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
@@ -19,62 +19,68 @@ namespace faiss {
|
|
19
19
|
/// The interface to this class itself is not thread safe
|
20
20
|
template <typename IndexT>
|
21
21
|
class ThreadedIndex : public IndexT {
|
22
|
-
|
23
|
-
|
24
|
-
|
25
|
-
|
26
|
-
|
27
|
-
|
28
|
-
|
29
|
-
|
30
|
-
|
31
|
-
|
32
|
-
|
33
|
-
|
34
|
-
|
35
|
-
|
36
|
-
|
37
|
-
|
38
|
-
|
39
|
-
|
40
|
-
|
41
|
-
|
42
|
-
|
43
|
-
|
44
|
-
|
45
|
-
|
46
|
-
|
47
|
-
|
48
|
-
|
49
|
-
|
50
|
-
|
51
|
-
|
52
|
-
|
53
|
-
|
54
|
-
|
55
|
-
|
56
|
-
|
57
|
-
|
58
|
-
|
59
|
-
|
60
|
-
|
61
|
-
|
62
|
-
|
63
|
-
|
64
|
-
|
65
|
-
|
66
|
-
|
67
|
-
|
68
|
-
|
69
|
-
|
70
|
-
|
71
|
-
|
72
|
-
|
73
|
-
|
74
|
-
|
75
|
-
|
22
|
+
public:
|
23
|
+
explicit ThreadedIndex(bool threaded);
|
24
|
+
explicit ThreadedIndex(int d, bool threaded);
|
25
|
+
|
26
|
+
~ThreadedIndex() override;
|
27
|
+
|
28
|
+
/// override an index that is managed by ourselves.
|
29
|
+
/// WARNING: once an index is added, it becomes unsafe to touch it from any
|
30
|
+
/// other thread than that on which is managing it, until we are shut
|
31
|
+
/// down. Use runOnIndex to perform work on it instead.
|
32
|
+
void addIndex(IndexT* index);
|
33
|
+
|
34
|
+
/// Remove an index that is managed by ourselves.
|
35
|
+
/// This will flush all pending work on that index, and then shut
|
36
|
+
/// down its managing thread, and will remove the index.
|
37
|
+
void removeIndex(IndexT* index);
|
38
|
+
|
39
|
+
/// Run a function on all indices, in the thread that the index is
|
40
|
+
/// managed in.
|
41
|
+
/// Function arguments are (index in collection, index pointer)
|
42
|
+
void runOnIndex(std::function<void(int, IndexT*)> f);
|
43
|
+
void runOnIndex(std::function<void(int, const IndexT*)> f) const;
|
44
|
+
|
45
|
+
/// faiss::Index API
|
46
|
+
/// All indices receive the same call
|
47
|
+
void reset() override;
|
48
|
+
|
49
|
+
/// Returns the number of sub-indices
|
50
|
+
int count() const {
|
51
|
+
return indices_.size();
|
52
|
+
}
|
53
|
+
|
54
|
+
/// Returns the i-th sub-index
|
55
|
+
IndexT* at(int i) {
|
56
|
+
return indices_[i].first;
|
57
|
+
}
|
58
|
+
|
59
|
+
/// Returns the i-th sub-index (const version)
|
60
|
+
const IndexT* at(int i) const {
|
61
|
+
return indices_[i].first;
|
62
|
+
}
|
63
|
+
|
64
|
+
/// Whether or not we are responsible for deleting our contained indices
|
65
|
+
bool own_fields;
|
66
|
+
|
67
|
+
protected:
|
68
|
+
/// Called just after an index is added
|
69
|
+
virtual void onAfterAddIndex(IndexT* index);
|
70
|
+
|
71
|
+
/// Called just after an index is removed
|
72
|
+
virtual void onAfterRemoveIndex(IndexT* index);
|
73
|
+
|
74
|
+
protected:
|
75
|
+
static void waitAndHandleFutures(std::vector<std::future<bool>>& v);
|
76
|
+
|
77
|
+
/// Collection of Index instances, with their managing worker thread if any
|
78
|
+
std::vector<std::pair<IndexT*, std::unique_ptr<WorkerThread>>> indices_;
|
79
|
+
|
80
|
+
/// Is this index multi-threaded?
|
81
|
+
bool isThreaded_;
|
76
82
|
};
|
77
83
|
|
78
|
-
} // namespace
|
84
|
+
} // namespace faiss
|
79
85
|
|
80
86
|
#include <faiss/impl/ThreadedIndex-inl.h>
|
@@ -9,11 +9,13 @@
|
|
9
9
|
|
10
10
|
#include <faiss/index_io.h>
|
11
11
|
|
12
|
+
#include <faiss/impl/io_macros.h>
|
13
|
+
|
12
14
|
#include <cstdio>
|
13
15
|
#include <cstdlib>
|
14
16
|
|
15
|
-
#include <sys/types.h>
|
16
17
|
#include <sys/stat.h>
|
18
|
+
#include <sys/types.h>
|
17
19
|
|
18
20
|
#include <faiss/impl/FaissAssert.h>
|
19
21
|
#include <faiss/impl/io.h>
|
@@ -22,343 +24,383 @@
|
|
22
24
|
|
23
25
|
#include <faiss/invlists/InvertedListsIOHook.h>
|
24
26
|
|
27
|
+
#include <faiss/Index2Layer.h>
|
25
28
|
#include <faiss/IndexFlat.h>
|
26
|
-
#include <faiss/
|
27
|
-
#include <faiss/IndexPreTransform.h>
|
28
|
-
#include <faiss/IndexLSH.h>
|
29
|
-
#include <faiss/IndexPQ.h>
|
29
|
+
#include <faiss/IndexHNSW.h>
|
30
30
|
#include <faiss/IndexIVF.h>
|
31
|
+
#include <faiss/IndexIVFFlat.h>
|
31
32
|
#include <faiss/IndexIVFPQ.h>
|
33
|
+
#include <faiss/IndexIVFPQFastScan.h>
|
32
34
|
#include <faiss/IndexIVFPQR.h>
|
33
|
-
#include <faiss/Index2Layer.h>
|
34
|
-
#include <faiss/IndexIVFFlat.h>
|
35
35
|
#include <faiss/IndexIVFSpectralHash.h>
|
36
|
-
#include <faiss/
|
37
|
-
#include <faiss/IndexScalarQuantizer.h>
|
38
|
-
#include <faiss/IndexHNSW.h>
|
36
|
+
#include <faiss/IndexLSH.h>
|
39
37
|
#include <faiss/IndexLattice.h>
|
38
|
+
#include <faiss/IndexNSG.h>
|
39
|
+
#include <faiss/IndexPQ.h>
|
40
40
|
#include <faiss/IndexPQFastScan.h>
|
41
|
-
#include <faiss/
|
41
|
+
#include <faiss/IndexPreTransform.h>
|
42
42
|
#include <faiss/IndexRefine.h>
|
43
|
+
#include <faiss/IndexResidual.h>
|
44
|
+
#include <faiss/IndexScalarQuantizer.h>
|
45
|
+
#include <faiss/MetaIndexes.h>
|
46
|
+
#include <faiss/VectorTransform.h>
|
43
47
|
|
44
48
|
#include <faiss/IndexBinaryFlat.h>
|
45
49
|
#include <faiss/IndexBinaryFromFloat.h>
|
46
50
|
#include <faiss/IndexBinaryHNSW.h>
|
47
|
-
#include <faiss/IndexBinaryIVF.h>
|
48
51
|
#include <faiss/IndexBinaryHash.h>
|
52
|
+
#include <faiss/IndexBinaryIVF.h>
|
49
53
|
|
50
54
|
namespace faiss {
|
51
55
|
|
52
|
-
|
53
56
|
/*************************************************************
|
54
57
|
* Read
|
55
58
|
**************************************************************/
|
56
59
|
|
57
|
-
static void read_index_header
|
58
|
-
READ1
|
59
|
-
READ1
|
60
|
+
static void read_index_header(Index* idx, IOReader* f) {
|
61
|
+
READ1(idx->d);
|
62
|
+
READ1(idx->ntotal);
|
60
63
|
Index::idx_t dummy;
|
61
|
-
READ1
|
62
|
-
READ1
|
63
|
-
READ1
|
64
|
-
READ1
|
64
|
+
READ1(dummy);
|
65
|
+
READ1(dummy);
|
66
|
+
READ1(idx->is_trained);
|
67
|
+
READ1(idx->metric_type);
|
65
68
|
if (idx->metric_type > 1) {
|
66
|
-
READ1
|
69
|
+
READ1(idx->metric_arg);
|
67
70
|
}
|
68
71
|
idx->verbose = false;
|
69
72
|
}
|
70
73
|
|
71
|
-
VectorTransform* read_VectorTransform
|
74
|
+
VectorTransform* read_VectorTransform(IOReader* f) {
|
72
75
|
uint32_t h;
|
73
|
-
READ1
|
74
|
-
VectorTransform
|
75
|
-
|
76
|
-
if (h == fourcc
|
77
|
-
h == fourcc
|
78
|
-
|
79
|
-
|
80
|
-
|
81
|
-
|
82
|
-
|
83
|
-
|
84
|
-
|
85
|
-
|
86
|
-
|
87
|
-
|
88
|
-
|
89
|
-
READVECTOR
|
90
|
-
READVECTOR (pca->eigenvalues);
|
91
|
-
READVECTOR (pca->PCAMat);
|
76
|
+
READ1(h);
|
77
|
+
VectorTransform* vt = nullptr;
|
78
|
+
|
79
|
+
if (h == fourcc("rrot") || h == fourcc("PCAm") || h == fourcc("LTra") ||
|
80
|
+
h == fourcc("PcAm") || h == fourcc("Viqm")) {
|
81
|
+
LinearTransform* lt = nullptr;
|
82
|
+
if (h == fourcc("rrot")) {
|
83
|
+
lt = new RandomRotationMatrix();
|
84
|
+
} else if (h == fourcc("PCAm") || h == fourcc("PcAm")) {
|
85
|
+
PCAMatrix* pca = new PCAMatrix();
|
86
|
+
READ1(pca->eigen_power);
|
87
|
+
READ1(pca->random_rotation);
|
88
|
+
if (h == fourcc("PcAm"))
|
89
|
+
READ1(pca->balanced_bins);
|
90
|
+
READVECTOR(pca->mean);
|
91
|
+
READVECTOR(pca->eigenvalues);
|
92
|
+
READVECTOR(pca->PCAMat);
|
92
93
|
lt = pca;
|
93
|
-
} else if (h == fourcc
|
94
|
-
ITQMatrix
|
95
|
-
READ1
|
96
|
-
READ1
|
94
|
+
} else if (h == fourcc("Viqm")) {
|
95
|
+
ITQMatrix* itqm = new ITQMatrix();
|
96
|
+
READ1(itqm->max_iter);
|
97
|
+
READ1(itqm->seed);
|
97
98
|
lt = itqm;
|
98
|
-
} else if (h == fourcc
|
99
|
-
lt = new LinearTransform
|
99
|
+
} else if (h == fourcc("LTra")) {
|
100
|
+
lt = new LinearTransform();
|
100
101
|
}
|
101
|
-
READ1
|
102
|
-
READVECTOR
|
103
|
-
READVECTOR
|
104
|
-
FAISS_THROW_IF_NOT
|
105
|
-
FAISS_THROW_IF_NOT
|
102
|
+
READ1(lt->have_bias);
|
103
|
+
READVECTOR(lt->A);
|
104
|
+
READVECTOR(lt->b);
|
105
|
+
FAISS_THROW_IF_NOT(lt->A.size() >= lt->d_in * lt->d_out);
|
106
|
+
FAISS_THROW_IF_NOT(!lt->have_bias || lt->b.size() >= lt->d_out);
|
106
107
|
lt->set_is_orthonormal();
|
107
108
|
vt = lt;
|
108
|
-
} else if (h == fourcc
|
109
|
-
RemapDimensionsTransform
|
110
|
-
READVECTOR
|
109
|
+
} else if (h == fourcc("RmDT")) {
|
110
|
+
RemapDimensionsTransform* rdt = new RemapDimensionsTransform();
|
111
|
+
READVECTOR(rdt->map);
|
111
112
|
vt = rdt;
|
112
|
-
} else if (h == fourcc
|
113
|
-
NormalizationTransform
|
114
|
-
READ1
|
113
|
+
} else if (h == fourcc("VNrm")) {
|
114
|
+
NormalizationTransform* nt = new NormalizationTransform();
|
115
|
+
READ1(nt->norm);
|
115
116
|
vt = nt;
|
116
|
-
} else if (h == fourcc
|
117
|
-
CenteringTransform
|
118
|
-
READVECTOR
|
117
|
+
} else if (h == fourcc("VCnt")) {
|
118
|
+
CenteringTransform* ct = new CenteringTransform();
|
119
|
+
READVECTOR(ct->mean);
|
119
120
|
vt = ct;
|
120
|
-
} else if (h == fourcc
|
121
|
-
ITQTransform
|
121
|
+
} else if (h == fourcc("Viqt")) {
|
122
|
+
ITQTransform* itqt = new ITQTransform();
|
122
123
|
|
123
|
-
READVECTOR
|
124
|
-
READ1
|
124
|
+
READVECTOR(itqt->mean);
|
125
|
+
READ1(itqt->do_pca);
|
125
126
|
{
|
126
|
-
ITQMatrix
|
127
|
-
(read_VectorTransform (f));
|
127
|
+
ITQMatrix* itqm = dynamic_cast<ITQMatrix*>(read_VectorTransform(f));
|
128
128
|
FAISS_THROW_IF_NOT(itqm);
|
129
129
|
itqt->itq = *itqm;
|
130
130
|
delete itqm;
|
131
131
|
}
|
132
132
|
{
|
133
|
-
LinearTransform
|
134
|
-
|
135
|
-
FAISS_THROW_IF_NOT
|
133
|
+
LinearTransform* pi =
|
134
|
+
dynamic_cast<LinearTransform*>(read_VectorTransform(f));
|
135
|
+
FAISS_THROW_IF_NOT(pi);
|
136
136
|
itqt->pca_then_itq = *pi;
|
137
137
|
delete pi;
|
138
138
|
}
|
139
139
|
vt = itqt;
|
140
140
|
} else {
|
141
141
|
FAISS_THROW_FMT(
|
142
|
-
|
143
|
-
|
144
|
-
|
142
|
+
"fourcc %ud (\"%s\") not recognized",
|
143
|
+
h,
|
144
|
+
fourcc_inv_printable(h).c_str());
|
145
145
|
}
|
146
|
-
READ1
|
147
|
-
READ1
|
148
|
-
READ1
|
146
|
+
READ1(vt->d_in);
|
147
|
+
READ1(vt->d_out);
|
148
|
+
READ1(vt->is_trained);
|
149
149
|
return vt;
|
150
150
|
}
|
151
151
|
|
152
|
-
|
153
|
-
|
154
|
-
|
155
|
-
{
|
152
|
+
static void read_ArrayInvertedLists_sizes(
|
153
|
+
IOReader* f,
|
154
|
+
std::vector<size_t>& sizes) {
|
156
155
|
uint32_t list_type;
|
157
156
|
READ1(list_type);
|
158
157
|
if (list_type == fourcc("full")) {
|
159
158
|
size_t os = sizes.size();
|
160
|
-
READVECTOR
|
161
|
-
FAISS_THROW_IF_NOT
|
159
|
+
READVECTOR(sizes);
|
160
|
+
FAISS_THROW_IF_NOT(os == sizes.size());
|
162
161
|
} else if (list_type == fourcc("sprs")) {
|
163
162
|
std::vector<size_t> idsizes;
|
164
|
-
READVECTOR
|
163
|
+
READVECTOR(idsizes);
|
165
164
|
for (size_t j = 0; j < idsizes.size(); j += 2) {
|
166
|
-
FAISS_THROW_IF_NOT
|
165
|
+
FAISS_THROW_IF_NOT(idsizes[j] < sizes.size());
|
167
166
|
sizes[idsizes[j]] = idsizes[j + 1];
|
168
167
|
}
|
169
168
|
} else {
|
170
169
|
FAISS_THROW_FMT(
|
171
|
-
|
172
|
-
|
173
|
-
|
170
|
+
"list_type %ud (\"%s\") not recognized",
|
171
|
+
list_type,
|
172
|
+
fourcc_inv_printable(list_type).c_str());
|
174
173
|
}
|
175
174
|
}
|
176
175
|
|
177
|
-
InvertedLists
|
176
|
+
InvertedLists* read_InvertedLists(IOReader* f, int io_flags) {
|
178
177
|
uint32_t h;
|
179
|
-
READ1
|
180
|
-
if (h == fourcc
|
181
|
-
fprintf(stderr,
|
178
|
+
READ1(h);
|
179
|
+
if (h == fourcc("il00")) {
|
180
|
+
fprintf(stderr,
|
181
|
+
"read_InvertedLists:"
|
182
182
|
" WARN! inverted lists not stored with IVF object\n");
|
183
183
|
return nullptr;
|
184
|
-
} else if (h == fourcc
|
185
|
-
auto ails = new ArrayInvertedLists
|
186
|
-
READ1
|
187
|
-
READ1
|
188
|
-
ails->ids.resize
|
189
|
-
ails->codes.resize
|
190
|
-
std::vector<size_t> sizes
|
191
|
-
read_ArrayInvertedLists_sizes
|
184
|
+
} else if (h == fourcc("ilar") && !(io_flags & IO_FLAG_SKIP_IVF_DATA)) {
|
185
|
+
auto ails = new ArrayInvertedLists(0, 0);
|
186
|
+
READ1(ails->nlist);
|
187
|
+
READ1(ails->code_size);
|
188
|
+
ails->ids.resize(ails->nlist);
|
189
|
+
ails->codes.resize(ails->nlist);
|
190
|
+
std::vector<size_t> sizes(ails->nlist);
|
191
|
+
read_ArrayInvertedLists_sizes(f, sizes);
|
192
192
|
for (size_t i = 0; i < ails->nlist; i++) {
|
193
|
-
ails->ids[i].resize
|
194
|
-
ails->codes[i].resize
|
193
|
+
ails->ids[i].resize(sizes[i]);
|
194
|
+
ails->codes[i].resize(sizes[i] * ails->code_size);
|
195
195
|
}
|
196
196
|
for (size_t i = 0; i < ails->nlist; i++) {
|
197
197
|
size_t n = ails->ids[i].size();
|
198
198
|
if (n > 0) {
|
199
|
-
READANDCHECK
|
200
|
-
READANDCHECK
|
199
|
+
READANDCHECK(ails->codes[i].data(), n * ails->code_size);
|
200
|
+
READANDCHECK(ails->ids[i].data(), n);
|
201
201
|
}
|
202
202
|
}
|
203
203
|
return ails;
|
204
204
|
|
205
|
-
} else if (h == fourcc
|
206
|
-
// code is always ilxx where xx is specific to the type of invlists we
|
207
|
-
// so we get the 16 high bits from the io_flag and the 16 low bits
|
205
|
+
} else if (h == fourcc("ilar") && (io_flags & IO_FLAG_SKIP_IVF_DATA)) {
|
206
|
+
// code is always ilxx where xx is specific to the type of invlists we
|
207
|
+
// want so we get the 16 high bits from the io_flag and the 16 low bits
|
208
|
+
// as "il"
|
208
209
|
int h2 = (io_flags & 0xffff0000) | (fourcc("il__") & 0x0000ffff);
|
209
210
|
size_t nlist, code_size;
|
210
|
-
READ1
|
211
|
-
READ1
|
212
|
-
std::vector<size_t> sizes
|
213
|
-
read_ArrayInvertedLists_sizes
|
211
|
+
READ1(nlist);
|
212
|
+
READ1(code_size);
|
213
|
+
std::vector<size_t> sizes(nlist);
|
214
|
+
read_ArrayInvertedLists_sizes(f, sizes);
|
214
215
|
return InvertedListsIOHook::lookup(h2)->read_ArrayInvertedLists(
|
215
216
|
f, io_flags, nlist, code_size, sizes);
|
216
217
|
} else {
|
217
218
|
return InvertedListsIOHook::lookup(h)->read(f, io_flags);
|
218
219
|
}
|
219
|
-
|
220
220
|
}
|
221
221
|
|
222
|
-
|
223
|
-
|
224
|
-
IndexIVF *ivf, IOReader *f, int io_flags) {
|
225
|
-
InvertedLists *ils = read_InvertedLists (f, io_flags);
|
222
|
+
static void read_InvertedLists(IndexIVF* ivf, IOReader* f, int io_flags) {
|
223
|
+
InvertedLists* ils = read_InvertedLists(f, io_flags);
|
226
224
|
if (ils) {
|
227
|
-
FAISS_THROW_IF_NOT
|
228
|
-
FAISS_THROW_IF_NOT
|
229
|
-
|
225
|
+
FAISS_THROW_IF_NOT(ils->nlist == ivf->nlist);
|
226
|
+
FAISS_THROW_IF_NOT(
|
227
|
+
ils->code_size == InvertedLists::INVALID_CODE_SIZE ||
|
228
|
+
ils->code_size == ivf->code_size);
|
230
229
|
}
|
231
230
|
ivf->invlists = ils;
|
232
231
|
ivf->own_invlists = true;
|
233
232
|
}
|
234
233
|
|
235
|
-
static void read_ProductQuantizer
|
236
|
-
READ1
|
237
|
-
READ1
|
238
|
-
READ1
|
239
|
-
pq->set_derived_values
|
240
|
-
READVECTOR
|
234
|
+
static void read_ProductQuantizer(ProductQuantizer* pq, IOReader* f) {
|
235
|
+
READ1(pq->d);
|
236
|
+
READ1(pq->M);
|
237
|
+
READ1(pq->nbits);
|
238
|
+
pq->set_derived_values();
|
239
|
+
READVECTOR(pq->centroids);
|
240
|
+
}
|
241
|
+
|
242
|
+
static void read_ResidualQuantizer(ResidualQuantizer* rq, IOReader* f) {
|
243
|
+
READ1(rq->d);
|
244
|
+
READ1(rq->M);
|
245
|
+
READVECTOR(rq->nbits);
|
246
|
+
rq->set_derived_values();
|
247
|
+
READ1(rq->is_trained);
|
248
|
+
READ1(rq->train_type);
|
249
|
+
READ1(rq->max_beam_size);
|
250
|
+
READVECTOR(rq->codebooks);
|
241
251
|
}
|
242
252
|
|
243
|
-
static void read_ScalarQuantizer
|
244
|
-
READ1
|
245
|
-
READ1
|
246
|
-
READ1
|
247
|
-
READ1
|
248
|
-
READ1
|
249
|
-
READVECTOR
|
250
|
-
ivsc->set_derived_sizes
|
253
|
+
static void read_ScalarQuantizer(ScalarQuantizer* ivsc, IOReader* f) {
|
254
|
+
READ1(ivsc->qtype);
|
255
|
+
READ1(ivsc->rangestat);
|
256
|
+
READ1(ivsc->rangestat_arg);
|
257
|
+
READ1(ivsc->d);
|
258
|
+
READ1(ivsc->code_size);
|
259
|
+
READVECTOR(ivsc->trained);
|
260
|
+
ivsc->set_derived_sizes();
|
251
261
|
}
|
252
262
|
|
263
|
+
static void read_HNSW(HNSW* hnsw, IOReader* f) {
|
264
|
+
READVECTOR(hnsw->assign_probas);
|
265
|
+
READVECTOR(hnsw->cum_nneighbor_per_level);
|
266
|
+
READVECTOR(hnsw->levels);
|
267
|
+
READVECTOR(hnsw->offsets);
|
268
|
+
READVECTOR(hnsw->neighbors);
|
269
|
+
|
270
|
+
READ1(hnsw->entry_point);
|
271
|
+
READ1(hnsw->max_level);
|
272
|
+
READ1(hnsw->efConstruction);
|
273
|
+
READ1(hnsw->efSearch);
|
274
|
+
READ1(hnsw->upper_beam);
|
275
|
+
}
|
253
276
|
|
254
|
-
static void
|
255
|
-
|
256
|
-
|
257
|
-
|
258
|
-
|
259
|
-
|
277
|
+
static void read_NSG(NSG* nsg, IOReader* f) {
|
278
|
+
READ1(nsg->ntotal);
|
279
|
+
READ1(nsg->R);
|
280
|
+
READ1(nsg->L);
|
281
|
+
READ1(nsg->C);
|
282
|
+
READ1(nsg->search_L);
|
283
|
+
READ1(nsg->enterpoint);
|
284
|
+
READ1(nsg->is_built);
|
285
|
+
|
286
|
+
if (!nsg->is_built) {
|
287
|
+
return;
|
288
|
+
}
|
260
289
|
|
261
|
-
|
262
|
-
|
263
|
-
|
264
|
-
|
265
|
-
|
290
|
+
constexpr int EMPTY_ID = -1;
|
291
|
+
int N = nsg->ntotal;
|
292
|
+
int R = nsg->R;
|
293
|
+
auto& graph = nsg->final_graph;
|
294
|
+
graph = std::make_shared<nsg::Graph<int>>(N, R);
|
295
|
+
std::fill_n(graph->data, N * R, EMPTY_ID);
|
296
|
+
|
297
|
+
int size = 0;
|
298
|
+
|
299
|
+
for (int i = 0; i < N; i++) {
|
300
|
+
for (int j = 0; j < R + 1; j++) {
|
301
|
+
int id;
|
302
|
+
READ1(id);
|
303
|
+
if (id != EMPTY_ID) {
|
304
|
+
graph->at(i, j) = id;
|
305
|
+
size += 1;
|
306
|
+
} else {
|
307
|
+
break;
|
308
|
+
}
|
309
|
+
}
|
310
|
+
}
|
266
311
|
}
|
267
312
|
|
268
|
-
ProductQuantizer
|
313
|
+
ProductQuantizer* read_ProductQuantizer(const char* fname) {
|
269
314
|
FileIOReader reader(fname);
|
270
315
|
return read_ProductQuantizer(&reader);
|
271
316
|
}
|
272
317
|
|
273
|
-
ProductQuantizer
|
274
|
-
|
275
|
-
|
318
|
+
ProductQuantizer* read_ProductQuantizer(IOReader* reader) {
|
319
|
+
ProductQuantizer* pq = new ProductQuantizer();
|
320
|
+
ScopeDeleter1<ProductQuantizer> del(pq);
|
276
321
|
|
277
|
-
|
278
|
-
|
279
|
-
|
322
|
+
read_ProductQuantizer(pq, reader);
|
323
|
+
del.release();
|
324
|
+
return pq;
|
280
325
|
}
|
281
326
|
|
282
|
-
static void read_direct_map
|
327
|
+
static void read_direct_map(DirectMap* dm, IOReader* f) {
|
283
328
|
char maintain_direct_map;
|
284
|
-
READ1
|
329
|
+
READ1(maintain_direct_map);
|
285
330
|
dm->type = (DirectMap::Type)maintain_direct_map;
|
286
|
-
READVECTOR
|
331
|
+
READVECTOR(dm->array);
|
287
332
|
if (dm->type == DirectMap::Hashtable) {
|
288
333
|
using idx_t = Index::idx_t;
|
289
334
|
std::vector<std::pair<idx_t, idx_t>> v;
|
290
|
-
READVECTOR
|
291
|
-
std::unordered_map<idx_t, idx_t
|
292
|
-
map.reserve
|
293
|
-
for (auto it: v) {
|
294
|
-
map
|
335
|
+
READVECTOR(v);
|
336
|
+
std::unordered_map<idx_t, idx_t>& map = dm->hashtable;
|
337
|
+
map.reserve(v.size());
|
338
|
+
for (auto it : v) {
|
339
|
+
map[it.first] = it.second;
|
295
340
|
}
|
296
341
|
}
|
297
|
-
|
298
342
|
}
|
299
343
|
|
300
|
-
|
301
|
-
|
302
|
-
|
303
|
-
|
304
|
-
|
305
|
-
|
306
|
-
READ1
|
307
|
-
|
308
|
-
ivf->quantizer = read_index (f);
|
344
|
+
static void read_ivf_header(
|
345
|
+
IndexIVF* ivf,
|
346
|
+
IOReader* f,
|
347
|
+
std::vector<std::vector<Index::idx_t>>* ids = nullptr) {
|
348
|
+
read_index_header(ivf, f);
|
349
|
+
READ1(ivf->nlist);
|
350
|
+
READ1(ivf->nprobe);
|
351
|
+
ivf->quantizer = read_index(f);
|
309
352
|
ivf->own_fields = true;
|
310
353
|
if (ids) { // used in legacy "Iv" formats
|
311
|
-
ids->resize
|
354
|
+
ids->resize(ivf->nlist);
|
312
355
|
for (size_t i = 0; i < ivf->nlist; i++)
|
313
|
-
READVECTOR
|
356
|
+
READVECTOR((*ids)[i]);
|
314
357
|
}
|
315
|
-
read_direct_map
|
358
|
+
read_direct_map(&ivf->direct_map, f);
|
316
359
|
}
|
317
360
|
|
318
361
|
// used for legacy formats
|
319
|
-
static ArrayInvertedLists
|
320
|
-
|
321
|
-
{
|
322
|
-
ArrayInvertedLists
|
323
|
-
|
324
|
-
std::swap
|
362
|
+
static ArrayInvertedLists* set_array_invlist(
|
363
|
+
IndexIVF* ivf,
|
364
|
+
std::vector<std::vector<Index::idx_t>>& ids) {
|
365
|
+
ArrayInvertedLists* ail =
|
366
|
+
new ArrayInvertedLists(ivf->nlist, ivf->code_size);
|
367
|
+
std::swap(ail->ids, ids);
|
325
368
|
ivf->invlists = ail;
|
326
369
|
ivf->own_invlists = true;
|
327
370
|
return ail;
|
328
371
|
}
|
329
372
|
|
330
|
-
static IndexIVFPQ
|
331
|
-
|
332
|
-
bool legacy = h == fourcc ("IvQR") || h == fourcc ("IvPQ");
|
373
|
+
static IndexIVFPQ* read_ivfpq(IOReader* f, uint32_t h, int io_flags) {
|
374
|
+
bool legacy = h == fourcc("IvQR") || h == fourcc("IvPQ");
|
333
375
|
|
334
|
-
IndexIVFPQR
|
335
|
-
|
336
|
-
|
337
|
-
IndexIVFPQ
|
376
|
+
IndexIVFPQR* ivfpqr = h == fourcc("IvQR") || h == fourcc("IwQR")
|
377
|
+
? new IndexIVFPQR()
|
378
|
+
: nullptr;
|
379
|
+
IndexIVFPQ* ivpq = ivfpqr ? ivfpqr : new IndexIVFPQ();
|
338
380
|
|
339
|
-
std::vector<std::vector<Index::idx_t
|
340
|
-
read_ivf_header
|
341
|
-
READ1
|
342
|
-
READ1
|
343
|
-
read_ProductQuantizer
|
381
|
+
std::vector<std::vector<Index::idx_t>> ids;
|
382
|
+
read_ivf_header(ivpq, f, legacy ? &ids : nullptr);
|
383
|
+
READ1(ivpq->by_residual);
|
384
|
+
READ1(ivpq->code_size);
|
385
|
+
read_ProductQuantizer(&ivpq->pq, f);
|
344
386
|
|
345
387
|
if (legacy) {
|
346
|
-
ArrayInvertedLists
|
388
|
+
ArrayInvertedLists* ail = set_array_invlist(ivpq, ids);
|
347
389
|
for (size_t i = 0; i < ail->nlist; i++)
|
348
|
-
READVECTOR
|
390
|
+
READVECTOR(ail->codes[i]);
|
349
391
|
} else {
|
350
|
-
read_InvertedLists
|
392
|
+
read_InvertedLists(ivpq, f, io_flags);
|
351
393
|
}
|
352
394
|
|
353
395
|
if (ivpq->is_trained) {
|
354
396
|
// precomputed table not stored. It is cheaper to recompute it
|
355
397
|
ivpq->use_precomputed_table = 0;
|
356
398
|
if (ivpq->by_residual)
|
357
|
-
ivpq->precompute_table
|
399
|
+
ivpq->precompute_table();
|
358
400
|
if (ivfpqr) {
|
359
|
-
read_ProductQuantizer
|
360
|
-
READVECTOR
|
361
|
-
READ1
|
401
|
+
read_ProductQuantizer(&ivfpqr->refine_pq, f);
|
402
|
+
READVECTOR(ivfpqr->refine_codes);
|
403
|
+
READ1(ivfpqr->k_factor);
|
362
404
|
}
|
363
405
|
}
|
364
406
|
return ivpq;
|
@@ -366,200 +408,216 @@ static IndexIVFPQ *read_ivfpq (IOReader *f, uint32_t h, int io_flags)
|
|
366
408
|
|
367
409
|
int read_old_fmt_hack = 0;
|
368
410
|
|
369
|
-
Index
|
370
|
-
Index
|
411
|
+
Index* read_index(IOReader* f, int io_flags) {
|
412
|
+
Index* idx = nullptr;
|
371
413
|
uint32_t h;
|
372
|
-
READ1
|
373
|
-
if (h == fourcc
|
374
|
-
IndexFlat
|
375
|
-
if (h == fourcc
|
376
|
-
idxf = new IndexFlatIP
|
414
|
+
READ1(h);
|
415
|
+
if (h == fourcc("IxFI") || h == fourcc("IxF2") || h == fourcc("IxFl")) {
|
416
|
+
IndexFlat* idxf;
|
417
|
+
if (h == fourcc("IxFI")) {
|
418
|
+
idxf = new IndexFlatIP();
|
377
419
|
} else if (h == fourcc("IxF2")) {
|
378
|
-
idxf = new IndexFlatL2
|
420
|
+
idxf = new IndexFlatL2();
|
379
421
|
} else {
|
380
|
-
idxf = new IndexFlat
|
422
|
+
idxf = new IndexFlat();
|
381
423
|
}
|
382
|
-
read_index_header
|
383
|
-
READVECTOR
|
384
|
-
FAISS_THROW_IF_NOT
|
424
|
+
read_index_header(idxf, f);
|
425
|
+
READVECTOR(idxf->xb);
|
426
|
+
FAISS_THROW_IF_NOT(idxf->xb.size() == idxf->ntotal * idxf->d);
|
385
427
|
// leak!
|
386
428
|
idx = idxf;
|
387
429
|
} else if (h == fourcc("IxHE") || h == fourcc("IxHe")) {
|
388
|
-
IndexLSH
|
389
|
-
read_index_header
|
390
|
-
READ1
|
391
|
-
READ1
|
392
|
-
READ1
|
393
|
-
READVECTOR
|
394
|
-
READ1
|
430
|
+
IndexLSH* idxl = new IndexLSH();
|
431
|
+
read_index_header(idxl, f);
|
432
|
+
READ1(idxl->nbits);
|
433
|
+
READ1(idxl->rotate_data);
|
434
|
+
READ1(idxl->train_thresholds);
|
435
|
+
READVECTOR(idxl->thresholds);
|
436
|
+
READ1(idxl->bytes_per_vec);
|
395
437
|
if (h == fourcc("IxHE")) {
|
396
|
-
FAISS_THROW_IF_NOT_FMT
|
397
|
-
|
398
|
-
|
399
|
-
|
438
|
+
FAISS_THROW_IF_NOT_FMT(
|
439
|
+
idxl->nbits % 64 == 0,
|
440
|
+
"can only read old format IndexLSH with "
|
441
|
+
"nbits multiple of 64 (got %d)",
|
442
|
+
(int)idxl->nbits);
|
400
443
|
// leak
|
401
444
|
idxl->bytes_per_vec *= 8;
|
402
445
|
}
|
403
446
|
{
|
404
|
-
RandomRotationMatrix
|
405
|
-
|
447
|
+
RandomRotationMatrix* rrot = dynamic_cast<RandomRotationMatrix*>(
|
448
|
+
read_VectorTransform(f));
|
406
449
|
FAISS_THROW_IF_NOT_MSG(rrot, "expected a random rotation");
|
407
450
|
idxl->rrot = *rrot;
|
408
451
|
delete rrot;
|
409
452
|
}
|
410
|
-
READVECTOR
|
411
|
-
FAISS_THROW_IF_NOT
|
412
|
-
|
413
|
-
FAISS_THROW_IF_NOT
|
414
|
-
|
453
|
+
READVECTOR(idxl->codes);
|
454
|
+
FAISS_THROW_IF_NOT(
|
455
|
+
idxl->rrot.d_in == idxl->d && idxl->rrot.d_out == idxl->nbits);
|
456
|
+
FAISS_THROW_IF_NOT(
|
457
|
+
idxl->codes.size() == idxl->ntotal * idxl->bytes_per_vec);
|
415
458
|
idx = idxl;
|
416
|
-
} else if (
|
417
|
-
|
459
|
+
} else if (
|
460
|
+
h == fourcc("IxPQ") || h == fourcc("IxPo") || h == fourcc("IxPq")) {
|
418
461
|
// IxPQ and IxPo were merged into the same IndexPQ object
|
419
|
-
IndexPQ
|
420
|
-
read_index_header
|
421
|
-
read_ProductQuantizer
|
422
|
-
READVECTOR
|
423
|
-
if (h == fourcc
|
424
|
-
READ1
|
425
|
-
READ1
|
426
|
-
READ1
|
462
|
+
IndexPQ* idxp = new IndexPQ();
|
463
|
+
read_index_header(idxp, f);
|
464
|
+
read_ProductQuantizer(&idxp->pq, f);
|
465
|
+
READVECTOR(idxp->codes);
|
466
|
+
if (h == fourcc("IxPo") || h == fourcc("IxPq")) {
|
467
|
+
READ1(idxp->search_type);
|
468
|
+
READ1(idxp->encode_signs);
|
469
|
+
READ1(idxp->polysemous_ht);
|
427
470
|
}
|
428
471
|
// Old versoins of PQ all had metric_type set to INNER_PRODUCT
|
429
472
|
// when they were in fact using L2. Therefore, we force metric type
|
430
473
|
// to L2 when the old format is detected
|
431
|
-
if (h == fourcc
|
474
|
+
if (h == fourcc("IxPQ") || h == fourcc("IxPo")) {
|
432
475
|
idxp->metric_type = METRIC_L2;
|
433
476
|
}
|
434
477
|
idx = idxp;
|
435
|
-
} else if (h == fourcc
|
436
|
-
|
437
|
-
|
438
|
-
|
478
|
+
} else if (h == fourcc("IxRQ")) {
|
479
|
+
IndexResidual* idxr = new IndexResidual();
|
480
|
+
read_index_header(idxr, f);
|
481
|
+
read_ResidualQuantizer(&idxr->rq, f);
|
482
|
+
READ1(idxr->search_type);
|
483
|
+
READ1(idxr->norm_min);
|
484
|
+
READ1(idxr->norm_max);
|
485
|
+
READ1(idxr->code_size);
|
486
|
+
READVECTOR(idxr->codes);
|
487
|
+
idx = idxr;
|
488
|
+
} else if (h == fourcc("ImRQ")) {
|
489
|
+
ResidualCoarseQuantizer* idxr = new ResidualCoarseQuantizer();
|
490
|
+
read_index_header(idxr, f);
|
491
|
+
read_ResidualQuantizer(&idxr->rq, f);
|
492
|
+
READ1(idxr->beam_factor);
|
493
|
+
idxr->set_beam_factor(idxr->beam_factor);
|
494
|
+
idx = idxr;
|
495
|
+
} else if (h == fourcc("IvFl") || h == fourcc("IvFL")) { // legacy
|
496
|
+
IndexIVFFlat* ivfl = new IndexIVFFlat();
|
497
|
+
std::vector<std::vector<Index::idx_t>> ids;
|
498
|
+
read_ivf_header(ivfl, f, &ids);
|
439
499
|
ivfl->code_size = ivfl->d * sizeof(float);
|
440
|
-
ArrayInvertedLists
|
500
|
+
ArrayInvertedLists* ail = set_array_invlist(ivfl, ids);
|
441
501
|
|
442
|
-
if (h == fourcc
|
502
|
+
if (h == fourcc("IvFL")) {
|
443
503
|
for (size_t i = 0; i < ivfl->nlist; i++) {
|
444
|
-
READVECTOR
|
504
|
+
READVECTOR(ail->codes[i]);
|
445
505
|
}
|
446
506
|
} else { // old format
|
447
507
|
for (size_t i = 0; i < ivfl->nlist; i++) {
|
448
508
|
std::vector<float> vec;
|
449
|
-
READVECTOR
|
509
|
+
READVECTOR(vec);
|
450
510
|
ail->codes[i].resize(vec.size() * sizeof(float));
|
451
|
-
memcpy(ail->codes[i].data(), vec.data(),
|
452
|
-
ail->codes[i].size());
|
511
|
+
memcpy(ail->codes[i].data(), vec.data(), ail->codes[i].size());
|
453
512
|
}
|
454
513
|
}
|
455
514
|
idx = ivfl;
|
456
|
-
} else if (h == fourcc
|
457
|
-
IndexIVFFlatDedup
|
458
|
-
read_ivf_header
|
515
|
+
} else if (h == fourcc("IwFd")) {
|
516
|
+
IndexIVFFlatDedup* ivfl = new IndexIVFFlatDedup();
|
517
|
+
read_ivf_header(ivfl, f);
|
459
518
|
ivfl->code_size = ivfl->d * sizeof(float);
|
460
519
|
{
|
461
520
|
std::vector<Index::idx_t> tab;
|
462
|
-
READVECTOR
|
521
|
+
READVECTOR(tab);
|
463
522
|
for (long i = 0; i < tab.size(); i += 2) {
|
464
|
-
std::pair<Index::idx_t, Index::idx_t>
|
465
|
-
|
466
|
-
ivfl->instances.insert (pair);
|
523
|
+
std::pair<Index::idx_t, Index::idx_t> pair(tab[i], tab[i + 1]);
|
524
|
+
ivfl->instances.insert(pair);
|
467
525
|
}
|
468
526
|
}
|
469
|
-
read_InvertedLists
|
527
|
+
read_InvertedLists(ivfl, f, io_flags);
|
470
528
|
idx = ivfl;
|
471
|
-
} else if (h == fourcc
|
472
|
-
IndexIVFFlat
|
473
|
-
read_ivf_header
|
529
|
+
} else if (h == fourcc("IwFl")) {
|
530
|
+
IndexIVFFlat* ivfl = new IndexIVFFlat();
|
531
|
+
read_ivf_header(ivfl, f);
|
474
532
|
ivfl->code_size = ivfl->d * sizeof(float);
|
475
|
-
read_InvertedLists
|
533
|
+
read_InvertedLists(ivfl, f, io_flags);
|
476
534
|
idx = ivfl;
|
477
|
-
} else if (h == fourcc
|
478
|
-
IndexScalarQuantizer
|
479
|
-
read_index_header
|
480
|
-
read_ScalarQuantizer
|
481
|
-
READVECTOR
|
535
|
+
} else if (h == fourcc("IxSQ")) {
|
536
|
+
IndexScalarQuantizer* idxs = new IndexScalarQuantizer();
|
537
|
+
read_index_header(idxs, f);
|
538
|
+
read_ScalarQuantizer(&idxs->sq, f);
|
539
|
+
READVECTOR(idxs->codes);
|
482
540
|
idxs->code_size = idxs->sq.code_size;
|
483
541
|
idx = idxs;
|
484
|
-
} else if (h == fourcc
|
542
|
+
} else if (h == fourcc("IxLa")) {
|
485
543
|
int d, nsq, scale_nbit, r2;
|
486
|
-
READ1
|
487
|
-
READ1
|
488
|
-
READ1
|
489
|
-
READ1
|
490
|
-
IndexLattice
|
491
|
-
read_index_header
|
492
|
-
READVECTOR
|
544
|
+
READ1(d);
|
545
|
+
READ1(nsq);
|
546
|
+
READ1(scale_nbit);
|
547
|
+
READ1(r2);
|
548
|
+
IndexLattice* idxl = new IndexLattice(d, nsq, scale_nbit, r2);
|
549
|
+
read_index_header(idxl, f);
|
550
|
+
READVECTOR(idxl->trained);
|
493
551
|
idx = idxl;
|
494
|
-
} else if(h == fourcc
|
495
|
-
IndexIVFScalarQuantizer
|
496
|
-
std::vector<std::vector<Index::idx_t
|
497
|
-
read_ivf_header
|
498
|
-
read_ScalarQuantizer
|
499
|
-
READ1
|
500
|
-
ArrayInvertedLists
|
501
|
-
for(int i = 0; i < ivsc->nlist; i++)
|
502
|
-
READVECTOR
|
552
|
+
} else if (h == fourcc("IvSQ")) { // legacy
|
553
|
+
IndexIVFScalarQuantizer* ivsc = new IndexIVFScalarQuantizer();
|
554
|
+
std::vector<std::vector<Index::idx_t>> ids;
|
555
|
+
read_ivf_header(ivsc, f, &ids);
|
556
|
+
read_ScalarQuantizer(&ivsc->sq, f);
|
557
|
+
READ1(ivsc->code_size);
|
558
|
+
ArrayInvertedLists* ail = set_array_invlist(ivsc, ids);
|
559
|
+
for (int i = 0; i < ivsc->nlist; i++)
|
560
|
+
READVECTOR(ail->codes[i]);
|
503
561
|
idx = ivsc;
|
504
|
-
} else if(h == fourcc
|
505
|
-
IndexIVFScalarQuantizer
|
506
|
-
read_ivf_header
|
507
|
-
read_ScalarQuantizer
|
508
|
-
READ1
|
509
|
-
if (h == fourcc
|
562
|
+
} else if (h == fourcc("IwSQ") || h == fourcc("IwSq")) {
|
563
|
+
IndexIVFScalarQuantizer* ivsc = new IndexIVFScalarQuantizer();
|
564
|
+
read_ivf_header(ivsc, f);
|
565
|
+
read_ScalarQuantizer(&ivsc->sq, f);
|
566
|
+
READ1(ivsc->code_size);
|
567
|
+
if (h == fourcc("IwSQ")) {
|
510
568
|
ivsc->by_residual = true;
|
511
569
|
} else {
|
512
|
-
READ1
|
570
|
+
READ1(ivsc->by_residual);
|
513
571
|
}
|
514
|
-
read_InvertedLists
|
572
|
+
read_InvertedLists(ivsc, f, io_flags);
|
515
573
|
idx = ivsc;
|
516
|
-
} else if(h == fourcc
|
517
|
-
IndexIVFSpectralHash
|
518
|
-
read_ivf_header
|
519
|
-
ivsp->vt = read_VectorTransform
|
574
|
+
} else if (h == fourcc("IwSh")) {
|
575
|
+
IndexIVFSpectralHash* ivsp = new IndexIVFSpectralHash();
|
576
|
+
read_ivf_header(ivsp, f);
|
577
|
+
ivsp->vt = read_VectorTransform(f);
|
520
578
|
ivsp->own_fields = true;
|
521
|
-
READ1
|
579
|
+
READ1(ivsp->nbit);
|
522
580
|
// not stored by write_ivf_header
|
523
581
|
ivsp->code_size = (ivsp->nbit + 7) / 8;
|
524
|
-
READ1
|
525
|
-
READ1
|
526
|
-
READVECTOR
|
527
|
-
read_InvertedLists
|
582
|
+
READ1(ivsp->period);
|
583
|
+
READ1(ivsp->threshold_type);
|
584
|
+
READVECTOR(ivsp->trained);
|
585
|
+
read_InvertedLists(ivsp, f, io_flags);
|
528
586
|
idx = ivsp;
|
529
|
-
} else if
|
530
|
-
|
531
|
-
|
532
|
-
idx = read_ivfpq
|
587
|
+
} else if (
|
588
|
+
h == fourcc("IvPQ") || h == fourcc("IvQR") || h == fourcc("IwPQ") ||
|
589
|
+
h == fourcc("IwQR")) {
|
590
|
+
idx = read_ivfpq(f, h, io_flags);
|
533
591
|
|
534
|
-
} else if(h == fourcc
|
535
|
-
IndexPreTransform
|
592
|
+
} else if (h == fourcc("IxPT")) {
|
593
|
+
IndexPreTransform* ixpt = new IndexPreTransform();
|
536
594
|
ixpt->own_fields = true;
|
537
|
-
read_index_header
|
595
|
+
read_index_header(ixpt, f);
|
538
596
|
int nt;
|
539
597
|
if (read_old_fmt_hack == 2) {
|
540
598
|
nt = 1;
|
541
599
|
} else {
|
542
|
-
READ1
|
600
|
+
READ1(nt);
|
543
601
|
}
|
544
602
|
for (int i = 0; i < nt; i++) {
|
545
|
-
ixpt->chain.push_back
|
603
|
+
ixpt->chain.push_back(read_VectorTransform(f));
|
546
604
|
}
|
547
|
-
ixpt->index = read_index
|
605
|
+
ixpt->index = read_index(f, io_flags);
|
548
606
|
idx = ixpt;
|
549
|
-
} else if(h == fourcc
|
550
|
-
MultiIndexQuantizer
|
551
|
-
read_index_header
|
552
|
-
read_ProductQuantizer
|
607
|
+
} else if (h == fourcc("Imiq")) {
|
608
|
+
MultiIndexQuantizer* imiq = new MultiIndexQuantizer();
|
609
|
+
read_index_header(imiq, f);
|
610
|
+
read_ProductQuantizer(&imiq->pq, f);
|
553
611
|
idx = imiq;
|
554
|
-
} else if(h == fourcc
|
555
|
-
IndexRefine
|
556
|
-
read_index_header
|
612
|
+
} else if (h == fourcc("IxRF")) {
|
613
|
+
IndexRefine* idxrf = new IndexRefine();
|
614
|
+
read_index_header(idxrf, f);
|
557
615
|
idxrf->base_index = read_index(f, io_flags);
|
558
616
|
idxrf->refine_index = read_index(f, io_flags);
|
559
|
-
READ1
|
617
|
+
READ1(idxrf->k_factor);
|
560
618
|
if (dynamic_cast<IndexFlat*>(idxrf->refine_index)) {
|
561
619
|
// then make a RefineFlat with it
|
562
|
-
IndexRefine
|
620
|
+
IndexRefine* idxrf_old = idxrf;
|
563
621
|
idxrf = new IndexRefineFlat();
|
564
622
|
*idxrf = *idxrf_old;
|
565
623
|
delete idxrf_old;
|
@@ -567,248 +625,260 @@ Index *read_index (IOReader *f, int io_flags) {
|
|
567
625
|
idxrf->own_fields = true;
|
568
626
|
idxrf->own_refine_index = true;
|
569
627
|
idx = idxrf;
|
570
|
-
} else if(h == fourcc
|
571
|
-
bool is_map2 = h == fourcc
|
572
|
-
IndexIDMap
|
573
|
-
read_index_header
|
574
|
-
idxmap->index = read_index
|
628
|
+
} else if (h == fourcc("IxMp") || h == fourcc("IxM2")) {
|
629
|
+
bool is_map2 = h == fourcc("IxM2");
|
630
|
+
IndexIDMap* idxmap = is_map2 ? new IndexIDMap2() : new IndexIDMap();
|
631
|
+
read_index_header(idxmap, f);
|
632
|
+
idxmap->index = read_index(f, io_flags);
|
575
633
|
idxmap->own_fields = true;
|
576
|
-
READVECTOR
|
634
|
+
READVECTOR(idxmap->id_map);
|
577
635
|
if (is_map2) {
|
578
|
-
static_cast<IndexIDMap2*>(idxmap)->construct_rev_map
|
636
|
+
static_cast<IndexIDMap2*>(idxmap)->construct_rev_map();
|
579
637
|
}
|
580
638
|
idx = idxmap;
|
581
|
-
} else if (h == fourcc
|
582
|
-
Index2Layer
|
583
|
-
read_index_header
|
584
|
-
idxp->q1.quantizer = read_index
|
585
|
-
READ1
|
586
|
-
READ1
|
587
|
-
read_ProductQuantizer
|
588
|
-
READ1
|
589
|
-
READ1
|
590
|
-
READ1
|
591
|
-
READVECTOR
|
639
|
+
} else if (h == fourcc("Ix2L")) {
|
640
|
+
Index2Layer* idxp = new Index2Layer();
|
641
|
+
read_index_header(idxp, f);
|
642
|
+
idxp->q1.quantizer = read_index(f, io_flags);
|
643
|
+
READ1(idxp->q1.nlist);
|
644
|
+
READ1(idxp->q1.quantizer_trains_alone);
|
645
|
+
read_ProductQuantizer(&idxp->pq, f);
|
646
|
+
READ1(idxp->code_size_1);
|
647
|
+
READ1(idxp->code_size_2);
|
648
|
+
READ1(idxp->code_size);
|
649
|
+
READVECTOR(idxp->codes);
|
592
650
|
idx = idxp;
|
593
|
-
} else if
|
594
|
-
|
595
|
-
|
596
|
-
|
597
|
-
if (h == fourcc("
|
598
|
-
|
599
|
-
if (h == fourcc("
|
600
|
-
|
601
|
-
|
602
|
-
|
651
|
+
} else if (
|
652
|
+
h == fourcc("IHNf") || h == fourcc("IHNp") || h == fourcc("IHNs") ||
|
653
|
+
h == fourcc("IHN2")) {
|
654
|
+
IndexHNSW* idxhnsw = nullptr;
|
655
|
+
if (h == fourcc("IHNf"))
|
656
|
+
idxhnsw = new IndexHNSWFlat();
|
657
|
+
if (h == fourcc("IHNp"))
|
658
|
+
idxhnsw = new IndexHNSWPQ();
|
659
|
+
if (h == fourcc("IHNs"))
|
660
|
+
idxhnsw = new IndexHNSWSQ();
|
661
|
+
if (h == fourcc("IHN2"))
|
662
|
+
idxhnsw = new IndexHNSW2Level();
|
663
|
+
read_index_header(idxhnsw, f);
|
664
|
+
read_HNSW(&idxhnsw->hnsw, f);
|
665
|
+
idxhnsw->storage = read_index(f, io_flags);
|
603
666
|
idxhnsw->own_fields = true;
|
604
667
|
if (h == fourcc("IHNp")) {
|
605
|
-
dynamic_cast<IndexPQ*>(idxhnsw->storage)->pq.compute_sdc_table
|
668
|
+
dynamic_cast<IndexPQ*>(idxhnsw->storage)->pq.compute_sdc_table();
|
606
669
|
}
|
607
670
|
idx = idxhnsw;
|
608
|
-
} else if(h == fourcc("
|
609
|
-
|
610
|
-
read_index_header
|
611
|
-
|
612
|
-
READ1
|
613
|
-
READ1
|
614
|
-
READ1
|
615
|
-
READ1
|
616
|
-
READ1
|
617
|
-
|
671
|
+
} else if (h == fourcc("INSf")) {
|
672
|
+
IndexNSG* idxnsg = new IndexNSGFlat();
|
673
|
+
read_index_header(idxnsg, f);
|
674
|
+
READ1(idxnsg->GK);
|
675
|
+
READ1(idxnsg->build_type);
|
676
|
+
READ1(idxnsg->nndescent_S);
|
677
|
+
READ1(idxnsg->nndescent_R);
|
678
|
+
READ1(idxnsg->nndescent_L);
|
679
|
+
READ1(idxnsg->nndescent_iter);
|
680
|
+
read_NSG(&idxnsg->nsg, f);
|
681
|
+
idxnsg->storage = read_index(f, io_flags);
|
682
|
+
idxnsg->own_fields = true;
|
683
|
+
idx = idxnsg;
|
684
|
+
} else if (h == fourcc("IPfs")) {
|
685
|
+
IndexPQFastScan* idxpqfs = new IndexPQFastScan();
|
686
|
+
read_index_header(idxpqfs, f);
|
687
|
+
read_ProductQuantizer(&idxpqfs->pq, f);
|
688
|
+
READ1(idxpqfs->implem);
|
689
|
+
READ1(idxpqfs->bbs);
|
690
|
+
READ1(idxpqfs->qbs);
|
691
|
+
READ1(idxpqfs->ntotal2);
|
692
|
+
READ1(idxpqfs->M2);
|
693
|
+
READVECTOR(idxpqfs->codes);
|
618
694
|
idx = idxpqfs;
|
619
695
|
|
620
696
|
} else if (h == fourcc("IwPf")) {
|
621
|
-
IndexIVFPQFastScan
|
622
|
-
read_ivf_header
|
623
|
-
READ1
|
624
|
-
READ1
|
625
|
-
READ1
|
626
|
-
READ1
|
627
|
-
READ1
|
628
|
-
READ1
|
629
|
-
read_ProductQuantizer
|
630
|
-
read_InvertedLists
|
697
|
+
IndexIVFPQFastScan* ivpq = new IndexIVFPQFastScan();
|
698
|
+
read_ivf_header(ivpq, f);
|
699
|
+
READ1(ivpq->by_residual);
|
700
|
+
READ1(ivpq->code_size);
|
701
|
+
READ1(ivpq->bbs);
|
702
|
+
READ1(ivpq->M2);
|
703
|
+
READ1(ivpq->implem);
|
704
|
+
READ1(ivpq->qbs2);
|
705
|
+
read_ProductQuantizer(&ivpq->pq, f);
|
706
|
+
read_InvertedLists(ivpq, f, io_flags);
|
631
707
|
ivpq->precompute_table();
|
632
708
|
idx = ivpq;
|
633
709
|
} else {
|
634
710
|
FAISS_THROW_FMT(
|
635
|
-
|
636
|
-
|
637
|
-
|
711
|
+
"Index type 0x%08x (\"%s\") not recognized",
|
712
|
+
h,
|
713
|
+
fourcc_inv_printable(h).c_str());
|
638
714
|
idx = nullptr;
|
639
715
|
}
|
640
716
|
return idx;
|
641
717
|
}
|
642
718
|
|
643
|
-
|
644
|
-
Index *read_index (FILE * f, int io_flags) {
|
719
|
+
Index* read_index(FILE* f, int io_flags) {
|
645
720
|
FileIOReader reader(f);
|
646
721
|
return read_index(&reader, io_flags);
|
647
722
|
}
|
648
723
|
|
649
|
-
Index
|
724
|
+
Index* read_index(const char* fname, int io_flags) {
|
650
725
|
FileIOReader reader(fname);
|
651
|
-
Index
|
726
|
+
Index* idx = read_index(&reader, io_flags);
|
652
727
|
return idx;
|
653
728
|
}
|
654
729
|
|
655
|
-
VectorTransform
|
730
|
+
VectorTransform* read_VectorTransform(const char* fname) {
|
656
731
|
FileIOReader reader(fname);
|
657
|
-
VectorTransform
|
732
|
+
VectorTransform* vt = read_VectorTransform(&reader);
|
658
733
|
return vt;
|
659
734
|
}
|
660
735
|
|
661
|
-
|
662
|
-
|
663
736
|
/*************************************************************
|
664
737
|
* Read binary indexes
|
665
738
|
**************************************************************/
|
666
739
|
|
667
|
-
static void read_InvertedLists
|
668
|
-
|
669
|
-
|
670
|
-
|
671
|
-
|
740
|
+
static void read_InvertedLists(IndexBinaryIVF* ivf, IOReader* f, int io_flags) {
|
741
|
+
InvertedLists* ils = read_InvertedLists(f, io_flags);
|
742
|
+
FAISS_THROW_IF_NOT(
|
743
|
+
!ils ||
|
744
|
+
(ils->nlist == ivf->nlist && ils->code_size == ivf->code_size));
|
672
745
|
ivf->invlists = ils;
|
673
746
|
ivf->own_invlists = true;
|
674
747
|
}
|
675
748
|
|
676
|
-
|
677
|
-
|
678
|
-
|
679
|
-
READ1
|
680
|
-
READ1
|
681
|
-
READ1
|
682
|
-
READ1 (idx->is_trained);
|
683
|
-
READ1 (idx->metric_type);
|
749
|
+
static void read_index_binary_header(IndexBinary* idx, IOReader* f) {
|
750
|
+
READ1(idx->d);
|
751
|
+
READ1(idx->code_size);
|
752
|
+
READ1(idx->ntotal);
|
753
|
+
READ1(idx->is_trained);
|
754
|
+
READ1(idx->metric_type);
|
684
755
|
idx->verbose = false;
|
685
756
|
}
|
686
757
|
|
687
|
-
static void read_binary_ivf_header
|
688
|
-
|
689
|
-
|
690
|
-
{
|
691
|
-
read_index_binary_header
|
692
|
-
READ1
|
693
|
-
READ1
|
694
|
-
ivf->quantizer = read_index_binary
|
758
|
+
static void read_binary_ivf_header(
|
759
|
+
IndexBinaryIVF* ivf,
|
760
|
+
IOReader* f,
|
761
|
+
std::vector<std::vector<Index::idx_t>>* ids = nullptr) {
|
762
|
+
read_index_binary_header(ivf, f);
|
763
|
+
READ1(ivf->nlist);
|
764
|
+
READ1(ivf->nprobe);
|
765
|
+
ivf->quantizer = read_index_binary(f);
|
695
766
|
ivf->own_fields = true;
|
696
767
|
if (ids) { // used in legacy "Iv" formats
|
697
|
-
ids->resize
|
768
|
+
ids->resize(ivf->nlist);
|
698
769
|
for (size_t i = 0; i < ivf->nlist; i++)
|
699
|
-
READVECTOR
|
770
|
+
READVECTOR((*ids)[i]);
|
700
771
|
}
|
701
|
-
read_direct_map
|
772
|
+
read_direct_map(&ivf->direct_map, f);
|
702
773
|
}
|
703
774
|
|
704
|
-
static void read_binary_hash_invlists
|
705
|
-
IndexBinaryHash::InvertedListMap
|
706
|
-
int b,
|
707
|
-
{
|
775
|
+
static void read_binary_hash_invlists(
|
776
|
+
IndexBinaryHash::InvertedListMap& invlists,
|
777
|
+
int b,
|
778
|
+
IOReader* f) {
|
708
779
|
size_t sz;
|
709
|
-
READ1
|
780
|
+
READ1(sz);
|
710
781
|
int il_nbit = 0;
|
711
|
-
READ1
|
782
|
+
READ1(il_nbit);
|
712
783
|
// buffer for bitstrings
|
713
784
|
std::vector<uint8_t> buf((b + il_nbit) * sz);
|
714
|
-
READVECTOR
|
715
|
-
BitstringReader rd
|
716
|
-
invlists.reserve
|
785
|
+
READVECTOR(buf);
|
786
|
+
BitstringReader rd(buf.data(), buf.size());
|
787
|
+
invlists.reserve(sz);
|
717
788
|
for (size_t i = 0; i < sz; i++) {
|
718
789
|
uint64_t hash = rd.read(b);
|
719
790
|
uint64_t ilsz = rd.read(il_nbit);
|
720
|
-
auto
|
721
|
-
READVECTOR
|
722
|
-
FAISS_THROW_IF_NOT
|
723
|
-
READVECTOR
|
791
|
+
auto& il = invlists[hash];
|
792
|
+
READVECTOR(il.ids);
|
793
|
+
FAISS_THROW_IF_NOT(il.ids.size() == ilsz);
|
794
|
+
READVECTOR(il.vecs);
|
724
795
|
}
|
725
796
|
}
|
726
797
|
|
727
798
|
static void read_binary_multi_hash_map(
|
728
|
-
IndexBinaryMultiHash::Map
|
729
|
-
int b,
|
730
|
-
|
731
|
-
{
|
799
|
+
IndexBinaryMultiHash::Map& map,
|
800
|
+
int b,
|
801
|
+
size_t ntotal,
|
802
|
+
IOReader* f) {
|
732
803
|
int id_bits;
|
733
804
|
size_t sz;
|
734
|
-
READ1
|
735
|
-
READ1
|
805
|
+
READ1(id_bits);
|
806
|
+
READ1(sz);
|
736
807
|
std::vector<uint8_t> buf;
|
737
|
-
READVECTOR
|
808
|
+
READVECTOR(buf);
|
738
809
|
size_t nbit = (b + id_bits) * sz + ntotal * id_bits;
|
739
|
-
FAISS_THROW_IF_NOT
|
740
|
-
BitstringReader rd
|
741
|
-
map.reserve
|
810
|
+
FAISS_THROW_IF_NOT(buf.size() == (nbit + 7) / 8);
|
811
|
+
BitstringReader rd(buf.data(), buf.size());
|
812
|
+
map.reserve(sz);
|
742
813
|
for (size_t i = 0; i < sz; i++) {
|
743
814
|
uint64_t hash = rd.read(b);
|
744
815
|
uint64_t ilsz = rd.read(id_bits);
|
745
|
-
auto
|
816
|
+
auto& il = map[hash];
|
746
817
|
for (size_t j = 0; j < ilsz; j++) {
|
747
|
-
il.push_back
|
818
|
+
il.push_back(rd.read(id_bits));
|
748
819
|
}
|
749
820
|
}
|
750
821
|
}
|
751
822
|
|
752
|
-
|
753
|
-
|
754
|
-
IndexBinary *read_index_binary (IOReader *f, int io_flags) {
|
755
|
-
IndexBinary * idx = nullptr;
|
823
|
+
IndexBinary* read_index_binary(IOReader* f, int io_flags) {
|
824
|
+
IndexBinary* idx = nullptr;
|
756
825
|
uint32_t h;
|
757
|
-
READ1
|
758
|
-
if (h == fourcc
|
759
|
-
IndexBinaryFlat
|
760
|
-
read_index_binary_header
|
761
|
-
READVECTOR
|
762
|
-
FAISS_THROW_IF_NOT
|
826
|
+
READ1(h);
|
827
|
+
if (h == fourcc("IBxF")) {
|
828
|
+
IndexBinaryFlat* idxf = new IndexBinaryFlat();
|
829
|
+
read_index_binary_header(idxf, f);
|
830
|
+
READVECTOR(idxf->xb);
|
831
|
+
FAISS_THROW_IF_NOT(idxf->xb.size() == idxf->ntotal * idxf->code_size);
|
763
832
|
// leak!
|
764
833
|
idx = idxf;
|
765
|
-
} else if (h == fourcc
|
766
|
-
IndexBinaryIVF
|
767
|
-
read_binary_ivf_header
|
768
|
-
read_InvertedLists
|
834
|
+
} else if (h == fourcc("IBwF")) {
|
835
|
+
IndexBinaryIVF* ivf = new IndexBinaryIVF();
|
836
|
+
read_binary_ivf_header(ivf, f);
|
837
|
+
read_InvertedLists(ivf, f, io_flags);
|
769
838
|
idx = ivf;
|
770
|
-
} else if (h == fourcc
|
771
|
-
IndexBinaryFromFloat
|
772
|
-
read_index_binary_header
|
839
|
+
} else if (h == fourcc("IBFf")) {
|
840
|
+
IndexBinaryFromFloat* idxff = new IndexBinaryFromFloat();
|
841
|
+
read_index_binary_header(idxff, f);
|
773
842
|
idxff->own_fields = true;
|
774
|
-
idxff->index = read_index
|
843
|
+
idxff->index = read_index(f, io_flags);
|
775
844
|
idx = idxff;
|
776
|
-
} else if (h == fourcc
|
777
|
-
IndexBinaryHNSW
|
778
|
-
read_index_binary_header
|
779
|
-
read_HNSW
|
780
|
-
idxhnsw->storage = read_index_binary
|
845
|
+
} else if (h == fourcc("IBHf")) {
|
846
|
+
IndexBinaryHNSW* idxhnsw = new IndexBinaryHNSW();
|
847
|
+
read_index_binary_header(idxhnsw, f);
|
848
|
+
read_HNSW(&idxhnsw->hnsw, f);
|
849
|
+
idxhnsw->storage = read_index_binary(f, io_flags);
|
781
850
|
idxhnsw->own_fields = true;
|
782
851
|
idx = idxhnsw;
|
783
|
-
} else if(h == fourcc
|
784
|
-
bool is_map2 = h == fourcc
|
785
|
-
IndexBinaryIDMap
|
786
|
-
|
787
|
-
read_index_binary_header
|
788
|
-
idxmap->index = read_index_binary
|
852
|
+
} else if (h == fourcc("IBMp") || h == fourcc("IBM2")) {
|
853
|
+
bool is_map2 = h == fourcc("IBM2");
|
854
|
+
IndexBinaryIDMap* idxmap =
|
855
|
+
is_map2 ? new IndexBinaryIDMap2() : new IndexBinaryIDMap();
|
856
|
+
read_index_binary_header(idxmap, f);
|
857
|
+
idxmap->index = read_index_binary(f, io_flags);
|
789
858
|
idxmap->own_fields = true;
|
790
|
-
READVECTOR
|
859
|
+
READVECTOR(idxmap->id_map);
|
791
860
|
if (is_map2) {
|
792
|
-
static_cast<IndexBinaryIDMap2*>(idxmap)->construct_rev_map
|
861
|
+
static_cast<IndexBinaryIDMap2*>(idxmap)->construct_rev_map();
|
793
862
|
}
|
794
863
|
idx = idxmap;
|
795
|
-
} else if(h == fourcc("IBHh")) {
|
796
|
-
IndexBinaryHash
|
797
|
-
read_index_binary_header
|
798
|
-
READ1
|
799
|
-
READ1
|
864
|
+
} else if (h == fourcc("IBHh")) {
|
865
|
+
IndexBinaryHash* idxh = new IndexBinaryHash();
|
866
|
+
read_index_binary_header(idxh, f);
|
867
|
+
READ1(idxh->b);
|
868
|
+
READ1(idxh->nflip);
|
800
869
|
read_binary_hash_invlists(idxh->invlists, idxh->b, f);
|
801
870
|
idx = idxh;
|
802
|
-
} else if(h == fourcc("IBHm")) {
|
803
|
-
IndexBinaryMultiHash* idxmh = new IndexBinaryMultiHash
|
804
|
-
read_index_binary_header
|
805
|
-
idxmh->storage = dynamic_cast<IndexBinaryFlat*>
|
806
|
-
FAISS_THROW_IF_NOT(
|
871
|
+
} else if (h == fourcc("IBHm")) {
|
872
|
+
IndexBinaryMultiHash* idxmh = new IndexBinaryMultiHash();
|
873
|
+
read_index_binary_header(idxmh, f);
|
874
|
+
idxmh->storage = dynamic_cast<IndexBinaryFlat*>(read_index_binary(f));
|
875
|
+
FAISS_THROW_IF_NOT(
|
876
|
+
idxmh->storage && idxmh->storage->ntotal == idxmh->ntotal);
|
807
877
|
idxmh->own_fields = true;
|
808
|
-
READ1
|
809
|
-
READ1
|
810
|
-
READ1
|
811
|
-
idxmh->maps.resize
|
878
|
+
READ1(idxmh->b);
|
879
|
+
READ1(idxmh->nhash);
|
880
|
+
READ1(idxmh->nflip);
|
881
|
+
idxmh->maps.resize(idxmh->nhash);
|
812
882
|
for (int i = 0; i < idxmh->nhash; i++) {
|
813
883
|
read_binary_multi_hash_map(
|
814
884
|
idxmh->maps[i], idxmh->b, idxmh->ntotal, f);
|
@@ -816,25 +886,23 @@ IndexBinary *read_index_binary (IOReader *f, int io_flags) {
|
|
816
886
|
idx = idxmh;
|
817
887
|
} else {
|
818
888
|
FAISS_THROW_FMT(
|
819
|
-
|
820
|
-
|
821
|
-
|
889
|
+
"Index type %08x (\"%s\") not recognized",
|
890
|
+
h,
|
891
|
+
fourcc_inv_printable(h).c_str());
|
822
892
|
idx = nullptr;
|
823
893
|
}
|
824
894
|
return idx;
|
825
895
|
}
|
826
896
|
|
827
|
-
IndexBinary
|
897
|
+
IndexBinary* read_index_binary(FILE* f, int io_flags) {
|
828
898
|
FileIOReader reader(f);
|
829
899
|
return read_index_binary(&reader, io_flags);
|
830
900
|
}
|
831
901
|
|
832
|
-
IndexBinary
|
902
|
+
IndexBinary* read_index_binary(const char* fname, int io_flags) {
|
833
903
|
FileIOReader reader(fname);
|
834
|
-
IndexBinary
|
904
|
+
IndexBinary* idx = read_index_binary(&reader, io_flags);
|
835
905
|
return idx;
|
836
906
|
}
|
837
907
|
|
838
|
-
|
839
|
-
|
840
908
|
} // namespace faiss
|