What's this

Implementation of Squeeze and Excitation Networks (SENets) [2] by chainer

Dependencies

git clone https://github.com/nutszebra/SENets.git
cd SENets
git submodule init
git submodule update
# Note: chainer==1.24.0

How to run

 # for SENets with compression rate 8
 python main_se_residual_net.py -g 0 -p ./result_senet_8 -e 250 -b 64 -lr 0.1 -k 1 -n 18 -multiplier 4 -r 8
  # for SENets with compression rate 16
 python main_se_residual_net.py -g 0 -p ./result_senet_16 -e 250 -b 64 -lr 0.1 -k 1 -n 18 -multiplier 4 -r 16
 # for resnet
 python main_residual_net.py -g 0 -p ./result_resnet -e 250 -b 64 -lr 0.1 -k 1 -n 18 -multiplier 4

Details about my implementation

Data Augmentation:

Train: Pictures are randomly resized in the range of [32, 36], then 32x32 patches are extracted randomly and are normalized locally. Horizontal flipping is applied with 0.5 probability. Cutout [3] is applied with 0.5 probability (16x16 window) before normalization.

Test: Pictures are resized to 32x32, then they are normalized locally. Single image test is used to calculate total accuracy.
Optimization: Momentum SGD (momentum is 0.9)
Scheduling: 0.1 is multiplied to learning rate at [150, 200] epochs.
Initial learning rate: 0.1, but warm-up learnig rate, 0.01, is only used at first epoch.
Weight decay: 0.0001

Cifar10 result

network	depth	Compression Rate: r	Parameters (M)	total accuracy (%)
SEResNet (my implementation) [2]	164 + 108	8	2.0	95.69
SEResNet (my implementation) [2]	164 + 108	16	1.8	95.91
ResNet [1]	164	1.6	1.7	94.54
ResNet (my implementation)[1]	164	1.6	1.7	95.48

Compression Rate: 8

Compression Rate: 16

ResNet:

References

Identity Mappings in Deep Residual Networks [1]

Squeeze-and-Excitation Networks [2]

Improved Regularization of Convolutional Neural Networks with Cutout [3]