A student-teacher variational autoencoder that utilizes a normal parameterization, coupled with the gumbel reparameterization in order to enforce consistency across different intervals.
An example usage of the Vanilla VAE:
python run_mnist_experiment.py --base_dir="." --device="/gpu:0" --sequential=False --device_percentage=0.9 --latent_size=14 --use_bn=True --epochs=100
An example usage of the Online VAE:
python run_mnist_experiment.py --base_dir="." --device="/gpu:0" --sequential=True --device_percentage=0.9 --latent_size=14 --use_bn=True --min_interval=12000 --max_dist_swaps=32
--sequential=True
utilizes the online VAE mentioned in the paper as opposed to the vanilla batch method