This is the official code repository for NeurIPS 2021 paper: CARMS: Categorical-Antithetic-REINFORCE Multi-Sample Gradient Estimator by Alek Dimitriev and Mingyuan Zhou.

To install the required packages run: pip install -r requirements.txt To reproduce the toy example run: python3 toy.py. Supported gradients: CARMS, LOORF, UNORD, ARSM. Supported datasets: Dynamic MNIST, Fashion MNIST, and Omniglot, with either a linear or nonlinear encoder/decoder pair. To run an experiment you can use the following template:

python3 -m main \
    --dataset=dynamic_mnist \
    --logdir=../logs \
    --ckptdir=../ckpts \
    --grad_type=carms \
    --num_samples=4 \
    --num_latent=10 \
    --num_categories=5 \
    --encoder_type=nonlinear \
    --repeat_idx=42 \
    --num_steps=1e6 \
    --demean_input \
    --initialize_with_bias \