faiss 0.1.0 → 0.1.1

data/vendor/faiss/tests/test_merge.cpp

@@ -0,0 +1,258 @@
+/**
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+
+#include <gtest/gtest.h>
+
+#include <faiss/IndexIVFFlat.h>
+#include <faiss/IndexIVFPQ.h>
+#include <faiss/IndexFlat.h>
+#include <faiss/MetaIndexes.h>
+#include <faiss/impl/FaissAssert.h>
+#include <faiss/IndexPreTransform.h>
+#include <faiss/OnDiskInvertedLists.h>
+#include <faiss/IVFlib.h>
+
+
+namespace {
+
+
+struct Tempfilename {
+
+    static pthread_mutex_t mutex;
+
+    std::string filename;
+
+    Tempfilename (const char *prefix = nullptr) {
+        pthread_mutex_lock (&mutex);
+        char *cfname = tempnam (nullptr, prefix);
+        filename = cfname;
+        free(cfname);
+        pthread_mutex_unlock (&mutex);
+    }
+
+    ~Tempfilename () {
+        if (access (filename.c_str(), F_OK)) {
+            unlink (filename.c_str());
+        }
+    }
+
+    const char *c_str() {
+        return filename.c_str();
+    }
+
+};
+
+pthread_mutex_t Tempfilename::mutex = PTHREAD_MUTEX_INITIALIZER;
+
+
+typedef faiss::Index::idx_t idx_t;
+
+// parameters to use for the test
+int d = 64;
+size_t nb = 1000;
+size_t nq = 100;
+int nindex = 4;
+int k = 10;
+int nlist = 40;
+
+struct CommonData {
+
+    std::vector <float> database;
+    std::vector <float> queries;
+    std::vector<idx_t> ids;
+    faiss::IndexFlatL2 quantizer;
+
+    CommonData(): database (nb * d), queries (nq * d), ids(nb), quantizer (d) {
+
+        for (size_t i = 0; i < nb * d; i++) {
+            database[i] = drand48();
+        }
+        for (size_t i = 0; i < nq * d; i++) {
+            queries[i] = drand48();
+        }
+        for (int i = 0; i < nb; i++) {
+            ids[i] = 123 + 456 * i;
+        }
+        { // just to train the quantizer
+            faiss::IndexIVFFlat iflat (&quantizer, d, nlist);
+            iflat.train(nb, database.data());
+        }
+    }
+};
+
+CommonData cd;
+
+/// perform a search on shards, then merge and search again and
+/// compare results.
+int compare_merged (faiss::IndexShards *index_shards, bool shift_ids,
+                    bool standard_merge = true)
+{
+
+    std::vector<idx_t> refI(k * nq);
+    std::vector<float> refD(k * nq);
+
+    index_shards->search(nq, cd.queries.data(), k, refD.data(), refI.data());
+    Tempfilename filename;
+
+    std::vector<idx_t> newI(k * nq);
+    std::vector<float> newD(k * nq);
+
+    if (standard_merge) {
+
+        for (int i = 1; i < nindex; i++) {
+            faiss::ivflib::merge_into(
+                   index_shards->at(0), index_shards->at(i),
+                   shift_ids);
+        }
+
+        index_shards->sync_with_shard_indexes();
+    } else {
+        std::vector<const faiss::InvertedLists *> lists;
+        faiss::IndexIVF *index0 = nullptr;
+        size_t ntotal = 0;
+        for (int i = 0; i < nindex; i++) {
+            auto index_ivf = dynamic_cast<faiss::IndexIVF*>(index_shards->at(i));
+            assert (index_ivf);
+            if (i == 0) {
+                index0 = index_ivf;
+            }
+            lists.push_back (index_ivf->invlists);
+            ntotal += index_ivf->ntotal;
+        }
+
+        auto il = new faiss::OnDiskInvertedLists(
+                        index0->nlist, index0->code_size,
+                        filename.c_str());
+
+        il->merge_from(lists.data(), lists.size());
+
+        index0->replace_invlists(il, true);
+        index0->ntotal = ntotal;
+    }
+    // search only on first index
+    index_shards->at(0)->search(nq, cd.queries.data(),
+                                k, newD.data(), newI.data());
+
+    size_t ndiff = 0;
+    for (size_t i = 0; i < k * nq; i++) {
+        if (refI[i] != newI[i]) {
+            ndiff ++;
+        }
+    }
+    return ndiff;
+}
+
+}  // namespace
+
+
+// test on IVFFlat with implicit numbering
+TEST(MERGE, merge_flat_no_ids) {
+    faiss::IndexShards index_shards(d);
+    index_shards.own_fields = true;
+    for (int i = 0; i < nindex; i++) {
+        index_shards.add_shard (
+            new faiss::IndexIVFFlat (&cd.quantizer, d, nlist));
+    }
+    EXPECT_TRUE(index_shards.is_trained);
+    index_shards.add(nb, cd.database.data());
+    size_t prev_ntotal = index_shards.ntotal;
+    int ndiff = compare_merged(&index_shards, true);
+    EXPECT_EQ (prev_ntotal, index_shards.ntotal);
+    EXPECT_EQ(0, ndiff);
+}
+
+
+// test on IVFFlat, explicit ids
+TEST(MERGE, merge_flat) {
+    faiss::IndexShards index_shards(d, false, false);
+    index_shards.own_fields = true;
+
+    for (int i = 0; i < nindex; i++) {
+        index_shards.add_shard (
+             new faiss::IndexIVFFlat (&cd.quantizer, d, nlist));
+    }
+
+    EXPECT_TRUE(index_shards.is_trained);
+    index_shards.add_with_ids(nb, cd.database.data(), cd.ids.data());
+    int ndiff = compare_merged(&index_shards, false);
+    EXPECT_GE(0, ndiff);
+}
+
+// test on IVFFlat and a VectorTransform
+TEST(MERGE, merge_flat_vt) {
+    faiss::IndexShards index_shards(d, false, false);
+    index_shards.own_fields = true;
+
+    // here we have to retrain because of the vectorTransform
+    faiss::RandomRotationMatrix rot(d, d);
+    rot.init(1234);
+    faiss::IndexFlatL2 quantizer (d);
+
+    { // just to train the quantizer
+        faiss::IndexIVFFlat iflat (&quantizer, d, nlist);
+        faiss::IndexPreTransform ipt (&rot, &iflat);
+        ipt.train(nb, cd.database.data());
+    }
+
+    for (int i = 0; i < nindex; i++) {
+        faiss::IndexPreTransform * ipt = new faiss::IndexPreTransform (
+             new faiss::RandomRotationMatrix (rot),
+             new faiss::IndexIVFFlat (&quantizer, d, nlist)
+        );
+        ipt->own_fields = true;
+        index_shards.add_shard (ipt);
+    }
+    EXPECT_TRUE(index_shards.is_trained);
+    index_shards.add_with_ids(nb, cd.database.data(), cd.ids.data());
+    size_t prev_ntotal = index_shards.ntotal;
+    int ndiff = compare_merged(&index_shards, false);
+    EXPECT_EQ (prev_ntotal, index_shards.ntotal);
+    EXPECT_GE(0, ndiff);
+}
+
+
+// put the merged invfile on disk
+TEST(MERGE, merge_flat_ondisk) {
+    faiss::IndexShards index_shards(d, false, false);
+    index_shards.own_fields = true;
+    Tempfilename filename;
+
+    for (int i = 0; i < nindex; i++) {
+        auto ivf = new faiss::IndexIVFFlat (&cd.quantizer, d, nlist);
+        if (i == 0) {
+            auto il = new faiss::OnDiskInvertedLists (
+                ivf->nlist, ivf->code_size,
+                filename.c_str());
+            ivf->replace_invlists(il, true);
+        }
+        index_shards.add_shard (ivf);
+    }
+
+    EXPECT_TRUE(index_shards.is_trained);
+    index_shards.add_with_ids(nb, cd.database.data(), cd.ids.data());
+    int ndiff = compare_merged(&index_shards, false);
+
+    EXPECT_EQ(ndiff, 0);
+}
+
+// now use ondisk specific merge
+TEST(MERGE, merge_flat_ondisk_2) {
+    faiss::IndexShards index_shards(d, false, false);
+    index_shards.own_fields = true;
+
+    for (int i = 0; i < nindex; i++) {
+        index_shards.add_shard (
+             new faiss::IndexIVFFlat (&cd.quantizer, d, nlist));
+    }
+    EXPECT_TRUE(index_shards.is_trained);
+    index_shards.add_with_ids(nb, cd.database.data(), cd.ids.data());
+    int ndiff = compare_merged(&index_shards, false, false);
+    EXPECT_GE(0, ndiff);
+}

data/vendor/faiss/tests/test_omp_threads.cpp

@@ -0,0 +1,14 @@
+/**
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <gtest/gtest.h>
+
+#include <faiss/utils/utils.h>
+
+TEST(Threading, openmp) {
+  EXPECT_TRUE(faiss::check_openmp());
+}

data/vendor/faiss/tests/test_ondisk_ivf.cpp

@@ -0,0 +1,220 @@
+/**
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+
+#include <omp.h>
+
+#include <unordered_map>
+#include <pthread.h>
+
+#include <gtest/gtest.h>
+
+#include <faiss/OnDiskInvertedLists.h>
+#include <faiss/IndexIVFFlat.h>
+#include <faiss/IndexFlat.h>
+#include <faiss/utils/random.h>
+#include <faiss/index_io.h>
+
+
+namespace {
+
+struct Tempfilename {
+
+    static pthread_mutex_t mutex;
+
+    std::string filename;
+
+    Tempfilename (const char *prefix = nullptr) {
+        pthread_mutex_lock (&mutex);
+        char *cfname = tempnam (nullptr, prefix);
+        filename = cfname;
+        free(cfname);
+        pthread_mutex_unlock (&mutex);
+    }
+
+    ~Tempfilename () {
+        if (access (filename.c_str(), F_OK)) {
+            unlink (filename.c_str());
+        }
+    }
+
+    const char *c_str() {
+        return filename.c_str();
+    }
+
+};
+
+pthread_mutex_t Tempfilename::mutex = PTHREAD_MUTEX_INITIALIZER;
+
+}  // namespace
+
+
+TEST(ONDISK, make_invlists) {
+    int nlist = 100;
+    int code_size = 32;
+    int nadd = 1000000;
+    std::unordered_map<int, int> listnos;
+
+    Tempfilename filename;
+
+    faiss::OnDiskInvertedLists ivf (
+                nlist, code_size,
+                filename.c_str());
+
+    {
+        std::vector<uint8_t> code(32);
+        for (int i = 0; i < nadd; i++) {
+            double d = drand48();
+            int list_no = int(nlist * d * d); // skewed distribution
+            int * ar = (int*)code.data();
+            ar[0] = i;
+            ar[1] = list_no;
+            ivf.add_entry (list_no, i, code.data());
+            listnos[i] = list_no;
+        }
+    }
+
+    int ntot = 0;
+    for (int i = 0; i < nlist; i++) {
+        int size = ivf.list_size(i);
+        const faiss::Index::idx_t *ids = ivf.get_ids (i);
+        const uint8_t *codes = ivf.get_codes (i);
+        for (int j = 0; j < size; j++) {
+            faiss::Index::idx_t id = ids[j];
+            const int * ar = (const int*)&codes[code_size * j];
+            EXPECT_EQ (ar[0], id);
+            EXPECT_EQ (ar[1], i);
+            EXPECT_EQ (listnos[id], i);
+            ntot ++;
+        }
+    }
+    EXPECT_EQ (ntot, nadd);
+};
+
+
+TEST(ONDISK, test_add) {
+    int d = 8;
+    int nlist = 30, nq = 200, nb = 1500, k = 10;
+    faiss::IndexFlatL2 quantizer(d);
+    {
+        std::vector<float> x(d * nlist);
+        faiss::float_rand(x.data(), d * nlist, 12345);
+        quantizer.add(nlist, x.data());
+    }
+    std::vector<float> xb(d * nb);
+    faiss::float_rand(xb.data(), d * nb, 23456);
+
+    faiss::IndexIVFFlat index(&quantizer, d, nlist);
+    index.add(nb, xb.data());
+
+    std::vector<float> xq(d * nb);
+    faiss::float_rand(xq.data(), d * nq, 34567);
+
+    std::vector<float> ref_D (nq * k);
+    std::vector<faiss::Index::idx_t> ref_I (nq * k);
+
+    index.search (nq, xq.data(), k,
+                  ref_D.data(), ref_I.data());
+
+    Tempfilename filename, filename2;
+
+    // test add + search
+    {
+        faiss::IndexIVFFlat index2(&quantizer, d, nlist);
+
+        faiss::OnDiskInvertedLists ivf (
+                index.nlist, index.code_size,
+                filename.c_str());
+
+        index2.replace_invlists(&ivf);
+
+        index2.add(nb, xb.data());
+
+        std::vector<float> new_D (nq * k);
+        std::vector<faiss::Index::idx_t> new_I (nq * k);
+
+        index2.search (nq, xq.data(), k,
+                       new_D.data(), new_I.data());
+
+        EXPECT_EQ (ref_D, new_D);
+        EXPECT_EQ (ref_I, new_I);
+
+        write_index(&index2, filename2.c_str());
+
+    }
+
+    // test io
+    {
+        faiss::Index *index3 = faiss::read_index(filename2.c_str());
+
+        std::vector<float> new_D (nq * k);
+        std::vector<faiss::Index::idx_t> new_I (nq * k);
+
+        index3->search (nq, xq.data(), k,
+                        new_D.data(), new_I.data());
+
+        EXPECT_EQ (ref_D, new_D);
+        EXPECT_EQ (ref_I, new_I);
+
+        delete index3;
+    }
+
+};
+
+
+
+// WARN this thest will run multithreaded only in opt mode
+TEST(ONDISK, make_invlists_threaded) {
+    int nlist = 100;
+    int code_size = 32;
+    int nadd = 1000000;
+
+    Tempfilename filename;
+
+    faiss::OnDiskInvertedLists ivf (
+                nlist, code_size,
+                filename.c_str());
+
+    std::vector<int> list_nos (nadd);
+
+    for (int i = 0; i < nadd; i++) {
+        double d = drand48();
+        list_nos[i] = int(nlist * d * d); // skewed distribution
+    }
+
+#pragma omp parallel
+    {
+        std::vector<uint8_t> code(32);
+#pragma omp for
+        for (int i = 0; i < nadd; i++) {
+            int list_no = list_nos[i];
+            int * ar = (int*)code.data();
+            ar[0] = i;
+            ar[1] = list_no;
+            ivf.add_entry (list_no, i, code.data());
+        }
+    }
+
+    int ntot = 0;
+    for (int i = 0; i < nlist; i++) {
+        int size = ivf.list_size(i);
+        const faiss::Index::idx_t *ids = ivf.get_ids (i);
+        const uint8_t *codes = ivf.get_codes (i);
+        for (int j = 0; j < size; j++) {
+            faiss::Index::idx_t id = ids[j];
+            const int * ar = (const int*)&codes[code_size * j];
+            EXPECT_EQ (ar[0], id);
+            EXPECT_EQ (ar[1], i);
+            EXPECT_EQ (list_nos[id], i);
+            ntot ++;
+        }
+    }
+    EXPECT_EQ (ntot, nadd);
+
+};