This python code allows you to generate results from NeurIPS 2020 Spotlight Paper: DisARM: An Antithetic Gradient Estimator for Binary Latent Variables, by Zhe Dong, Andriy Mnih, George Tucker.

Paper link: https://arxiv.org/abs/2006.10680 NeurIPS link: https://papers.nips.cc/paper/2020/hash/d880e783834172e5ebd1868d84463d93-Abstract.html

Please find the required packages in requirements.txt.

The python binary for experiments on VAE with a single stochastic layer is in experiment_launcher_singlelayer.py, with multiple stochastic layers is in experiment_launcher_multilayer.py. There are 5 supported datasets are supported: static_mnist, dynamic_mnist, fashion_mnist, and omniglot.

The following grad_types are supported: ARM, REINFORCE LOO, DisARM, RELAX and etc.

To launch experiments, call the binary with:

python -m disarm.experiment_launcher_singlelayer.py \
  --dataset=dynamic_mnist \
  --logdir=/tmp/disarm/dynamic_mnist \
  --grad_type=disarm \
  --encoder_type=nonlinear \
  --num_steps=1000000 \
  --demean_input \
  --initialize_with_bias

For multi-layer case, replace experiment_launcher_singlelayer.py with experiment_launcher_multilayer.py. The encoder and decoder structure of the VAE model, e.g. number of layers etc., can be specified in experiment_launcher_multilayer.py.