Google Professional Machine Learning Engineer Question 287
Select 2Google Cloud PlatformYou are designing an image recognition model for a production environment. The model training involves a large dataset with millions of high-resolution images, and the inference will be deployed on edge devices with limited computational resources. Which combination of compute and accelerator options is the most suitable for this scenario?
- A
Use GPUs for training and optimize the model for edge devices during deployment.
- B
Use TPUs for training and deploy the model on edge devices with TensorFlow Lite.
- C
Use CPUs for both training and deployment to maintain consistency.
- D
Use GPUs for training and deploy the model on edge devices with a full TensorFlow runtime.
- E
Use TPUs for training and deploy the model on edge devices using AutoML Edge.
Show answer and explanation
Correct answers: A, B
Explanation
For training, GPUs and TPUs are the most suitable options for handling large datasets and complex models due to their optimized parallel processing capabilities. For deployment on edge devices, TensorFlow Lite is a lightweight, efficient runtime that caters to the limited computing resources of such devices. Combining GPUs or TPUs for training with TensorFlow Lite for deployment ensures both efficiency and scalability.
- A. Correct.
Using GPUs for training is a common choice when dealing with large datasets and complex models, as GPUs are optimized for parallel processing. Optimizing the model for edge devices ensures it can run efficiently on hardware with limited resources.
- B. Correct.
TPUs are highly efficient for training large-scale machine learning models, particularly for TensorFlow-based workflows. TensorFlow Lite provides a lightweight runtime specifically designed for deployment on edge devices, making this a viable choice.
- C. Incorrect.
CPUs are not ideal for training large-scale models due to their relatively lower parallel processing capabilities. While they can be used for deployment, they are less efficient for training tasks in this scenario.
- D. Incorrect.
Deploying the model with a full TensorFlow runtime on edge devices is not resource-efficient, as edge devices typically have constrained computational and memory resources. TensorFlow Lite or other optimized runtimes are more appropriate.
- E. Incorrect.
While AutoML Edge simplifies the deployment process for edge devices, this approach assumes the use of AutoML for training. The question specifies designing and training the model directly, so this option is not relevant.