TAO Toolkit Quick Start Guide

This page provides a video and text-based quick start guide for installing and running TAO Toolkit.



The following system configuration is recommended to achieve reasonable training performance with TAO Toolkit and supported models provided:

  • 32 GB system RAM

  • 32 GB of GPU RAM

  • 8 core CPU


  • 100 GB of SSD space

TAO Toolkit is supported on discrete GPUs, such as A100, A40, A30, A2, A16, A100x, A30x, V100, T4, Titan-RTX and Quadro-RTX.


TAO Toolkit is not supported on GPU’s before the Pascal generation.

Software Requirements



** Comment**

Ubuntu LTS




Not needed if you use TAO toolkit API



Not needed if you use TAO toolkit API



Not needed if you use TAO toolkit API



Not needed if you use TAO toolkit API



Not needed if you use TAO toolkit API



Not needed if you use TAO toolkit API



Not needed if you use TAO toolkit API



Not needed if you use TAO toolkit API

Package Content

Download the TAO package which contains startup scripts, Jupyter notebooks and config files. TAO is supported on Google Colab; if you want to try on Colab, you can skip this step and directly scroll down to the running in colab section.

wget --content-disposition https://api.ngc.nvidia.com/v2/resources/nvidia/tao/tao-getting-started/versions/4.0.0/zip -O getting_started_v4.0.0.zip
unzip -u getting_started_v4.0.0.zip  -d ./getting_started_v4.0.0 && rm -rf getting_started_v4.0.0.zip && cd ./getting_started_v4.0.0

Running TAO Toolkit

TAO toolkit is available as a docker container or a collection of python wheels. There are 4 ways to run TAO Toolkit depending on your preference and setup, through

  1. the launcher CLI

  2. the containers directly

  3. the tao toolkit apis

  4. python wheels

Launcher CLI

The TAO Toolkit launcher is a lightweight Python based command-line interface. The launcher acts as a front-end for TAO Toolkit containers built on top of PyTorch and TensorFlow. The CLI abstracts the user from the information about which network is implemented in what container. The respective container gets launcher automatically based on the model you plan to use.

To get started with the launcher, follow the instructions below to install the required pre-requisite software.

Installing the Pre-requisites

The TAO Toolkit launcher is strictly a python3 only package, capable of running on python versions >= 3.6.9.

  1. Install docker-ce by following the official instructions.

    Once you have installed docker-ce, follow the post-installation steps to ensure that the docker can be run without sudo.

  2. Install nvidia-container-toolkit by following the install-guide.

  3. Get an NGC account and API key:

    1. Go to NGC and click the TAO Toolkit container in the Catalog tab. This message is displayed: “Sign in to access the PULL feature of this repository”.

    2. Enter your Email address and click Next, or click Create an Account.

    3. Choose your organization when prompted for Organization/Team.

    4. Click Sign In.

  4. Log in to the NGC docker registry (nvcr.io) using the command docker login nvcr.io and enter the following credentials:

    a. Username: "$oauthtoken"
    b. Password: "YOUR_NGC_API_KEY"

    where YOUR_NGC_API_KEY corresponds to the key you generated from step 3.


DeepStream 6.0 - NVIDIA SDK for IVA inference is recommended.

  1. Setup your python environment with python version >= 3.6.9


    NVIDIA recommends setting up a python environment using miniconda. The following instructions show how to setup a python conda environment.

    1. Follow the instructions in this link to set up a conda environment using a miniconda.

    2. Once you have installed miniconda, create a new environment by setting the Python version to 3.6.

      conda create -n launcher python=3.6
    3. Activate the conda environment that you have just created.

      conda activate launcher
    4. Once you have activated your conda environment, the command prompt should show the name of your conda environment.

      (launcher) py-3.6.9 desktop:
    5. When you are done with you session, you may deactivate your conda environment using the deactivate command:

      conda deactivate
    6. You may re-instantiate this created conda environment using the following command.

    conda activate launcher

Installing TAO Launcher

Once you have installed the required pre-requisites.

  1. Install the CLI launcher via the quick start script downloaded with the getting_started NGC package from here

    bash setup/quickstart_launcher.sh --install
  2. You can also use this script to update the launcher to the latest version of TAO Toolkit by running the following command

    bash setup/quickstart_launcher.sh --upgrade
  3. Invoke the entrypoints using the tao command.

    tao --help

    The sample output of the above command is:

    usage: tao [-h]
    Launcher for TAO
    optional arguments:
    -h, --help            show this help message and exit

    Note that under tasks you can see all the launcher-invokable tasks. The following are the specific tasks that help with handling the launched commands using the TAO Launcher:

    • list

    • stop

    • info


When installing the TAO Toolkit Launcher to your host machine’s native python3 as opposed to the recommended route of using virtual environment, you may get an error saying that tao binary wasn’t found. This is because the path to your tao binary installed by pip wasn’t added to the PATH environment variable in your local machine. In this case, please run the following command:

export PATH=$PATH:~/.local/bin

Running from container

Users have option to also run TAO directly using the docker container. To use container directly, user needs to know which container to pull. There are multiple containers under TAO, and depending on the model that you want to train you will need to pull the appropriate container. This is not required when using the Launcher CLI.

export DOCKER_REGISTRY="nvcr.io"
export DOCKER_NAME="nvidia/tao/tao-toolkit"
export DOCKER_TAG="4.0.0-tf1.15.5" ## for TensorFlow docker

docker run -it --rm --gpus all -v /path/in/host:/path/in/docker $DOCKER_CONTAINER \
       detectnet_v2 train -e /path/to/experiment/spec.txt -r /path/to/results/dir -k $KEY --gpus 4

For detailed instructions on how to run directly from containers, refer to this section

Running TAO Toolkit APIs

TAO Toolkit API is a Kubernetes service that enables building end-to-end AI models using REST APIs. The API service can be installed on a Kubernetes cluster (local / AWS EKS) using a Helm chart along with minimal dependencies. TAO toolkit jobs can be run using GPUs available on the cluster and can scale to a multi-node setting. Users can use a TAO client CLI to interact with TAO services remotely or can integrate it in their own apps and services directly using REST APIs.

To get started, use the provided one-click deploy script to deploy either on bare-metal setup or on managed Kubernetes service like Amazon EKS. Jupyter notebooks to train using the APIs directly or using the client app is provided under notebooks/api_starter_kit.

bash setup/quickstart_api_bare_metal
bash setup/quickstart_api_aws_eks

More information about setting up the API services and the API is provided in here.

Running from python wheels

You can also run TAO directly on bare-metal without docker or K8s by using the python wheels. As a sample tutorial, TAO Toolkit provides sample tutorials to deploy directly on Google Colab without having to configure your infrastructure. The full instructions on how to work with google colab are provided here.

Run sample jupyter notebooks

As a part of the TAO toolkit package, example jupyter notebooks are available for all the tasks that supported in TAO.

Computer vision

All computer vision samples are included in the getting started resource on NGC.

Purpose-Built Pre-trained Models

The following is a list of purpose-built pretrained models mapped with their corresponding samples.

Model Name

Jupyter Notebook











License Plate Recognition


License Plate Detection


Heart Rate Estimation


Gesture Recognition


Gaze Estimation


Facial Landmark






Emotion Recognition














Open model architecture

Network Architecture

Jupyter Notebook





















Classification (TF1)


Classification (TF2)


EfficientDet (TF1)


EfficientDet (TF2)




Conversational AI

The TAO Conversational AI package, provides several end to end sample workflows to train conversational AI models using TAO Toolkit and subsequently deploying them to Riva. You can find these samples at:

Conversational AI Task

Jupyter Notebooks

Speech to Text Citrinet

Speech to Text Citrinet Notebook

Speech to Text Conformer

Speech to Text Conformer Notebook

Question Answering

Question Answering Notebook

Text Classification

Text Classification Notebook

Token Classification

Token Classification Notebook

Punctuation and Capitalization

Punctuation Capitalization Notebook

Intent and Slot Classification

Intent Slot Classification Notebook

NGram Language Model

NGram Language Model Notebook

Text to Speech

Text to Speech Notebook

You can download these resources, by using the NGC CLI command available at the NGC resource page. Once you download the respective tutorial resource, you may instantiate the jupyter notebook server.

pip3 install jupyter
jupyter notebook --ip --allow-root --port 8888

Copy and paste the link produced from this command into your browser to access the notebook. The /workspace/examples folder will contain a demo notebook. Feel free to use any free port available to host the notebook if port 8888 is unavailable.

Downloading the Models

The TAO Toolkit Docker gives you access to a repository of pretrained models that can serve as a starting point when training deep neural networks. These models are hosted on the NGC. To download the models, please download the NGC CLI and install it. More information about the NGC Catalog CLI is available here. Once you have installed the CLI, you may follow the instructions below to configure the NGC CLI and download the models.

Listing all available models

Use this command to get a list of models that are hosted in the NGC model registry:

ngc registry model list <model_glob_string>

Here is an example of using this command for the computer vision models:

ngc registry model list nvidia/tao/pretrained_*

This command gives us a list of the pretrained backbones available for different tasks:

  • Classification

  • Object Detection with Detectnet_v2

  • Object Detection with SSD/DSSD/YOLOv3/YOLOv4/YOLOv4-Tiny/FasterRCNN/RetinaNet

  • Object Detection with EfficientDet

  • Instance Segmentation

  • Semantic Segmentation


All our classification models have names based on this template: nvidia/tao/pretrained_classification:<template>.

To view the full list of computer vision and conversational AI models, use the following command:

ngc registry model list nvidia/tao/*

Downloading a model

Use this command to download the model you have chosen from the NGC model registry:

ngc registry model download-version <org/team/model_name:version> -dest <path_to_download_dir>

For example, use this command to download the resnet 18 classification model to the $USER_EXPERIMENT_DIR directory:

ngc registry model download-version
nvidia/tao/pretrained_classification:resnet18 --dest $USER_EXPERIMENT_DIR/pretrained_resnet18
Downloaded 82.41 MB in 9s, Download speed: 9.14 MB/s
Transfer id: pretrained_classification_vresnet18 Download status: Completed.
Downloaded local path: /workspace/tao-experiments/pretrained_resnet18/
Total files downloaded: 2
Total downloaded size: 82.41 MB
Started at: 2019-07-16 01:29:53.028400
Completed at: 2019-07-16 01:30:02.053016
Duration taken: 9s seconds

Once you have downloaded the notebook samples and required pretrained models, you can start the respective sample notebook with the following command:

jupyter notebook --ip --port 8888 --allow-root

Open an internet browser on localhost and navigate to the following URL:


If you want to run the notebook from a remote server then please follow these steps.

Execute the cells in the notebook to train a model using TAO Toolkit.

Once training is complete, follow these instructions to deploy a computer vision model to DeepStream. For Conversational AI models, follow the instructions in this section.