Introduction

The source code of the proposed method is available online at Github.

Abstract

This project contains the source code of the paper Cellular Binary Neural Network for Accurate Classification.

The demo consists of two parts. The first one contains the training implementation of CBN-Net model on the CIFAR-10 and ImageNet dataset. The second one is about the evaluation of the CBN-Net on these two datasets.

For easy evaluation, we also provide the pre-trained model parameters.

Prerequisites

• Ubuntu 16.04
• Python 3.6
• Pytorch >1.7.0

Structure

For both the CIFAR-10 and ImageNet dataset, the main folder should be in the following structure:

CBN-Net	
  Evaluation
    Cifar10
    	ResNet20
        	eval.py
    	VGGSmall
        	eval.py
    ImageNet
    	eval.py
    
  Training
    Cifar10
    	ResNet20
        	train_kl.py
    	VGGSmall
        	train_kl.py
    ImageNet
    	train_kl.py

The datasets should be put into the ‘data’ folder or point to the place where the data is stored. The log files are stored in the ‘log’ folder.

Pre-trained Model

The pre-trained model file on the ImageNet dataset can be downloaded from https://pan.baidu.com/s/1AhgGIBNFc4R-E15KeA-98w (extracted code: nips), and one should put it in the ‘save’ folder for easy evaluation.

Training

Run the training script to train the CBN-Net with different backbones on both datasets:

python train_kl.py

Evaluation

Run the evaluating script to evaluate the trained models with different backbones on both datasets:

python eval.py