Code for the paper "The Flip Side of the Reweighted Coin: Duality of Adaptive Dropout and Regularization," arXiv:2106.07769.
To generate the plots for the variational dropout case study, first run experiments/Sparse MNIST.ipynb. To generate both the computed effective penalties and the case study comarison plots, run experiments/Figures.ipynb.
The primary dependencies are pytorch 1.8.1 with CUDA 11.1 and skorch 0.10.1 (https://skorch.readthedocs.io/en/stable/), but we provide a complete description of the conda environment in environment.yml for completeness.