Google Professional Machine Learning Engineer Question 230
Select 3Google Cloud PlatformYou are building a machine learning pipeline to classify images of animals using Google Cloud. The dataset consists of thousands of high-resolution images stored in a local file system. Your goal is to organize the training data efficiently for model training while ensuring scalability and compatibility with Google Cloud ML services. Which of the following steps should you take?
- A
Upload the images to a Google Cloud Storage bucket organized by class labels in folder names.
- B
Convert the images to a tabular format and store them in BigQuery for faster access.
- C
Use the Google Cloud Storage bucket path to create a TFRecord file for efficient training data input.
- D
Store the images directly in a Compute Engine instance and access them from there during training.
- E
Ensure that the dataset is sharded into smaller chunks for parallel processing during training.
Show answer and explanation
Correct answers: A, C, E
Explanation
To organize training data for image classification on Google Cloud, you should upload the images to Cloud Storage, as it provides scalability and integration with ML services. Using TFRecord files helps optimize data loading during training. Sharding the dataset ensures efficient parallel processing, which is crucial for handling large-scale datasets. These steps together align with best practices for organizing and optimizing training data in Google Cloud.
- A. Correct.
Uploading images to a Google Cloud Storage bucket organized by class labels is a best practice as it ensures scalability and compatibility with Google Cloud ML services. It also enables easy integration with ML pipelines.
- B. Incorrect.
Converting images to a tabular format and storing them in BigQuery is not suitable for image data. BigQuery is optimized for structured, tabular data rather than high-resolution image files.
- C. Correct.
Creating a TFRecord file from the images is recommended as it provides an efficient way to store and load data for TensorFlow-based training. This is particularly useful for large datasets.
- D. Incorrect.
Storing the images directly in a Compute Engine instance is not scalable or cost-effective for large datasets and does not take advantage of Google Cloud's managed storage solutions.
- E. Correct.
Sharding the dataset into smaller chunks is important for parallel processing and distributed training, which improves training efficiency for large datasets.