A few questions regarding TAO 5.0

I’ve been trying official TAO sample codes, which utilized TAO 4.0 on Google Colab as I still have no access to a machine atm.

Saw the introduction and video clips regarding TAO 5.0 a bit. I’d like to further see a few things clarified.

AI-assisted data annotation
I saw the following video showing this feature.

A few questions I’m wondering:
1. Do users have to draw the bounding boxes by themselves before hand or can the bounding boxes be AUTOMATICALLY generated during the process?

I was expecting that the input for the AI-assisted data annotation is raw data, which refer to the images WITHOUT any annotations. This means that those images first need to be used as the input for some object detection model such as Yolo so as to generate the bounding boxes and then what’s displayed in the video is done.

If what I expected is incorrect, does that mean users have to first generate the bounding boxes for each of the image by either drawing them with some other tools such as labelme or doing inference using object detection models and saving the results and then converting them into COCO format so that it can then be used by the AI-assisted annotation tool?

2. Is it possible to try TAO 5.0 on google colab at the moment?

You can refer to Access the Latest in Vision AI Model Development Workflows with NVIDIA TAO Toolkit 5.0 | NVIDIA Technical Blog. It is MAL(Mask Auto Labeler) . MAL is a transformer-based, mask auto-labeling framework for instance segmentation using only box annotations. MAL takes box-cropped images as inputs and conditionally generates the mask pseudo-labels.

You can find MAL notebook in latest 5.0 version. TAO Toolkit Getting Started | NVIDIA NGCGPU-optimized AI, Machine Learning, & HPC Software | NVIDIA NGC

For running TAO5.0 in colab, please check coming user guide. Thanks.

1 Like

Is there an ETA for the TAO 5.0 colab :) ?

The latest user guide for running Colab is available in Running TAO Toolkit on Google Colab - NVIDIA Docs.

Thanks a lot for the response, but at least checking the Yolov4 , ssd and detectnev2 notebooks it still pulls this repository ( GitHub - NVIDIA-AI-IOT/nvidia-tao ) and it sets up the environment using this script: https://github.com/NVIDIA-AI-IOT/nvidia-tao/blob/main/tensorflow/setup_env.sh , and that installs python3.6 -m pip install nvidia-tao==4.0.0 not 5.0.0

EDIT: I can also open a new thread for this but it seems closely related to this

Will check further.
For the wheel, actually there is 5.0.0 version. nvidia-tao · PyPI

That is correct, can we assume it is safe to just run:

pip install nvidia-tao==5.0.0

And overwrite the old version?

Yes, you can run it for colab.

1 Like

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

After checking internally, for the wheel, it is needed to stick with 4.0.

More, for colab , actually there are only some changes like bug fixes and switching object detection data from kitti to synthetic has been made.
The newer models of TAO 5.0 are not added.
In short, there are not major changes for colab.

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.