Vision for the Future

You know the feeling — when inspiration strikes at the least expected time. Your idea is, if not the greatest idea ever, at least in the top ten ideas in all of human history by your estimation. You need to make your idea a reality now — right now! But sometimes we have no idea how to go from inspiration to finished project because some of the aspects of it are just way outside of our areas of expertise.

For many of us, we hit a brick wall when our ideas involve machine learning, especially where computer vision is involved. Between the toolkits, models, frameworks, libraries, dependencies, hardware, and everything else that goes into these systems, it is not hard to make even an experienced engineer’s head start to swim. Whether that idea involves gesture recognition, object detection, or a navigational system for an autonomous vehicle, getting bogged down in the details can easily prevent the idea from taking flight.

One option that can serve both hackers and industry well in such cases is NVIDIA’s TAO. This open-source toolkit hides many of the complexities of building a complex vision AI application, allowing developers to focus instead on the system or device that they have in their minds. TAO accomplishes this by giving developers access to over 100 pre-trained vision AI models. They can then leverage the knowledge already contained in these models, and also tune them to suit their specific needs with their own data via transfer learning. This can all be done without writing a single line of code, and the retrained model can be deployed in popular formats.

NVIDIA recently updated TAO, and some significant enhancements were included in the latest version. As of the release of version 5.2, TAO has the ability to export trained models in the popular ONNX format. Using ONNX, models can be deployed virtually anywhere, from GPUs and CPUs to microcontrollers. Also available with this release is a new batch of state-of-the-art vision transformers that can tackle even very challenging computer vision problems. And don’t worry about the hard work of tuning the model’s hyperparameters to get everything running smoothly — integration with AutoML techniques will deal with those challenges for you.

If you are looking for an easy way to get started with a vision AI project, you might want to consider running through some tutorials on Google Colab.