Interleaved Group Convolutions

This project contains the code implementation used for the experiments in the paper:

Interleaved Group Convolutions. Ting Zhang, Guo-Jun Qi, Bin Xiao, and Jingdong Wang. In International Conference on Computer Vision (ICCV), 2017. arXiv preprint arXIV:1707.02725 (2017)

Introduction

In this work, we present a simple and modularized neural network architecture, named interleaved group convolutional neural networks (IGCNets). The main point lies in a novel building block, a pair of two successive interleaved group convolutions: primary group convolution and secondary group convolution. The two group convolutions are complementary.

Illustrating the interleaved group convolution, with L = 2 primary partitions and M = 3 secondary partitions. The convolution for each primary partition in primary group convolution is spatial. The convolution for each secondary partition in secondary group convolution is point-wise (1 × 1).

Our motivation comes from the four branch presentation of regular convolution illustrated in the following picture.

(a) Regular convolution. (b) Four-branch representation of the regular convolution. The shaded part in (b), we call crosssummation, is equivalent to a three-step transformation: permutation, secondary group convolution, and permutation back.

Results

	#Params (M)	FLOPs (10^9)	training error		test error
			top-1	top-5	top-1	top-5
ResNet(C=64)	11.151	1.8	22.41	6.53	31.06	11.38
ResNet(C=69)	11.333	2.1	21.43	5.96	30.58	10.77
IGC-L4M32+Ident.	11.205	1.9	21.71	6.21	30.77	10.99
IGC-L16M16+Ident.	11.329	2.2	19.97	5.44	29.40	10.32
IGC-L100M2+Ident.	8.61	1.3	13.93	2.75	26.95	8.92

Imagenet classification results of a ResNet of depth 18 and our approach. Both ResNets and our networks contain four stages, and when down-sampling is performed, the channel number is doubled. For ResNets, C is the channel number at the first stage. For our networks except IGC-L100M2+Ident., we double the channel number by doubling M and keeping L unchanged. For IGCL100M2+Ident., we double the channel number by doubling L and keeping M unchanged.

More results can be found in the paper.

Requirements

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

• Add the code in src/ to MXNet src/operator/

• Build MXNet

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

How to Train

Current code supports training IGCNets on Cifar-10, Cifar-100 and SVHN, such as plain, resnet, plain_igc,resnet_igc. All the networks are contained in the network folder.

For example, running the following command can train the plain_igc network on Cifar-10.

python train_model.py --network=plain_igc --depth=38 --gpus=0,1 --primary-partition=24 --secondary-partition=2 --batch-size=64 --data-dir=<dataset location> --dataset=cifar10

For example, running the following command can train the resnet_igc_imgnet_d18 network on ImageNet.

python train_imagenet.py --network=resnet_igc_imgnet_d18 --depth=18 --gpus=0,1,2,3,4,5,6,7 --primary-partition=100 --batch-size=256 --data-dir=<dataset location>

Citation

Please cite our papers 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}
}

@article{DBLP:journals/corr/ZhangQ0W17,
  author    = {Ting Zhang and
               Guo{-}Jun Qi and
               Bin Xiao and
               Jingdong Wang},
  title     = {Interleaved Group Convolutions for Deep Neural Networks},
  journal   = {CoRR},
  volume    = {abs/1707.02725},
  year      = {2017},
  url       = {http://arxiv.org/abs/1707.02725}
}