Repo update

Author: Dreipfelt

?

@@ -0,0 +1,52 @@
+# ─────────────────────────────────────────────────────────
+# OASIS MLflow — Hugging Face Space
+# MLflow Tracking Server
+# Backend  : NeonDB (PostgreSQL)
+# Artifacts: HuggingFace Dataset (hf://datasets/Dreipfelt/oasis-mlflow-artifacts)
+# ─────────────────────────────────────────────────────────
+FROM continuumio/miniconda3
+
+LABEL maintainer="Dreipfelt" \
+      project="OASIS Security" \
+      component="MLflow Tracking Server" \
+      version="2.0"
+
+# System dependencies
+RUN apt-get update && apt-get install -y \
+    curl \
+    unzip \
+    nano \
+    && rm -rf /var/lib/apt/lists/*
+
+# HuggingFace Spaces: non-root user required
+RUN useradd -m -u 1000 user
+USER user
+
+ENV HOME=/home/user \
+    PATH=/home/user/.local/bin:$PATH
+
+WORKDIR $HOME/app
+
+# Copy files with correct ownership
+COPY --chown=user . $HOME/app
+
+# Install Python dependencies
+RUN pip install --no-cache-dir --upgrade pip \
+    && pip install --no-cache-dir -r requirements.txt
+
+# Authenticate to HuggingFace Hub using the HF_TOKEN secret
+# This allows MLflow to push artefacts to the HF Dataset
+RUN --mount=type=secret,id=HF_TOKEN,mode=0444 \
+    pip install --no-cache-dir huggingface_hub && \
+    echo "HF auth configured"
+
+# Launch MLflow tracking server
+# All env vars are injected as HF Space Secrets — never hardcoded
+CMD huggingface-cli login --token $HF_TOKEN --add-to-git-credential && \
+    mlflow server \
+    --host 0.0.0.0 \
+    --port $PORT \
+    --backend-store-uri $BACKEND_STORE_URI \
+    --default-artifact-root $ARTIFACT_STORE_URI \
+    --serve-artifacts \
+    --workers 2

MlFlow/Dockerfile

@@ -0,0 +1,28 @@
+---
+title: Oasis MLflow
+emoji: 📊
+colorFrom: blue
+colorTo: indigo
+sdk: docker
+pinned: false
+---
+
+# OASIS MLflow — Experiment Tracking Server
+
+MLflow tracking server for the OASIS Security project.
+
+| Component | Technology |
+|---|---|
+| Metadata backend | NeonDB (PostgreSQL) |
+| Artefact store | HuggingFace Dataset |
+| UI | MLflow 2.12 |
+
+## Tracked experiments
+
+Gradient Boosting · XGBoost · Random Forest · Ridge · LightGBM · Prophet · Holt-Winters
+
+## Champion model
+
+**Gradient Boosting — R² = 0.979 · RMSE = 48.84 · MAE = 29.95**
+
+*CDSD Certification Project — RNCP35288*

MlFlow/README.md

@@ -0,0 +1,4 @@
+mlflow==2.12.1
+psycopg2-binary==2.9.9
+huggingface_hub==0.23.0
+hf-transfer==0.1.6

MlFlow/requirements.txt

@@ -0,0 +1,192 @@
+# models/crime_predictor/src/train.py
+# Training pipeline with MLflow tracking
+# Artefacts pushed to hf://datasets/Dreipfelt/oasis-mlflow-artifacts
+
+import os
+import json
+import joblib
+import numpy as np
+import pandas as pd
+import mlflow
+import mlflow.sklearn
+from sklearn.ensemble import GradientBoostingRegressor, RandomForestRegressor
+from sklearn.linear_model import Ridge
+from sklearn.model_selection import TimeSeriesSplit, cross_val_score
+from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
+from pathlib import Path
+
+# ── MLflow remote tracking server (HF Space) ──────────────
+MLFLOW_TRACKING_URI = os.getenv(
+    "MLFLOW_TRACKING_URI",
+    "https://dreipfelt-oasis-mlflow.hf.space"
+)
+mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
+mlflow.set_experiment("oasis-security-crime-prediction")
+
+# ── Data source ────────────────────────────────────────────
+DATA_URL = os.getenv(
+    "DATA_URL",
+    "https://static.data.gouv.fr/resources/"
+    "bases-statistiques-communale-departementale-et-regionale-de-la-delinquance-"
+    "enregistree-par-la-police-et-la-gendarmerie-nationales/"
+    "20260129-160256/donnee-reg-data.gouv-2025-geographie2025-produit-le2026-01-22.csv"
+)
+
+# ── Models to benchmark ────────────────────────────────────
+MODELS = {
+    "GradientBoosting": GradientBoostingRegressor(
+        n_estimators=300, max_depth=4, learning_rate=0.08,
+        random_state=42
+    ),
+    "RandomForest": RandomForestRegressor(
+        n_estimators=200, max_depth=6, random_state=42, n_jobs=-1
+    ),
+    "Ridge": Ridge(alpha=1.0),
+}
+
+
+def load_data(url: str) -> pd.DataFrame:
+    """Load and clean data from data.gouv.fr"""
+    print(f"📥 Loading data from {url[:60]}...")
+    df = pd.read_csv(url, sep=";", encoding="utf-8", low_memory=False)
+    df = df[df["unite_de_compte"] == "nombre"].copy()
+    df["taux_100k"] = df["nombre"] / df["insee_pop"] * 100_000
+    return df.dropna(subset=["taux_100k"])
+
+
+def engineer_features(df: pd.DataFrame) -> pd.DataFrame:
+    """Production-grade feature engineering"""
+    df = df.copy()
+
+    # Cyclic temporal encoding
+    df["year_sin"] = np.sin(2 * np.pi * df["annee"] / 10)
+    df["year_cos"] = np.cos(2 * np.pi * df["annee"] / 10)
+    df["year_trend"] = (
+        (df["annee"] - df["annee"].min())
+        / (df["annee"].max() - df["annee"].min())
+    )
+
+    # Lag features (per indicator × region group)
+    grp = df.groupby(["indicateur", "Code_region"])["taux_100k"]
+    df["lag1"] = grp.shift(1).fillna(grp.transform("mean"))
+    df["lag2"] = grp.shift(2).fillna(grp.transform("mean"))
+    df["roll_mean_3"] = (
+        grp.rolling(3, min_periods=1).mean().reset_index(0, drop=True)
+    )
+
+    # Regional aggregate
+    df["region_mean"] = df.groupby("Code_region")["taux_100k"].transform("mean")
+
+    # Categorical encoding
+    df["ind_code"] = pd.Categorical(df["indicateur"]).codes
+    df["reg_code"] = pd.Categorical(df["Code_region"]).codes
+
+    feature_cols = [
+        "year_sin", "year_cos", "year_trend",
+        "lag1", "lag2", "roll_mean_3",
+        "region_mean", "ind_code", "reg_code",
+    ]
+    return df[feature_cols + ["taux_100k"]].dropna()
+
+
+def evaluate(model, X_test, y_test) -> dict:
+    """Compute test set metrics"""
+    preds = model.predict(X_test)
+    return {
+        "r2_test":   round(r2_score(y_test, preds), 4),
+        "rmse_test": round(np.sqrt(mean_squared_error(y_test, preds)), 4),
+        "mae_test":  round(mean_absolute_error(y_test, preds), 4),
+    }
+
+
+def train_and_log(model_name: str, model, X_train, X_test, y_train, y_test):
+    """Train one model and log everything to MLflow"""
+    print(f"\n🔧 Training {model_name}...")
+
+    with mlflow.start_run(run_name=model_name):
+
+        # Cross-validation (TimeSeriesSplit)
+        tscv = TimeSeriesSplit(n_splits=3)
+        cv_scores = cross_val_score(
+            model, X_train, y_train, cv=tscv, scoring="r2", n_jobs=-1
+        )
+
+        # Final fit on full train set
+        model.fit(X_train, y_train)
+        r2_train = model.score(X_train, y_train)
+
+        # Test metrics
+        metrics = evaluate(model, X_test, y_test)
+        metrics["r2_train"]    = round(r2_train, 4)
+        metrics["cv_r2_mean"]  = round(cv_scores.mean(), 4)
+        metrics["cv_r2_std"]   = round(cv_scores.std(), 4)
+
+        # Log to MLflow
+        mlflow.log_param("model", model_name)
+        mlflow.log_metrics(metrics)
+
+        # Log model artefact → pushed to HF Dataset
+        mlflow.sklearn.log_model(
+            sk_model=model,
+            artifact_path="model",
+            registered_model_name=f"crime_predictor_{model_name.lower()}",
+        )
+
+        print(f"   R² test={metrics['r2_test']} · "
+              f"RMSE={metrics['rmse_test']} · "
+              f"CV R²={metrics['cv_r2_mean']}±{metrics['cv_r2_std']}")
+
+    return metrics
+
+
+def main():
+    # ── Load & prepare data ────────────────────────────────
+    df = load_data(DATA_URL)
+    df_features = engineer_features(df)
+
+    FEATURE_COLS = [
+        "year_sin", "year_cos", "year_trend",
+        "lag1", "lag2", "roll_mean_3",
+        "region_mean", "ind_code", "reg_code",
+    ]
+    X = df_features[FEATURE_COLS]
+    y = df_features["taux_100k"]
+
+    # Temporal train/test split — last 20% as test
+    split = int(len(X) * 0.8)
+    X_train, X_test = X.iloc[:split], X.iloc[split:]
+    y_train, y_test = y.iloc[:split], y.iloc[split:]
+
+    print(f"📊 Train: {len(X_train)} rows · Test: {len(X_test)} rows")
+
+    # ── Benchmark all models ───────────────────────────────
+    all_metrics = {}
+    for name, model in MODELS.items():
+        all_metrics[name] = train_and_log(
+            name, model, X_train, X_test, y_train, y_test
+        )
+
+    # ── Select champion ────────────────────────────────────
+    best_name = max(all_metrics, key=lambda k: all_metrics[k]["r2_test"])
+    best_metrics = all_metrics[best_name]
+    print(f"\n🏆 Champion: {best_name} (R²={best_metrics['r2_test']})")
+
+    # ── Save champion locally ──────────────────────────────
+    champion = MODELS[best_name]
+    champion.fit(X, y)  # Retrain on full dataset
+
+    artifacts_dir = Path(__file__).parent.parent / "artifacts"
+    artifacts_dir.mkdir(exist_ok=True)
+
+    joblib.dump(champion, artifacts_dir / "crime_predictor.pkl")
+
+    metrics_out = {"best_model": best_name, **best_metrics, "all_models": all_metrics}
+    with open(artifacts_dir / "metrics.json", "w") as f:
+        json.dump(metrics_out, f, indent=2)
+
+    print(f"✅ Artefacts saved to {artifacts_dir}")
+    print(f"📊 metrics.json: R²={best_metrics['r2_test']}")
+
+
+if __name__ == "__main__":
+    main()