Skip to content

README.md

Latest commit

 

History

History
100 lines (75 loc) · 2.93 KB

README.md

File metadata and controls

100 lines (75 loc) · 2.93 KB

Elastic Weight Consolidation (EWC)

Overview

  • Focuses on protecting prior knowledge while sequentially learning Split CIFAR-100 tasks.
  • Provides both continual learning training and an explicit unlearning workflow.
  • Results from earlier experiments are archived under results/.

Layout

  • src/ewc/ – package with training entrypoints (main.py, ewc_unlearning.py) plus helpers.
  • results/ – text summaries collected from previous experiment runs.
  • requirements.txt – minimal dependency set for the method.

Quickstart

  1. cd repositories/Regularization-Continual-Learning-PyTorch
  2. python -m venv .venv && source .venv/bin/activate
  3. pip install -r requirements.txt
  4. export PYTHONPATH=src
  5. Run continual training: python -m ewc.main

Unlearning demo

export PYTHONPATH=src
python -m ewc.ewc_unlearning

Elastic Weight Consolidation (EWC)

Overview

This repository provides a reference implementation of Elastic Weight Consolidation (EWC) for continual learning experiments on Split CIFAR-100. It includes training scripts, an explicit unlearning workflow, helper utilities for dataset splits, and example experiment outputs in results/.

Repository layout

  • src/ewc/ – Python package with training and unlearning entry points (main.py, ewc_unlearning.py) and supporting modules (utils.py, strategies.py).
  • results/ – experiment output summaries and logs.
  • requirements.txt – Python dependencies used for development and experiments.

Quick start

  1. Create and activate a virtual environment:
python -m venv .venv
source .venv/bin/activate
  1. Install dependencies:
pip install -r requirements.txt
  1. Make the package importable and run the training example:
export PYTHONPATH=src
python -m ewc.main

Unlearning demo

export PYTHONPATH=src
python -m ewc.ewc_unlearning

Data

By default the code downloads CIFAR-100 into ./data. If you prefer a different location, update the root argument in src/ewc/utils.py or set up a data directory and point the scripts to it.

Contributing and development

  • Tests: A minimal smoke test is included under tests/ that verifies the package import. Run tests with pytest (install pytest in your environment).
  • Packaging: Basic pyproject.toml and setup.cfg are provided for local installation using pip install -e ..
  • Style and linting: The project does not enforce a style guide currently; adding pre-commit and flake8/ruff is recommended for collaborators.

Notes

  • The implementation is intended as a research reference and is not production hardened. Review training loops, device placement, and checkpointing before using it for large-scale experiments.
  • If you want, I can add CI (GitHub Actions) for tests and linting, and a more complete example notebook demonstrating an experiment run.