SustainaML AutoML

A lightweight, energy‑aware AutoML toolkit. A Flask back‑end trains multiple ML frameworks while tracking CO₂/energy with , and a Streamlit front‑end lets you explore results, feature importance, and hyper‑parameter impact — all in your browser.

Table of Contents

1. Features

2. Project Structure

3. Quick Start

4. API Reference

5. Package Versions

6. Troubleshooting

7. Acknowledgement

1. Features

• AutoML search across frameworks — FLAML, “H2O”, MLJAR

• Time‑budgeted training with incremental / warm‑start loops

• Energy & CO₂ tracking via CodeCarbon (offline mode)

• Interactive Streamlit UI ▸ upload CSV ▸ pick frameworks & algorithms ▸ edit hyper‑parameters ▸ visualise leaderboards, scatter plots, feature importance, pipeline diagrams, time‑budget trade‑offs

• REST API (/run_automl) for programmatic access

2. Project Structure

├── backend.py # Flask API – training & carbon tracking

├── frontend.py # Main Streamlit interface (rich UI)

├── app.py # Minimal Streamlit demo (optional)

└── README.md # You are here

Tip : frontend.py and app.py do not share state; keep one running at a time.

3. Quick Start

1 · Clone & create env

git clone <your‑repo‑url> sustainaml-automl

cd sustainaml-automl

python -m venv venv

source venv/bin/activate # Windows: venv\Scripts\activate_ 2 · Install requirements

pip install -r requirements.txt

#OR – for the exact versions we tested

#pip install -r requirements_locked.txt

3 · Run the stack

In two terminals or two split panes 

#Terminal 1 – REST back‑end

python backend.py

#Terminal 2 – UI front‑end

streamlit run frontend.py # or streamlit run app.py

Visit the URL Streamlit prints (default http://localhost:8501). Upload a CSV where the last column is the target label.

4. API Reference

POST /run_automl

{

"frameworks": ["FLAML", "H2O", "MLJAR"],

"algorithms": {

  "FLAML": {"RF": true, "XGBoost": true},
  
  "H2O":   {"GLM": true}
  
  ..............................

},

"hyperparams": {

  "FLAML": {"RF": {"n_estimators": 150}}
  
  .................................

},

"time_budget":10, 30, 60, 120

"data": "<pandas.DataFrame>.to_json()" // last col = y

}

Returns per‑algorithm metrics, carbon/energy figures, hyper‑params & feature importances.

5. Package Versions

6. Troubleshooting

Symptom Fix

-ModuleNotFoundError: Re‑run pip install -r requirements.txt inside the activated venv

• Tracker raises No GPU found: Ignore – CodeCarbon falls back to CPU measurement

• Streamlit shows blank page: Refresh browser; check that backend.py is still running

7. Acknowledgements

CodeCarbon Library for carbon estimation & the open‑source community • 🌱