NVIDIA has launched a new version of its TAO Toolkit, designed to offer a low-code entry point for its train, adapt, and optimize (TAO) framework for machine and deep learning model creation — now boasting new pre-trained models for computer vision and speech applications.
"With TAO," says NVIDIA's Akhil Docca, "developers can use the power of transfer learning to create production-ready models customized and optimized for many use-cases. These include detecting defects, translating languages, or managing traffic—without the need for massive amounts of data. This version boosts developer productivity with new pre-trained vision and speech models. It also includes key new features such as ONNX model weights import, REST APIs, and TensorBoard integration."
The latter feature, which ties the TAO Toolkit into TensorFlow's TensorBoard visualization and experimentation platform, is aimed at making training performance easier to understand — as well as permitting direct comparison between different experiments through hyper-parameter modification.
NVIDIA's latest pre-trained models, meanwhile, get developers up and running as quickly as possible — but use the company's transfer learning capabilities to allow for fine-tuning for specific tasks. Among the new models are one designed for LIDAR sensor data in robotics and automotive applications, a pose classification system for public safety, retail, and worker safety applications, keypoint estimation for action or object shape definition on humans, animals, and objects, and a voice synthesis model capable of working with just 30 minutes of recorded data and targeting the smart device, gaming, and service sectors.
Other features of the new TAO Toolkit release include REST application programming interface (APIs), allowing for what NVIDIA calls "TAO Toolkit-as-a-Service" deployment or integration into an existing service with Kubernetes management and orchestration, and the ability to import pre-trained weights from ONNX while retaining features like pruning and quantization — though only for image classification and segmentation tasks for now.
The latest release is available to download now, free of charge, from the NVIDIA Developer site; for enterprise users, the new version will be rolled into NVIDIA AI Enterprise in the next quarterly update.