CoolFace
Apppublic

EnYa32/steel-plate-defect-prediction

sourceHugging Facemitupdated 9mo agoView on Hugging Face
0likes
App README

🧱 Steel Plate Defect Predictor (XGBoost)

This Streamlit app predicts the probability of 7 steel plate defect types:

  • —Pastry
  • —Z_Scratch
  • —K_Scatch
  • —Stains
  • —Dirtiness
  • —Bumps
  • —Other_Faults

It uses 7 saved XGBoost models (one model per target) and outputs probabilities in the same spirit as the Kaggle submission format.

✅ Files you must have in the same folder as app.py

Models (7 files):

  • —xgb_Pastry.json
  • —xgb_Z_Scratch.json
  • —xgb_K_Scatch.json
  • —xgb_Stains.json
  • —xgb_Dirtiness.json
  • —xgb_Bumps.json
  • —xgb_Other_Faults.json

Metadata:

  • —feature_names.pkl (list of feature columns in the exact training order)
  • —target_names.pkl (optional but recommended; list of targets)
  • —feature_engineering.json (optional; documentation only)

🧠 Feature Engineering inside the app (same as training)

The app recomputes these engineered features automatically:

  • —area_perimeter_ratio = Pixels_Areas / (X_Perimeter + Y_Perimeter + 1)
  • —luminosity_range = Maximum_of_Luminosity - Minimum_of_Luminosity

So your input CSV should contain the raw columns required to compute them (e.g. PixelsAreas, XPerimeter, YPerimeter, MaximumofLuminosity, Minimumof_Luminosity).

▶️ Run locally

bash
pip install -r requirements.txt
streamlit run app.py
📄 How to use
Option 1: Upload CSV
Upload a CSV containing the raw feature columns (and optionally an id column).
The app will:

compute engineered features

select/reorder features using feature_names.pkl

output probabilities for all 7 targets

allow you to download predictions.csv

Option 2: Single row (manual)
Enter values for the raw features and get a probability table + bar chart.

🧪 Test input example (quick CSV idea)
Create a small CSV with the same raw feature columns as your dataset (like test.csv).
If your file includes id, the app will keep it in the output.

📌 Notes
Probabilities are between 0 and 1.

The model expects the exact same feature order as training (feature_names.pkl ensures this).

If you change feature engineering in the notebook, you must update the Streamlit app accordingly.