quantize

This package allows to train regular torch neural networks with quantized weights.

How to use it?

local q = require 'quantize'
model:quantize(q.fixed(1, 4))

To enable call :quantize() on a model and pass a quantizer function. For now there's one provided with the package.

q.fixed(integer, reminder) rounds all weights to a range representable using integer bits for integer part and reminder bits for remainder part (it also adds one implicit bit for the sign).