Image classification - Deep Learning for Default Detection
What you’ll learn
Being able to analyze factory default in real time is a critical task to increase production line quality and reducing defaults. Implementing such a use case with deep learning for computer vision can be challenging at scale, especially when it comes to data preprocessing and building production-grade pipelines. Databricks simplifies this process end to end, making all the operational tasks simple so that you can focus on improving the model performance. In this demo, we will cover how to implement a complete deep learning pipeline to detect printed circuit board (PCB) defaults, from the image ingestion to real-time inferences (over REST API):
- Simplify data and image ingestions using Databricks Auto Loader and Delta Lake
- Learn how to do image preprocessing at scale
- Train and deploy a computer vision pipeline with Hugging Face and the new Spark DataFrame data set for transformers
- Deploy the pipeline for batch or streaming inferences and real-time serving with Databricks Serverless model endpoints
- Understand which pixels are flagged as damaged PCBs to highlight potential default
- A complete training and inference example using PyTorch Lightning if the Hugging Face library isn’t enough for your requirements, including deltatorch and distributed training with TorchDistributor
To install the demo, get a free Databricks workspace and execute the following two commands in a Python notebook
Dbdemos is a Python library that installs complete Databricks demos in your workspaces. Dbemos will load and start notebooks, Delta Live Tables pipelines, clusters, Databricks SQL dashboards, warehouse models … See how to use dbdemos
Dbdemos is distributed as a GitHub project.
For more details, please view the GitHub README.md file and follow the documentation.
Dbdemos is provided as is. See the License and Notice for more information.
Databricks does not offer official support for dbdemos and the associated assets.
For any issue, please open a ticket and the demo team will have a look on a best-effort basis.
