Delta tables and YOLO computer vision tasks

[Andrewcon]

Hi all,

I would really appreciate if someone could help me out. I feel it’s both a data engineering and ML question.

One thing we use at wo is YOLO for object detection. I’ve managed to run YOLO by loading data from the blob storage, but I’ve seen that the best way to do deep learning tasks in Databricks is to train your ML models on Delta Live Tables.

I currently have my training dataset as a Delta table, and I was wondering if anyone has managed to train computer vision models on Delta tables.

I’ve read the documentations and have seen repos such as petastorm that try to implement training on delta tables, but I can’t for the life of me understand how to actually run yolo this way, especially since YOLO uses yaml for config.

Thank in advance for your help! 😇

[Kaniz]

Hi @Andrewcon, Training computer vision models on Delta Live Tables in Databricks is an interesting challenge. Let’s break it down:

1. Delta Live Tables:

   • Delta Live Tables is a declarative framework for building reliable, maintainable, and testable data processing pipelines. It manages task orchestration, cluster management, monitoring, data quality, and error handling¹.
   • You define transformations on your data, and Delta Live Tables takes care of the rest.

2. Training Computer Vision Models:

   • While YOLO (You Only Look Once) is a powerful object detection algorithm, integrating it with Delta Live Tables requires some additional steps.
   • Here’s a high-level approach to train YOLO using your Delta table:

3. Steps:

   • Data Preparation:

     • Ensure your training dataset in the Delta table is well-structured. You might need to convert it into the format expected by YOLO (e.g., bounding box annotations, image paths, class labels).
     • Consider using tools like petastorm to read data from Delta tables efficiently².

   • Model Configuration:

     • YOLO uses a yaml configuration file to define model architecture, hyperparameters, and other settings.
     • Adapt your YOLO configuration to work with Delta Live Tables. You might need to modify the data input pipeline to read from Delta tables instead of blob storage.

   • Training:

     • Set up a Databricks cluster with appropriate resources for training.
     • Write custom code (likely in Python) that:
       • Reads data from the Delta table.
       • Parses the yaml configuration.
       • Initializes the YOLO model.
       • Trains the model using your data.

   • Evaluation and Deployment:

     • After training, evaluate the model’s performance using validation data.
     • Once satisfied, deploy the model for inference.

4. Challenges:

   • YOLO’s yaml configuration: Adapting it to work with Delta tables might involve custom code to load data dynamically.
   • Ensuring efficient data access: Delta Live Tables provides optimizations, but you’ll need to handle data loading efficiently.

5. Resources:

   • Explore the Databricks documentation on Delta Live Tables and the Python language reference for deta...³.

Remember that this integration might require some experimentation and custom development. Good luck, and feel free to ask if you need further assistance! 😊