Link & code is an indexing method that combines HNSW indexing with compression and exploits the neighborhood structure of the similarity graph to improve the reconstruction. It is described in
@inproceedings{link_and_code,
author = {Matthijs Douze and Alexandre Sablayrolles and Herv\'e J\'egou},
title = {Link and code: Fast indexing with graphs and compact regression codes},
booktitle = {CVPR},
year = {2018}
}
ArXiV here
The test runs with 3 files:
-
bench_link_and_code.py
: driver script -
datasets.py
: code to load the datasets. The example code runs on the deep1b and bigann datasets. See the toplevel README on how to downlod them. They should be put in a directory, edit datasets.py to set the path. -
neighbor_codec.py
: this is where the representation is trained.
The code runs on top of Faiss. The HNSW index can be extended with a
ReconstructFromNeighbors
C++ object that refines the distances. The
training is implemented in Python.
The results of table 2 (accuracy on deep100M) in the paper can be obtained with:
python bench_link_and_code.py \
--db deep100M \
--M0 6 \
--indexkey OPQ36_144,HNSW32_PQ36 \
--indexfile $bdir/deep100M_PQ36_L6.index \
--beta_nsq 4 \
--beta_centroids $bdir/deep100M_PQ36_L6_nsq4.npy \
--neigh_recons_codes $bdir/deep100M_PQ36_L6_nsq4_codes.npy \
--k_reorder 0,5 --efSearch 1,1024
Set bdir
to a scratch directory.
Explanation of the flags:
-
--db deep1M
: dataset to process -
--M0 6
: number of links on the base level (L6) -
--indexkey OPQ36_144,HNSW32_PQ36
: Faiss index key to construct the HNSW structure. It means that vectors are transformed by OPQ and encoded with PQ 36x8 (with an intermediate size of 144D). The HNSW level>0 nodes have 32 links (theses ones are "cheap" to store because there are fewer nodes in the upper levels. -
--indexfile $bdir/deep1M_PQ36_M6.index
: name of the index file (without information for the L&C extension) -
--beta_nsq 4
: number of bytes to allocate for the codes (M in the paper) -
--beta_centroids $bdir/deep1M_PQ36_M6_nsq4.npy
: filename to store the trained beta centroids -
--neigh_recons_codes $bdir/deep1M_PQ36_M6_nsq4_codes.npy
: filename for the encoded weights (beta) of the combination -
--k_reorder 0,5
: number of restults to reorder. 0 = baseline without reordering, 5 = value used throughout the paper -
--efSearch 1,1024
: number of nodes to visit (T in the paper)
The script will proceed with the following steps:
-
load dataset (and possibly compute the ground-truth if the ground-truth file is not provided)
-
train the OPQ encoder
-
build the index and store it
-
compute the residuals and train the beta vocabulary to do the reconstuction
-
encode the vertices
-
search and evaluate the search results.
With option --exhaustive
the results of the exhaustive column can be
obtained.
The run above should output:
...
setting k_reorder=5
...
efSearch=1024 0.3132 ms per query, R@1: 0.4283 R@10: 0.6337 R@100: 0.6520 ndis 40941919 nreorder 50000
which matches the paper's table 2.
Note that in multi-threaded mode, the building of the HNSW strcuture is not deterministic. Therefore, the results across runs may not be exactly the same.
Figure 5 just evaluates the combination of HNSW and PQ. For example, the operating point L6&OPQ40 can be obtained with
python bench_link_and_code.py \
--db deep1M \
--M0 6 \
--indexkey OPQ40_160,HNSW32_PQ40 \
--indexfile $bdir/deep1M_PQ40_M6.index \
--beta_nsq 1 --beta_k 1 \
--beta_centroids $bdir/deep1M_PQ40_M6_nsq0.npy \
--neigh_recons_codes $bdir/deep1M_PQ36_M6_nsq0_codes.npy \
--k_reorder 0 --efSearch 16,64,256,1024
The arguments are similar to the previous table. Note that nsq = 0 is simulated by setting beta_nsq = 1 and beta_k = 1 (ie a code with a single reproduction value).
The output should look like:
setting k_reorder=0
efSearch=16 0.0147 ms per query, R@1: 0.3409 R@10: 0.4388 R@100: 0.4394 ndis 2629735 nreorder 0
efSearch=64 0.0122 ms per query, R@1: 0.4836 R@10: 0.6490 R@100: 0.6509 ndis 4623221 nreorder 0
efSearch=256 0.0344 ms per query, R@1: 0.5730 R@10: 0.7915 R@100: 0.7951 ndis 11090176 nreorder 0
efSearch=1024 0.2656 ms per query, R@1: 0.6212 R@10: 0.8722 R@100: 0.8765 ndis 33501951 nreorder 0
The results with k_reorder=5 are not reported in the paper, they represent the performance of a "free coding" version of the algorithm.