NVIDIA Launches Transfer Learning Toolkit 3.0, Offers New Free-Use Models
NVIDIA's Transfer Learning Toolkit 3.0 launches today, bringing a wealth of pre-trained models — including a heart rate estimation model.
NVIDIA has announced the launch of the third iteration of its Transfer Learning Toolkit, its zero-code approach to machine learning — and with it a dramatic increase in the number of models it's making available for use and reuse, free of charge.
NVIDIA's Transfer Learning Toolkit is designed to provide developers with a means to add a variety of machine learning models to their systems without the need to be a machine learning expert — or even to write any code of their own. The system works by taking pre-existing models and transferring them to a new dataset, coupled with a period of re-training and pruning that improves accuracy while reducing the size of the resulting model.
The ability to transfer learning from one model to another is only of use if you can find a base model, of course, and it's here that NVIDIA's Transfer Learning Toolkit 3.0 aims to shine: The company is making a wide variety of models, including new conversational AI models and a heart rate estimation model which operates from a standard webcam with a claimed 0.7 beats-per-minute accuracy, available completely free of charge.
Using NVIDIA's new TLT 3.0 is a simple three-stage process: A pre-trained model is selected, then applied to the target dataset to create a customized model; the new model is evaluated and pruned, which reduces the size of the customized model while boosting performance; finally, the new model is retrained on the target dataset to recover any accuracy lost in pruning.
As with previous versions of the NVIDIA Transfer Learning Toolkit, TLT 3.0 is a Linux-exclusive application - though, NVIDIA's Chintan Shah told us during a pre-release briefing, Windows support is on the company's roadmap — and only runs on x86 systems with one or more NVIDIA GPUs.
The resulting models, however, are compatible with both x86 machines and the company's Jetson family of Arm-based embedded systems, with NVIDIA measuring the performance of its automated number plate recognition (ANPR) model at nine frames per second on a single stream when running on the entry-level Jetson Nano and rising to 147 FPS across five streams on the Jetson AGX Xavier.
NVIDIA TLT 3.0 is now available to download from the official website.
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.