This is a simple autoencoder model for dataset mnist.
It can get data feature by unsupervised learning, and then reduce the feature dimension by tsne algorithm.
Meanwhile, global average pooling can not help get the middle feature, which should be replaced by dense layer.