Google Professional Machine Learning Engineer Question 261
Select 3Google Cloud PlatformYou are building a machine learning pipeline on Google Cloud to train a large-scale image classification model. The training data is stored in Cloud Storage, and the model requires distributed training across multiple GPUs. You want to ensure the training pipeline is reliable, scalable, and minimizes potential bottlenecks. Which combination of practices should you implement?
- A
Use TensorFlow's tf.data API to create an efficient input pipeline that reads data in parallel and prefetches batches.
- B
Manually partition the dataset into multiple chunks and assign each chunk to a specific worker node.
- C
Leverage AI Platform Training with distributed training strategies, such as MirroredStrategy or MultiWorkerMirroredStrategy.
- D
Use Cloud Pub/Sub to orchestrate communication between worker nodes during training.
- E
Ensure model checkpoints are saved to Cloud Storage at regular intervals to recover from potential failures.
Show answer and explanation
Correct answers: A, C, E
Explanation
To ensure a reliable and scalable distributed training pipeline on Google Cloud, it is essential to use efficient input pipelines (e.g., tf.data API), leverage managed distributed training tools like AI Platform Training with appropriate strategies, and implement mechanisms like checkpointing to recover from potential failures. Manually partitioning datasets or using tools like Cloud Pub/Sub for worker communication is error-prone and not optimized for distributed machine learning workflows.
- A. Correct.
Using TensorFlow's tf.data API helps create an efficient and scalable input pipeline by reading data in parallel, performing transformations, and prefetching data to avoid bottlenecks during training.
- B. Incorrect.
Manually partitioning the dataset can lead to errors, inefficiencies, and difficulty scaling to larger datasets or additional worker nodes. It is better to use automated tools like tf.data.
- C. Correct.
AI Platform Training supports distributed training strategies like MirroredStrategy or MultiWorkerMirroredStrategy, enabling reliable and scalable distributed training across multiple GPUs or nodes.
- D. Incorrect.
Cloud Pub/Sub is designed for messaging and event-driven architectures, not for managing communication between distributed training workers. TensorFlow handles this communication internally.
- E. Correct.
Saving checkpoints to Cloud Storage ensures that the training process can recover from failures without losing progress, which is critical for reliability in distributed training pipelines.