Competitive-Inner-Imaging-SENet

Source code of paper:

(not availbale now)

Architecture

Competitive Squeeze-Exciation Architecutre for Residual block

SE-ResNet module and CMPE-SE-ResNet modules:

Normal SE	Double FC squeezes	Conv 2x1 pair-view	Conv 1x1 pair-view

The Novel Inner-Imaging Mechanism for Channel Relation Modeling in Channel-wise Attention of ResNets (even All CNNs):

Basic Inner-Imaing Mode	Folded Inner-Imaging Mode

Requirements

• MXNet 1.2.0
• Python 2.7
• CUDA 8.0+(for GPU)

Citation

not available now

Essential Results

Best record of this novel model on CIFAR-10 and CIFAR-100 (used "mixup" (https://arxiv.org/abs/1710.09412)) can achieve: 97.55% and 84.38%.

The test result on Kaggle: CIFAR-10 - Object Recognition in Images

Inner-Imaging Examples & Channel-wise Attention Outputs