/resnet-cifar10

ResNet for Cifar10

Residual Networks Test

============================

Original Paper: "Deep Residual Learning for Image Recognition"(http://arxiv.org/abs/1512.03385) and "Identity Mappings in Deep Residual Networks"(http://arxiv.org/abs/1603.05027)

By Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.

Microsoft Research Asia (MSRA).

Introduction

This repository contains the models for testing resnet networks on cifar-10

  • 50000 samples for training, batch size 250, 200 iterations for 1 epoch, 64000 iterations in total
  • 10000 samples for testing, test batch size 100 and 100 test iterations reduce learning rate after 32000 iters by factor of 10 then another factor of 10 after anohter 16000 iters

Notes

Data augmentation applied (please find the data augmentation fork in https://github.com/twtygqyy/caffe-augmentation):

max_color_shift = 5

contrast_variation = 0.8 ~ 1.2

max_brightness_shift = 5 

zero-padding with 2 pixels for each side and crop with 32x32

Please download the training images with zero-padding here Google Drive

Result

  • Resnet-20: best model achieved 0.927 accuracy on test datasets for single round and single crop

  • Resnet-32: best model achieved 0.9364 accuracy on test datasets for single round and single crop

  • Resnet-56: best model achieved 0.9418 accuracy on test datasets for single round and single crop

  • Resnet-56: 0.944 accuracy on test datasets with LSUV initializer (https://github.com/ducha-aiki/LSUVinit/blob/master/tools/extra/lsuv_init.py/)

  • WRN-28-10 ["Wide Residual Networks" (http://arxiv.org/abs/1605.07146)]: best model achieved 0.958 accuracy on test datasets for single round and single crop