masabdi/multi-resnet

Multi-Residual Networks

Multi-Residual Networks

By Masoud Abdi and Saeid Nahavandi

Implementation of Multi-Residual Networks (http://arxiv.org/abs/1609.05672).

CIFAR Test Error Rates

Network	depth	k	w	Parameters	CIFAR-10 (%)	CIFAR-100 (%)
Pre-Resnet	1001	1	1	10.2M	4.62(4.69+/-0.20)	22.71(22.68+/-0.22)
Multi-Resnet	200	5	1	10.2M	4.35(4.36+/-0.04)	20.42(20.44+/-0.15)
Multi-Resnet	398	5	1	20.4M	3.92	20.59
Multi-Resnet	26	2	10	72M	3.96	19.45
Multi-Resnet	26	4	10	154M	3.73	19.60

##Usage:

th main.lua -netType multi-resnet -depth 200 -k 5 -batchSize 64 -nGPU 2 -nThreads 4 -dataset cifar10 -nEpochs 200

##Notes:

In order to see the effect of model parallelism use modelParallel (Tested on K80 GPU):

for model parallelism on 2 GPUs use:

th main.lua -netType mpreresnet -dataset cifar10 -batchSize 128 -depth 110 -k 4 -modelParallel true

for data parallelism on 2 GPUs use:

th main.lua -netType mpreresnet -dataset cifar10 -batchSize 128 -depth 434 -k 1 -nGPU 2

It achieves up to 15% speed up and there is room for improvement.

The code is based on https://github.com/facebook/fb.resnet.torch and https://github.com/KaimingHe/resnet-1k-layers

##Contact

Please contact me on mabdi{at}deakin.edu.au

I appreciate any discussion, suggestion or contribution.