Google Professional Machine Learning Engineer Question 249
Single answerGoogle Cloud PlatformYou are tasked with training a custom machine learning model on Google Cloud. The model requires the use of a specialized Python library that is not natively available on pre-configured environments. Additionally, you need to ensure reproducibility for the training process. Which approach is the most appropriate for this requirement?
- A
Use Vertex AI custom training with a custom container that includes the specialized Python library.
- B
Use Kubeflow Pipelines with pre-built TensorFlow components to handle the training process.
- C
Use Vertex AI AutoML to automate the training process without needing custom configurations.
- D
Deploy the model training script on a Compute Engine instance and manually install the specialized library.
Show answer and explanation
Correct answer: A
Explanation
Vertex AI custom training with a custom container is the best approach for this scenario because it provides a controlled and reproducible environment for training. By creating a Docker container that includes all necessary dependencies, you can ensure consistency across training runs and simplify deployment to Vertex AI. This approach is ideal when custom libraries or specific configurations are required.
- A. Correct.
This is the correct choice because Vertex AI custom training with a custom container allows you to package the specialized Python library in a Docker container, ensuring that the training environment is fully customized and reproducible.
- B. Incorrect.
This is incorrect because while Kubeflow Pipelines is a powerful orchestration tool, using pre-built TensorFlow components will not allow for the inclusion of a specialized library unless you create and use custom components, which adds unnecessary complexity.
- C. Incorrect.
This is incorrect because Vertex AI AutoML is designed for automated training and does not support custom configurations or the use of specialized libraries.
- D. Incorrect.
This is incorrect because manually installing the library on a Compute Engine instance does not ensure reproducibility and requires additional effort to manage dependencies and training scripts.