QINCo2
A neural vector compression method for large-scale similarity search that predicts each quantization step's codebook from a neural network instead of using a fixed one.
QINCo, short for Quantization with Implicit Neural Codebooks, is a residual vector quantization method in which a small neural network predicts the codebook to use at each quantization step, conditioned on the vector as quantized so far, rather than relying on one fixed codebook per step. This increases the compression system's capacity without storing more codebook vectors, and because each step is trained with its own quantization error, a system trained for one compression rate can also be used at lower rates.
QINCo2, the version in this repository, adds a faster approximate encoding method with similar error to the original at much lower training and encoding cost, beam search during encoding for lower compression error, and an optional pairwise decoder that improves retrieval accuracy.
It is written in Python and accompanies papers at ICML and ICLR.