Monotone operator equilibrium networks
Repository to hold the code of final project of Course "Optimization Methods for Machine Learning". The code is based on the original forked repository.
See the tutorial notebook for an introduction to the original paper and the code in this repo.
Code to replicate the experiments can be found in experiments.ipynb.
Experiment Results
Training time (per epoch) of different operator splitting methods on Multi-tier network without data augmentation.
| Forward-Backward | Peaceman | Douglas | ProxDecomp | |
|---|---|---|---|---|
| Time/s | 202.97 | 87.33 | 129.57 | 140.82 |
Model accuracy:
| Single conv | Multi-tier | Single conv lg. | Multi-tier lg. | |
|---|---|---|---|---|
| ReLU | 74.10 | 72.14 | 81.40 | 88.62 |
| LeakyReLU | 74.13 | 72.70 | 81.60 | 88.87 |
Requirements
Compatible with python 3.5+ and known to work with pytorch 1.4, torchvision 0.5, and numpy 1.18. Can install with pip install -r requirements.txt.