Service de modèles

Model serving is the process of hosting a trained machine learning model and making its functionality available to software applications via a network interface. It acts as the bridge between a static model file saved on a disk and a live system that processes real-world data. Once a model has completed the machine learning (ML) training phase, it must be integrated into a production environment where it can receive inputs—such as images, text, or tabular data—and return predictions. This is typically achieved by wrapping the model in an Application Programming Interface (API), allowing it to communicate with web servers, mobile apps, or IoT devices.

The Role of Model Serving in AI

The primary goal of model serving is to operationalize predictive modeling capabilities effectively. While training focuses on accuracy and loss minimization, serving focuses on performance metrics like latency (how fast a prediction is returned) and throughput (how many requests can be handled per second). Robust serving infrastructure ensures that computer vision (CV) systems remain reliable under heavy loads. It often involves technologies like containerization using tools such as Docker, which packages the model with its dependencies to ensure consistent behavior across different computing environments.

Applications concrètes

Model serving alimente des fonctions d'IA omniprésentes dans diverses industries en permettant une prise de décision immédiate basée sur des données. données.

• Fabrication intelligente : dans les environnements industriels, l'IA dans les systèmes de fabrication utilise des modèles pour inspecter les chaînes de montage. Des images haute résolution des composants sont envoyées à un serveur local, où un modèle YOLO26 détecte les défauts tels que les rayures ou les désalignements, déclenchant des alertes immédiates pour retirer les articles défectueux.
• Automatisation du commerce de détail : les détaillants utilisent l'IA dans le commerce de détail pour améliorer l'expérience client. Des caméras équipées de modèles de détection d'objets identifient les produits dans une zone de caisse et calculent automatiquement le coût total sans qu'il soit nécessaire de scanner manuellement les codes-barres .

Mise en œuvre pratique

To serve a model effectively, it is often beneficial to export models to a standardized format like ONNX, which promotes interoperability between different training frameworks and serving engines. The following example demonstrates how to load a model and run inference, simulating the logic that would exist inside a serving endpoint using Python.

from ultralytics import YOLO

# Load the YOLO26 model (this typically happens once when the server starts)
model = YOLO("yolo26n.pt")

# Simulate an incoming API request with an image source URL
image_source = "https://ultralytics.com/images/bus.jpg"

# Run inference to generate predictions for the user
results = model.predict(source=image_source)

# Process results (e.g., simulating a JSON response to a client)
print(f"Detected {len(results[0].boxes)} objects in the image.")

Choisir la bonne stratégie

The choice of serving strategy depends heavily on the specific use case. Online Serving provides immediate responses via protocols like REST or gRPC, which is essential for user-facing web applications. Conversely, Batch Serving processes large volumes of data offline, suitable for tasks like nightly report generation. For applications requiring privacy or low latency without internet dependence, Edge AI moves the serving process directly to the device, utilizing optimized formats like TensorRT to maximize performance on constrained hardware. Many organizations leverage the Ultralytics Platform to simplify the deployment of these models to various endpoints, including cloud APIs and edge devices.

Distinction par rapport aux termes apparentés

While closely related, "Model Serving" is distinct from Model Deployment and Inference.

• Model Deployment: This refers to the broader lifecycle stage of releasing a model into a production environment. Serving is the specific mechanism or software (like NVIDIA Triton Inference Server or TorchServe) used to execute the deployed model.
• Inference: This is the mathematical act of calculating a prediction from an input. Model serving provides the infrastructure (networking, scalability, and security) that allows inference to happen reliably for end-users.
• Microservices: Serving is often architected as a set of microservices, where the model runs as an independent service that other parts of an application can query, often exchanging data in lightweight formats like JSON.