| 1 |
# Copyright (c) Meta Platforms, Inc. and affiliates. |
| 2 |
# |
| 3 |
# This source code is licensed under the MIT license found in the |
| 4 |
# LICENSE file in the root directory of this source tree. |
| 5 |
|
| 6 |
import faiss |
| 7 |
|
| 8 |
from faiss.contrib.evaluation import knn_intersection_measure |
| 9 |
from faiss.contrib import datasets |
| 10 |
|
| 11 |
# 64-dim vectors, 50000 vectors in the training, 100000 in database, |
| 12 |
# 10000 in queries, dtype ('float32') |
| 13 |
ds = datasets.SyntheticDataset(64, 50000, 100000, 10000) |
| 14 |
d = 64 # dimension |
| 15 |
|
| 16 |
# Constructing the refine PQ index with SQfp16 with index factory |
| 17 |
index_fp16 = faiss.index_factory(d, 'PQ32x4fs,Refine(SQfp16)') |
| 18 |
index_fp16.train(ds.get_train()) |
| 19 |
index_fp16.add(ds.get_database()) |
| 20 |
|
| 21 |
# Constructing the refine PQ index with SQ8 |
| 22 |
index_sq8 = faiss.index_factory(d, 'PQ32x4fs,Refine(SQ8)') |
| 23 |
index_sq8.train(ds.get_train()) |
| 24 |
index_sq8.add(ds.get_database()) |
| 25 |
|
| 26 |
# Parameterization on k factor while doing search for index refinement |
| 27 |
k_factor = 3.0 |
| 28 |
params = faiss.IndexRefineSearchParameters(k_factor=k_factor) |
| 29 |
|
| 30 |
# Perform index search using different index refinement |
| 31 |
D_fp16, I_fp16 = index_fp16.search(ds.get_queries(), 100, params=params) |
| 32 |
D_sq8, I_sq8 = index_sq8.search(ds.get_queries(), 100, params=params) |
| 33 |
|
| 34 |
# Calculating knn intersection measure for different index types on refinement |
| 35 |
KIM_fp16 = knn_intersection_measure(I_fp16, ds.get_groundtruth()) |
| 36 |
KIM_sq8 = knn_intersection_measure(I_sq8, ds.get_groundtruth()) |
| 37 |
|
| 38 |
# KNN intersection measure accuracy shows that choosing SQ8 impacts accuracy |
| 39 |
assert (KIM_fp16 > KIM_sq8) |
| 40 |
|
| 41 |
print(I_sq8[:5]) |
| 42 |
print(I_fp16[:5]) |