Google Professional Machine Learning Engineer Question 262
Select 3Google Cloud PlatformYou are working on training a deep learning model using TensorFlow on a large dataset that exceeds the memory and processing capabilities of a single VM. You decide to use distributed training on Google Cloud. Which steps are critical to ensure a reliable and scalable distributed training pipeline?
- A
Use Cloud Storage to store your dataset and ensure it is accessible to all training workers.
- B
Configure a Distributed Training Strategy, such as TensorFlow MirroredStrategy or MultiWorkerMirroredStrategy.
- C
Manually split the dataset into smaller chunks and distribute them to each worker node using custom scripts.
- D
Leverage AI Platform Training for managed distributed training and scaling.
- E
Use VM instances with local SSDs to store the dataset for faster access by each worker node.
Show answer and explanation
Correct answers: A, B, D
Explanation
To organize a reliable distributed training pipeline, it is crucial to use scalable and managed solutions that simplify the complexities of data access, worker synchronization, and resource management. Cloud Storage provides a centralized and scalable storage solution, while TensorFlow's distributed training strategies and AI Platform Training ensure seamless coordination and scalability. Manually splitting datasets or relying on local SSDs can introduce errors and are less reliable in a distributed setup.
- A. Correct.
Correct. Using Cloud Storage ensures that your dataset can be easily accessed by all worker nodes in a distributed training setup. This is crucial for a reliable pipeline.
- B. Correct.
Correct. Configuring a distributed training strategy such as TensorFlow's MirroredStrategy or MultiWorkerMirroredStrategy is essential for synchronizing training across multiple nodes.
- C. Incorrect.
Incorrect. While manually splitting the dataset is an option, it is error-prone and less reliable compared to using established frameworks or services that handle data distribution automatically.
- D. Correct.
Correct. AI Platform Training provides a managed environment for distributed training, automatically handling the setup, scaling, and management of worker nodes, which improves reliability.
- E. Incorrect.
Incorrect. While local SSDs may provide faster storage access, they are not suitable for distributed training as they are not shared across worker nodes, making them unreliable for this purpose.