faiss 0.2.0 → 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 +4 -0
- 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 +20 -2
@@ -15,41 +15,39 @@
|
|
15
15
|
|
16
16
|
namespace faiss {
|
17
17
|
|
18
|
-
|
19
18
|
/** Class for the clustering parameters. Can be passed to the
|
20
19
|
* constructor of the Clustering object.
|
21
20
|
*/
|
22
21
|
struct ClusteringParameters {
|
23
|
-
int niter;
|
24
|
-
int nredo;
|
22
|
+
int niter; ///< clustering iterations
|
23
|
+
int nredo; ///< redo clustering this many times and keep best
|
25
24
|
|
26
25
|
bool verbose;
|
27
|
-
bool spherical;
|
28
|
-
bool int_centroids;
|
29
|
-
bool update_index;
|
30
|
-
bool frozen_centroids;
|
26
|
+
bool spherical; ///< do we want normalized centroids?
|
27
|
+
bool int_centroids; ///< round centroids coordinates to integer
|
28
|
+
bool update_index; ///< re-train index after each iteration?
|
29
|
+
bool frozen_centroids; ///< use the centroids provided as input and do not
|
30
|
+
///< change them during iterations
|
31
31
|
|
32
32
|
int min_points_per_centroid; ///< otherwise you get a warning
|
33
|
-
int max_points_per_centroid;
|
33
|
+
int max_points_per_centroid; ///< to limit size of dataset
|
34
34
|
|
35
35
|
int seed; ///< seed for the random number generator
|
36
36
|
|
37
|
-
size_t decode_block_size;
|
37
|
+
size_t decode_block_size; ///< how many vectors at a time to decode
|
38
38
|
|
39
39
|
/// sets reasonable defaults
|
40
|
-
ClusteringParameters
|
40
|
+
ClusteringParameters();
|
41
41
|
};
|
42
42
|
|
43
|
-
|
44
43
|
struct ClusteringIterationStats {
|
45
|
-
float obj;
|
46
|
-
double time;
|
44
|
+
float obj; ///< objective values (sum of distances reported by index)
|
45
|
+
double time; ///< seconds for iteration
|
47
46
|
double time_search; ///< seconds for just search
|
48
47
|
double imbalance_factor; ///< imbalance factor of iteration
|
49
48
|
int nsplit; ///< number of cluster splits
|
50
49
|
};
|
51
50
|
|
52
|
-
|
53
51
|
/** K-means clustering based on assignment - centroid update iterations
|
54
52
|
*
|
55
53
|
* The clustering is based on an Index object that assigns training
|
@@ -62,21 +60,22 @@ struct ClusteringIterationStats {
|
|
62
60
|
* initialization.
|
63
61
|
*
|
64
62
|
*/
|
65
|
-
struct Clustering: ClusteringParameters {
|
63
|
+
struct Clustering : ClusteringParameters {
|
66
64
|
typedef Index::idx_t idx_t;
|
67
|
-
size_t d;
|
68
|
-
size_t k;
|
65
|
+
size_t d; ///< dimension of the vectors
|
66
|
+
size_t k; ///< nb of centroids
|
69
67
|
|
70
68
|
/** centroids (k * d)
|
71
|
-
* if centroids are set on input to train, they will be used as
|
69
|
+
* if centroids are set on input to train, they will be used as
|
70
|
+
* initialization
|
72
71
|
*/
|
73
72
|
std::vector<float> centroids;
|
74
73
|
|
75
74
|
/// stats at every iteration of clustering
|
76
75
|
std::vector<ClusteringIterationStats> iteration_stats;
|
77
76
|
|
78
|
-
Clustering
|
79
|
-
Clustering
|
77
|
+
Clustering(int d, int k);
|
78
|
+
Clustering(int d, int k, const ClusteringParameters& cp);
|
80
79
|
|
81
80
|
/** run k-means training
|
82
81
|
*
|
@@ -84,9 +83,11 @@ struct Clustering: ClusteringParameters {
|
|
84
83
|
* @param index index used for assignment
|
85
84
|
* @param x_weights weight associated to each vector: NULL or size n
|
86
85
|
*/
|
87
|
-
virtual void train
|
88
|
-
|
89
|
-
|
86
|
+
virtual void train(
|
87
|
+
idx_t n,
|
88
|
+
const float* x,
|
89
|
+
faiss::Index& index,
|
90
|
+
const float* x_weights = nullptr);
|
90
91
|
|
91
92
|
/** run with encoded vectors
|
92
93
|
*
|
@@ -96,17 +97,69 @@ struct Clustering: ClusteringParameters {
|
|
96
97
|
* @param codec codec used to decode the vectors (nullptr =
|
97
98
|
* vectors are in fact floats) *
|
98
99
|
*/
|
99
|
-
void train_encoded
|
100
|
-
|
101
|
-
|
100
|
+
void train_encoded(
|
101
|
+
idx_t nx,
|
102
|
+
const uint8_t* x_in,
|
103
|
+
const Index* codec,
|
104
|
+
Index& index,
|
105
|
+
const float* weights = nullptr);
|
102
106
|
|
103
107
|
/// Post-process the centroids after each centroid update.
|
104
108
|
/// includes optional L2 normalization and nearest integer rounding
|
105
|
-
void post_process_centroids
|
109
|
+
void post_process_centroids();
|
106
110
|
|
107
111
|
virtual ~Clustering() {}
|
108
112
|
};
|
109
113
|
|
114
|
+
struct ProgressiveDimClusteringParameters : ClusteringParameters {
|
115
|
+
int progressive_dim_steps; ///< number of incremental steps
|
116
|
+
bool apply_pca; ///< apply PCA on input
|
117
|
+
|
118
|
+
ProgressiveDimClusteringParameters();
|
119
|
+
};
|
120
|
+
|
121
|
+
/** generates an index suitable for clustering when called */
|
122
|
+
struct ProgressiveDimIndexFactory {
|
123
|
+
/// ownership transferred to caller
|
124
|
+
virtual Index* operator()(int dim);
|
125
|
+
|
126
|
+
virtual ~ProgressiveDimIndexFactory() {}
|
127
|
+
};
|
128
|
+
|
129
|
+
/** K-means clustering with progressive dimensions used
|
130
|
+
*
|
131
|
+
* The clustering first happens in dim 1, then with exponentially increasing
|
132
|
+
* dimension until d (I steps). This is typically applied after a PCA
|
133
|
+
* transformation (optional). Reference:
|
134
|
+
*
|
135
|
+
* "Improved Residual Vector Quantization for High-dimensional Approximate
|
136
|
+
* Nearest Neighbor Search"
|
137
|
+
*
|
138
|
+
* Shicong Liu, Hongtao Lu, Junru Shao, AAAI'15
|
139
|
+
*
|
140
|
+
* https://arxiv.org/abs/1509.05195
|
141
|
+
*/
|
142
|
+
struct ProgressiveDimClustering : ProgressiveDimClusteringParameters {
|
143
|
+
using idx_t = Index::idx_t;
|
144
|
+
size_t d; ///< dimension of the vectors
|
145
|
+
size_t k; ///< nb of centroids
|
146
|
+
|
147
|
+
/** centroids (k * d) */
|
148
|
+
std::vector<float> centroids;
|
149
|
+
|
150
|
+
/// stats at every iteration of clustering
|
151
|
+
std::vector<ClusteringIterationStats> iteration_stats;
|
152
|
+
|
153
|
+
ProgressiveDimClustering(int d, int k);
|
154
|
+
ProgressiveDimClustering(
|
155
|
+
int d,
|
156
|
+
int k,
|
157
|
+
const ProgressiveDimClusteringParameters& cp);
|
158
|
+
|
159
|
+
void train(idx_t n, const float* x, ProgressiveDimIndexFactory& factory);
|
160
|
+
|
161
|
+
virtual ~ProgressiveDimClustering() {}
|
162
|
+
};
|
110
163
|
|
111
164
|
/** simplified interface
|
112
165
|
*
|
@@ -117,13 +170,13 @@ struct Clustering: ClusteringParameters {
|
|
117
170
|
* @param centroids output centroids (size k * d)
|
118
171
|
* @return final quantization error
|
119
172
|
*/
|
120
|
-
float kmeans_clustering
|
121
|
-
|
122
|
-
|
123
|
-
|
124
|
-
|
125
|
-
|
126
|
-
|
127
|
-
|
173
|
+
float kmeans_clustering(
|
174
|
+
size_t d,
|
175
|
+
size_t n,
|
176
|
+
size_t k,
|
177
|
+
const float* x,
|
178
|
+
float* centroids);
|
179
|
+
|
180
|
+
} // namespace faiss
|
128
181
|
|
129
182
|
#endif
|
@@ -12,108 +12,95 @@
|
|
12
12
|
#include <memory>
|
13
13
|
|
14
14
|
#include <faiss/IndexPreTransform.h>
|
15
|
-
#include <faiss/impl/FaissAssert.h>
|
16
15
|
#include <faiss/MetaIndexes.h>
|
16
|
+
#include <faiss/impl/FaissAssert.h>
|
17
17
|
#include <faiss/utils/utils.h>
|
18
18
|
|
19
|
+
namespace faiss {
|
20
|
+
namespace ivflib {
|
19
21
|
|
20
|
-
|
21
|
-
|
22
|
-
|
23
|
-
|
24
|
-
void check_compatible_for_merge (const Index * index0,
|
25
|
-
const Index * index1)
|
26
|
-
{
|
27
|
-
|
28
|
-
const faiss::IndexPreTransform *pt0 =
|
29
|
-
dynamic_cast<const faiss::IndexPreTransform *>(index0);
|
22
|
+
void check_compatible_for_merge(const Index* index0, const Index* index1) {
|
23
|
+
const faiss::IndexPreTransform* pt0 =
|
24
|
+
dynamic_cast<const faiss::IndexPreTransform*>(index0);
|
30
25
|
|
31
26
|
if (pt0) {
|
32
|
-
const faiss::IndexPreTransform
|
33
|
-
|
34
|
-
FAISS_THROW_IF_NOT_MSG
|
27
|
+
const faiss::IndexPreTransform* pt1 =
|
28
|
+
dynamic_cast<const faiss::IndexPreTransform*>(index1);
|
29
|
+
FAISS_THROW_IF_NOT_MSG(pt1, "both indexes should be pretransforms");
|
35
30
|
|
36
|
-
FAISS_THROW_IF_NOT
|
31
|
+
FAISS_THROW_IF_NOT(pt0->chain.size() == pt1->chain.size());
|
37
32
|
for (int i = 0; i < pt0->chain.size(); i++) {
|
38
|
-
FAISS_THROW_IF_NOT
|
33
|
+
FAISS_THROW_IF_NOT(typeid(pt0->chain[i]) == typeid(pt1->chain[i]));
|
39
34
|
}
|
40
35
|
|
41
36
|
index0 = pt0->index;
|
42
37
|
index1 = pt1->index;
|
43
38
|
}
|
44
|
-
FAISS_THROW_IF_NOT
|
45
|
-
FAISS_THROW_IF_NOT
|
46
|
-
|
39
|
+
FAISS_THROW_IF_NOT(typeid(index0) == typeid(index1));
|
40
|
+
FAISS_THROW_IF_NOT(
|
41
|
+
index0->d == index1->d &&
|
42
|
+
index0->metric_type == index1->metric_type);
|
47
43
|
|
48
|
-
const faiss::IndexIVF
|
44
|
+
const faiss::IndexIVF* ivf0 = dynamic_cast<const faiss::IndexIVF*>(index0);
|
49
45
|
if (ivf0) {
|
50
|
-
const faiss::IndexIVF
|
51
|
-
|
52
|
-
FAISS_THROW_IF_NOT
|
46
|
+
const faiss::IndexIVF* ivf1 =
|
47
|
+
dynamic_cast<const faiss::IndexIVF*>(index1);
|
48
|
+
FAISS_THROW_IF_NOT(ivf1);
|
53
49
|
|
54
|
-
ivf0->check_compatible_for_merge
|
50
|
+
ivf0->check_compatible_for_merge(*ivf1);
|
55
51
|
}
|
56
52
|
|
57
53
|
// TODO: check as thoroughfully for other index types
|
58
|
-
|
59
54
|
}
|
60
55
|
|
61
|
-
const IndexIVF
|
62
|
-
{
|
63
|
-
if (auto *pt =
|
64
|
-
dynamic_cast<const IndexPreTransform *>(index)) {
|
56
|
+
const IndexIVF* try_extract_index_ivf(const Index* index) {
|
57
|
+
if (auto* pt = dynamic_cast<const IndexPreTransform*>(index)) {
|
65
58
|
index = pt->index;
|
66
59
|
}
|
67
60
|
|
68
|
-
if (auto
|
69
|
-
dynamic_cast<const IndexIDMap *>(index)) {
|
61
|
+
if (auto* idmap = dynamic_cast<const IndexIDMap*>(index)) {
|
70
62
|
index = idmap->index;
|
71
63
|
}
|
72
|
-
if (auto
|
73
|
-
dynamic_cast<const IndexIDMap2 *>(index)) {
|
64
|
+
if (auto* idmap = dynamic_cast<const IndexIDMap2*>(index)) {
|
74
65
|
index = idmap->index;
|
75
66
|
}
|
76
67
|
|
77
|
-
auto
|
68
|
+
auto* ivf = dynamic_cast<const IndexIVF*>(index);
|
78
69
|
|
79
70
|
return ivf;
|
80
71
|
}
|
81
72
|
|
82
|
-
IndexIVF
|
83
|
-
return const_cast<IndexIVF*>
|
73
|
+
IndexIVF* try_extract_index_ivf(Index* index) {
|
74
|
+
return const_cast<IndexIVF*>(try_extract_index_ivf((const Index*)(index)));
|
84
75
|
}
|
85
76
|
|
86
|
-
const IndexIVF
|
87
|
-
|
88
|
-
|
89
|
-
FAISS_THROW_IF_NOT (ivf);
|
77
|
+
const IndexIVF* extract_index_ivf(const Index* index) {
|
78
|
+
const IndexIVF* ivf = try_extract_index_ivf(index);
|
79
|
+
FAISS_THROW_IF_NOT(ivf);
|
90
80
|
return ivf;
|
91
81
|
}
|
92
82
|
|
93
|
-
IndexIVF
|
94
|
-
return const_cast<IndexIVF*>
|
83
|
+
IndexIVF* extract_index_ivf(Index* index) {
|
84
|
+
return const_cast<IndexIVF*>(extract_index_ivf((const Index*)(index)));
|
95
85
|
}
|
96
86
|
|
87
|
+
void merge_into(faiss::Index* index0, faiss::Index* index1, bool shift_ids) {
|
88
|
+
check_compatible_for_merge(index0, index1);
|
89
|
+
IndexIVF* ivf0 = extract_index_ivf(index0);
|
90
|
+
IndexIVF* ivf1 = extract_index_ivf(index1);
|
97
91
|
|
98
|
-
|
99
|
-
|
100
|
-
check_compatible_for_merge (index0, index1);
|
101
|
-
IndexIVF * ivf0 = extract_index_ivf (index0);
|
102
|
-
IndexIVF * ivf1 = extract_index_ivf (index1);
|
103
|
-
|
104
|
-
ivf0->merge_from (*ivf1, shift_ids ? ivf0->ntotal : 0);
|
92
|
+
ivf0->merge_from(*ivf1, shift_ids ? ivf0->ntotal : 0);
|
105
93
|
|
106
94
|
// useful for IndexPreTransform
|
107
95
|
index0->ntotal = ivf0->ntotal;
|
108
96
|
index1->ntotal = ivf1->ntotal;
|
109
97
|
}
|
110
98
|
|
111
|
-
|
112
|
-
|
113
|
-
|
114
|
-
|
115
|
-
|
116
|
-
{
|
99
|
+
void search_centroid(
|
100
|
+
faiss::Index* index,
|
101
|
+
const float* x,
|
102
|
+
int n,
|
103
|
+
idx_t* centroid_ids) {
|
117
104
|
std::unique_ptr<float[]> del;
|
118
105
|
if (auto index_pre = dynamic_cast<faiss::IndexPreTransform*>(index)) {
|
119
106
|
x = index_pre->apply_chain(n, x);
|
@@ -125,19 +112,17 @@ void search_centroid(faiss::Index *index,
|
|
125
112
|
index_ivf->quantizer->assign(n, x, centroid_ids);
|
126
113
|
}
|
127
114
|
|
128
|
-
|
129
|
-
|
130
|
-
|
131
|
-
|
132
|
-
|
133
|
-
|
134
|
-
|
135
|
-
|
136
|
-
|
137
|
-
|
138
|
-
|
139
|
-
const float *x = xin;
|
140
|
-
std::unique_ptr<float []> del;
|
115
|
+
void search_and_return_centroids(
|
116
|
+
faiss::Index* index,
|
117
|
+
size_t n,
|
118
|
+
const float* xin,
|
119
|
+
long k,
|
120
|
+
float* distances,
|
121
|
+
idx_t* labels,
|
122
|
+
idx_t* query_centroid_ids,
|
123
|
+
idx_t* result_centroid_ids) {
|
124
|
+
const float* x = xin;
|
125
|
+
std::unique_ptr<float[]> del;
|
141
126
|
if (auto index_pre = dynamic_cast<faiss::IndexPreTransform*>(index)) {
|
142
127
|
x = index_pre->apply_chain(n, x);
|
143
128
|
del.reset((float*)x);
|
@@ -147,19 +132,18 @@ void search_and_return_centroids(faiss::Index *index,
|
|
147
132
|
assert(index_ivf);
|
148
133
|
|
149
134
|
size_t nprobe = index_ivf->nprobe;
|
150
|
-
std::vector<idx_t> cent_nos
|
151
|
-
std::vector<float> cent_dis
|
135
|
+
std::vector<idx_t> cent_nos(n * nprobe);
|
136
|
+
std::vector<float> cent_dis(n * nprobe);
|
152
137
|
index_ivf->quantizer->search(
|
153
|
-
|
138
|
+
n, x, nprobe, cent_dis.data(), cent_nos.data());
|
154
139
|
|
155
140
|
if (query_centroid_ids) {
|
156
141
|
for (size_t i = 0; i < n; i++)
|
157
142
|
query_centroid_ids[i] = cent_nos[i * nprobe];
|
158
143
|
}
|
159
144
|
|
160
|
-
index_ivf->search_preassigned
|
161
|
-
|
162
|
-
distances, labels, true);
|
145
|
+
index_ivf->search_preassigned(
|
146
|
+
n, x, k, cent_nos.data(), cent_dis.data(), distances, labels, true);
|
163
147
|
|
164
148
|
for (size_t i = 0; i < n * k; i++) {
|
165
149
|
idx_t label = labels[i];
|
@@ -167,8 +151,8 @@ void search_and_return_centroids(faiss::Index *index,
|
|
167
151
|
if (result_centroid_ids)
|
168
152
|
result_centroid_ids[i] = -1;
|
169
153
|
} else {
|
170
|
-
long list_no = lo_listno
|
171
|
-
long list_index = lo_offset
|
154
|
+
long list_no = lo_listno(label);
|
155
|
+
long list_index = lo_offset(label);
|
172
156
|
if (result_centroid_ids)
|
173
157
|
result_centroid_ids[i] = list_no;
|
174
158
|
labels[i] = index_ivf->invlists->get_single_id(list_no, list_index);
|
@@ -176,61 +160,61 @@ void search_and_return_centroids(faiss::Index *index,
|
|
176
160
|
}
|
177
161
|
}
|
178
162
|
|
179
|
-
|
180
|
-
SlidingIndexWindow::SlidingIndexWindow (Index *index): index (index) {
|
163
|
+
SlidingIndexWindow::SlidingIndexWindow(Index* index) : index(index) {
|
181
164
|
n_slice = 0;
|
182
|
-
IndexIVF* index_ivf = const_cast<IndexIVF*>(extract_index_ivf
|
183
|
-
ils = dynamic_cast<ArrayInvertedLists
|
184
|
-
FAISS_THROW_IF_NOT_MSG
|
185
|
-
|
165
|
+
IndexIVF* index_ivf = const_cast<IndexIVF*>(extract_index_ivf(index));
|
166
|
+
ils = dynamic_cast<ArrayInvertedLists*>(index_ivf->invlists);
|
167
|
+
FAISS_THROW_IF_NOT_MSG(
|
168
|
+
ils, "only supports indexes with ArrayInvertedLists");
|
186
169
|
nlist = ils->nlist;
|
187
170
|
sizes.resize(nlist);
|
188
171
|
}
|
189
172
|
|
190
|
-
template<class T>
|
191
|
-
static void shift_and_add
|
192
|
-
|
193
|
-
|
194
|
-
{
|
173
|
+
template <class T>
|
174
|
+
static void shift_and_add(
|
175
|
+
std::vector<T>& dst,
|
176
|
+
size_t remove,
|
177
|
+
const std::vector<T>& src) {
|
195
178
|
if (remove > 0)
|
196
|
-
memmove
|
197
|
-
|
179
|
+
memmove(dst.data(),
|
180
|
+
dst.data() + remove,
|
181
|
+
(dst.size() - remove) * sizeof(T));
|
198
182
|
size_t insert_point = dst.size() - remove;
|
199
|
-
dst.resize
|
200
|
-
memcpy
|
183
|
+
dst.resize(insert_point + src.size());
|
184
|
+
memcpy(dst.data() + insert_point, src.data(), src.size() * sizeof(T));
|
201
185
|
}
|
202
186
|
|
203
|
-
template<class T>
|
204
|
-
static void remove_from_begin
|
205
|
-
size_t remove)
|
206
|
-
{
|
187
|
+
template <class T>
|
188
|
+
static void remove_from_begin(std::vector<T>& v, size_t remove) {
|
207
189
|
if (remove > 0)
|
208
|
-
v.erase
|
190
|
+
v.erase(v.begin(), v.begin() + remove);
|
209
191
|
}
|
210
192
|
|
211
|
-
void SlidingIndexWindow::step(const Index
|
212
|
-
|
213
|
-
|
214
|
-
|
193
|
+
void SlidingIndexWindow::step(const Index* sub_index, bool remove_oldest) {
|
194
|
+
FAISS_THROW_IF_NOT_MSG(
|
195
|
+
!remove_oldest || n_slice > 0,
|
196
|
+
"cannot remove slice: there is none");
|
215
197
|
|
216
|
-
const ArrayInvertedLists
|
217
|
-
if(sub_index) {
|
218
|
-
check_compatible_for_merge
|
198
|
+
const ArrayInvertedLists* ils2 = nullptr;
|
199
|
+
if (sub_index) {
|
200
|
+
check_compatible_for_merge(index, sub_index);
|
219
201
|
ils2 = dynamic_cast<const ArrayInvertedLists*>(
|
220
|
-
|
221
|
-
FAISS_THROW_IF_NOT_MSG
|
202
|
+
extract_index_ivf(sub_index)->invlists);
|
203
|
+
FAISS_THROW_IF_NOT_MSG(ils2, "supports only ArrayInvertedLists");
|
222
204
|
}
|
223
|
-
IndexIVF
|
205
|
+
IndexIVF* index_ivf = extract_index_ivf(index);
|
224
206
|
|
225
207
|
if (remove_oldest && ils2) {
|
226
208
|
for (int i = 0; i < nlist; i++) {
|
227
|
-
std::vector<size_t
|
209
|
+
std::vector<size_t>& sizesi = sizes[i];
|
228
210
|
size_t amount_to_remove = sizesi[0];
|
229
211
|
index_ivf->ntotal += ils2->ids[i].size() - amount_to_remove;
|
230
212
|
|
231
|
-
shift_and_add
|
232
|
-
shift_and_add
|
233
|
-
|
213
|
+
shift_and_add(ils->ids[i], amount_to_remove, ils2->ids[i]);
|
214
|
+
shift_and_add(
|
215
|
+
ils->codes[i],
|
216
|
+
amount_to_remove * ils->code_size,
|
217
|
+
ils2->codes[i]);
|
234
218
|
for (int j = 0; j + 1 < n_slice; j++) {
|
235
219
|
sizesi[j] = sizesi[j + 1] - amount_to_remove;
|
236
220
|
}
|
@@ -239,8 +223,8 @@ void SlidingIndexWindow::step(const Index *sub_index, bool remove_oldest) {
|
|
239
223
|
} else if (ils2) {
|
240
224
|
for (int i = 0; i < nlist; i++) {
|
241
225
|
index_ivf->ntotal += ils2->ids[i].size();
|
242
|
-
shift_and_add
|
243
|
-
shift_and_add
|
226
|
+
shift_and_add(ils->ids[i], 0, ils2->ids[i]);
|
227
|
+
shift_and_add(ils->codes[i], 0, ils2->codes[i]);
|
244
228
|
sizes[i].push_back(ils->ids[i].size());
|
245
229
|
}
|
246
230
|
n_slice++;
|
@@ -248,74 +232,72 @@ void SlidingIndexWindow::step(const Index *sub_index, bool remove_oldest) {
|
|
248
232
|
for (int i = 0; i < nlist; i++) {
|
249
233
|
size_t amount_to_remove = sizes[i][0];
|
250
234
|
index_ivf->ntotal -= amount_to_remove;
|
251
|
-
remove_from_begin
|
252
|
-
remove_from_begin
|
253
|
-
amount_to_remove * ils->code_size);
|
235
|
+
remove_from_begin(ils->ids[i], amount_to_remove);
|
236
|
+
remove_from_begin(ils->codes[i], amount_to_remove * ils->code_size);
|
254
237
|
for (int j = 0; j + 1 < n_slice; j++) {
|
255
238
|
sizes[i][j] = sizes[i][j + 1] - amount_to_remove;
|
256
239
|
}
|
257
|
-
sizes[i].pop_back
|
240
|
+
sizes[i].pop_back();
|
258
241
|
}
|
259
242
|
n_slice--;
|
260
243
|
} else {
|
261
|
-
FAISS_THROW_MSG
|
244
|
+
FAISS_THROW_MSG("nothing to do???");
|
262
245
|
}
|
263
246
|
index->ntotal = index_ivf->ntotal;
|
264
247
|
}
|
265
248
|
|
266
|
-
|
267
|
-
|
268
249
|
// Get a subset of inverted lists [i0, i1). Works on IndexIVF's and
|
269
250
|
// IndexIVF's embedded in a IndexPreTransform
|
270
251
|
|
271
|
-
ArrayInvertedLists *
|
272
|
-
|
273
|
-
{
|
274
|
-
const IndexIVF *ivf = extract_index_ivf (index);
|
252
|
+
ArrayInvertedLists* get_invlist_range(const Index* index, long i0, long i1) {
|
253
|
+
const IndexIVF* ivf = extract_index_ivf(index);
|
275
254
|
|
276
|
-
FAISS_THROW_IF_NOT
|
255
|
+
FAISS_THROW_IF_NOT(0 <= i0 && i0 <= i1 && i1 <= ivf->nlist);
|
277
256
|
|
278
|
-
const InvertedLists
|
257
|
+
const InvertedLists* src = ivf->invlists;
|
279
258
|
|
280
|
-
ArrayInvertedLists
|
259
|
+
ArrayInvertedLists* il = new ArrayInvertedLists(i1 - i0, src->code_size);
|
281
260
|
|
282
261
|
for (long i = i0; i < i1; i++) {
|
283
|
-
il->add_entries(
|
284
|
-
|
285
|
-
|
262
|
+
il->add_entries(
|
263
|
+
i - i0,
|
264
|
+
src->list_size(i),
|
265
|
+
InvertedLists::ScopedIds(src, i).get(),
|
266
|
+
InvertedLists::ScopedCodes(src, i).get());
|
286
267
|
}
|
287
268
|
return il;
|
288
269
|
}
|
289
270
|
|
271
|
+
void set_invlist_range(
|
272
|
+
Index* index,
|
273
|
+
long i0,
|
274
|
+
long i1,
|
275
|
+
ArrayInvertedLists* src) {
|
276
|
+
IndexIVF* ivf = extract_index_ivf(index);
|
290
277
|
|
278
|
+
FAISS_THROW_IF_NOT(0 <= i0 && i0 <= i1 && i1 <= ivf->nlist);
|
291
279
|
|
292
|
-
|
293
|
-
|
294
|
-
|
295
|
-
|
296
|
-
|
297
|
-
FAISS_THROW_IF_NOT (0 <= i0 && i0 <= i1 && i1 <= ivf->nlist);
|
298
|
-
|
299
|
-
ArrayInvertedLists *dst = dynamic_cast<ArrayInvertedLists *>(ivf->invlists);
|
300
|
-
FAISS_THROW_IF_NOT_MSG (dst, "only ArrayInvertedLists supported");
|
301
|
-
FAISS_THROW_IF_NOT (src->nlist == i1 - i0 &&
|
302
|
-
dst->code_size == src->code_size);
|
280
|
+
ArrayInvertedLists* dst = dynamic_cast<ArrayInvertedLists*>(ivf->invlists);
|
281
|
+
FAISS_THROW_IF_NOT_MSG(dst, "only ArrayInvertedLists supported");
|
282
|
+
FAISS_THROW_IF_NOT(
|
283
|
+
src->nlist == i1 - i0 && dst->code_size == src->code_size);
|
303
284
|
|
304
285
|
size_t ntotal = index->ntotal;
|
305
|
-
for (long i = i0
|
306
|
-
ntotal -= dst->list_size
|
307
|
-
ntotal += src->list_size
|
308
|
-
std::swap
|
309
|
-
std::swap
|
286
|
+
for (long i = i0; i < i1; i++) {
|
287
|
+
ntotal -= dst->list_size(i);
|
288
|
+
ntotal += src->list_size(i - i0);
|
289
|
+
std::swap(src->codes[i - i0], dst->codes[i]);
|
290
|
+
std::swap(src->ids[i - i0], dst->ids[i]);
|
310
291
|
}
|
311
292
|
ivf->ntotal = index->ntotal = ntotal;
|
312
293
|
}
|
313
294
|
|
314
|
-
static size_t count_ndis(
|
315
|
-
|
316
|
-
|
295
|
+
static size_t count_ndis(
|
296
|
+
const IndexIVF* index_ivf,
|
297
|
+
size_t n_list_scan,
|
298
|
+
const idx_t* Iq) {
|
317
299
|
size_t nb_dis = 0;
|
318
|
-
const InvertedLists
|
300
|
+
const InvertedLists* il = index_ivf->invlists;
|
319
301
|
for (idx_t i = 0; i < n_list_scan; i++) {
|
320
302
|
if (Iq[i] >= 0) {
|
321
303
|
nb_dis += il->list_size(Iq[i]);
|
@@ -324,23 +306,24 @@ static size_t count_ndis(const IndexIVF * index_ivf, size_t n_list_scan,
|
|
324
306
|
return nb_dis;
|
325
307
|
}
|
326
308
|
|
327
|
-
|
328
|
-
|
329
|
-
|
330
|
-
|
331
|
-
|
332
|
-
|
333
|
-
|
334
|
-
|
335
|
-
|
336
|
-
|
337
|
-
|
309
|
+
void search_with_parameters(
|
310
|
+
const Index* index,
|
311
|
+
idx_t n,
|
312
|
+
const float* x,
|
313
|
+
idx_t k,
|
314
|
+
float* distances,
|
315
|
+
idx_t* labels,
|
316
|
+
const IVFSearchParameters* params,
|
317
|
+
size_t* nb_dis_ptr,
|
318
|
+
double* ms_per_stage) {
|
319
|
+
FAISS_THROW_IF_NOT(params);
|
320
|
+
const float* prev_x = x;
|
338
321
|
ScopeDeleter<float> del;
|
339
322
|
|
340
323
|
double t0 = getmillisecs();
|
341
324
|
|
342
|
-
if (auto ip = dynamic_cast<const IndexPreTransform
|
343
|
-
x = ip->apply_chain
|
325
|
+
if (auto ip = dynamic_cast<const IndexPreTransform*>(index)) {
|
326
|
+
x = ip->apply_chain(n, x);
|
344
327
|
if (x != prev_x) {
|
345
328
|
del.set(x);
|
346
329
|
}
|
@@ -352,21 +335,19 @@ void search_with_parameters (const Index *index,
|
|
352
335
|
std::vector<idx_t> Iq(params->nprobe * n);
|
353
336
|
std::vector<float> Dq(params->nprobe * n);
|
354
337
|
|
355
|
-
const IndexIVF
|
356
|
-
FAISS_THROW_IF_NOT
|
338
|
+
const IndexIVF* index_ivf = dynamic_cast<const IndexIVF*>(index);
|
339
|
+
FAISS_THROW_IF_NOT(index_ivf);
|
357
340
|
|
358
|
-
index_ivf->quantizer->search(n, x, params->nprobe,
|
359
|
-
Dq.data(), Iq.data());
|
341
|
+
index_ivf->quantizer->search(n, x, params->nprobe, Dq.data(), Iq.data());
|
360
342
|
|
361
343
|
if (nb_dis_ptr) {
|
362
|
-
*nb_dis_ptr = count_ndis
|
344
|
+
*nb_dis_ptr = count_ndis(index_ivf, n * params->nprobe, Iq.data());
|
363
345
|
}
|
364
346
|
|
365
347
|
double t2 = getmillisecs();
|
366
348
|
|
367
|
-
index_ivf->search_preassigned(
|
368
|
-
|
369
|
-
false, params);
|
349
|
+
index_ivf->search_preassigned(
|
350
|
+
n, x, k, Iq.data(), Dq.data(), distances, labels, false, params);
|
370
351
|
double t3 = getmillisecs();
|
371
352
|
if (ms_per_stage) {
|
372
353
|
ms_per_stage[0] = t1 - t0;
|
@@ -375,21 +356,23 @@ void search_with_parameters (const Index *index,
|
|
375
356
|
}
|
376
357
|
}
|
377
358
|
|
378
|
-
void range_search_with_parameters
|
379
|
-
|
380
|
-
|
381
|
-
|
382
|
-
|
383
|
-
|
384
|
-
|
385
|
-
|
386
|
-
|
359
|
+
void range_search_with_parameters(
|
360
|
+
const Index* index,
|
361
|
+
idx_t n,
|
362
|
+
const float* x,
|
363
|
+
float radius,
|
364
|
+
RangeSearchResult* result,
|
365
|
+
const IVFSearchParameters* params,
|
366
|
+
size_t* nb_dis_ptr,
|
367
|
+
double* ms_per_stage) {
|
368
|
+
FAISS_THROW_IF_NOT(params);
|
369
|
+
const float* prev_x = x;
|
387
370
|
ScopeDeleter<float> del;
|
388
371
|
|
389
372
|
double t0 = getmillisecs();
|
390
373
|
|
391
|
-
if (auto ip = dynamic_cast<const IndexPreTransform
|
392
|
-
x = ip->apply_chain
|
374
|
+
if (auto ip = dynamic_cast<const IndexPreTransform*>(index)) {
|
375
|
+
x = ip->apply_chain(n, x);
|
393
376
|
if (x != prev_x) {
|
394
377
|
del.set(x);
|
395
378
|
}
|
@@ -401,22 +384,19 @@ void range_search_with_parameters (const Index *index,
|
|
401
384
|
std::vector<idx_t> Iq(params->nprobe * n);
|
402
385
|
std::vector<float> Dq(params->nprobe * n);
|
403
386
|
|
404
|
-
const IndexIVF
|
405
|
-
FAISS_THROW_IF_NOT
|
387
|
+
const IndexIVF* index_ivf = dynamic_cast<const IndexIVF*>(index);
|
388
|
+
FAISS_THROW_IF_NOT(index_ivf);
|
406
389
|
|
407
|
-
index_ivf->quantizer->search(n, x, params->nprobe,
|
408
|
-
Dq.data(), Iq.data());
|
390
|
+
index_ivf->quantizer->search(n, x, params->nprobe, Dq.data(), Iq.data());
|
409
391
|
|
410
392
|
if (nb_dis_ptr) {
|
411
|
-
*nb_dis_ptr = count_ndis
|
393
|
+
*nb_dis_ptr = count_ndis(index_ivf, n * params->nprobe, Iq.data());
|
412
394
|
}
|
413
395
|
|
414
396
|
double t2 = getmillisecs();
|
415
397
|
|
416
398
|
index_ivf->range_search_preassigned(
|
417
|
-
n, x, radius, Iq.data(), Dq.data(),
|
418
|
-
result, false, params
|
419
|
-
);
|
399
|
+
n, x, radius, Iq.data(), Dq.data(), result, false, params);
|
420
400
|
|
421
401
|
double t3 = getmillisecs();
|
422
402
|
if (ms_per_stage) {
|
@@ -426,9 +406,5 @@ void range_search_with_parameters (const Index *index,
|
|
426
406
|
}
|
427
407
|
}
|
428
408
|
|
429
|
-
|
430
|
-
|
431
|
-
|
432
|
-
|
433
|
-
|
434
|
-
} } // namespace faiss::ivflib
|
409
|
+
} // namespace ivflib
|
410
|
+
} // namespace faiss
|