faiss/tests/test_index_accuracy.py

733 lines
22 KiB
Python

# 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.
from __future__ import absolute_import, division, print_function
# noqa E741
# translation of test_knn.lua
import numpy as np
import unittest
import faiss
from common import Randu10k, get_dataset_2, Randu10kUnbalanced
ev = Randu10k()
d = ev.d
# Parameters inverted indexes
ncentroids = int(4 * np.sqrt(ev.nb))
kprobe = int(np.sqrt(ncentroids))
# Parameters for LSH
nbits = d
# Parameters for indexes involving PQ
M = int(d / 8) # for PQ: #subquantizers
nbits_per_index = 8 # for PQ
class IndexAccuracy(unittest.TestCase):
def test_IndexFlatIP(self):
q = faiss.IndexFlatIP(d) # Ask inner product
res = ev.launch('FLAT / IP', q)
e = ev.evalres(res)
assert e[1] == 1.0
def test_IndexFlatL2(self):
q = faiss.IndexFlatL2(d)
res = ev.launch('FLAT / L2', q)
e = ev.evalres(res)
assert e[1] == 1.0
def test_ivf_kmeans(self):
ivfk = faiss.IndexIVFFlat(faiss.IndexFlatL2(d), d, ncentroids)
ivfk.nprobe = kprobe
res = ev.launch('IndexIVFFlat', ivfk)
e = ev.evalres(res)
# should give 0.260 0.260 0.260
assert e[1] > 0.2
# test parallel mode
Dref, Iref = ivfk.search(ev.xq, 100)
ivfk.parallel_mode = 1
Dnew, Inew = ivfk.search(ev.xq, 100)
print((Iref != Inew).sum(), Iref.size)
assert (Iref != Inew).sum() < Iref.size / 5000.0
assert np.all(Dref == Dnew)
def test_indexLSH(self):
q = faiss.IndexLSH(d, nbits)
res = ev.launch('FLAT / LSH Cosine', q)
e = ev.evalres(res)
# should give 0.070 0.250 0.580
assert e[10] > 0.2
def test_IndexLSH_32_48(self):
# CHECK: the difference between 32 and 48 does not make much sense
for nbits2 in 32, 48:
q = faiss.IndexLSH(d, nbits2)
res = ev.launch('LSH half size', q)
e = ev.evalres(res)
# should give 0.003 0.019 0.108
assert e[10] > 0.018
def test_IndexPQ(self):
q = faiss.IndexPQ(d, M, nbits_per_index)
res = ev.launch('FLAT / PQ L2', q)
e = ev.evalres(res)
# should give 0.070 0.230 0.260
assert e[10] > 0.2
# Approximate search module: PQ with inner product distance
def test_IndexPQ_ip(self):
q = faiss.IndexPQ(d, M, nbits_per_index, faiss.METRIC_INNER_PRODUCT)
res = ev.launch('FLAT / PQ IP', q)
e = ev.evalres(res)
# should give 0.070 0.230 0.260
#(same result as regular PQ on normalized distances)
assert e[10] > 0.2
def test_IndexIVFPQ(self):
ivfpq = faiss.IndexIVFPQ(faiss.IndexFlatL2(d), d, ncentroids, M, 8)
ivfpq.nprobe = kprobe
res = ev.launch('IVF PQ', ivfpq)
e = ev.evalres(res)
# should give 0.070 0.230 0.260
assert e[10] > 0.2
# TODO: translate evaluation of nested
# Approximate search: PQ with full vector refinement
def test_IndexPQ_refined(self):
q = faiss.IndexPQ(d, M, nbits_per_index)
res = ev.launch('PQ non-refined', q)
e = ev.evalres(res)
q.reset()
rq = faiss.IndexRefineFlat(q)
res = ev.launch('PQ refined', rq)
e2 = ev.evalres(res)
assert e2[10] >= e[10]
rq.k_factor = 4
res = ev.launch('PQ refined*4', rq)
e3 = ev.evalres(res)
assert e3[10] >= e2[10]
def test_polysemous(self):
index = faiss.IndexPQ(d, M, nbits_per_index)
index.do_polysemous_training = True
# reduce nb iterations to speed up training for the test
index.polysemous_training.n_iter = 50000
index.polysemous_training.n_redo = 1
res = ev.launch('normal PQ', index)
e_baseline = ev.evalres(res)
index.search_type = faiss.IndexPQ.ST_polysemous
index.polysemous_ht = int(M / 16. * 58)
stats = faiss.cvar.indexPQ_stats
stats.reset()
res = ev.launch('Polysemous ht=%d' % index.polysemous_ht,
index)
e_polysemous = ev.evalres(res)
print(e_baseline, e_polysemous, index.polysemous_ht)
print(stats.n_hamming_pass, stats.ncode)
# The randu dataset is difficult, so we are not too picky on
# the results. Here we assert that we have < 10 % loss when
# computing full PQ on fewer than 20% of the data.
assert stats.n_hamming_pass < stats.ncode / 5
# Test disabled because difference is 0.17 on aarch64
# TODO check why???
# assert e_polysemous[10] > e_baseline[10] - 0.1
def test_ScalarQuantizer(self):
quantizer = faiss.IndexFlatL2(d)
ivfpq = faiss.IndexIVFScalarQuantizer(
quantizer, d, ncentroids,
faiss.ScalarQuantizer.QT_8bit)
ivfpq.nprobe = kprobe
res = ev.launch('IVF SQ', ivfpq)
e = ev.evalres(res)
# should give 0.234 0.236 0.236
assert e[10] > 0.235
class TestSQFlavors(unittest.TestCase):
""" tests IP in addition to L2, non multiple of 8 dimensions
"""
def add2columns(self, x):
return np.hstack((
x, np.zeros((x.shape[0], 2), dtype='float32')
))
def subtest_add2col(self, xb, xq, index, qname):
"""Test with 2 additional dimensions to take also the non-SIMD
codepath. We don't retrain anything but add 2 dims to the
queries, the centroids and the trained ScalarQuantizer.
"""
nb, d = xb.shape
d2 = d + 2
xb2 = self.add2columns(xb)
xq2 = self.add2columns(xq)
nlist = index.nlist
quantizer = faiss.downcast_index(index.quantizer)
quantizer2 = faiss.IndexFlat(d2, index.metric_type)
centroids = faiss.vector_to_array(quantizer.xb).reshape(nlist, d)
centroids2 = self.add2columns(centroids)
quantizer2.add(centroids2)
index2 = faiss.IndexIVFScalarQuantizer(
quantizer2, d2, index.nlist, index.sq.qtype,
index.metric_type)
index2.nprobe = 4
if qname in ('8bit', '4bit'):
trained = faiss.vector_to_array(index.sq.trained).reshape(2, -1)
nt = trained.shape[1]
# 2 lines: vmins and vdiffs
new_nt = int(nt * d2 / d)
trained2 = np.hstack((
trained,
np.zeros((2, new_nt - nt), dtype='float32')
))
trained2[1, nt:] = 1.0 # set vdiff to 1 to avoid div by 0
faiss.copy_array_to_vector(trained2.ravel(), index2.sq.trained)
else:
index2.sq.trained = index.sq.trained
index2.is_trained = True
index2.add(xb2)
return index2.search(xq2, 10)
# run on Sept 18, 2018 with nprobe=4 + 4 bit bugfix
ref_results = {
(0, '8bit'): 984,
(0, '4bit'): 978,
(0, '8bit_uniform'): 985,
(0, '4bit_uniform'): 979,
(0, 'fp16'): 985,
(1, '8bit'): 979,
(1, '4bit'): 973,
(1, '8bit_uniform'): 979,
(1, '4bit_uniform'): 972,
(1, 'fp16'): 979,
# added 2019-06-26
(0, '6bit'): 985,
(1, '6bit'): 987,
}
def subtest(self, mt):
d = 32
xt, xb, xq = get_dataset_2(d, 2000, 1000, 200)
nlist = 64
gt_index = faiss.IndexFlat(d, mt)
gt_index.add(xb)
gt_D, gt_I = gt_index.search(xq, 10)
quantizer = faiss.IndexFlat(d, mt)
for qname in '8bit 4bit 8bit_uniform 4bit_uniform fp16 6bit'.split():
qtype = getattr(faiss.ScalarQuantizer, 'QT_' + qname)
index = faiss.IndexIVFScalarQuantizer(
quantizer, d, nlist, qtype, mt)
index.train(xt)
index.add(xb)
index.nprobe = 4 # hopefully more robust than 1
D, I = index.search(xq, 10)
ninter = faiss.eval_intersection(I, gt_I)
print('(%d, %s): %d, ' % (mt, repr(qname), ninter))
assert abs(ninter - self.ref_results[(mt, qname)]) <= 10
if qname == '6bit':
# the test below fails triggers ASAN. TODO check what's wrong
continue
D2, I2 = self.subtest_add2col(xb, xq, index, qname)
assert np.all(I2 == I)
# also test range search
if mt == faiss.METRIC_INNER_PRODUCT:
radius = float(D[:, -1].max())
else:
radius = float(D[:, -1].min())
print('radius', radius)
lims, D3, I3 = index.range_search(xq, radius)
ntot = ndiff = 0
for i in range(len(xq)):
l0, l1 = lims[i], lims[i + 1]
Inew = set(I3[l0:l1])
if mt == faiss.METRIC_INNER_PRODUCT:
mask = D2[i] > radius
else:
mask = D2[i] < radius
Iref = set(I2[i, mask])
ndiff += len(Inew ^ Iref)
ntot += len(Iref)
print('ndiff %d / %d' % (ndiff, ntot))
assert ndiff < ntot * 0.01
for pm in 1, 2:
print('parallel_mode=%d' % pm)
index.parallel_mode = pm
lims4, D4, I4 = index.range_search(xq, radius)
print('sizes', lims4[1:] - lims4[:-1])
for qno in range(len(lims) - 1):
Iref = I3[lims[qno]: lims[qno+1]]
Inew = I4[lims4[qno]: lims4[qno+1]]
assert set(Iref) == set(Inew), "q %d ref %s new %s" % (
qno, Iref, Inew)
def test_SQ_IP(self):
self.subtest(faiss.METRIC_INNER_PRODUCT)
def test_SQ_L2(self):
self.subtest(faiss.METRIC_L2)
def test_parallel_mode(self):
d = 32
xt, xb, xq = get_dataset_2(d, 2000, 1000, 200)
index = faiss.index_factory(d, "IVF64,SQ8")
index.train(xt)
index.add(xb)
index.nprobe = 4 # hopefully more robust than 1
Dref, Iref = index.search(xq, 10)
for pm in 1, 2, 3:
index.parallel_mode = pm
Dnew, Inew = index.search(xq, 10)
np.testing.assert_array_equal(Iref, Inew)
np.testing.assert_array_equal(Dref, Dnew)
class TestSQByte(unittest.TestCase):
def subtest_8bit_direct(self, metric_type, d):
xt, xb, xq = get_dataset_2(d, 500, 1000, 30)
# rescale everything to get integer
tmin, tmax = xt.min(), xt.max()
def rescale(x):
x = np.floor((x - tmin) * 256 / (tmax - tmin))
x[x < 0] = 0
x[x > 255] = 255
return x
xt = rescale(xt)
xb = rescale(xb)
xq = rescale(xq)
gt_index = faiss.IndexFlat(d, metric_type)
gt_index.add(xb)
Dref, Iref = gt_index.search(xq, 10)
index = faiss.IndexScalarQuantizer(
d, faiss.ScalarQuantizer.QT_8bit_direct, metric_type)
index.add(xb)
D, I = index.search(xq, 10)
assert np.all(I == Iref)
assert np.all(D == Dref)
# same, with IVF
nlist = 64
quantizer = faiss.IndexFlat(d, metric_type)
gt_index = faiss.IndexIVFFlat(quantizer, d, nlist, metric_type)
gt_index.nprobe = 4
gt_index.train(xt)
gt_index.add(xb)
Dref, Iref = gt_index.search(xq, 10)
index = faiss.IndexIVFScalarQuantizer(
quantizer, d, nlist,
faiss.ScalarQuantizer.QT_8bit_direct, metric_type)
index.nprobe = 4
index.by_residual = False
index.train(xt)
index.add(xb)
D, I = index.search(xq, 10)
assert np.all(I == Iref)
assert np.all(D == Dref)
def test_8bit_direct(self):
for d in 13, 16, 24:
for metric_type in faiss.METRIC_L2, faiss.METRIC_INNER_PRODUCT:
self.subtest_8bit_direct(metric_type, d)
class TestPQFlavors(unittest.TestCase):
# run on Dec 14, 2018
ref_results = {
(1, True): 800,
(1, True, 20): 794,
(1, False): 769,
(0, True): 831,
(0, True, 20): 828,
(0, False): 829,
}
def test_IVFPQ_IP(self):
self.subtest(faiss.METRIC_INNER_PRODUCT)
def test_IVFPQ_L2(self):
self.subtest(faiss.METRIC_L2)
def subtest(self, mt):
d = 32
xt, xb, xq = get_dataset_2(d, 2000, 1000, 200)
nlist = 64
gt_index = faiss.IndexFlat(d, mt)
gt_index.add(xb)
gt_D, gt_I = gt_index.search(xq, 10)
quantizer = faiss.IndexFlat(d, mt)
for by_residual in True, False:
index = faiss.IndexIVFPQ(
quantizer, d, nlist, 4, 8)
index.metric_type = mt
index.by_residual = by_residual
if by_residual:
# perform cheap polysemous training
index.do_polysemous_training = True
pt = faiss.PolysemousTraining()
pt.n_iter = 50000
pt.n_redo = 1
index.polysemous_training = pt
index.train(xt)
index.add(xb)
index.nprobe = 4
D, I = index.search(xq, 10)
ninter = faiss.eval_intersection(I, gt_I)
print('(%d, %s): %d, ' % (mt, by_residual, ninter))
assert abs(ninter - self.ref_results[mt, by_residual]) <= 3
index.use_precomputed_table = 0
D2, I2 = index.search(xq, 10)
assert np.all(I == I2)
if by_residual:
index.use_precomputed_table = 1
index.polysemous_ht = 20
D, I = index.search(xq, 10)
ninter = faiss.eval_intersection(I, gt_I)
print('(%d, %s, %d): %d, ' % (
mt, by_residual, index.polysemous_ht, ninter))
# polysemous behaves bizarrely on ARM
assert (ninter >= self.ref_results[
mt, by_residual, index.polysemous_ht] - 4)
# also test range search
if mt == faiss.METRIC_INNER_PRODUCT:
radius = float(D[:, -1].max())
else:
radius = float(D[:, -1].min())
print('radius', radius)
lims, D3, I3 = index.range_search(xq, radius)
ntot = ndiff = 0
for i in range(len(xq)):
l0, l1 = lims[i], lims[i + 1]
Inew = set(I3[l0:l1])
if mt == faiss.METRIC_INNER_PRODUCT:
mask = D2[i] > radius
else:
mask = D2[i] < radius
Iref = set(I2[i, mask])
ndiff += len(Inew ^ Iref)
ntot += len(Iref)
print('ndiff %d / %d' % (ndiff, ntot))
assert ndiff < ntot * 0.02
def test_IVFPQ_non8bit(self):
d = 16
xt, xb, xq = get_dataset_2(d, 10000, 2000, 200)
nlist = 64
gt_index = faiss.IndexFlat(d)
gt_index.add(xb)
gt_D, gt_I = gt_index.search(xq, 10)
quantizer = faiss.IndexFlat(d)
ninter = {}
for v in '2x8', '8x2':
if v == '8x2':
index = faiss.IndexIVFPQ(
quantizer, d, nlist, 2, 8)
else:
index = faiss.IndexIVFPQ(
quantizer, d, nlist, 8, 2)
index.train(xt)
index.add(xb)
index.npobe = 16
D, I = index.search(xq, 10)
ninter[v] = faiss.eval_intersection(I, gt_I)
print('ninter=', ninter)
# this should be the case but we don't observe
# that... Probavly too few test points
# assert ninter['2x8'] > ninter['8x2']
# ref numbers on 2019-11-02
assert abs(ninter['2x8'] - 458) < 4
assert abs(ninter['8x2'] - 465) < 4
class TestFlat1D(unittest.TestCase):
def test_flat_1d(self):
rs = np.random.RandomState(123545)
k = 10
xb = rs.uniform(size=(100, 1)).astype('float32')
# make sure to test below and above
xq = rs.uniform(size=(1000, 1)).astype('float32') * 1.1 - 0.05
ref = faiss.IndexFlatL2(1)
ref.add(xb)
ref_D, ref_I = ref.search(xq, k)
new = faiss.IndexFlat1D()
new.add(xb)
new_D, new_I = new.search(xq, 10)
ndiff = (np.abs(ref_I - new_I) != 0).sum()
assert(ndiff < 100)
new_D = new_D ** 2
max_diff_D = np.abs(ref_D - new_D).max()
assert(max_diff_D < 1e-5)
class OPQRelativeAccuracy(unittest.TestCase):
# translated from test_opq.lua
def test_OPQ(self):
M = 4
ev = Randu10kUnbalanced()
d = ev.d
index = faiss.IndexPQ(d, M, 8)
res = ev.launch('PQ', index)
e_pq = ev.evalres(res)
index_pq = faiss.IndexPQ(d, M, 8)
opq_matrix = faiss.OPQMatrix(d, M)
# opq_matrix.verbose = true
opq_matrix.niter = 10
opq_matrix.niter_pq = 4
index = faiss.IndexPreTransform(opq_matrix, index_pq)
res = ev.launch('OPQ', index)
e_opq = ev.evalres(res)
print('e_pq=%s' % e_pq)
print('e_opq=%s' % e_opq)
# verify that OPQ better than PQ
for r in 1, 10, 100:
assert(e_opq[r] > e_pq[r])
def test_OIVFPQ(self):
# Parameters inverted indexes
ncentroids = 50
M = 4
ev = Randu10kUnbalanced()
d = ev.d
quantizer = faiss.IndexFlatL2(d)
index = faiss.IndexIVFPQ(quantizer, d, ncentroids, M, 8)
index.nprobe = 5
res = ev.launch('IVFPQ', index)
e_ivfpq = ev.evalres(res)
quantizer = faiss.IndexFlatL2(d)
index_ivfpq = faiss.IndexIVFPQ(quantizer, d, ncentroids, M, 8)
index_ivfpq.nprobe = 5
opq_matrix = faiss.OPQMatrix(d, M)
opq_matrix.niter = 10
index = faiss.IndexPreTransform(opq_matrix, index_ivfpq)
res = ev.launch('O+IVFPQ', index)
e_oivfpq = ev.evalres(res)
# verify same on OIVFPQ
for r in 1, 10, 100:
print(e_oivfpq[r], e_ivfpq[r])
assert(e_oivfpq[r] >= e_ivfpq[r])
class TestRoundoff(unittest.TestCase):
def test_roundoff(self):
# params that force use of BLAS implementation
nb = 100
nq = 25
d = 4
xb = np.zeros((nb, d), dtype='float32')
xb[:, 0] = np.arange(nb) + 12345
xq = xb[:nq] + 0.3
index = faiss.IndexFlat(d)
index.add(xb)
D, I = index.search(xq, 1)
# this does not work
assert not np.all(I.ravel() == np.arange(nq))
index = faiss.IndexPreTransform(
faiss.CenteringTransform(d),
faiss.IndexFlat(d))
index.train(xb)
index.add(xb)
D, I = index.search(xq, 1)
# this works
assert np.all(I.ravel() == np.arange(nq))
class TestSpectralHash(unittest.TestCase):
# run on 2019-04-02
ref_results = {
(32, 'global', 10): 505,
(32, 'centroid', 10): 524,
(32, 'centroid_half', 10): 21,
(32, 'median', 10): 510,
(32, 'global', 1): 8,
(32, 'centroid', 1): 20,
(32, 'centroid_half', 1): 26,
(32, 'median', 1): 14,
(64, 'global', 10): 768,
(64, 'centroid', 10): 767,
(64, 'centroid_half', 10): 21,
(64, 'median', 10): 765,
(64, 'global', 1): 28,
(64, 'centroid', 1): 21,
(64, 'centroid_half', 1): 20,
(64, 'median', 1): 29,
(128, 'global', 10): 968,
(128, 'centroid', 10): 945,
(128, 'centroid_half', 10): 21,
(128, 'median', 10): 958,
(128, 'global', 1): 271,
(128, 'centroid', 1): 279,
(128, 'centroid_half', 1): 171,
(128, 'median', 1): 253,
}
def test_sh(self):
d = 32
xt, xb, xq = get_dataset_2(d, 2000, 1000, 200)
nlist, nprobe = 1, 1
gt_index = faiss.IndexFlatL2(d)
gt_index.add(xb)
gt_D, gt_I = gt_index.search(xq, 10)
for nbit in 32, 64, 128:
quantizer = faiss.IndexFlatL2(d)
index_lsh = faiss.IndexLSH(d, nbit, True)
index_lsh.add(xb)
D, I = index_lsh.search(xq, 10)
ninter = faiss.eval_intersection(I, gt_I)
print('LSH baseline: %d' % ninter)
for period in 10.0, 1.0:
for tt in 'global centroid centroid_half median'.split():
index = faiss.IndexIVFSpectralHash(quantizer, d, nlist,
nbit, period)
index.nprobe = nprobe
index.threshold_type = getattr(
faiss.IndexIVFSpectralHash,
'Thresh_' + tt
)
index.train(xt)
index.add(xb)
D, I = index.search(xq, 10)
ninter = faiss.eval_intersection(I, gt_I)
key = (nbit, tt, period)
print('(%d, %s, %g): %d, ' % (nbit, repr(tt), period, ninter))
assert abs(ninter - self.ref_results[key]) <= 12
class TestRefine(unittest.TestCase):
def do_test(self, metric):
d = 32
xt, xb, xq = get_dataset_2(d, 2000, 1000, 200)
index1 = faiss.index_factory(d, "PQ4x4np", metric)
Dref, Iref = faiss.knn(xq, xb, 10, metric)
index1.train(xt)
index1.add(xb)
D1, I1 = index1.search(xq, 100)
recall1 = (I1 == Iref[:, :1]).sum()
# add refine index on top
index_flat = faiss.IndexFlat(d, metric)
index_flat.add(xb)
index2 = faiss.IndexRefine(index1, index_flat)
index2.k_factor = 10.0
D2, I2 = index2.search(xq, 10)
# check distance is computed properly
for i in range(len(xq)):
x1 = xq[i]
x2 = xb[I2[i, 5]]
if metric == faiss.METRIC_L2:
dref = ((x1 - x2) ** 2).sum()
else:
dref = np.dot(x1, x2)
np.testing.assert_almost_equal(dref, D2[i, 5], decimal=5)
# check that with refinement, the recall@10 is the same as
# the original recall@100
recall2 = (I2 == Iref[:, :1]).sum()
# print("recalls", recall1, recall2)
self.assertEquals(recall1, recall2)
def test_IP(self):
self.do_test(faiss.METRIC_INNER_PRODUCT)
def test_L2(self):
self.do_test(faiss.METRIC_L2)