jlm429 / pyperch

Randomized opt networks with PyTorch

pyperch

Getting Started

About

Pyperch is a neural network weight optimization package developed to support students taking Georgia Tech’s graduate machine learning course CS7641. Three random optimization algorithms—randomized hill climbing, simulated annealing, and a genetic algorithm—can be used as drop-in replacements for traditional gradient-based optimizers using PyTorch.

Install

pip install pyperch

Examples

For starter code and examples, see

Backprop Neural Net

RHC Optimized Neural Net

SA Optimized Neural Net

GA Optimized Neural Net

Regression Examples

Contributing

Pull requests are welcome.

• Fork pyperch.
• Create a branch (git checkout -b branch_name)
• Commit changes (git commit -m "Comments")
• Push to branch (git push origin branch_name)
• Open a pull request

pyperch's Issues

dropout bug

dropouts not being applied to all NN layers

re-add sources to example/starter code notebooks

When splitting up notebooks into separate opt networks, sources were accidentally cut. Re-add citataions for sklearn and skorch starter code in notebooks starter code/examples.

https://github.com/jlm429/pyperch/blob/b1b6ed435ef57c9e2b8623009ceaaa280a1ca38a/notebooks/rhc_opt_network.ipynb