Nice article that explains product quantization very well, thanks! PQ is really a nice compression technique. I implemented PQ and Optimized PQ [1] a while bac

Nice article that explains product quantization very well, thanks! PQ is really a nice compression technique. I implemented PQ and Optimized PQ [1] a while back in our word embedding package for Rust: https://github.com/finalfusion/finalfusion-rust/ https://github.com/finalfusion/finalfusion-rust/ https://github.com/finalfusion/reductive/ https://github.com/finalfusion/reductive/ Particularly Optimized PQ was effective in reducing vector sizes ~10 times with virtually no reconstruction loss. This made it much easier to ship models (no more 3GB embedding matrix with a neural net that is just a few megabytes large). [1] http://kaiminghe.com/publications/pami13opq.pdf http://kaiminghe.com/publications/pami13opq.pdf