Source code of paper:
(not availbale now)
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 |
---|---|
not available now
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