Google Professional Machine Learning Engineer Question 232
Single answerGoogle Cloud PlatformYou are designing a machine learning pipeline to train a model on a large dataset of labeled images stored in Google Cloud Storage. The dataset is organized into folders, where each folder represents a class label. You need to optimize the data organization for efficient training and scalability while ensuring compatibility with Google Cloud AI tools. What is the recommended way to organize and manage the training data in this scenario?
- A
Store the images in a single folder and use a CSV file to map each image's filename to its class label.
- B
Organize the images into separate folders for each class and name each folder after the class label.
- C
Store all images in a Cloud SQL database and query the database during training to fetch image data.
- D
Use Cloud Datastore to store image metadata and retrieve the images dynamically during training.
Show answer and explanation
Correct answer: B
Explanation
When working with large datasets of labeled images, organizing the data into separate folders by class label is a best practice. This structure is natively supported by TensorFlow and other Google Cloud AI tools, enabling efficient data loading and processing during training. Other approaches, such as using databases to store images or dynamically retrieving metadata, introduce unnecessary complexity and are less efficient for training large-scale models.
- A. Incorrect.
This approach is valid and commonly used when working with large image datasets in Google Cloud AI tools. However, while it allows for flexibility in mapping labels, it may be less efficient and harder to scale compared to organizing by folders.
- B. Correct.
Organizing images into folders by class label is a best practice for image classification tasks, as it simplifies data management and is natively supported by TensorFlow and other Google Cloud AI tools.
- C. Incorrect.
Storing images in a Cloud SQL database is not an efficient practice for machine learning tasks, as it introduces significant latency and complexity during training. Databases are not optimized for serving large binary files like images in this context.
- D. Incorrect.
Using Cloud Datastore for image metadata and dynamic retrieval is not a recommended approach for training ML models, as it adds unnecessary overhead and complexity. Direct access to organized image data is more efficient.