Google Professional Machine Learning Engineer Question 354
Select 3Google Cloud PlatformYou are deploying a deep learning model on Google Cloud that processes real-time video streams. The model is deployed on Vertex AI and uses a custom-trained TensorFlow model. During the initial testing phase, you observe high latency and insufficient throughput, which affects the application's performance. What steps should you take to optimize the model for increased performance and reduced latency?
- A
Use TensorFlow Lite to convert the model for optimized inference.
- B
Enable GPU acceleration for model serving in Vertex AI.
- C
Increase the batch size for inference requests to process more data at once.
- D
Use a smaller, pre-trained model from TensorFlow Hub with fewer parameters.
- E
Switch from Vertex AI to a Compute Engine VM for more control over the serving environment.
Show answer and explanation
Correct answers: A, B, D
Explanation
To optimize the model's performance for real-time video processing, you should focus on reducing latency and improving throughput. TensorFlow Lite provides optimized inference capabilities, GPU acceleration can leverage hardware for faster computations, and selecting a smaller model reduces computational overhead. Increasing batch size is not suitable for real-time scenarios, and switching to Compute Engine does not inherently optimize the model's performance.
- A. Correct.
TensorFlow Lite is designed for optimized inference and can help reduce latency and improve performance, especially for real-time applications.
- B. Correct.
Enabling GPU acceleration for model serving can significantly reduce latency and improve throughput by leveraging the parallel processing capabilities of GPUs.
- C. Incorrect.
While increasing batch size can improve throughput, it may increase latency for real-time applications, making it unsuitable for this scenario.
- D. Correct.
Using a smaller, pre-trained model with fewer parameters can reduce computational requirements, decreasing latency and improving throughput.
- E. Incorrect.
Switching to a Compute Engine VM may provide more control over the environment but does not directly optimize the model's performance or address latency issues.