Faiss: How to use OPQ to reduce the dimensions
I have read the paper：Polysemous Codes in ECCV2016
I have an question：In the paper, you say “Alexandre Sablayrolles had the idea of extending the OPQ method to reduce the number of dimensions”. The original OPQ can not reduce the dimensions.
I want to know how to use OPQ to reduce the dimensions?
Thank you!
question
zhaokang0826
>All comments
If you look at the formulation in the original paper "Optimized Product Quantization for Approximate Nearest Neighbor Search" (CVPR'13), it turns out you can use Algorithm 1 with a non-square R matrix (say d*p). You just need to initialize it so that R^T R = I and you follow the steps described, as all the subsequent matrices will have the right number of dimensions.
alexandresablayrolles
on 17 Nov 2017
👍2
Was this page helpful?
0 / 5 - 0 ratings
Related issues
zoe-cheung
·
3Comments
hipitt
·
3Comments
jukaradayi
·
3Comments
zjjott
·
3Comments
cherryPotter
·
3Comments
Most helpful comment
If you look at the formulation in the original paper "Optimized Product Quantization for Approximate Nearest Neighbor Search" (CVPR'13), it turns out you can use Algorithm 1 with a non-square R matrix (say d*p). You just need to initialize it so that R^T R = I and you follow the steps described, as all the subsequent matrices will have the right number of dimensions.