Genetic algorithms for neural-network pruning

This project studies how the mask representation changes genetic-algorithm pruning of an MLP (784 → 256 → 256 → 256 → 10) trained on FashionMNIST. The final structured approach evolves one bit per hidden neuron and performs much better at high sparsity.

Present the final report

Requirements are Python 3.13 and uv. The report reads the three saved JSON files under artifacts/results/; it never reruns training or pruning.

uv sync --frozen
uv run marimo run notebooks/report.py

The live app contains the optional sparsity/method explorer. Create a static, non-reactive backup after copying in the final artifacts:

uv run marimo export html notebooks/report.py -o report.html --no-include-code -f

Run the experiments

uv run python main.py train
uv run python main.py structured
uv run python main.py unstructured
uv run python main.py ablation

These commands produce structured.json, unstructured.json, and ablation.json. Run the offline tests with uv run pytest -q. For the cluster workflow, see docs/ORFEO.md; for the complete methodology and fairness rules, see docs/PROJECT_GUIDE.md.

Repository map

Path	Purpose
main.py	training and the three final experiment commands
src/gamo/ga/	unstructured and structured genetic searches
src/gamo/run/	canonical experiment protocol and result generation
notebooks/report.py	artifact-only Marimo presentation
scripts/run_orfeo.sh	three-job final Slurm submission