How to use custom Tensorflow code (Python) with DeepStream?

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hi,

I have an application where I use tensorflow backend with Keras for inference. While I understand how to use the pre-trained models by way of changing the configuration file for DeepStream, how do I port my tensorflow / keras code in Python so that I can leverage the power of DeepStream for such custom applications? Couldn’t find any samples that illustrate this scenario.

2

Hi,

Currently, Deepstream doesn’t support TensorFlow model yet.

You will need to convert the model into .uff file and feed it into Deepstream.
Here is an uff sample for your refernece:
/opt/nvidia/deepstream/deepstream-4.0/sources/objectDetector_SSD/

Thanks.

3

Hi,

I have the same query that i am using jetson nano, i have python program in that implementing tiny-yolo. I am having performance issue.

So i saw deepstream SDK,Is there any way to implement deepstream in python? or Is there any way to customise deepstream as programmed in python?

Cause in my python script i did some calculation like distance between points and mid point of square at every frame using opencv, Is it possible to program in deepstream?.

Thanks.

4

Hi,

Deepstream starts to support python recently.
For more information, please check this GitHub for details:
https://github.com/NVIDIA-AI-IOT/deepstream_python_apps/blob/master/HOWTO.md

Thanks.