/nn8

Low precision Torch nn library using uint8_t GEMM (experiment)

Primary LanguageCMIT LicenseMIT

Low precision (8-bit) Torch nn library

This experimental work uses Google's low precision GEMM and only supports few modules.

Install

git clone https://github.com/jhjin/nn8 --recursive
cd nn8
luarocks make rocks/nn8-scm-1.rockspec

Test

th test-precision.lua   # small model
th test-speed.lua       # large model