intro_dfc

Introspective Deep Feature Consistent Variational Autoencoder

My attempt to implement a Deep Feature Consistent Variational Autoencoder but in the introspective style of this paper. I call it, The Introspective Deep Feature Consistent Variational Autoencoder, or if you like word salads, Autoencoding Variational Bayes Using Self-Supervised High Level Latent Features, or if you don't, Introspective DFC VAE.

This project makes use of the new TensorFlow 2.0 beta using a custom training loop. Man oh man things are easier now!

Model defined in model.py, data input done in data.py, training functions defined in train_ops.py, and the jupyter notebook is for testing things locally with a scaled down model.

todo:

• Clean the code more or less done
• learn how to take advantage of multiple GPUs eh, that was overrated anyways
• Train my normal VAE Done!
• Actually implement it Done! It doesn't work very well yet. But it works.

resources used:

• https://www.tensorflow.org/beta/tutorials/generative/cvae
• https://www.tensorflow.org/tutorials/load_data/images
• https://nbviewer.jupyter.org/github/krasserm/bayesian-machine-learning/blob/master/variational_autoencoder.ipynb
• In depth step by step on the math of VAEs: https://arxiv.org/abs/1606.05908