ResNet-on-Cifar10

by junyuseu

Reimplementation ResNet on cifar10 with caffe

122 Stars 82 Forks Last release: Not found 42 Commits 0 Releases

Available items

No Items, yet!

The developer of this repository has not created any items for sale yet. Need a bug fixed? Help with integration? A different license? Create a request here:

ResNets experiments on cifar10 with caffe

Citation

@article{He2015,
    author = {Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun},
    title = {Deep Residual Learning for Image Recognition},
    journal = {arXiv preprint arXiv:1512.03385},
    year = {2015}
    }

Introduction

This repository reimplements resnet experiments on cifar10 with caffe according to the paper "Deep Residual Learning for Image Recognition" (http://arxiv.org/abs/1512.03385). The data augmentation means 4 pixels are padded on each side for every images during training. You can make datasets prepared by using the scripts.

Structure

The network structure is here(we only list the network of 20 depth):
ResNet_20
PlainNet_20

Usage

First, you should make sure that your caffe is correctly installed. You can follow this blog's instructions if you use windows.(https://zhuanlan.zhihu.com/p/22129880)

for training

  caffe train -solver=solver.prototxt -gpu 0

for testing

  caffe test -model=res20_cifar_train_test.prototxt -weights=ResNet_20.caffemodel -iterations=100 -gpu 0

Result

Result with data augmentation:

model

Repeated Reference
20 lyaers 91.94% 91.25%
32 layers 92.70% 92.49%
44 layers 93.01% 92.83%
56 layers 93.19% 93.03%
110 layers 93.56% 93.39%

notice:'Repeated' means reimplementation results and 'Reference' means result in the paper.We got even better results than the original paper

Compare result(without data augmentation):

model

PlainNet ResNet
20 lyaers 90.10% 91.74%
32 layers 86.96% 92.23%
44 layers 84.45% 92.67%
56 layers 85.26% 93.09%
110 layers X 93.27%

Blog address


zhihu

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.