Last active
March 1, 2017 09:30
-
-
Save maheshakya/b7bcf915c9d5bab89d8d to your computer and use it in GitHub Desktop.
Comparison of indexing, query time and accury among FLANN, ANNOY and LSH Forest
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import time | |
import numpy as np | |
from sklearn.datasets.samples_generator import make_blobs | |
from sklearn.neighbors import LSHForest | |
from sklearn.neighbors import NearestNeighbors | |
from sklearn.preprocessing import normalize | |
from annoy import AnnoyIndex | |
from pyflann import FLANN | |
n_iter = 100 | |
n_neighbors = 10 | |
rng = np.random.RandomState(42) | |
n_samples_n_features_pairs = [(1000, 100), (1000, 500), (10000, 100), (10000, 500), | |
(10000, 1000), (10000, 5000)] | |
annoy_n_trees = 10 | |
average_query_times_lshf = [] | |
average_query_times_flann = [] | |
average_query_times_annoy = [] | |
accuracies_lshf = [] | |
accuracies_annoy = [] | |
accuracies_flann = [] | |
build_time_lshf = 0 | |
build_time_flann = 0 | |
build_time_annoy = 0 | |
# Calculate the average query time | |
for j, pair in enumerate(n_samples_n_features_pairs): | |
print "----------------------------------------------------------------------------------------" | |
print "n_samples: ", pair[0], "n_features: ", pair[1] | |
X, labels_true = make_blobs(n_samples=pair[0]+n_iter, n_features=pair[1], | |
centers=10, cluster_std=5, | |
random_state=0) | |
#Initialize NearestNeighbors | |
nbrs = NearestNeighbors(n_neighbors=n_neighbors, metric='cosine', algorithm='brute') | |
nbrs.fit(X) | |
# Initialize LSHForest | |
lshf = LSHForest(n_candidates=50, n_neighbors=n_neighbors) | |
t0 = time.time() | |
lshf.fit(X[:pair[0]]) | |
build_time_lshf = time.time() - t0 | |
print "LSHF index build time: ", build_time_lshf | |
# Initialize ANNOY | |
annoy = AnnoyIndex(pair[1], metric = 'angular') | |
t0 = time.time() | |
for i in range(pair[0]): | |
annoy.add_item(i, X[i].tolist()) | |
annoy.build(annoy_n_trees) | |
build_time_annoy = time.time() - t0 | |
print "ANNOY index build time: ", build_time_annoy | |
# Initialize FLANN | |
X_normed = normalize(X, axis=1, norm='l2') | |
flann = FLANN(target_precision=0.8, algorithm='autotuned') | |
t0 = time.time() | |
flann.build_index(X_normed) | |
build_time_flann = time.time() - t0 | |
print "FLANN index build time: ", build_time_flann | |
average_time_lshf = 0 | |
average_time_annoy = 0 | |
average_time_flann = 0 | |
accuracy_lshf = 0 | |
accuracy_annoy = 0 | |
accuracy_flann = 0 | |
queries = X[pair[0]:] | |
for i in range(n_iter): | |
query = queries[i] | |
# LSHF query | |
t0 = time.time() | |
approx_neighbors_lshf = lshf.kneighbors(query, | |
return_distance=False) | |
T = time.time() - t0 | |
average_time_lshf = average_time_lshf + T | |
# ANNOY query | |
t0 = time.time() | |
approx_neighbors_annoy = annoy.get_nns_by_vector(query.tolist(), n_neighbors) | |
T = time.time() - t0 | |
average_time_annoy = average_time_annoy + T | |
# FLANN query | |
query_normed = normalize(query, axis=1, norm='l2')[0] | |
t0 = time.time() | |
approx_neighbors_flann, distance = flann.nn_index(query_normed, n_neighbors) | |
T = time.time() - t0 | |
average_time_flann = average_time_flann + T | |
# NearestNeighbors query | |
neighbors_exact = nbrs.kneighbors(query, return_distance=False) | |
# Calculate accuracies | |
intersection = np.intersect1d(approx_neighbors_lshf, | |
neighbors_exact).shape[0] | |
ratio = intersection/float(n_neighbors) | |
accuracy_lshf = accuracy_lshf + ratio | |
intersection = np.intersect1d(approx_neighbors_annoy, | |
neighbors_exact).shape[0] | |
ratio = intersection/float(n_neighbors) | |
accuracy_annoy = accuracy_annoy + ratio | |
intersection = np.intersect1d(approx_neighbors_flann[0], | |
neighbors_exact).shape[0] | |
ratio = intersection/float(n_neighbors) | |
accuracy_flann = accuracy_flann + ratio | |
average_query_times_lshf.append(average_time_lshf/float(n_iter)) | |
accuracies_lshf.append(accuracy_lshf/float(n_iter)) | |
average_query_times_annoy.append(average_time_annoy/float(n_iter)) | |
accuracies_annoy.append(accuracy_annoy/float(n_iter)) | |
average_query_times_flann.append(average_time_flann/float(n_iter)) | |
accuracies_flann.append(accuracy_flann/float(n_iter)) | |
print "LSHF average query time: ", average_query_times_lshf[j], ", Average accuracy: ", accuracies_lshf[j] | |
print "ANNOY average query time: ", average_query_times_annoy[j], ", Average accuracy: ", accuracies_annoy[j] | |
print "FLANN average query time: ", average_query_times_flann[j], ", Average accuracy: ", accuracies_flann[j] |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment