implement of DoReFaNet with tensorflow based on cifar10 dataset

the link of the paper is:https://arxiv.org/abs/1606.06160

My implementation is based on work in : https://github.com/AngusG/tensorflow-xnor-bnn & https://github.com/skoppula/dorefa-net.git

Data

This implementation supports cifar10/cifar100

Dependencies

tensorflow version 1.2.1

Training

• Train cifar10 model using gpu:

  Full presion:

  python main_full.py

  if you want to train your own DoReFanet

  python main_for_DoReFa.py

  Dorefanet:

  weight: 1 bit output:2 bits epoches: 100 learning rate:0.001 accuracy:85.3%

  weight: 2 bits output:2 bits epoches: 100 learning rate:0.001 accuracy:85.5%

  weight: 1 bit output:3 bits epoches: 100 learning rate:0.001 accuracy:85.5%

BNN:

weight: 1 bit(-1,1) output:1 bits(-1,1)
epoches: 100
learning rate:0.001
accuracy:84.1%

• Train cifar10 model using cpu:

  if you did not own a GPU which can speed up the training, you just need to change the GPU in main.py into True

Results

Cifar10 should reach at least 88% top-1 accuracy