MLA-C01 Question 292
Select 3You are building a machine learning workflow to classify images of damaged and undamaged cars for an insurance company. The dataset contains millions of labeled images. To optimize training time and cost, you decide to use Amazon SageMaker. Which of the following steps should you take to manage the data and training process effectively?
- A
Use Amazon S3 for storing the image dataset and enable S3 Transfer Acceleration to optimize data transfer speeds.
- B
Use Amazon SageMaker's built-in image classification algorithm to pre-process the images and automatically split the dataset into training, validation, and test sets.
- C
Use Amazon SageMaker Pipe mode to stream data directly from Amazon S3 to the training instances, reducing the need to download the entire dataset.
- D
Use Spot Instances in SageMaker training jobs to reduce cost while specifying checkpoints to save intermediate model states.
- E
Manually download the dataset from Amazon S3 to a local EC2 instance, pre-process the images locally, and then upload the processed dataset back to S3 for training.
Show answer and explanation
Correct answers: A, C, D
Explanation
To efficiently manage a large-scale machine learning workflow on Amazon SageMaker, you should leverage AWS services and features designed for scalability, cost efficiency, and performance. Storing the dataset in Amazon S3 with Transfer Acceleration ensures fast and reliable access. Using SageMaker Pipe mode allows you to stream data directly to training instances, avoiding the need to download the entire dataset. Spot Instances, coupled with checkpointing, further optimize costs without risking data loss. Avoid manual processes that increase complexity and are not aligned with AWS best practices.
- A. Correct.
Correct: Using Amazon S3 for storing the dataset ensures scalability and reliability, and enabling S3 Transfer Acceleration can improve transfer speeds when uploading or accessing large datasets.
- B. Incorrect.
Incorrect: While SageMaker offers built-in algorithms, pre-processing and dataset splitting are tasks that typically require explicit configuration or external tools. The built-in algorithm does not automatically handle these tasks.
- C. Correct.
Correct: SageMaker Pipe mode streams data directly from S3 to the training instances, which reduces the overhead of downloading large datasets and improves efficiency for large-scale training.
- D. Correct.
Correct: Using Spot Instances can significantly reduce training costs. Specifying checkpoints ensures that progress is saved, mitigating the risk of losing work when a Spot Instance is terminated.
- E. Incorrect.
Incorrect: Manually downloading and pre-processing the dataset on a local EC2 instance introduces unnecessary complexity and is not aligned with best practices for SageMaker workflows, which are designed to streamline data management and training.