pyperch 0.1.5

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

PyTorch-Only (No Skorch Required)

• rhc_adam_hybrid.ipynb
  New notebook demonstrating PyTorch training using pyperch's RHC optimizer and Adam together on different layers - no Skorch dependency required. Ideal for hybrid workflows and experimentation.

---

Classic Skorch Examples

• backprop_network.ipynb
  Standard neural network training using backpropagation with Skorch.

• rhc_opt_network.ipynb
  Neural network trained using Randomized Hill Climbing (RHC) via Skorch.

• sa_opt_network.ipynb
  Neural network trained using Simulated Annealing (SA) via Skorch.

• ga_opt_network.ipynb
  Neural network trained using a Genetic Algorithm (GA) via Skorch.

• regression_examples.ipynb
  Regression tasks using randomized optimization.

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