This directory contains additional module altid to the Faiss indexes useful for ID compression of NSG indexes.
altid contains 3 functions for alternative id storage:
- a function that traces the nodes visited during NSG search
- three classes that contain the graph structure (alternatives of the regular uncompressed one)
Note that this module needs a Faiss compiled after 2024-10-22 to compile (because replacing the NSG graph was not supported before). To install this Faiss version on conda, do:
conda install -c pytorch/label/nightly -c nvidia faiss-gpu=1.9.0 pytorch=*=*cuda*
This also installs pytorch in case you'd need it...
To compile, just run make.
This automatically runs the test below.
Run
python -m unittest test_altid.py
See also that file on how to use the module.
The C++ implementation is in altid_impl.{h,cpp}.