Mise à jour du 25/09/2024

src/app/prediction_service/Dockerfile

@@ -1 +1,12 @@
+FROM python:3.9-slim
 
+WORKDIR /app
+
+COPY requirements.txt .
+
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY . .
+
+EXPOSE 8001
+CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8001"]

src/app/prediction_service/main.py

@@ -0,0 +1,96 @@
+from fastapi import FastAPI, HTTPException
+from pydantic import BaseModel
+import os
+import mlflow
+import mlflow.sklearn
+from mlflow.tracking import MlflowClient
+
+
+# Définir l'application FastAPI
+app = FastAPI()
+
+# Charger le modèle
+mlflow.set_tracking_uri("http://mlflow_service:5000") 
+client = MlflowClient()
+
+def load_production_model_and_evaluate(model_name="model_rf_clf"):
+    # Charger toutes les  versions du modele et chercher le tag "is_production"
+    try:
+        versions = client.search_model_versions(f"name='{model_name}'")
+    except mlflow.exceptions.RestException as e:
+        print(f"Model '{model_name}' not found in the registry. No production model to evaluate.")
+        return None, None
+
+
+    prod_model_info = None
+    for version in versions:
+        if version.tags.get("is_production") == "true":
+            prod_model_info = version
+            break
+
+    if not prod_model_info:
+        print(f"No model tagged as 'is_production=true' exists for '{model_name}'.")
+        return None, None
+
+    prod_model_version = prod_model_info.version
+    print(f"Loading model version {prod_model_version} (tagged as production).")
+
+    model_uri = f"models:/{model_name}/{prod_model_version}"
+
+    prod_model = mlflow.pyfunc.load_model(model_uri)
+    return prod_model
+
+
+class DonneesAccident(BaseModel):
+    place: int
+    catu: int
+    trajet: float
+    an_nais: int
+    catv: int
+    choc: float
+    manv: float
+    mois: int
+    jour: int
+    lum: int
+    agg: int
+    int: int
+    col: float
+    com: int
+    dep: int
+    hr: int
+    mn: int
+    catr: int
+    circ: float
+    nbv: int
+    prof: float
+    plan: float
+    lartpc: int
+    larrout: int
+    situ: float
+
+@app.post("/predict")
+def predict(accident: DonneesAccident):
+    """
+    Endpoint pour prédire la gravité de l'accident.
+
+    Args:
+    - accident : Les données de l'accident défini selon le BaseModel
+
+    Returns:
+    - dict: La prédiction de la gravité de l'accident.
+    """
+    model=load_production_model_and_evaluate(model_name="model_rf_clf")
+    try:
+        features = [
+            accident.place, accident.catu, accident.trajet,
+            accident.an_nais, accident.catv, accident.choc, accident.manv,
+            accident.mois, accident.jour, accident.lum, accident.agg,
+            accident.int, accident.col, accident.com, accident.dep,
+            accident.hr, accident.mn, accident.catr, accident.circ,
+            accident.nbv, accident.prof, accident.plan, accident.lartpc,
+            accident.larrout, accident.situ
+    ]
+        prediction = model.predict([features])
+        return {"Cet accident est de niveau de gravité": int(prediction[0])}
+    except Exception as e:
+        raise HTTPException(status_code=400, detail=str(e))

src/app/prediction_service/requirements.txt

@@ -0,0 +1,34 @@
+# local package
+#-e .
+
+# external requirements
+click
+Sphinx
+coverage
+#awscli>=1.29.0
+flake8
+pathlib 
+joblib
+
+# pour le modele
+pandas
+numpy
+imblearn
+scikit-learn==1.3.2
+imbalanced-learn
+mlflow
+
+# pour l api
+uvicorn
+fastapi 
+pydantic 
+#logging
+passlib
+evidently
+
+#pour test
+pytest
+httpx
+
+#pour la base de données
+psycopg2-binary