BVAE-tf

Disentangled Variational Auto-Encoder in TensorFlow (Beta-VAE)

Example Reconstructed Image

alt text

darknet 19 (fully convolutional & fast) encoder and decoder
Custom keras sampling layer for sampling the distribution of variational autoencoders
Custom loss in sampling layer for latent space regularization
- Options are no reg, vae reg (kl divergence), or bvae reg (beta*kl-divergence)
- You can also set a target capacity for dimension usage of the latent space
Simple interface for setting up your own VAE or B-VAE
- See the test function in ae.py for usage information

I am using conda to ensure the enviroment is easy to install

Install Anaconda or Miniconda (the python 3 version) for your platform
Recreate the conda environment from the yml: conda env create -f environment.yml
Active the enviroment
1. Windows: go to cmd and activate bvae-tf
2. Linux: source activate bvae-tf
If you want to use CPU only, run pip uninstall tensorflow-gpu followed by pip install tensorflow==1.4.0 after you activate the environment.

If you do not want to / cannot use conda, I am using tensorflow 1.4.0; see the environment.yml for more package info.

For a simple overfitting demonstration, run ae.py in your terminal. This will cause the autoencoder to run on the included demo image.

Note: The demo takes a few minutes on my 1060 6GB, so it will take a while on a CPU...