The latest version of Nvidia’s TAO toolkit boosts developer productivity with the addition of AutoML capability, integration with third-party MLOPs (machine learning operations) services and new pre-trained vision AI models. The enterprise version now includes access to the full source code and model weights for pretrained models.
The toolkit is designed for efficient model training for vision and conversational AI. By simplifying complex AI models and deep learning frameworks, Nvidia said even developers without AI expertise can use the toolkit to produce AI models. Users can optimise model inference throughput with AI expertise or large training datasets using transfer learning to fine-tune Nvidia pre-trained models with developer’s own data. Models can be trained without the hassle of manually fine-tuning hundreds of parameters.
Other enhancements are access to virtual models from leading cloud providers and Kubernetes services like Amazon EKS or Azure AKS. It is also possible to simplify infrastructure management and scaling on cloud machine learning services such as Google Colab, Google Vertex AI, and Microsoft Azure Machine Learning.
Nvidia also announced new cloud integrations and third-party MLOps services, such as W&B and ClearML, to provide developers and enterprises with an optimised AI workflow. It is possible to build a new AI service or integrate into an existing one with REST APIs.
Developers can also create custom production-ready models optimised for specific environments and scenarios with TAO. Transformer-based pre-trained models (CitySemSegformer, Peoplenet Transformer) and retail-specific pre-trained models (RetailObjectDetection, RetailObjectRecognition, and ReIdentificationNet) can be accessed.
Nvidia highlighted a new feature which helps developers build object detection models without massive amounts of data. The use cases include detecting assembly line defects, translating particular phrases across languages, or managing city traffic.
