Introduction

Tensorflow (Keras) implementation of ICLR 2018 paper Learn To Pay Attention.
Including two attention method (dot product and parametrise) and visualization of attention map.
CIFAR-10:

Due to the limited computational power, the VGG-type CNN net is condensed as follow:

C64*3*3-C128*3*3-C256*3*3-P2*2-(Att1)-C512*3*3-P2*2-(Att2)-C512*3*3-P2*2-(Att3)-C512*3*3-P2*2-C512*3*3-P2*2-L512-L10

Require

tensorflow == 2.0.0

Run

1. run vgg.py for pretraining the VGG net.
2. run vgg_att.py for training the proposed net and visualizing the attention map.