Google Professional Machine Learning Engineer Question 263
Select 3Google Cloud PlatformYou are building a machine learning model on Google Cloud that requires training a large dataset with millions of records. To ensure efficient and reliable training, you decide to use distributed training on Google Cloud AI Platform. Which steps should you take to organize a reliable distributed training pipeline?
- A
Use AI Platform Training with a custom training container that supports distributed training frameworks like TensorFlow or PyTorch.
- B
Set up a managed instance group to manually distribute data shards across virtual machines.
- C
Configure a hyperparameter tuning job to automatically scale the infrastructure during training.
- D
Employ Vertex AI Pipelines to orchestrate preprocessing, training, and evaluation tasks.
- E
Use a shared Cloud Storage bucket to store training data and model checkpoints.
Show answer and explanation
Correct answers: A, D, E
Explanation
To organize a reliable distributed training pipeline, you need tools that support distributed training (like AI Platform Training with custom containers), orchestration mechanisms (like Vertex AI Pipelines), and shared storage to synchronize data and checkpoints across workers. Managed instance groups and hyperparameter tuning are not specific or directly relevant to distributed training workflows, making them less suited for this scenario.
- A. Correct.
Correct. AI Platform Training supports distributed training using custom containers and frameworks like TensorFlow or PyTorch. This is a critical step for leveraging distributed training effectively.
- B. Incorrect.
Incorrect. While managed instance groups can distribute workloads, they are not specifically designed for ML training pipelines, and this approach would require significant manual effort and lacks reliability.
- C. Incorrect.
Incorrect. Hyperparameter tuning jobs are used to optimize model parameters but do not handle scaling infrastructure for distributed training directly.
- D. Correct.
Correct. Vertex AI Pipelines enable orchestration of ML workflows, which is essential for building reliable pipelines that include distributed training.
- E. Correct.
Correct. A shared Cloud Storage bucket ensures that all distributed training workers can access the same data and save checkpoints, which is vital for distributed training reliability.