canesin / bnn_kernel

This is a BNN_Kernel on PyTorch for 1-bit networks in image data processing

Description

In theory, Binarized Neural Networks boast of x32 compression and x10 acceleration on GPUs. However, all variables are stored and manipulated in the form of float-32 on PyTorch. In this case, if we want to implement 1-bit networks on PyTorch, we have to change the kernel.

This repository contains a wrapped CUDA kernel that supports both 1-bit storage and computation on PyTorch

How to use

check the environment

GCC:     >=5.0
PyTorch: <1.0
PyThon:  >=3.5
CUDA:    >=8.0

build the kernel

GPU

cd ./csrc/binop
make

CPU

cd ./csrc/binop_cpu
make

check the API and use

GPU

import torch
import binop
...

CPU

import torch
import binop_cpu

References

How to wrap a CUDA and use its functions in PyTorch:

Pytorch Custom CUDA kernel Tutorial

Examples

You can see the GPU example in test

You can see the CPU example in test