Autoencoder

Image Autoencoder trained on FFHQ image dataset

Trained from scratch using Pytorch

Check my Variational Autoencoder repo for a better version of this.

Data

The images used to train the model come from the Flickr-Faces-HQ dataset

The model

Simple Autoencoder 128x128x3 -> latent_space -> 128x128x3, with latent_space = 200

Both Encoder and Decoder contain 3 convolutional layers (kernel_size = 4, stride = 2), 3 maxpool layers (kernel_size = 2, stride = 2) and ReLU activation

It was trained overnight on a laptop GPU, more training should improve the results significantly.

Results

At epoch = 200, ground truth images :

Reconstructions :

Code

Models are created, loaded, saved using a model number. Saving is done automatically every save_every epoch, when training ends or when program in interrupted

except KeyboardInterruption:
  print("Training stopped.")
  save(...)