Vector quantization is a large area that use many algorithms and theoretical results (from Fejes Toth's hexagonal grid to Peter Gruber's quantization on manifolds, including non-uniform and non-isotropic source results, KNN, SOM, different metrics, etc).
I know that:
- in tensorflow the layer could be quantized as a tensor object,
- in pytorch there is useful methods for standart quantization.
But what if I'm interested to implemet my own algorithm, some special grids?
Should I rewrite layer in my own way, or there is interface for such tasks?