CNN AutoEncoder on MNIST

A simple autoencoder to recover MNIST data using convolutional and de-convolutional layers.

Usage

Train an AutoEncoder, generate recoverd images, and do t-sne on embeddings.

    python main.py

Recovered Image

The dimension of embedding is 10.

Fig.1 and Fig3 in each row are real images, Fig.2 and Fig.4 in each row are recovered images.

Epoch 0	Epoch 5	Epoch 9

t-sne visualization of embeddings

Use t-sne to reduce embeddings' dimension 10 down to 2, so as to scatter in a coordinate system.