Software Feature Detection using Source Code Minimaps as Visual Signatures

This repository trains deep learning classifiers on grayscale 128×128 minimap images, using GPU when available.
It automatically runs three classification tasks in one execution, using the label columns from the provided CSV:

• type (≈63 classes)
• project (≈99 classes)
• author (≈97 classes)

The pipeline is clarity-first, modular, and reproducible (fixed seed). It logs progress,
saves confusion matrices, ROC curves (micro/macro AUC), classification reports, checkpoints, and XAI (Integrated Gradients).

Authors

Dataset & CSV

Place your images under data/minimaps_root/ (recursively) and the CSV at data/catalogo.csv (or set paths via CLI/YAML).
The CSV must include:

• image → key column that matches image basenames (without extension).
• One of the label columns: type, project, or author.
• Other descriptive columns such as file_path and file_name can be used for traceability.

Quickstart (PyCharm + venv)

python3 -m venv .venv
source .venv/bin/activate    # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# Train all three tasks (type, project, author):
python train_minimaps.py --config configs/resnet18.yaml

Outputs

A new run folder is created under outputs/run-YYYYmmdd-HHMMSS/. For each task, you will find:

• metrics/metrics.json, metrics.csv, classification_report.txt
• confusion_matrix.png
• roc_curves_micro_macro.png (micro/macro ROC curves), plus per-class AUC in metrics
• checkpoints/best.pt, last.pt
• xai/integrated_gradients/ (attributions on a few validation samples per class)

Configuration

Edit configs/default.yaml or override via CLI (--root-dir, --csv-path, --csv-label-columns etc.).

How to cite (Zenodo)

After creating a Zenodo record for this dataset/code, replace the DOI placeholder below.

BibTeX:

@dataset{minimaps-zenodo, title = {Software Feature Detection using Source Code Minimaps as Visual Signatures},

year = {2025}, publisher = {Zenodo}, version = {1.1}, doi = {10.5281/zenodo.16929672}, url = {https://zenodo.org/records/16894040](https://zenodo.org/records/16929672} }

License

This code is provided for research purposes. Check dataset/image licenses before redistribution.