Test Drive the NVIDIA TAO Toolkit for Free on NVIDIA LaunchPad. Sign up now

NVIDIA TAO Toolkit

Create highly accurate, customized, and production-ready AI models to power your speech and computer vision AI applications.

Download TAO     Get started

What Is the NVIDIA TAO Toolkit?

Creating an AI/machine learning model from scratch requires mountains of data and an army of data scientists. Now, you can speed up the model development process with transfer learning —a popular technique that extracts learned features from an existing neural network model to a new customized one.

The NVIDIA TAO Toolkit, built on TensorFlow and PyTorch, is a low-code version of the NVIDIA TAO framework that accelerates the model training process by abstracting away the AI/deep learning framework complexity. The TAO Toolkit lets you use the power of transfer learning to fine-tune NVIDIA pretrained models with your own data and optimize for inference—without AI expertise or large training datasets.

What is TAO Toolkit and how does it fit into AI model development workflow?

Key Benefits

TAO toolkit lets you train models with Jupyter notebooks easily.

Train Models Easily

The TAO toolkit is a low-code solution that lets you train models with Jupyter notebooks, eliminating the need for AI framework expertise.

TAO toolkit helps you build highly accurate AI models for your use-case.

Build Highly Accurate AI

Use NVIDIA pretrained models and model architectures to create highly accurate and custom AI models for your use-case.

TAO toolkit allows you to optimize the model for inference.

Optimize For Inference

Go beyond customization and achieve up to 4X performance by optimizing the model for inference.

TAO toolkit helps you deploy optimized models with ease.

Deploy With Ease

Deploy optimized models using NVIDIA DeepStream for vision AI, Riva for speech AI, and Triton Inference Server™.

What Are Pretrained AI Models?

Pretrained AI/deep learning models have been trained on representative datasets and fine-tuned with weights and biases. You can quickly and easily customize these models with only a fraction of real-world or synthetic data, compared to training from scratch.

Pretrained Models For Vision AI

Create custom deep learning models for computer vision tasks like image processing and classification, object detection, and semantic segmentation using 100+ NVIDIA-optimized model architectures.

You can also use task-based models to recognize human actions and poses, detect people in crowded spaces, detect and classify vehicles and license plates, and much more.

Download Jupyter notebooks from NGC View performance chart

Pretrained Models for Conversational AI

Design personalized real-time call center experiences, smart kiosks, and other conversation AI services by fine-tuning Automatic Speech Recognition (ASR) , Natural Language Processing (NLP), and Text-to-Speech (TTS)-based modes.

Download Jupyter notebooks from NGC

Key Features

Integrate TAO Toolkit in Your Application with Rest APIs

Deploy TAO Toolkit in a modern cloud-native infrastructure on Kubernetes and integrate it in your application with REST APIs. Build a new AI service or integrate the TAO Toolkit into your existing service and enable automation between disparate tools.

Learn more

Import ONNX Model Weights

Simply import your own classification and segmentation model weights from an ONNX model into TAO Toolkit for fine-tuning and optimization.

See demo

Augment Your Data

Improve your model performance with data augmentation features such as spatial and color transformations to diversify your existing image data set.

Learn more

Visualize with TensorBoard

Understand how your model converges by visualizing scalars such as training and validation loss, accuracy, model weights, and predicted images.

Learn more

Optimize for Inference

Achieve up to 4X speed up in inference with pruning and INT8 quantization while maintaining comparable accuracy.

Read blog

Power DeepStream and Riva

The optimized models produced by TAO Toolkit can be easily integrated with NVIDIA DeepStream SDK for vision AI and Riva for speech AI, letting you unlock greater performance and accelerating your development for your applications.

Integrate with DeepStream Integrate with Riva

Create AI Anywhere with NVIDIA AI Enterprise

The TAO toolkit is now an integral part of the NVIDIA AI Enterprise, an end-to-end, cloud-native suite of AI and data analytics software. It’s optimized, certified, and supported by NVIDIA so enterprises can create and deploy AI—certified to deploy anywhere—with enterprise support to keep AI projects on track.

You can now experience all the features of the TAO Toolkit for free through NVIDIA LaunchPad

Sign up here

Inference Performance

Unlock peak inference performance with NVIDIA pre-trained models across platforms from the edge with NVIDIA Jetsons to the cloud with NVIDIA Ampere architecture GPUs. For more details on batch size and other models, check the detailed performance datasheet.

Xavier NX
AGX Xavier
AGX Orin
T4
A2
A30
A100
Model Arch
Inference Resolution
Precision
Model Accuracy
GPU + 2*DLA(FPS)
GPU + 2*DLA(FPS)
GPU + 2*DLA(FPS)
GPU(FPS)
GPU (FPS)
GPU (FPS)
GPU (FPS)
PeopleNet
960x544x3
INT8
80%mAP
263
418
1294
1064 
581
3160
6245
3D Pose Estimation
256x192x3
FP16
8 pixel error
147
235
711
713 
471
2242
4179
Pose Action Classification
3x300x34x1
FP16
90%
87
150
262
376
211
1122
2145
DashCamNet
960x544x3
INT8
84% mAP
423
670
1895
1676
865
4900
9500
License Plate Recognition
96x48x3
FP16
98% mAP
706
1190
4118
3959 
2180
12400
27200
Action Recognition 2D
224x224x96
FP16
83%
245
471
1577
1897
1044
6000
12600
People Semantic Segmentation
960x544x3
FP16
87% mIoU
199
356
673
1027
631
2862
5745

Customer Stories

OneCup AI Customer Story

OneCup AI

OneCup AI’s computer vision system tracks and classifies animal activity using NVIDIA pretrained models and TAO Toolkit, significantly reducing their development time from months to weeks.

Learn more
RocketBoots Customer Story

RocketBoots

RocketBoots uses NVIDIA TAO toolkit and pretrained models to improve the inference performance in their automated workforce management solution.

Learn more
Floatbot Customer Story

Floatbot

Floatbot uses NVIDIA Riva and NVIDIA TAO for their customized Singaporean English voice AI applications, automating call centers for insurance carriers and finance clients globally.

Learn more

Leading Adopters

booz allen
Kion group
Inex tech
Lexmark ventures
One cup AI
Rocketboots
SmartCow
appen
cvedia
hasty
lexset
lightly
Rendered AI
Sky Engine
Innotescus
Yuva AI

General FAQ

NVIDIA TAO is a framework for training, adapting and optimizing computer vision and conversational AI models with your custom data, in a fraction of time without large training datasets or AI expertise.

There are two versions of TAO.

  1. TAO Toolkit, which is the low-code version of TAO
  2. A no-code, GUI-based solution currently in development. Apply for early access here.
Transfer learning is the process of transferring learned features from one application to another. It’s a commonly used training technique where a model trained on one task is re-trained for use on a different task. You can apply transfer learning on vision, speech, and language-understanding models.
Yes. For specific licensing terms, refer to model EULA.
The TAO Toolkit supports 100+ permutations of NVIDIA-optimized model architectures and backbones such as EfficientNet, YOLOv3/v4, RetinaNet, FasterRCNN, UNET, and many more.

The full matrix of supported model architectures can be found here.
TAO Toolkit uses TensorFlow and PyTorch framework completely abstracted away from the user. Users operate TAO Toolkit through documented spec files and do not have to learn about DL framework.
  • Vision models can be deployed through DeepStream
  • Speech Models can deployed through Riva
  • You can also deploy the models through Triton Inference Server.

Refer to the documentation section for deployment details.

With the TAO Toolkit, you can use a model that detects people, and add a new “helmet class” to that model. Fine-tune it with the dataset that contains classes for both people and helmet. Read more about it in the white paper.
Yes, you can deploy the TAO Toolkit in the cloud. Please refer to the TAO Toolkit documentation to learn more about running the TAO Toolkit on AWS, Azure, or GCP.
You can only train with TAO toolkit on an x86 system. You can, however, deploy the optimized models on a Jetson solution.
NVIDIA Metropolis is an application framework, set of developer tools, and partner ecosystem that brings visual data and AI together to improve operational efficiency and safety across a broad range of industries. Learn more here.

Resources

New Developer Blog

Developing and deploying AI-powered robots with NVIDIA Isaac Sim and NVIDIA TAO.

Read blog

Whitepaper

Learn endless ways to adapt and supercharge your AI workflows with transfer learning.

Read whitepaper

GTC Webinar

Learn how TAO Toolkit can help developers overcome their data challenges and fine-tune models with minimum coding, and more.

Watch On-Demand webinar

Simplify and speed-up AI training with TAO Toolkit

Get started