alexxnica/resnet-1k-layers

Deep Residual Networks with 1K Layers

Deep Residual Networks with 1K Layers

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

Microsoft Research Asia (MSRA).

Table of Contents

0. Introduction
1. Notes
2. Usage

Introduction

This repository contains re-implemented code for the paper "Identity Mappings in Deep Residual Networks" (http://arxiv.org/abs/1603.05027). This work enables training quality 1k-layer neural networks in a super simple way.

Acknowledgement: This code is re-implemented by Xiang Ming from Xi'an Jiaotong Univeristy for the ease of release.

Seel Also: Re-implementations of ResNet-200 [a] on ImageNet from Facebook AI Research (FAIR): https://github.com/facebook/fb.resnet.torch/tree/master/pretrained

Related papers:

[a]	@article{He2016,
		author = {Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun},
		title = {Identity Mappings in Deep Residual Networks},
		journal = {arXiv preprint arXiv:1603.05027},
		year = {2016}
	}

[b] @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}
	}

Notes

0. This code is based on the implementation of Torch ResNets (https://github.com/facebook/fb.resnet.torch).

1. The experiments in the paper were conducted in Caffe, whereas this code is re-implemented in Torch. We observed similar results within reasonable statistical variations.

2. To fit the 1k-layer models into memory without modifying much code, we simply reduced the mini-batch size to 64, noting that results in the paper were obtained with a mini-batch size of 128. Less expectedly, the results with the mini-batch size of 64 are slightly better:

   mini-batch	CIFAR-10 test error (%): (median (mean+/-std))
   128 (as in [a])	4.92 (4.89+/-0.14)
   64 (as in this code)	4.62 (4.69+/-0.20)

3. Curves obtained by running this code with a mini-batch size of 64 (training loss: y-axis on the left; test error: y-axis on the right):

Usage

0. Install Torch ResNets (https://github.com/facebook/fb.resnet.torch) following instructions therein.
1. Add the file resnet-pre-act.lua from this repository to ./models.
2. To train ResNet-1001 as of the form in [a]:

th main.lua -netType resnet-pre-act -depth 1001 -batchSize 64 -nGPU 2 -nThreads 4 -dataset cifar10 -nEpochs 200 -shareGradInput false

Note: ``shareGradInput=true'' is not valid for this model yet.