mirror of
https://github.com/facebookresearch/faiss.git
synced 2026-10-12 02:13:36 +00:00
Alibek Zhakubayev
350a50b02e
Summary: Pull Request resolved: https://github.com/facebookresearch/faiss/pull/5415 The post-admission prune loop in `OperatingPoints::add` used a strict `<` comparison on time to detect dominated points: if (a[i].t < a[i - 1].t) When a new point B has strictly better performance than an existing point A at exactly the same search time, B is admitted (perf > a.back().perf), pushed to the back, and the prune loop runs. The condition evaluates `B.t < A.t` — false when times are equal — so A is left on the frontier. The optimal set then contains A which B strictly dominates (B.perf > A.perf, B.t == A.t), breaking the Pareto invariant. Fix: change `<` to `<=`. This is safe: the admission branches prevent duplicate (perf, t) pairs from being stored, so the prune loop never encounters equal time AND equal perf simultaneously. The `<=` condition only activates for the genuinely dominated case: higher-perf point at lower-or-equal time. D110116217 (2026-06-30) fixed the identical bug in the Python port of this function (`contrib/evaluation.py::add_operating_point`), changing `t < t2` to `t <= t2` in its prune loop. This diff fixes the C++ source that the Python version was documented to port. Reviewed By: mdouze Differential Revision: D110586087 fbshipit-source-id: 3bd48e6ce254ceef81478c966620741ed9faa089
126 lines
4.9 KiB
Python
126 lines
4.9 KiB
Python
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
|
#
|
|
# This source code is licensed under the MIT license found in the
|
|
# LICENSE file in the root directory of this source tree.
|
|
|
|
|
|
import unittest
|
|
import numpy as np
|
|
import faiss
|
|
|
|
from common_faiss_tests import for_all_simd_levels
|
|
|
|
|
|
@for_all_simd_levels
|
|
class TestParameterSpace(unittest.TestCase):
|
|
|
|
def test_nprobe(self):
|
|
index = faiss.index_factory(32, "IVF32,Flat")
|
|
ps = faiss.ParameterSpace()
|
|
ps.set_index_parameter(index, "nprobe", 5)
|
|
self.assertEqual(index.nprobe, 5)
|
|
|
|
def test_nprobe_2(self):
|
|
index = faiss.index_factory(32, "IDMap,IVF32,Flat")
|
|
ps = faiss.ParameterSpace()
|
|
ps.set_index_parameter(index, "nprobe", 5)
|
|
index2 = faiss.downcast_index(index.index)
|
|
self.assertEqual(index2.nprobe, 5)
|
|
|
|
def test_nprobe_3(self):
|
|
index = faiss.index_factory(32, "IVF32,SQ8,RFlat")
|
|
ps = faiss.ParameterSpace()
|
|
ps.set_index_parameter(index, "nprobe", 5)
|
|
index2 = faiss.downcast_index(index.base_index)
|
|
self.assertEqual(index2.nprobe, 5)
|
|
|
|
def test_nprobe_4(self):
|
|
index = faiss.index_factory(32, "PCAR32,IVF32,SQ8,RFlat")
|
|
ps = faiss.ParameterSpace()
|
|
|
|
ps.set_index_parameter(index, "nprobe", 5)
|
|
index2 = faiss.downcast_index(index.base_index)
|
|
index2 = faiss.downcast_index(index2.index)
|
|
self.assertEqual(index2.nprobe, 5)
|
|
|
|
def test_efSearch(self):
|
|
index = faiss.index_factory(32, "IVF32_HNSW32,SQ8")
|
|
ps = faiss.ParameterSpace()
|
|
ps.set_index_parameter(index, "quantizer_efSearch", 5)
|
|
index2 = faiss.downcast_index(index.quantizer)
|
|
self.assertEqual(index2.hnsw.efSearch, 5)
|
|
|
|
|
|
class TestAutoTuneCriterion(unittest.TestCase):
|
|
|
|
def test_set_groundtruth_with_distances(self):
|
|
nq, gt_nnn = 10, 5
|
|
crit = faiss.OneRecallAtRCriterion(nq, gt_nnn)
|
|
rng = np.random.default_rng(42)
|
|
gt_I = rng.integers(0, 100, size=(nq, gt_nnn)).astype("int64")
|
|
gt_D = rng.random(size=(nq, gt_nnn)).astype("float32")
|
|
crit.set_groundtruth(gt_D, gt_I)
|
|
self.assertEqual(crit.gt_nnn, gt_nnn)
|
|
self.assertEqual(crit.gt_I.size(), nq * gt_nnn)
|
|
|
|
def test_set_groundtruth_null_distances(self):
|
|
# gt_D is documented as optional (None skips the distances copy).
|
|
nq, gt_nnn = 10, 5
|
|
crit = faiss.OneRecallAtRCriterion(nq, gt_nnn)
|
|
rng = np.random.default_rng(42)
|
|
gt_I = rng.integers(0, 100, size=(nq, gt_nnn)).astype("int64")
|
|
crit.set_groundtruth(None, gt_I)
|
|
self.assertEqual(crit.gt_nnn, gt_nnn)
|
|
self.assertEqual(crit.gt_I.size(), nq * gt_nnn)
|
|
|
|
|
|
class TestOperatingPoints(unittest.TestCase):
|
|
def _keys(self, ops):
|
|
n = ops.optimal_pts.size()
|
|
return [ops.optimal_pts.at(i).key for i in range(n)]
|
|
|
|
def test_equal_time_better_perf_removes_dominated(self):
|
|
# B (perf=0.9, t=1.0) dominates A (perf=0.5, t=1.0) — same time,
|
|
# strictly better perf. A must be pruned from the optimal frontier.
|
|
ops = faiss.OperatingPoints()
|
|
self.assertTrue(ops.add(0.5, 1.0, "A"))
|
|
self.assertTrue(ops.add(0.9, 1.0, "B"))
|
|
keys = self._keys(ops)
|
|
# origin ("") + B only — A must not survive
|
|
self.assertEqual(len(keys), 2)
|
|
self.assertNotIn("A", keys, "A dominated by B; must not remain")
|
|
self.assertIn("B", keys)
|
|
|
|
def test_middle_insert_equal_time_removes_dominated(self):
|
|
# Frontier has P1=(0.5, t=1.0) and P2=(0.9, t=2.0). Adding
|
|
# P3=(0.7, t=1.0) inserts P3 in the middle and must prune P1 (same
|
|
# time, worse perf) — exercises the else-branch insertion path.
|
|
ops = faiss.OperatingPoints()
|
|
self.assertTrue(ops.add(0.5, 1.0, "P1"))
|
|
self.assertTrue(ops.add(0.9, 2.0, "P2"))
|
|
self.assertTrue(ops.add(0.7, 1.0, "P3"))
|
|
keys = self._keys(ops)
|
|
# origin ("") + P3 + P2 — P1 dominated by P3 at equal time
|
|
self.assertEqual(len(keys), 3)
|
|
self.assertNotIn("P1", keys, "P1 dominated by P3; must not remain")
|
|
self.assertIn("P3", keys)
|
|
self.assertIn("P2", keys)
|
|
|
|
def test_normal_pareto_frontier_preserved(self):
|
|
# Two points where neither dominates the other: higher perf at higher
|
|
# cost vs lower perf at lower cost. Both must remain on the frontier.
|
|
ops = faiss.OperatingPoints()
|
|
self.assertTrue(ops.add(0.5, 0.5, "approximate"))
|
|
self.assertTrue(ops.add(0.9, 2.0, "accurate"))
|
|
keys = self._keys(ops)
|
|
self.assertIn("approximate", keys)
|
|
self.assertIn("accurate", keys)
|
|
|
|
def test_worse_perf_at_equal_time_rejected(self):
|
|
# A point with worse perf and equal time is not admitted.
|
|
ops = faiss.OperatingPoints()
|
|
self.assertTrue(ops.add(0.9, 1.0, "good"))
|
|
admitted = ops.add(0.5, 1.0, "bad")
|
|
self.assertFalse(admitted)
|
|
self.assertNotIn("bad", self._keys(ops))
|