alexklibisz / elastiknn

Elasticsearch plugin for nearest neighbor search. Store vectors and run similarity search using exact and approximate algorithms.

https://alexklibisz.github.io/elastiknn

Apache License 2.0

371 stars 48 forks source link

Performance: picked better parameters for Fashion MNIST dataset (239QPS @ 0.96 recall) #727

Closed alexklibisz closed 1 month ago

alexklibisz commented 1 month ago

Related Issue

611

Changes

Ran a big gridsearch over some new parameters and found ones that drastically performance by about ~25% and get the 0.96 QPS up to ~240, well above the goal of 200.

The full gridsearch results are visible in this commit: https://github.com/alexklibisz/elastiknn/pull/727/commits/c7efcf14a59c4cdcd15636042aeb6f4e381634ac

I selected one of the better parameters to keep in main.

Testing and Validation

Standard CI