Google Professional Machine Learning Engineer Question 259
Select 3Google Cloud PlatformYou are developing a machine learning model that requires distributed training due to the size of the dataset and complexity of the model. You want to ensure that the pipeline is reliable and scalable while leveraging Google Cloud services. Which actions should you take to organize a reliable distributed training pipeline?
- A
Use AI Platform (Vertex AI) Training with custom containers to enable distributed training.
- B
Configure a distributed training strategy such as TensorFlow's MultiWorkerMirroredStrategy or ParameterServerStrategy.
- C
Disable checkpoints during training to improve performance and reduce storage costs.
- D
Use Cloud Storage to store training data and model checkpoints to ensure scalability and durability.
- E
Manually configure virtual machines for training instead of using managed services to have more control over resources.
Show answer and explanation
Correct answers: A, B, D
Explanation
To organize a reliable distributed training pipeline, you should leverage managed services like Vertex AI Training for scalability and ease of use. Configuring an appropriate distributed training strategy ensures synchronization across workers, while storing data and checkpoints in Cloud Storage enhances durability and fault tolerance. Avoid disabling checkpoints, as they are critical for recovery, and prefer managed services over manual infrastructure management for reliability and scalability.
- A. Correct.
Using AI Platform (Vertex AI) Training with custom containers allows you to run distributed training jobs efficiently on Google Cloud while benefiting from managed infrastructure and scalability.
- B. Correct.
Configuring a distributed training strategy like MultiWorkerMirroredStrategy or ParameterServerStrategy is critical for ensuring the training process is reliable and properly synchronized across multiple workers.
- C. Incorrect.
Disabling checkpoints is not recommended since checkpoints are essential for ensuring the reliability of the pipeline by saving intermediate model states in case of failures.
- D. Correct.
Using Cloud Storage to store training data and model checkpoints ensures durability and scalability, making it a reliable choice for distributed training pipelines.
- E. Incorrect.
Manually configuring virtual machines for training is unnecessary when managed services like Vertex AI are available, as they simplify resource management and improve reliability.