Code to show the simulation results in [ARSM: Augment-REINFORCE-Swap-Merge Estimator for Gradient Backpropagation Through Categorical Variables]
The MNIST data is self-contained and the Omniglot data is in the repository.
RL data is from OpenAI Gym
Below are the paper to cite if you find the algorithms in this repository useful in your own research:
@inproceedings{ARSM_ICML2019,
title={{ARSM}: {A}ugment-{REINFORCE}-swap-merge estimator for gradient backpropagation through categorical variables},
author={Mingzhang Yin and Yuguang Yue and Mingyuan Zhou},
booktitle={ICML},
year={2019}
}
This code is offered under the MIT License.