Google Professional Machine Learning Engineer Question 305
Single answerGoogle Cloud PlatformYou are deploying a machine learning model built with PyTorch to production on Google Cloud. The model is designed for real-time inference. Which of the following approaches should you take to efficiently serve the model while ensuring scalability and low latency?
- A
Use TensorFlow Serving to deploy the PyTorch model directly without conversion.
- B
Deploy the PyTorch model on Vertex AI and use a custom container to handle inference.
- C
Convert the PyTorch model to TensorFlow format and use TensorFlow Serving for deployment.
- D
Deploy the PyTorch model using AI Platform Prediction with built-in PyTorch serving support.
Show answer and explanation
Correct answer: B
Explanation
Vertex AI with a custom container is the recommended approach for deploying a PyTorch model on Google Cloud for real-time inference. It provides flexibility to package the model and dependencies while benefiting from autoscaling, low latency, and seamless integration with Google Cloud services. TensorFlow Serving is specific to TensorFlow models, and converting the PyTorch model to TensorFlow format introduces unnecessary complexity. AI Platform Prediction does not natively support PyTorch models, making a custom container essential for PyTorch deployment.
- A. Incorrect.
TensorFlow Serving is designed for TensorFlow models, not PyTorch models. Using TensorFlow Serving without converting the model is not possible.
- B. Correct.
Using Vertex AI with a custom container allows you to package your PyTorch model with the necessary runtime dependencies and scaling capabilities, making it an efficient solution for real-time inference.
- C. Incorrect.
While converting the PyTorch model to TensorFlow format and using TensorFlow Serving might work, this approach introduces an unnecessary conversion step, which can lead to inaccuracies and is not the most efficient option.
- D. Incorrect.
AI Platform Prediction does not have built-in support for serving PyTorch models directly. Instead, custom containers are required for such use cases.