@jegou Sorry to the link incorrect. I did the experiment based on your script. The code I added to bench_gpu_sift1m_falconn.py is as follows:

#################################################################
# Falconn search experiment
#################################################################
print "============ Falconn search"

import falconn
params_cp = falconn.LSHConstructionParameters()
params_cp.dimension = d
params_cp.lsh_family = 'cross_polytope'
params_cp.distance_function = 'negative_inner_product'
params_cp.storage_hash_table = 'flat_hash_table'
params_cp.k = 3
params_cp.l = 10
params_cp.num_setup_threads = 0
params_cp.last_cp_dimension = 16
params_cp.num_rotations = 3
seed = 119417657
params_cp.seed = seed ^ 833840234
cp_table = falconn.LSHIndex(params_cp)
cp_table.setup(xb)
print "falconn hashing table finished..."

nprobes = [16, 32, 64, 128, 256, 512]
for i, nprobe in enumerate(nprobes):
    I = []
    delta_t = 0.0
    cp_table.set_num_probes(nprobe)
    for i in range(xq.shape[0]):
        t0 = time.time()
        idxs = cp_table.find_k_nearest_neighbors(xq[i, :], 100)
        t1 = time.time()
        delta_t += t1 - t0
        I.append(idxs)
    I = np.array(I)
    print "nprobe=%4d %.3f s recalls=" % (nprobe, delta_t / float(nq)),
    for rank in 1, 10, 100:
        n_ok = (I[:, :rank] == gt[:, :1]).sum() # recall rate @ 1, 10, 100 nearest neighbors
        print "%.4f" % (n_ok / float(nq)),
    print

I didn't take much time to tune the number of hashing function k and the hashing table l. but the result shows ok.

So for hundred millions to billions of vectors (very large-scale), it's a good option to choose Faiss. For small-scale dataset, Falconn and Nmslib are good choice. Did I get the point?

benchmark #23

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions