model.py

import pandas as pd
import plotly.express as px
import matplotlib.pyplot as plt
import plotly.io as pio
import sklearn
import warnings
from scipy.special import expit, logit
import sksurv.datasets
import numpy as np
import joblib
import xgboost as xgb
from xgboost import XGBRegressor
from xgboost import XGBClassifier
from xgboost import DMatrix
from xgboost import train
from lifelines import CoxPHFitter
from itertools import product
from tqdm import tqdm
from xgbse import XGBSEKaplanNeighbors
from xgbse.converters import convert_to_structured
from sklearn.metrics import roc_auc_score
from sklearn.preprocessing import StandardScaler
from sklearn.impute import SimpleImputer
from sklearn.pipeline import Pipeline
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report, confusion_matrix, accuracy_score
from sklearn.exceptions import UndefinedMetricWarning
from sklearn import set_config
from sklearn.model_selection import GridSearchCV, KFold
from sklearn.pipeline import make_pipeline
from sklearn.model_selection import ParameterGrid
from sksurv.datasets import load_breast_cancer
from sksurv.metrics import cumulative_dynamic_auc
from sksurv.metrics import concordance_index_censored
from sksurv.linear_model import CoxnetSurvivalAnalysis, CoxPHSurvivalAnalysis
from sksurv.preprocessing import OneHotEncoder
from sksurv.util import Surv
from dotenv import load_dotenv
import boto3
import mlflow
import os
import io

from sksurv.ensemble import GradientBoostingSurvivalAnalysis


load_dotenv(dotenv_path=".secrets")

mlflow.set_tracking_uri(os.getenv('BACKEND_STORE_URI=postgresql+psycopg2://neondb_owner:npg_GZ5FuPYjaf3b@ep-fancy-lab-adrdogpa-pooler.c-2.us-east-1.aws.neon.tech/neondb?sslmode=require&channel_binding=require'))
os.environ['AWS_ACCESS_KEY_ID'] = os.getenv('AWS_ACCESS_KEY_ID')
os.environ['AWS_SECRET_ACCESS_KEY'] = os.getenv('AWS_SECRET_ACCESS_KEY')
os.environ['MLFLOW_DEFAULT_ARTIFACT_ROOT'] = os.getenv('MLFLOW_DEFAULT_ARTIFACT_ROOT')
os.environ['S3_BUCKET'] = os.getenv('S3_BUCKET')

# Log configurations au démarrage
print("=== Configuration MLflow ===")
print(f"Tracking URI: {mlflow.get_tracking_uri()}")
print(f"Artifact Store: {os.getenv('MLFLOW_DEFAULT_ARTIFACT_ROOT')}")
print(f"AWS Access: {'Configuré' if os.getenv('AWS_ACCESS_KEY_ID') else 'Manquant'}")

s3 = boto3.client('s3')
try:
   response = s3.list_objects_v2(Bucket=os.getenv('S3_BUCKET'))
   print("S3 contents:", response.get('Contents', []))
except Exception as e:
   print("S3 error:", e)


warnings.filterwarnings("ignore", category=UndefinedMetricWarning)
set_config(display="text")

df=pd.read_csv('https://projet-feux-fullstack.s3.eu-west-3.amazonaws.com/datas/dataset_modele_decompte2.csv', sep=';', low_memory=False)
mask = df.Année == 2025
df = df[~mask]
df['Feu prévu'] = df['Feu prévu'].astype(bool)
df_clean = df.copy()

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_clean.columns]

# Nous mettons à 0 les NAN de la colonne décompte
df_clean["décompte"] = df_clean["décompte"].fillna(0)


# 🔹 Préparation des données réelles
df_clean = df_clean.rename(columns={"Feu prévu": "event", "décompte": "duration"})
y_structured = Surv.from_dataframe("event", "duration", df_clean)

X = df_clean[features]
y = y_structured

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)

event_train = y_train["event"]
duration_train = y_train["duration"]
event_test = y_test["event"]
duration_test = y_test["duration"]

# 🔹 Pipeline XGBoost survie avec StandardScaler
pipeline = 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",
        device="cuda",
        random_state=42
    ))
])



def train_evaluate_model_with_mlflow(model, X_train, X_test, y_train, y_test, model_name):
   print(f"\n=== Démarrage entraînement {model_name} ===")
   print(f"Tracking URI: {mlflow.get_tracking_uri()}")
   print(f"Registry URI: {mlflow.get_registry_uri()}")
   
   mlflow.set_experiment("fire_survival")
   print(f"Experiment: fire_survival")
   s3 = boto3.client('s3')

   with mlflow.start_run() as run:
        print(f"Run ID: {run.info.run_id}")
        
        print("Entraînement du modèle...")
        model.fit(X_train, duration_train, xgb__sample_weight=event_train)
        
        #save model to S3
        print("Enregistrement du modèle sur S3...")
        model_path = f"mlflow/models/{model_name}_{run.info.run_id}.joblib"

        # mlflow.sklearn.log_model(model, "model")

        buffer = io.BytesIO()
        joblib.dump(model, buffer)
        s3.put_object(
            Bucket=os.getenv('S3_BUCKET'),
            Key=model_path,
            Body=buffer.getvalue()
        )
        print("Modèle enregistré")
       
       # 🔹 Prédictions réelles (log(HR)) sur données test
        log_hr_test = model.predict(X_test)

        # 🔹 Jeu factice pour estimer le modèle de Cox
        df_fake = pd.DataFrame({
            "duration": duration_train,
            "event": event_train,
            "const": 1
        })
        dtrain_fake = DMatrix(df_fake[["const"]])
        dtrain_fake.set_float_info("label", df_fake["duration"])
        dtrain_fake.set_float_info("label_lower_bound", df_fake["duration"])
        dtrain_fake.set_float_info("label_upper_bound", df_fake["duration"])
        dtrain_fake.set_float_info("weight", df_fake["event"])

        params = {
            "objective": "survival:cox",
            "eval_metric": "cox-nloglik",
            "learning_rate": 0.1,
            "max_depth": 1,
            "verbosity": 0
        }
        bst_fake = train(params, dtrain_fake, num_boost_round=100)

        log_hr_fake = bst_fake.predict(dtrain_fake)
        df_risque = pd.DataFrame({
            "duration": duration_train,
            "event": event_train,
            "log_risque": log_hr_fake
        })
        # insertion de bruit pour aider le modèle à converger
        df_risque["log_risque"] += np.random.normal(0, 1e-4, size=len(df_risque))

        # 🔹 Modèle de Cox factice
        cph = CoxPHFitter()
        cph.fit(df_risque, duration_col="duration", event_col="event", show_progress=False)

        # 🔹 Évaluation avec le c-index
        c_index = concordance_index_censored(event_test, duration_test, log_hr_test)[0]
        print(f"\nC-index (test) : {c_index:.3f}")

        
        print("\nEnregistrement des métriques...")
        mlflow.log_metric("c_index", c_index)


        # mlflow.register_model(
        #     f"runs:/{run.info.run_id}/model",
        #     "fire_survival"
        # )
        # Exemple : une ligne de ton jeu de données
        input_example = X_train.iloc[:1]
        mlflow.sklearn.log_model(
            sk_model=model,
            artifact_path="model",
            input_example=input_example
        )

        # 🔹 Enregistrer dans le Registry
        result = mlflow.register_model(
            model_uri=f"runs:/{run.info.run_id}/model",
            name="fire_survival"
        )
        return model, run.info.run_id

if __name__ == "__main__":
    xgb_final = pipeline
    _, run_id = train_evaluate_model_with_mlflow(
        xgb_final, X_train, X_test, y_train, y_test, "xgboost_survivalCOX_model"
    )
    print(f"Run ID: {run_id}")