Training AutoML Models: A Multi-Modal Guide

Training AutoML Models

AutoML on Vertex AI provides a streamlined process for training high-quality models on a variety of data types. The general workflow is the same for all data types, but the specifics of data preparation and model evaluation differ.

1. Tabular Data

• Use Cases: Classification and regression on structured data.
• Data Preparation:
  • Provide data in a BigQuery table or a CSV file in Cloud Storage.
  • Ensure your data is clean and that the target column is clearly identified.
  • AutoML will automatically handle feature engineering, but you can also apply manual transformations.
• Training:
  • Select your target column and the features to use for training.
  • Choose a budget (in node hours) for training. AutoML will automatically search for the best model architecture within that budget.

2. Text Data

• Use Cases: Text classification, entity extraction, and sentiment analysis.
• Data Preparation:
  • Provide data in a CSV file or a JSONL file in Cloud Storage.
  • For classification, each document should have a corresponding label.
  • For entity extraction, you need to provide annotations that identify the entities in your text.
• Training:
  • AutoML will automatically fine-tune a pre-trained language model on your data.

3. Image Data

• Use Cases: Image classification and object detection.
• Data Preparation:
  • Provide images in a Cloud Storage bucket.
  • For classification, each image should be in a folder named after its label.
  • For object detection, you need to provide bounding box annotations for the objects in your images.
• Training:
  • AutoML will use transfer learning to adapt a pre-trained vision model to your data.

4. Video Data

• Use Cases: Video classification and object tracking.
• Data Preparation:
  • Provide videos in a Cloud Storage bucket.
  • Similar to image data, you need to provide labels for classification or bounding box annotations for object tracking.
• Training:
  • AutoML will analyze the video frames and train a model to recognize patterns over time.

Knowledge Check