Merge pull request #5 from erikbern/master

upmerge

.travis.yml

@@ -12,14 +12,14 @@ env:
   - LIBRARY=datasketch
   - LIBRARY=dolphinn
   - LIBRARY=faiss
-  - LIBRARY=falconn
   - LIBRARY=flann
   - LIBRARY=kgraph
   - LIBRARY=nearpy
   - LIBRARY=nmslib
   - LIBRARY=panns
   - LIBRARY=rpforest
   - LIBRARY=sklearn
+  - LIBRARY=pynndescent
 
 before_install:
   - pip install -r requirements.txt
@@ -28,3 +28,5 @@ before_install:
 script:
   - python run.py --docker-tag ann-benchmarks-${LIBRARY} --max-n-algorithms 5 --dataset random-xs-20-angular
   - python plot.py --dataset random-xs-20-angular --output plot.png
+  - python -m unittest test/test-metrics.py
+  - python create_website.py --outputdir . --scatter --latex

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2018 Erik Bernhardsson
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -7,8 +7,6 @@ Doing fast searching of nearest neighbors in high dimensional spaces is an incre
 
 This project contains some tools to benchmark various implementations of approximate nearest neighbor (ANN) search for different metrics. We have pregenerated datasets (in HDF5) formats and we also have Docker containers for each algorithm. There's a [test suite](https://travis-ci.org/erikbern/ann-benchmarks) that makes sure every algorithm works.
 
-See [the results of this benchmark](http://sss.projects.itu.dk/ann-benchmarks).
-
 Evaluated
 =========
 
@@ -20,10 +18,10 @@ Evaluated
 * [KGraph](https://github.com/aaalgo/kgraph)
 * [NMSLIB (Non-Metric Space Library)](https://github.com/searchivarius/nmslib): SWGraph, HNSW, BallTree, MPLSH
 * [RPForest](https://github.com/lyst/rpforest)
-* [FALCONN](http://falconn-lib.org/)
 * [FAISS](https://github.com/facebookresearch/faiss.git)
 * [DolphinnPy](https://github.com/ipsarros/DolphinnPy)
 * [Datasketch](https://github.com/ekzhu/datasketch)
+* [PyNNDescent](https://github.com/lmcinnes/pynndescent)
 
 Data sets
 =========
@@ -33,15 +31,45 @@ We have a number of precomputed data sets for this. All data sets are pre-split
 | Dataset                                                           | Dimensions | Train size | Test size | Neighbors | Distance  | Download                                                                     |
 | ----------------------------------------------------------------- | ---------: | ---------: | --------: | --------: | --------- | ---------------------------------------------------------------------------- |
 | [Fashion-MNIST](https://github.com/zalandoresearch/fashion-mnist) |        784 |     60,000 |    10,000 |       100 | Euclidean | [HDF5](http://vectors.erikbern.com/fashion-mnist-784-euclidean.hdf5) (217MB) |
-| [GIST](http://corpus-texmex.irisa.fr/)                            |        960 |  1,000,000 |     1,000 |       100 | Euclidean | [HDF5](http://vectors.erikbern.com/gist-960-euclidean.hdf5) (121MB)          |
-| [GloVe](http://nlp.stanford.edu/projects/glove/)                  |         25 |  1,133,628 |    59,886 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/glove-25-angular.hdf5) (121MB)            |
-| GloVe                                                             |         50 |  1,133,628 |    59,886 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/glove-50-angular.hdf5) (235MB)            |
-| GloVe                                                             |        100 |  1,133,628 |    59,886 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/glove-100-angular.hdf5) (463MB)           |
-| GloVe                                                             |        200 |  1,133,628 |    59,886 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/glove-200-angular.hdf5) (918MB)           |
+| [GIST](http://corpus-texmex.irisa.fr/)                            |        960 |  1,000,000 |     1,000 |       100 | Euclidean | [HDF5](http://vectors.erikbern.com/gist-960-euclidean.hdf5) (3.6GB)          |
+| [GloVe](http://nlp.stanford.edu/projects/glove/)                  |         25 |  1,183,514 |    10,000 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/glove-25-angular.hdf5) (121MB)            |
+| GloVe                                                             |         50 |  1,183,514 |    10,000 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/glove-50-angular.hdf5) (235MB)            |
+| GloVe                                                             |        100 |  1,183,514 |    10,000 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/glove-100-angular.hdf5) (463MB)           |
+| GloVe                                                             |        200 |  1,183,514 |    10,000 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/glove-200-angular.hdf5) (918MB)           |
 | [MNIST](http://yann.lecun.com/exdb/mnist/)                        |        784 |     60,000 |    10,000 |       100 | Euclidean | [HDF5](http://vectors.erikbern.com/mnist-784-euclidean.hdf5) (217MB)         |
 | [NYTimes](https://archive.ics.uci.edu/ml/datasets/bag+of+words)   |        256 |    290,000 |    10,000 |       100 | Angular   | [HDF5](http://vectors.erikbern.com/nytimes-256-angular.hdf5) (301MB)         |
 | [SIFT](https://corpus-texmex.irisa.fr/)                           |        128 |  1,000,000 |    10,000 |       100 | Euclidean | [HDF5](http://vectors.erikbern.com/sift-128-euclidean.hdf5) (501MB)          |
 
+Results
+=======
+
+
+glove-100-angular
+
+![glove-100-angular](https://raw.github.com/erikbern/ann-benchmarks/master/results/glove-100-angular.png)
+
+sift-128-euclidean
+
+![glove-100-angular](https://raw.github.com/erikbern/ann-benchmarks/master/results/sift-128-euclidean.png)
+
+fashion-mnist-784-euclidean
+
+![fashion-mnist-784-euclidean](https://raw.github.com/erikbern/ann-benchmarks/master/results/fashion-mnist-784-euclidean.png)
+
+gist-960-euclidean
+
+![gist-960-euclidean](https://raw.github.com/erikbern/ann-benchmarks/master/results/gist-960-euclidean.png)
+
+nytimes-256-angular
+
+![nytimes-256-angular](https://raw.github.com/erikbern/ann-benchmarks/master/results/nytimes-256-angular.png)
+
+glove-25-angular
+
+![glove-25-angular](https://raw.github.com/erikbern/ann-benchmarks/master/results/glove-25-angular.png)
+
+Results as of Feb 2018-02-05, running all benchmarks on a c5.4xlarge machine on AWS.
+
 Install
 =======
 
@@ -85,3 +113,8 @@ Principles
 * We currently support CPU-based ANN algorithms. GPU support is planned as future work.
 * Do proper train/test set of index data and query points.
 * Note that Hamming distance and set similarity was supported in the past. This might hopefully be added back soon.
+
+Authors
+=======
+
+Built by [Erik Bernhardsson](https://erikbern.com) with significant contributions from [Martin Aumüller](http://itu.dk/people/maau/) and [Alexander Faithfull](https://github.com/ale-f).

algos.yaml

@@ -135,6 +135,14 @@ float:
           # {"tuneK": 10, "desiredRecall": 0.97}), and so on up to
           # NmslibNewIndex("angular", "vptree", {"tuneK": 10, "desiredRecall":
           # 0.1}).
+    pynndescent:
+      docker-tag: ann-benchmarks-pynndescent
+      module: ann_benchmarks.algorithms.pynndescent
+      constructor: PyNNDescent
+      base-args: ["@metric"]
+      run-groups:
+        pynndescent:
+          args: [[5, 10, 15, 20], [4, 8, 16], [40], [1.0, 2.0, 4.0, 8.0]]
   euclidean:
     kgraph:
       docker-tag: ann-benchmarks-kgraph
@@ -279,28 +287,17 @@ float:
         base:
           args: [[3, 5, 10, 20, 40, 100, 200, 400],
                 [3, 5, 10, 20, 40, 100, 200, 400]]
-    falconn:
-      docker-tag: ann-benchmarks-falconn
-      module: ann_benchmarks.algorithms.falconn
-      constructor: FALCONN
-      base-args: ["@metric", 16]
+    pynndescent:
+      docker-tag: ann-benchmarks-pynndescent
+      module: ann_benchmarks.algorithms.pynndescent
+      constructor: PyNNDescent
+      base-args: ["@metric"]
       run-groups:
-        base:
-          L: &l [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 21, 24, 27,
-             30, 33, 37, 41, 46, 51, 57, 63, 70, 77, 85, 94, 104, 115, 127,
-             140, 154, 170, 188, 207, 228, 251, 277, 305, 336, 370, 408, 449,
-             494, 544, 599, 659, 725, 798, 878, 966, 1063, 1170, 1287, 1416]
-          args: [*l]
+        pynndescent:
+          args: [[10, 20, 30, 40], [8, 16, 32], [40], [2.0, 4.0, 8.0, 16.0, 32.0]]
+
 bit:
   hamming:
-    falconn:
-      docker-tag: ann-benchmarks-falconn
-      module: ann_benchmarks.algorithms.falconn
-      constructor: FALCONN
-      base-args: ["@metric", 16]
-      run-groups:
-        base:
-          args: [*l]
     kgraph:
       docker-tag: ann-benchmarks-kgraph
       module: ann_benchmarks.algorithms.kgraph

ann_benchmarks/algorithms/definitions.py

@@ -8,6 +8,7 @@
 import sys
 import traceback
 import yaml
+from enum import Enum
 from itertools import product
 
 
@@ -20,6 +21,20 @@ def instantiate_algorithm(definition):
     constructor = getattr(module, definition.constructor)
     return constructor(*definition.arguments)
 
+class InstantiationStatus(Enum):
+    AVAILABLE = 0
+    NO_CONSTRUCTOR = 1
+    NO_MODULE = 2
+
+def algorithm_status(definition):
+    try:
+        module = importlib.import_module(definition.module)
+        if hasattr(module, definition.constructor):
+            return InstantiationStatus.AVAILABLE
+        else:
+            return InstantiationStatus.NO_CONSTRUCTOR
+    except ImportError:
+        return InstantiationStatus.NO_MODULE
 
 def get_result_filename(dataset, count, definition):
     d = ['results',

ann_benchmarks/algorithms/nmslib.py

@@ -4,10 +4,12 @@
 from ann_benchmarks.constants import INDEX_DIR
 from ann_benchmarks.algorithms.base import BaseANN
 
+
 class NmslibReuseIndex(BaseANN):
     @staticmethod
     def encode(d):
         return ["%s=%s" % (a, b) for (a, b) in d.iteritems()]
+
     def __init__(self, metric, method_name, index_param, save_index, query_param):
         self._nmslib_metric = {'angular': 'cosinesimil', 'euclidean': 'l2'}[metric]
         self._method_name = method_name
@@ -19,7 +21,7 @@ def __init__(self, metric, method_name, index_param, save_index, query_param):
 
         d = os.path.dirname(self._index_name)
         if not os.path.exists(d):
-          os.makedirs(d)
+            os.makedirs(d)
 
     def fit(self, X):
         if self._method_name == 'vptree':
@@ -28,28 +30,24 @@ def fit(self, X):
             # what():  The data size is too small or the bucket size is too big. Select the parameters so that <total # of records> is NOT less than <bucket size> * 1000
             # Aborted (core dumped)
             self._index_param.append('bucketSize=%d' % min(int(X.shape[0] * 0.0005), 1000))
-
-        self._index = nmslib.init(space=self._nmslib_metric,
-                                  method=self._method_name)
 
-        for i, x in enumerate(X):
-            nmslib.addDataPoint(self._index, i, x.tolist())
+        self._index = nmslib.init(space=self._nmslib_metric, method=self._method_name)
+        self._index.addDataPointBatch(X)
 
         if os.path.exists(self._index_name):
             print('Loading index from file')
-            nmslib.loadIndex(self._index, self._index_name)
+            self._index.loadIndex(self._index_name)
         else:
-            nmslib.createIndex(self._index, self._index_param)
-            if self._save_index: 
-              nmslib.saveIndex(self._index, self._index_name)
+            self._index.createIndex(self._index_param)
+            if self._save_index:
+                self._index.saveIndex(self._index_name)
 
-        nmslib.setQueryTimeParams(self._index, self._query_param)
+        self._index.setQueryTimeParams(self._query_param)
 
     def query(self, v, n):
-        return nmslib.knnQuery(self._index, n, v.tolist())
+        ids, distances = self._index.knnQuery(v, n)
+        return ids
 
-    def freeIndex(self):
-        nmslib.freeIndex(self._index)
 
 class NmslibNewIndex(BaseANN):
     def __init__(self, metric, method_name, method_param):
@@ -65,16 +63,12 @@ def fit(self, X):
             # what():  The data size is too small or the bucket size is too big. Select the parameters so that <total # of records> is NOT less than <bucket size> * 1000
             # Aborted (core dumped)
             self._method_param.append('bucketSize=%d' % min(int(X.shape[0] * 0.0005), 1000))
-
-        self._index = nmslib.init(self._nmslib_metric, [], self._method_name, nmslib.DataType.DENSE_VECTOR, nmslib.DistType.FLOAT)
-
-        for i, x in enumerate(X):
-            nmslib.addDataPoint(self._index, i, x.tolist())
+
+        self._index = nmslib.init(space=self._nmslib_metric, method=self._method_name)
+        self._index.addDataPointBatch(X)
 
         nmslib.createIndex(self._index, self._method_param)
 
     def query(self, v, n):
-        return nmslib.knnQuery(self._index, n, v.tolist())
-
-    def freeIndex(self):
-        nmslib.freeIndex(self._index)
+        ids, distances = self._index.knnQuery(v, n)
+        return ids

ann_benchmarks/algorithms/pynndescent.py

@@ -0,0 +1,28 @@
+from __future__ import absolute_import
+import pynndescent
+import numpy as np
+from ann_benchmarks.algorithms.base import BaseANN
+
+class PyNNDescent(BaseANN):
+    def __init__(self, metric, n_neighbors=10, n_trees=8, leaf_size=40, queue_size=2.0):
+        self._n_neighbors = int(n_neighbors)
+        self._n_trees = int(n_trees)
+        self._leaf_size = int(leaf_size)
+        self._queue_size = float(queue_size)
+        self._pynnd_metric = {'angular': 'cosine', 'euclidean': 'euclidean'}[metric]
+        self.name = 'PyNNDescent(n_neighbors=%d,n_trees=%d,leaf_size=%d,queue_size=%.2f)' % \
+                        (self._n_neighbors, self._n_trees, self._leaf_size, self._queue_size)
+
+    def fit(self, X):
+        self._index = pynndescent.NNDescent(X,
+                        n_neighbors=self._n_neighbors,
+                        n_trees=self._n_trees,
+                        leaf_size=self._leaf_size,
+                        metric=self._pynnd_metric)
+
+    def query(self, v, n):
+        ind, dist = self._index.query(np.array([v]), k=n, queue_size=self._queue_size)
+        return ind[0]
+
+    def use_threads(self):
+        return False