facebook, Inc. and its affiliates.
|
3
|
+
*
|
4
|
+
* This source code is licensed under the MIT license found in the
|
5
|
+
* LICENSE file in the root directory of this source tree.
|
6
|
+
*/
|
7
|
+
|
8
|
+
#pragma once
|
9
|
+
|
10
|
+
#include <stdint.h>
|
11
|
+
|
12
|
+
#include <random>
|
13
|
+
#include <string>
|
14
|
+
#include <unordered_map>
|
15
|
+
#include <vector>
|
16
|
+
|
17
|
+
#include <faiss/impl/AdditiveQuantizer.h>
|
18
|
+
#include <faiss/utils/utils.h>
|
19
|
+
|
20
|
+
namespace faiss {
|
21
|
+
|
22
|
+
/** Implementation of LSQ/LSQ++ described in the following two papers:
|
23
|
+
*
|
24
|
+
* Revisiting additive quantization
|
25
|
+
* Julieta Martinez, et al. ECCV 2016
|
26
|
+
*
|
27
|
+
* LSQ++: Lower running time and higher recall in multi-codebook quantization
|
28
|
+
* Julieta Martinez, et al. ECCV 2018
|
29
|
+
*
|
30
|
+
* This implementation is mostly translated from the Julia implementations
|
31
|
+
* by Julieta Martinez:
|
32
|
+
* (https://github.com/una-dinosauria/local-search-quantization,
|
33
|
+
* https://github.com/una-dinosauria/Rayuela.jl)
|
34
|
+
*
|
35
|
+
* The trained codes are stored in `codebooks` which is called
|
36
|
+
* `centroids` in PQ and RQ.
|
37
|
+
*/
|
38
|
+
|
39
|
+
struct LocalSearchQuantizer : AdditiveQuantizer {
|
40
|
+
size_t K; ///< number of codes per codebook
|
41
|
+
|
42
|
+
size_t train_iters; ///< number of iterations in training
|
43
|
+
|
44
|
+
size_t encode_ils_iters; ///< iterations of local search in encoding
|
45
|
+
size_t train_ils_iters; ///< iterations of local search in training
|
46
|
+
size_t icm_iters; ///< number of iterations in icm
|
47
|
+
|
48
|
+
float p; ///< temperature factor
|
49
|
+
float lambd; ///< regularization factor
|
50
|
+
|
51
|
+
size_t chunk_size; ///< nb of vectors to encode at a time
|
52
|
+
|
53
|
+
int random_seed; ///< seed for random generator
|
54
|
+
size_t nperts; ///< number of perturbation in each code
|
55
|
+
|
56
|
+
LocalSearchQuantizer(
|
57
|
+
size_t d, /* dimensionality of the input vectors */
|
58
|
+
size_t M, /* number of subquantizers */
|
59
|
+
size_t nbits); /* number of bit per subvector index */
|
60
|
+
|
61
|
+
// Train the local search quantizer
|
62
|
+
void train(size_t n, const float* x) override;
|
63
|
+
|
64
|
+
/** Encode a set of vectors
|
65
|
+
*
|
66
|
+
* @param x vectors to encode, size n * d
|
67
|
+
* @param codes output codes, size n * code_size
|
68
|
+
*/
|
69
|
+
void compute_codes(const float* x, uint8_t* codes, size_t n) const override;
|
70
|
+
|
71
|
+
/** Update codebooks given encodings
|
72
|
+
*
|
73
|
+
* @param x training vectors, size n * d
|
74
|
+
* @param codes encoded training vectors, size n * M
|
75
|
+
*/
|
76
|
+
void update_codebooks(const float* x, const int32_t* codes, size_t n);
|
77
|
+
|
78
|
+
/** Encode vectors given codebooks using iterative conditional mode (icm).
|
79
|
+
*
|
80
|
+
* @param x vectors to encode, size n * d
|
81
|
+
* @param codes output codes, size n * M
|
82
|
+
* @param ils_iters number of iterations of iterative local search
|
83
|
+
*/
|
84
|
+
void icm_encode(
|
85
|
+
const float* x,
|
86
|
+
int32_t* codes,
|
87
|
+
size_t n,
|
88
|
+
size_t ils_iters,
|
89
|
+
std::mt19937& gen) const;
|
90
|
+
|
91
|
+
void icm_encode_partial(
|
92
|
+
size_t index,
|
93
|
+
const float* x,
|
94
|
+
int32_t* codes,
|
95
|
+
size_t n,
|
96
|
+
const float* binaries,
|
97
|
+
size_t ils_iters,
|
98
|
+
std::mt19937& gen) const;
|
99
|
+
|
100
|
+
void icm_encode_step(
|
101
|
+
const float* unaries,
|
102
|
+
const float* binaries,
|
103
|
+
int32_t* codes,
|
104
|
+
size_t n) const;
|
105
|
+
|
106
|
+
/** Add some perturbation to codebooks
|
107
|
+
*
|
108
|
+
* @param T temperature of simulated annealing
|
109
|
+
* @param stddev standard derivations of each dimension in training data
|
110
|
+
*/
|
111
|
+
void perturb_codebooks(
|
112
|
+
float T,
|
113
|
+
const std::vector<float>& stddev,
|
114
|
+
std::mt19937& gen);
|
115
|
+
|
116
|
+
/** Add some perturbation to codes
|
117
|
+
*
|
118
|
+
* @param codes codes to be perturbed, size n * M
|
119
|
+
*/
|
120
|
+
void perturb_codes(int32_t* codes, size_t n, std::mt19937& gen) const;
|
121
|
+
|
122
|
+
/** Compute binary terms
|
123
|
+
*
|
124
|
+
* @param binaries binary terms, size M * M * K * K
|
125
|
+
*/
|
126
|
+
void compute_binary_terms(float* binaries) const;
|
127
|
+
|
128
|
+
/** Compute unary terms
|
129
|
+
*
|
130
|
+
* @param x vectors to encode, size n * d
|
131
|
+
* @param unaries unary terms, size n * M * K
|
132
|
+
*/
|
133
|
+
void compute_unary_terms(const float* x, float* unaries, size_t n) const;
|
134
|
+
|
135
|
+
/** Helper function to compute reconstruction error
|
136
|
+
*
|
137
|
+
* @param x vectors to encode, size n * d
|
138
|
+
* @param codes encoded codes, size n * M
|
139
|
+
* @param objs if it is not null, store reconstruction
|
140
|
+
error of each vector into it, size n
|
141
|
+
*/
|
142
|
+
float evaluate(
|
143
|
+
const int32_t* codes,
|
144
|
+
const float* x,
|
145
|
+
size_t n,
|
146
|
+
float* objs = nullptr) const;
|
147
|
+
};
|
148
|
+
|
149
|
+
/** A helper struct to count consuming time during training.
|
150
|
+
* It is NOT thread-safe.
|
151
|
+
*/
|
152
|
+
struct LSQTimer {
|
153
|
+
std::unordered_map<std::string, double> duration;
|
154
|
+
std::unordered_map<std::string, double> t0;
|
155
|
+
std::unordered_map<std::string, bool> started;
|
156
|
+
|
157
|
+
LSQTimer() {
|
158
|
+
reset();
|
159
|
+
}
|
160
|
+
|
161
|
+
double get(const std::string& name);
|
162
|
+
|
163
|
+
void start(const std::string& name);
|
164
|
+
|
165
|
+
void end(const std::string& name);
|
166
|
+
|
167
|
+
void reset();
|
168
|
+
};
|
169
|
+
|
170
|
+
FAISS_API extern LSQTimer lsq_timer; ///< timer to count consuming time
|
171
|
+
|
172
|
+
} // namespace faiss
|
@@ -0,0 +1,487 @@
|
|
1
|
+
/**
|
2
|
+
* Copyright (c) Facebook, Inc. and its affiliates.
|
3
|
+
*
|
4
|
+
* This source code is licensed under the MIT license found in the
|
5
|
+
* LICENSE file in the root directory of this source tree.
|
6
|
+
*/
|
7
|
+
|
8
|
+
// -*- c++ -*-
|
9
|
+
|
10
|
+
#include <faiss/impl/NNDescent.h>
|
11
|
+
|
12
|
+
#include <mutex>
|
13
|
+
#include <string>
|
14
|
+
|
15
|
+
#include <faiss/impl/AuxIndexStructures.h>
|
16
|
+
|
17
|
+
namespace faiss {
|
18
|
+
|
19
|
+
using LockGuard = std::lock_guard<std::mutex>;
|
20
|
+
|
21
|
+
namespace nndescent {
|
22
|
+
|
23
|
+
void gen_random(std::mt19937& rng, int* addr, const int size, const int N);
|
24
|
+
|
25
|
+
Nhood::Nhood(int l, int s, std::mt19937& rng, int N) {
|
26
|
+
M = s;
|
27
|
+
nn_new.resize(s * 2);
|
28
|
+
gen_random(rng, nn_new.data(), (int)nn_new.size(), N);
|
29
|
+
}
|
30
|
+
|
31
|
+
/// Copy operator
|
32
|
+
Nhood& Nhood::operator=(const Nhood& other) {
|
33
|
+
M = other.M;
|
34
|
+
std::copy(
|
35
|
+
other.nn_new.begin(),
|
36
|
+
other.nn_new.end(),
|
37
|
+
std::back_inserter(nn_new));
|
38
|
+
nn_new.reserve(other.nn_new.capacity());
|
39
|
+
pool.reserve(other.pool.capacity());
|
40
|
+
return *this;
|
41
|
+
}
|
42
|
+
|
43
|
+
/// Copy constructor
|
44
|
+
Nhood::Nhood(const Nhood& other) {
|
45
|
+
M = other.M;
|
46
|
+
std::copy(
|
47
|
+
other.nn_new.begin(),
|
48
|
+
other.nn_new.end(),
|
49
|
+
std::back_inserter(nn_new));
|
50
|
+
nn_new.reserve(other.nn_new.capacity());
|
51
|
+
pool.reserve(other.pool.capacity());
|
52
|
+
}
|
53
|
+
|
54
|
+
/// Insert a point into the candidate pool
|
55
|
+
void Nhood::insert(int id, float dist) {
|
56
|
+
LockGuard guard(lock);
|
57
|
+
if (dist > pool.front().distance)
|
58
|
+
return;
|
59
|
+
for (int i = 0; i < pool.size(); i++) {
|
60
|
+
if (id == pool[i].id)
|
61
|
+
return;
|
62
|
+
}
|
63
|
+
if (pool.size() < pool.capacity()) {
|
64
|
+
pool.push_back(Neighbor(id, dist, true));
|
65
|
+
std::push_heap(pool.begin(), pool.end());
|
66
|
+
} else {
|
67
|
+
std::pop_heap(pool.begin(), pool.end());
|
68
|
+
pool[pool.size() - 1] = Neighbor(id, dist, true);
|
69
|
+
std::push_heap(pool.begin(), pool.end());
|
70
|
+
}
|
71
|
+
}
|
72
|
+
|
73
|
+
/// In local join, two objects are compared only if at least
|
74
|
+
/// one of them is new.
|
75
|
+
template <typename C>
|
76
|
+
void Nhood::join(C callback) const {
|
77
|
+
for (int const i : nn_new) {
|
78
|
+
for (int const j : nn_new) {
|
79
|
+
if (i < j) {
|
80
|
+
callback(i, j);
|
81
|
+
}
|
82
|
+
}
|
83
|
+
for (int j : nn_old) {
|
84
|
+
callback(i, j);
|
85
|
+
}
|
86
|
+
}
|
87
|
+
}
|
88
|
+
|
89
|
+
void gen_random(std::mt19937& rng, int* addr, const int size, const int N) {
|
90
|
+
for (int i = 0; i < size; ++i) {
|
91
|
+
addr[i] = rng() % (N - size);
|
92
|
+
}
|
93
|
+
std::sort(addr, addr + size);
|
94
|
+
for (int i = 1; i < size; ++i) {
|
95
|
+
if (addr[i] <= addr[i - 1]) {
|
96
|
+
addr[i] = addr[i - 1] + 1;
|
97
|
+
}
|
98
|
+
}
|
99
|
+
int off = rng() % N;
|
100
|
+
for (int i = 0; i < size; ++i) {
|
101
|
+
addr[i] = (addr[i] + off) % N;
|
102
|
+
}
|
103
|
+
}
|
104
|
+
|
105
|
+
// Insert a new point into the candidate pool in ascending order
|
106
|
+
int insert_into_pool(Neighbor* addr, int size, Neighbor nn) {
|
107
|
+
// find the location to insert
|
108
|
+
int left = 0, right = size - 1;
|
109
|
+
if (addr[left].distance > nn.distance) {
|
110
|
+
memmove((char*)&addr[left + 1], &addr[left], size * sizeof(Neighbor));
|
111
|
+
addr[left] = nn;
|
112
|
+
return left;
|
113
|
+
}
|
114
|
+
if (addr[right].distance < nn.distance) {
|
115
|
+
addr[size] = nn;
|
116
|
+
return size;
|
117
|
+
}
|
118
|
+
while (left < right - 1) {
|
119
|
+
int mid = (left + right) / 2;
|
120
|
+
if (addr[mid].distance > nn.distance)
|
121
|
+
right = mid;
|
122
|
+
else
|
123
|
+
left = mid;
|
124
|
+
}
|
125
|
+
// check equal ID
|
126
|
+
|
127
|
+
while (left > 0) {
|
128
|
+
if (addr[left].distance < nn.distance)
|
129
|
+
break;
|
130
|
+
if (addr[left].id == nn.id)
|
131
|
+
return size + 1;
|
132
|
+
left--;
|
133
|
+
}
|
134
|
+
if (addr[left].id == nn.id || addr[right].id == nn.id)
|
135
|
+
return size + 1;
|
136
|
+
memmove((char*)&addr[right + 1],
|
137
|
+
&addr[right],
|
138
|
+
(size - right) * sizeof(Neighbor));
|
139
|
+
addr[right] = nn;
|
140
|
+
return right;
|
141
|
+
}
|
142
|
+
|
143
|
+
} // namespace nndescent
|
144
|
+
|
145
|
+
using namespace nndescent;
|
146
|
+
|
147
|
+
constexpr int NUM_EVAL_POINTS = 100;
|
148
|
+
|
149
|
+
NNDescent::NNDescent(const int d, const int K) : K(K), random_seed(2021), d(d) {
|
150
|
+
ntotal = 0;
|
151
|
+
has_built = false;
|
152
|
+
S = 10;
|
153
|
+
R = 100;
|
154
|
+
L = K + 50;
|
155
|
+
iter = 10;
|
156
|
+
search_L = 0;
|
157
|
+
}
|
158
|
+
|
159
|
+
NNDescent::~NNDescent() {}
|
160
|
+
|
161
|
+
void NNDescent::join(DistanceComputer& qdis) {
|
162
|
+
#pragma omp parallel for default(shared) schedule(dynamic, 100)
|
163
|
+
for (int n = 0; n < ntotal; n++) {
|
164
|
+
graph[n].join([&](int i, int j) {
|
165
|
+
if (i != j) {
|
166
|
+
float dist = qdis.symmetric_dis(i, j);
|
167
|
+
graph[i].insert(j, dist);
|
168
|
+
graph[j].insert(i, dist);
|
169
|
+
}
|
170
|
+
});
|
171
|
+
}
|
172
|
+
}
|
173
|
+
|
174
|
+
/// Sample neighbors for each node to peform local join later
|
175
|
+
/// Store them in nn_new and nn_old
|
176
|
+
void NNDescent::update() {
|
177
|
+
// Step 1.
|
178
|
+
// Clear all nn_new and nn_old
|
179
|
+
#pragma omp parallel for
|
180
|
+
for (int i = 0; i < ntotal; i++) {
|
181
|
+
std::vector<int>().swap(graph[i].nn_new);
|
182
|
+
std::vector<int>().swap(graph[i].nn_old);
|
183
|
+
}
|
184
|
+
|
185
|
+
// Step 2.
|
186
|
+
// Compute the number of neighbors which is new i.e. flag is true
|
187
|
+
// in the candidate pool. This must not exceed the sample number S.
|
188
|
+
// That means We only select S new neighbors.
|
189
|
+
#pragma omp parallel for
|
190
|
+
for (int n = 0; n < ntotal; ++n) {
|
191
|
+
auto& nn = graph[n];
|
192
|
+
std::sort(nn.pool.begin(), nn.pool.end());
|
193
|
+
|
194
|
+
if (nn.pool.size() > L)
|
195
|
+
nn.pool.resize(L);
|
196
|
+
nn.pool.reserve(L); // keep the pool size be L
|
197
|
+
|
198
|
+
int maxl = std::min(nn.M + S, (int)nn.pool.size());
|
199
|
+
int c = 0;
|
200
|
+
int l = 0;
|
201
|
+
|
202
|
+
while ((l < maxl) && (c < S)) {
|
203
|
+
if (nn.pool[l].flag)
|
204
|
+
++c;
|
205
|
+
++l;
|
206
|
+
}
|
207
|
+
nn.M = l;
|
208
|
+
}
|
209
|
+
|
210
|
+
// Step 3.
|
211
|
+
// Find reverse links for each node
|
212
|
+
// Randomly choose R reverse links.
|
213
|
+
#pragma omp parallel
|
214
|
+
{
|
215
|
+
std::mt19937 rng(random_seed * 5081 + omp_get_thread_num());
|
216
|
+
#pragma omp for
|
217
|
+
for (int n = 0; n < ntotal; ++n) {
|
218
|
+
auto& node = graph[n];
|
219
|
+
auto& nn_new = node.nn_new;
|
220
|
+
auto& nn_old = node.nn_old;
|
221
|
+
|
222
|
+
for (int l = 0; l < node.M; ++l) {
|
223
|
+
auto& nn = node.pool[l];
|
224
|
+
auto& other = graph[nn.id]; // the other side of the edge
|
225
|
+
|
226
|
+
if (nn.flag) { // the node is inserted newly
|
227
|
+
// push the neighbor into nn_new
|
228
|
+
nn_new.push_back(nn.id);
|
229
|
+
// push itself into other.rnn_new if it is not in
|
230
|
+
// the candidate pool of the other side
|
231
|
+
if (nn.distance > other.pool.back().distance) {
|
232
|
+
LockGuard guard(other.lock);
|
233
|
+
if (other.rnn_new.size() < R) {
|
234
|
+
other.rnn_new.push_back(n);
|
235
|
+
} else {
|
236
|
+
int pos = rng() % R;
|
237
|
+
other.rnn_new[pos] = n;
|
238
|
+
}
|
239
|
+
}
|
240
|
+
nn.flag = false;
|
241
|
+
|
242
|
+
} else { // the node is old
|
243
|
+
// push the neighbor into nn_old
|
244
|
+
nn_old.push_back(nn.id);
|
245
|
+
// push itself into other.rnn_old if it is not in
|
246
|
+
// the candidate pool of the other side
|
247
|
+
if (nn.distance > other.pool.back().distance) {
|
248
|
+
LockGuard guard(other.lock);
|
249
|
+
if (other.rnn_old.size() < R) {
|
250
|
+
other.rnn_old.push_back(n);
|
251
|
+
} else {
|
252
|
+
int pos = rng() % R;
|
253
|
+
other.rnn_old[pos] = n;
|
254
|
+
}
|
255
|
+
}
|
256
|
+
}
|
257
|
+
}
|
258
|
+
// make heap to join later (in join() function)
|
259
|
+
std::make_heap(node.pool.begin(), node.pool.end());
|
260
|
+
}
|
261
|
+
}
|
262
|
+
|
263
|
+
// Step 4.
|
264
|
+
// Combine the forward and the reverse links
|
265
|
+
// R = 0 means no reverse links are used.
|
266
|
+
#pragma omp parallel for
|
267
|
+
for (int i = 0; i < ntotal; ++i) {
|
268
|
+
auto& nn_new = graph[i].nn_new;
|
269
|
+
auto& nn_old = graph[i].nn_old;
|
270
|
+
auto& rnn_new = graph[i].rnn_new;
|
271
|
+
auto& rnn_old = graph[i].rnn_old;
|
272
|
+
|
273
|
+
nn_new.insert(nn_new.end(), rnn_new.begin(), rnn_new.end());
|
274
|
+
nn_old.insert(nn_old.end(), rnn_old.begin(), rnn_old.end());
|
275
|
+
if (nn_old.size() > R * 2) {
|
276
|
+
nn_old.resize(R * 2);
|
277
|
+
nn_old.reserve(R * 2);
|
278
|
+
}
|
279
|
+
|
280
|
+
std::vector<int>().swap(graph[i].rnn_new);
|
281
|
+
std::vector<int>().swap(graph[i].rnn_old);
|
282
|
+
}
|
283
|
+
}
|
284
|
+
|
285
|
+
void NNDescent::nndescent(DistanceComputer& qdis, bool verbose) {
|
286
|
+
int num_eval_points = std::min(NUM_EVAL_POINTS, ntotal);
|
287
|
+
std::vector<int> eval_points(num_eval_points);
|
288
|
+
std::vector<std::vector<int>> acc_eval_set(num_eval_points);
|
289
|
+
std::mt19937 rng(random_seed * 6577 + omp_get_thread_num());
|
290
|
+
gen_random(rng, eval_points.data(), eval_points.size(), ntotal);
|
291
|
+
generate_eval_set(qdis, eval_points, acc_eval_set, ntotal);
|
292
|
+
for (int it = 0; it < iter; it++) {
|
293
|
+
join(qdis);
|
294
|
+
update();
|
295
|
+
|
296
|
+
if (verbose) {
|
297
|
+
float recall = eval_recall(eval_points, acc_eval_set);
|
298
|
+
printf("Iter: %d, recall@%d: %lf\n", it, K, recall);
|
299
|
+
}
|
300
|
+
}
|
301
|
+
}
|
302
|
+
|
303
|
+
/// Sample a small number of points to evaluate the quality of KNNG built
|
304
|
+
void NNDescent::generate_eval_set(
|
305
|
+
DistanceComputer& qdis,
|
306
|
+
std::vector<int>& c,
|
307
|
+
std::vector<std::vector<int>>& v,
|
308
|
+
int N) {
|
309
|
+
#pragma omp parallel for
|
310
|
+
for (int i = 0; i < c.size(); i++) {
|
311
|
+
std::vector<Neighbor> tmp;
|
312
|
+
for (int j = 0; j < N; j++) {
|
313
|
+
if (i == j)
|
314
|
+
continue; // skip itself
|
315
|
+
float dist = qdis.symmetric_dis(c[i], j);
|
316
|
+
tmp.push_back(Neighbor(j, dist, true));
|
317
|
+
}
|
318
|
+
|
319
|
+
std::partial_sort(tmp.begin(), tmp.begin() + K, tmp.end());
|
320
|
+
for (int j = 0; j < K; j++) {
|
321
|
+
v[i].push_back(tmp[j].id);
|
322
|
+
}
|
323
|
+
}
|
324
|
+
}
|
325
|
+
|
326
|
+
/// Evaluate the quality of KNNG built
|
327
|
+
float NNDescent::eval_recall(
|
328
|
+
std::vector<int>& eval_points,
|
329
|
+
std::vector<std::vector<int>>& acc_eval_set) {
|
330
|
+
float mean_acc = 0.0f;
|
331
|
+
for (size_t i = 0; i < eval_points.size(); i++) {
|
332
|
+
float acc = 0;
|
333
|
+
std::vector<Neighbor>& g = graph[eval_points[i]].pool;
|
334
|
+
std::vector<int>& v = acc_eval_set[i];
|
335
|
+
for (size_t j = 0; j < g.size(); j++) {
|
336
|
+
for (size_t k = 0; k < v.size(); k++) {
|
337
|
+
if (g[j].id == v[k]) {
|
338
|
+
acc++;
|
339
|
+
break;
|
340
|
+
}
|
341
|
+
}
|
342
|
+
}
|
343
|
+
mean_acc += acc / v.size();
|
344
|
+
}
|
345
|
+
return mean_acc / eval_points.size();
|
346
|
+
}
|
347
|
+
|
348
|
+
/// Initialize the KNN graph randomly
|
349
|
+
void NNDescent::init_graph(DistanceComputer& qdis) {
|
350
|
+
graph.reserve(ntotal);
|
351
|
+
{
|
352
|
+
std::mt19937 rng(random_seed * 6007);
|
353
|
+
for (int i = 0; i < ntotal; i++) {
|
354
|
+
graph.push_back(Nhood(L, S, rng, (int)ntotal));
|
355
|
+
}
|
356
|
+
}
|
357
|
+
#pragma omp parallel
|
358
|
+
{
|
359
|
+
std::mt19937 rng(random_seed * 7741 + omp_get_thread_num());
|
360
|
+
#pragma omp for
|
361
|
+
for (int i = 0; i < ntotal; i++) {
|
362
|
+
std::vector<int> tmp(S);
|
363
|
+
|
364
|
+
gen_random(rng, tmp.data(), S, ntotal);
|
365
|
+
|
366
|
+
for (int j = 0; j < S; j++) {
|
367
|
+
int id = tmp[j];
|
368
|
+
if (id == i)
|
369
|
+
continue;
|
370
|
+
float dist = qdis.symmetric_dis(i, id);
|
371
|
+
|
372
|
+
graph[i].pool.push_back(Neighbor(id, dist, true));
|
373
|
+
}
|
374
|
+
std::make_heap(graph[i].pool.begin(), graph[i].pool.end());
|
375
|
+
graph[i].pool.reserve(L);
|
376
|
+
}
|
377
|
+
}
|
378
|
+
}
|
379
|
+
|
380
|
+
void NNDescent::build(DistanceComputer& qdis, const int n, bool verbose) {
|
381
|
+
FAISS_THROW_IF_NOT_MSG(L >= K, "L should be >= K in NNDescent.build");
|
382
|
+
|
383
|
+
if (verbose) {
|
384
|
+
printf("Parameters: K=%d, S=%d, R=%d, L=%d, iter=%d\n",
|
385
|
+
K,
|
386
|
+
S,
|
387
|
+
R,
|
388
|
+
L,
|
389
|
+
iter);
|
390
|
+
}
|
391
|
+
|
392
|
+
ntotal = n;
|
393
|
+
init_graph(qdis);
|
394
|
+
nndescent(qdis, verbose);
|
395
|
+
|
396
|
+
final_graph.resize(ntotal * K);
|
397
|
+
|
398
|
+
// Store the neighbor link structure into final_graph
|
399
|
+
// Clear the old graph
|
400
|
+
for (int i = 0; i < ntotal; i++) {
|
401
|
+
std::sort(graph[i].pool.begin(), graph[i].pool.end());
|
402
|
+
for (int j = 0; j < K; j++) {
|
403
|
+
FAISS_ASSERT(graph[i].pool[j].id < ntotal);
|
404
|
+
final_graph[i * K + j] = graph[i].pool[j].id;
|
405
|
+
}
|
406
|
+
}
|
407
|
+
std::vector<Nhood>().swap(graph);
|
408
|
+
has_built = true;
|
409
|
+
|
410
|
+
if (verbose) {
|
411
|
+
printf("Addes %d points into the index\n", ntotal);
|
412
|
+
}
|
413
|
+
}
|
414
|
+
|
415
|
+
void NNDescent::search(
|
416
|
+
DistanceComputer& qdis,
|
417
|
+
const int topk,
|
418
|
+
idx_t* indices,
|
419
|
+
float* dists,
|
420
|
+
VisitedTable& vt) const {
|
421
|
+
FAISS_THROW_IF_NOT_MSG(has_built, "The index is not build yet.");
|
422
|
+
int L = std::max(search_L, topk);
|
423
|
+
|
424
|
+
// candidate pool, the K best items is the result.
|
425
|
+
std::vector<Neighbor> retset(L + 1);
|
426
|
+
|
427
|
+
// Randomly choose L points to intialize the candidate pool
|
428
|
+
std::vector<int> init_ids(L);
|
429
|
+
std::mt19937 rng(random_seed);
|
430
|
+
|
431
|
+
gen_random(rng, init_ids.data(), L, ntotal);
|
432
|
+
for (int i = 0; i < L; i++) {
|
433
|
+
int id = init_ids[i];
|
434
|
+
float dist = qdis(id);
|
435
|
+
retset[i] = Neighbor(id, dist, true);
|
436
|
+
}
|
437
|
+
|
438
|
+
// Maintain the candidate pool in ascending order
|
439
|
+
std::sort(retset.begin(), retset.begin() + L);
|
440
|
+
|
441
|
+
int k = 0;
|
442
|
+
|
443
|
+
// Stop until the smallest position updated is >= L
|
444
|
+
while (k < L) {
|
445
|
+
int nk = L;
|
446
|
+
|
447
|
+
if (retset[k].flag) {
|
448
|
+
retset[k].flag = false;
|
449
|
+
int n = retset[k].id;
|
450
|
+
|
451
|
+
for (int m = 0; m < K; ++m) {
|
452
|
+
int id = final_graph[n * K + m];
|
453
|
+
if (vt.get(id))
|
454
|
+
continue;
|
455
|
+
|
456
|
+
vt.set(id);
|
457
|
+
float dist = qdis(id);
|
458
|
+
if (dist >= retset[L - 1].distance)
|
459
|
+
continue;
|
460
|
+
|
461
|
+
Neighbor nn(id, dist, true);
|
462
|
+
int r = insert_into_pool(retset.data(), L, nn);
|
463
|
+
|
464
|
+
if (r < nk)
|
465
|
+
nk = r;
|
466
|
+
}
|
467
|
+
}
|
468
|
+
if (nk <= k)
|
469
|
+
k = nk;
|
470
|
+
else
|
471
|
+
++k;
|
472
|
+
}
|
473
|
+
for (size_t i = 0; i < topk; i++) {
|
474
|
+
indices[i] = retset[i].id;
|
475
|
+
dists[i] = retset[i].distance;
|
476
|
+
}
|
477
|
+
|
478
|
+
vt.advance();
|
479
|
+
}
|
480
|
+
|
481
|
+
void NNDescent::reset() {
|
482
|
+
has_built = false;
|
483
|
+
ntotal = 0;
|
484
|
+
final_graph.resize(0);
|
485
|
+
}
|
486
|
+
|
487
|
+
} // namespace faiss
|