Awesome

We plan to release these two modified architectures implemented by MXNet for image classification.

Modified CRUNet.mxnet

A MXNet implementation of modified CRUNet.

In this part, we implement a modified CRU-Net (CRU-Net 56, 116) architecture via MXNet. The original one is described in the paper Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks proposed by Yunpeng Chen, Xiaojie Jin, Bingyi Kang, Jiashi Feng, Shuicheng Yan. This paper is accepted by IJCAI 2018.

Original architecture

This is an illustration of a collective residual unit (CRU) building block.

The original CRU Networks (CRUNet 56 and CRUNet 116) are based on the following table:

We implement the modified CRU Networks based on original CRUNet 56 and CRUNet 116.

What's the difference between modified version and original version ?

1. The size of input data is 112x112 not 224x224. In order to preserve higher feature map resolution, we follow the setting of input in [2]. Specifically, The first convolution layer with 7x7 kernel size and 2 stride is replaced by 3x3 kernel size and 1 stride. Moreover, we remove the following max pooling layer with 3x3 kernel size and 2 stride.

2. We adopt the improved residual unit mentioned in [2]. Specifically, the improved residual unit is constructed by BN-Conv-BN-PReLu-Conv-BN, where BN denotes batch normalization layer, PReLu is Parametric Rectified Linear Unit activation layer and Conv means convolution layer.

3. We replace all ReLu activation layers with PReLu activation layers in our whole architecture.

4. We follow the output setting mentioned in [2]. Specifically, we choose Option-E with structure of BN-Dropout-FC-BN after the last convolutional layer, where Dropout means dropout layer and FC denotes fully connected layer.

This modified Collective Residual Unit Network architecture can be directly integrated into the library of insightface.

In order to verify the performance of our modified CRUNet 116, we train it with arcface loss on data given by deepinsight and deploy data mirror, data color and random crop augmentation and adopt 80 batch size on 8 GPUs. Finally,
LFW: 0.99867+-0.00180;
CFP_FP: 0.98386+-0.00542;
CFP_FF: 0.99729+-0.00216;
CALFW: 0.95933+-0.01119;
CPLFW: 0.92833+-0.01624;
AgeDB30: 0.97983+-0.00765.
All results come from the model trained at epoch 36. Please see this link.

Reference

[1] Yunpeng Chen, Xiaojie Jin, Bingyi Kang, Jiashi Feng, Shuicheng Yan. "Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks" IJCAI 2018.

ImageNet Pre-trained models can be downloaded from the author's github

[2] Jiankang Deng, Jia Guo, Stefanos Zafeiriou. "ArcFace: Additive Angular Margin Loss for Deep Face Recognition"

Modified Residual-Attention-Network.mxnet

A MXNet implementation of modified Residual Attention Network

In this part, we implement a modified Residual Attention Network (Attention 92) architecture via MXNet. The original one is described in the paper Residual Attention Network for Image Classification proposed by Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yan, Chen Li, Honggang Zhang, Xiaogang Wang, Xiaoou Tang. This paper is accepted by CVPR 2017 Spotlight.

Original architecture

This is an example architecture of Residual Attention Network for ImageNet.

The receptive field comparison between mask branch and trunk branch is illustraded by the following figure:

The original Residual Attention Network (Attention 92) for ImageNet is based on the following table:

We implement the modified Residual Attention Network based on original Attention 92. This MXNet implementation is refered to tengshaofeng's PyTorch version.

What's the difference between modified version and original version ?

1. The size of input data is 112x112 not 224x224. In order to preserve higher feature map resolution, we follow the setting of input in [2]. Specifically, The first convolution layer with 7x7 kernel size and 2 stride is replaced by 3x3 kernel size and 1 stride. Moreover, we remove the following max pooling layer with 3x3 kernel size and 2 stride.

2. We adopt the improved residual unit mentioned in [2]. Specifically, the improved residual unit is constructed by BN-Conv-BN-PReLu-Conv-BN, where BN denotes batch normalization layer, PReLu is Parametric Rectified Linear Unit activation layer and Conv means convolution layer.

3. We replace all ReLu activation layers with PReLu activation layers in our whole architecture.

4. We follow the output setting mentioned in [2]. Specifically, we choose Option-E with structure of BN-Dropout-FC-BN after the last convolutional layer, where Dropout means dropout layer and FC denotes fully connected layer.

This modified Residual Attention Network architecture can be directly integrated into the library of insightface.

:fire:Note! If you want to train Residual Attention Network, please make sure the MXNet built from source.:fire:

Reference

[1] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yan, Chen Li, Honggang Zhang, Xiaogang Wang, Xiaoou Tang. "Residual Attention Network for Image Classification" CVPR 2017 Spotlight.

Caffe implementation

Pytorch implementation

[2] Jiankang Deng, Jia Guo, Stefanos Zafeiriou. "ArcFace: Additive Angular Margin Loss for Deep Face Recognition"