Deep Fusion Project

This project contains the models used for the experiments in the paper:

Liming Zhao, Jingdong Wang, Xi Li, Zhuowen Tu, and Wenjun Zeng. "On the Connection of Deep Fusion to Ensembling." arXiv preprint arXiv:1611.07718 (2016).

Contact: Liming Zhao ([email protected])

Introduction

In this work, we provide a systematic study to the prevailing ResNet architecture by showing a connection from a general deeply-fused net view to ensembling.

Our empirical results uncover that the deepest network among the ensemble components does not contribute the most significantly to the overall performance and instead it provides a manner to introduce many layers and thus guarantee the ensemble size.

Guided by the above study and observation, we develop a new deeply-fused network that combines two networks in a merge-and-run fusion manner.

Our approach demonstrates consistent improvements over the ResNet with the comparable setup on CIFAR-10, CIFAR-100, SVHN, and ImageNet.

Results

• Test error (%) on CIFAR (flip/translation augmentation) and SVHN (no augmentation):

Method	Depth	#Params	CIFAR-10	CIFAR-100	SVHN
DFN-MR1	56	1.7M	4.94	24.46	1.66
DFN-MR2	32	14.9M	3.94	19.25	1.51
DFN-MR3	50	24.8M	3.57	19.00	1.55

• Empirical results on ImageNet:

Requirements

• Install MXNet on a machine (Windows, Linux, and Mac OS) with CUDA GPU and optional cuDNN.

• Apply my modified data processing patch on the latest MXNet by merging the pull request:

git pull origin pull/3936/head master ```

• If you fail to apply the above patch, you can simply use my MXNet repository:

git clone --recursive -b fusenet https://github.com/zlmzju/mxnet.git ```

How to Train

Step by step tutorial with jupyter notebook will be available soon.

dataset

You can prepare the *.rec file by yourself, or simply download the Cifar dataset from data.dmlc.ml or my google drive, which includes both Cifar and SVHN datasets.

training

Current code supports training different deeply-fused nets on Cifar-10, Cifar-100 and SVHN, such as plain network, resnet, cross (dfn-mr),half (dfn-il), side (dfn-il without identities), fuse3 (three fusions), fuse6 (three fusions), and ensemble (with sharing weights, training code will come later). All the networks are contained in the network folder.

Note that the codes for training on ImageNet are available in the network/imagenet folder, but they still need refactoring to merge into the main codes.

For example, running the following command can train the DFN-MR network (we call it cross in the coding stage) on Cifar-10.

python train_model.py --dataset=cifar10 --network=cross --depth=56 --gpus=0,1 --dataset=<dataset location>

Other usages

visualization

If you wish to show the network architecture, run the following command to visualize the network.

python show_model.py --network=half --depth=26 --widen-factor=1

You will obtain a picture half_d26.png in the visualize folder, and more examples can be found there.

Note that you may need to install graphviz for visualization.

show results

The training log are saved to snapshot folder, and you can use get_results.py to obtain the final result of multiple runs in the format of median (mean +/- std, best).

Citation

Please cite our papers on deep fusion in your publications if it helps your research:

@article{WangWZZ16,
  author    = {Jingdong Wang and
               Zhen Wei and
               Ting Zhang and
               Wenjun Zeng},
  title     = {Deeply-Fused Nets},
  journal   = {CoRR},
  volume    = {abs/1605.07716},
  year      = {2016},
  url       = {http://arxiv.org/abs/1605.07716},
}

@article{ZhaoWLTZ16,
  author    = {Liming Zhao and
               Jingdong Wang and
               Xi Li and
               Zhuowen Tu and
               Wenjun Zeng},
  title     = {On the Connection of Deep Fusion to Ensembling},
  journal   = {CoRR},
  volume    = {abs/1611.07718},
  year      = {2016},
  url       = {http://arxiv.org/abs/1611.07718},
}