Deep Competitive Pathway Network (CoPaNet)

This repository contains the code for CoPaNet introduced in the paper "Deep Competitive Pathway Network" by Jia-Ren Chang and Yong-Sheng Chen.

This paper is accepted by Asian Conference on Machine Learning (ACML) 2017.

The code is built on fb.resnet.torch.

Introduction

CoPaNet is a network architecture where multiple pathways compete with each other. This network architecture yields a novel phenomenon which we called "pathway encoding". The pathway encoding means that the routing patterns of features can represent object semantic. The CoPaNet peforms state-of-the-art accuracy on CIFAR-10 and SVHN. On the large scale ILSVRC 2012 (ImageNet) dataset, CoPaNet achieves a similar accuracy as ResNet, but using less amount of parameters.

Figure 1: The concept of pathway encoding.

Figure 2: The pathway encoding on CIFAR-10 test set.

Usage

0. Install Torch and required dependencies like cuDNN. See the instructions here for a step-by-step guide.
1. Clone this repo: https://github.com/JiaRenChang/CoPaNet.git

We also provide our implementation of "CMaxTable".
It runs above 2x faster than naive implementation in Torch's nn.

As an example, the following command trains a CoPaNet with depth 164 on CIFAR-10:

th main.lua -netType CoPaNet -dataset cifar10 -batchSize 128 -nEpochs 300 -depth 164

As another example, the following command trains a CoPaNet with depth 26 on ImageNet:

th main.lua -netType CoPaNet -dataset imagenet -data [dataFolder] -batchSize 256 -nEpochs 100 -depth 26 -nGPU 4

Please refer to fb.resnet.torch for data preparation.

Contact

followwar at gmail.com
Any discussions, suggestions and questions are welcome!