VIVelev / SENet-flow

Squeeze-and-Excitation Network - implementation in TensorFlow

SENet-flow

Squeeze-and-Excitation Network (SENet) implementation in TensorFlow

You can find the original paper here
written by Jie Hu, Li Shen, Gang Sun

If you want to see the original author's code, please refer to this link

Requirements

• Python 3.x
• Tensorflow 1.x

Idea

What is SENet?

Figure 1: Diagram of a Squeeze-and-Excitation building block

How do you integrate it in existing powerful architectures? (Inception Network, ResNet)

 

Figure 2: Schema of SE-Inception and SE-ResNet modules

Hyper-parameters

• Reduction Ratio - controls the bottleneck size (the number of units in the bottleneck dense layer)

 

Figure 3: Different choices for the Reduction Ratio hyperparameter and the consequent results

Why should you use Squeeze-and-Excitation Netwroks

State-of-the-art performace on ILSVRC 2017 (ImageNet 2017 dataset)

 

Figure 4: State-of-the-art performace on ILSVRC 2017