grossemodif

app.py

@@ -130,10 +130,18 @@ def load_raw_data() -> pd.DataFrame:
 # ────────────────────────────────────────────────
 # 2) FONCTION D’ENTRAÎNEMENT + PRÉDICTIONS
 # ────────────────────────────────────────────────
-@st.cache_resource(show_spinner="⚙️ Entraînement du modèle…", ttl=None)
+
+
+import mlflow.sklearn
+
+# ────────────────────────────────────────────────
+# 2) FONCTION DE PREDICTIONS (sans entraînement)
+# ────────────────────────────────────────────────
+@st.cache_resource(show_spinner="⚙️ Chargement du modèle…", ttl=None)
 def train_model_and_predict(df_raw: pd.DataFrame) -> pd.DataFrame:
     """Retourne df_map prêt pour la carte avec les colonnes
        proba_7j, proba_30j, …, proba_180j."""
+
     # a) Nettoyage
     df = df_raw.copy()
     df = df.rename(columns={"Feu prévu": "event", "décompte": "duration"})
@@ -153,35 +161,14 @@ def train_model_and_predict(df_raw: pd.DataFrame) -> pd.DataFrame:
     ]
     features = [f for f in features if f in df.columns]
 
-    # c) split + Surv
-    y_struct = Surv.from_dataframe("event", "duration", df)
-    X_train, X_test, y_train, y_test = train_test_split(
-        df[features], y_struct, test_size=0.3, random_state=42
-    )
-    ev_train, du_train = y_train["event"], y_train["duration"]
-    ev_test, du_test = y_test["event"], y_test["duration"]
-
-    # d) Pipeline XGBSurv
-    pipe = Pipeline([
-        ("imputer", SimpleImputer(strategy="median")),
-        ("scaler", StandardScaler()),
-        ("xgb", XGBRegressor(
-            objective="survival:cox",
-            n_estimators=100,
-            learning_rate=0.05,
-            max_depth=3,
-            tree_method="hist",
-            random_state=42,
-        )),
-    ])
-    pipe.fit(X_train, du_train, xgb__sample_weight=ev_train)
+    # c) Chargement du modèle depuis MLflow
+    model_uri = "models:/fire_survival/Production"  # ou /1 pour la v1
+    pipe = mlflow.sklearn.load_model(model_uri)
 
-    # e) Affiche C-index dans la sidebar
-    log_hr_test = pipe.predict(X_test)
-    c_index = concordance_index_censored(ev_test, du_test, log_hr_test)[0]
-    st.sidebar.write(f"**C-index (test)** : {c_index:.3f}")
+    # d) Estimation baseline hazard avec Cox factice (comme avant)
+    y_struct = Surv.from_dataframe("event", "duration", df)
+    ev_train, du_train = y_struct["event"], y_struct["duration"]
 
-    # f) Estimation du baseline hazard (Cox factice)
     df_fake = pd.DataFrame({
         "duration": du_train,
         "event": ev_train,
@@ -192,6 +179,7 @@ def train_model_and_predict(df_raw: pd.DataFrame) -> pd.DataFrame:
     dmat.set_float_info("label_lower_bound", df_fake["duration"])
     dmat.set_float_info("label_upper_bound", df_fake["duration"])
     dmat.set_float_info("weight", df_fake["event"])
+
     bst_fake = xgb_train(
         params={
             "objective": "survival:cox",
@@ -212,7 +200,6 @@ def train_model_and_predict(df_raw: pd.DataFrame) -> pd.DataFrame:
     })
     cph = CoxPHFitter()
     cph.fit(df_risque, duration_col="duration", event_col="event", show_progress=False)
-
     baseline_cumhaz = cph.baseline_cumulative_hazard_
 
     def S0(t: int) -> float:
@@ -224,6 +211,7 @@ def train_model_and_predict(df_raw: pd.DataFrame) -> pd.DataFrame:
             H0 = baseline_cumhaz.loc[idx[idx <= t]].iloc[-1, 0]
         return float(np.exp(-H0))
 
+    # e) Calcul proba sur plusieurs horizons
     horizons = {7: "proba_7j", 30: "proba_30j", 60: "proba_60j",
                 90: "proba_90j", 180: "proba_180j"}
 
@@ -236,6 +224,118 @@ def train_model_and_predict(df_raw: pd.DataFrame) -> pd.DataFrame:
     df_map = df[["latitude", "longitude", "ville"] + list(horizons.values())].copy()
     return df_map
 
+
+
+
+
+
+
+# @st.cache_resource(show_spinner="⚙️ Entraînement du modèle…", ttl=None)
+# def train_model_and_predict(df_raw: pd.DataFrame) -> pd.DataFrame:
+#     """Retourne df_map prêt pour la carte avec les colonnes
+#        proba_7j, proba_30j, …, proba_180j."""
+#     # a) Nettoyage
+#     df = df_raw.copy()
+#     df = df.rename(columns={"Feu prévu": "event", "décompte": "duration"})
+#     df["event"] = df["event"].astype(bool)
+#     df["duration"] = df["duration"].fillna(0)
+
+#     # b) Features
+#     features = [
+#         "moyenne precipitations mois", "moyenne temperature mois",
+#         "moyenne evapotranspiration mois", "moyenne vitesse vent année",
+#         "moyenne vitesse vent mois", "moyenne temperature année",
+#         "RR", "UM", "ETPMON", "TN", "TX", "Nombre de feu par an",
+#         "Nombre de feu par mois", "jours_sans_pluie", "jours_TX_sup_30",
+#         "ETPGRILLE_7j", "compteur jours vers prochain feu",
+#         "compteur feu log", "Année", "Mois",
+#         "moyenne precipitations année", "moyenne evapotranspiration année",
+#     ]
+#     features = [f for f in features if f in df.columns]
+
+#     # c) split + Surv
+#     y_struct = Surv.from_dataframe("event", "duration", df)
+#     X_train, X_test, y_train, y_test = train_test_split(
+#         df[features], y_struct, test_size=0.3, random_state=42
+#     )
+#     ev_train, du_train = y_train["event"], y_train["duration"]
+#     ev_test, du_test = y_test["event"], y_test["duration"]
+
+#     # d) Pipeline XGBSurv
+#     pipe = Pipeline([
+#         ("imputer", SimpleImputer(strategy="median")),
+#         ("scaler", StandardScaler()),
+#         ("xgb", XGBRegressor(
+#             objective="survival:cox",
+#             n_estimators=100,
+#             learning_rate=0.05,
+#             max_depth=3,
+#             tree_method="hist",
+#             random_state=42,
+#         )),
+#     ])
+#     pipe.fit(X_train, du_train, xgb__sample_weight=ev_train)
+
+#     # e) Affiche C-index dans la sidebar
+#     log_hr_test = pipe.predict(X_test)
+#     c_index = concordance_index_censored(ev_test, du_test, log_hr_test)[0]
+#     st.sidebar.write(f"**C-index (test)** : {c_index:.3f}")
+
+#     # f) Estimation du baseline hazard (Cox factice)
+#     df_fake = pd.DataFrame({
+#         "duration": du_train,
+#         "event": ev_train,
+#         "const": 1,
+#     })
+#     dmat = DMatrix(df_fake[["const"]])
+#     dmat.set_float_info("label", df_fake["duration"])
+#     dmat.set_float_info("label_lower_bound", df_fake["duration"])
+#     dmat.set_float_info("label_upper_bound", df_fake["duration"])
+#     dmat.set_float_info("weight", df_fake["event"])
+#     bst_fake = xgb_train(
+#         params={
+#             "objective": "survival:cox",
+#             "eval_metric": "cox-nloglik",
+#             "learning_rate": 0.1,
+#             "max_depth": 1,
+#             "verbosity": 0,
+#         },
+#         dtrain=dmat,
+#         num_boost_round=100,
+#     )
+#     log_hr_fake = bst_fake.predict(dmat)
+
+#     df_risque = pd.DataFrame({
+#         "duration": du_train,
+#         "event": ev_train,
+#         "log_risque": log_hr_fake + np.random.normal(0, 1e-4, size=len(log_hr_fake)),
+#     })
+#     cph = CoxPHFitter()
+#     cph.fit(df_risque, duration_col="duration", event_col="event", show_progress=False)
+
+#     baseline_cumhaz = cph.baseline_cumulative_hazard_
+
+#     def S0(t: int) -> float:
+#         """Survie de base S0(t) = exp(-H0(t))."""
+#         idx = baseline_cumhaz.index
+#         if t in idx:
+#             H0 = baseline_cumhaz.loc[t].values[0]
+#         else:
+#             H0 = baseline_cumhaz.loc[idx[idx <= t]].iloc[-1, 0]
+#         return float(np.exp(-H0))
+
+#     horizons = {7: "proba_7j", 30: "proba_30j", 60: "proba_60j",
+#                 90: "proba_90j", 180: "proba_180j"}
+
+#     log_hr_all = pipe.predict(df[features])
+#     HR = np.exp(log_hr_all)
+
+#     for t, col in horizons.items():
+#         df[col] = 1 - (S0(t) ** HR)   # P(event ≤ t)
+
+#     df_map = df[["latitude", "longitude", "ville"] + list(horizons.values())].copy()
+#     return df_map
+
 # ────────────────────────────────────────────────
 # 3) AFFICHAGE SUR LA PAGE « Accueil »
 # ────────────────────────────────────────────────