how to deploy a new onnx model or trt model for detection in DeepStream ?

Hello,I have trained a custom object detection network and it has been saved as onnx model and transformed to trt model.
I’d like to deploy this network with deepstream on Jetson Xavier.
I am confused about the config file.In another word, is there any example config file for deploy a onnx or trt detection model?

Hi,

onnx model is supported by TensorRT/Deepstream directly.
The procedure is similar to the TensorFlow based model.
You can check this sample and give it a try:
/opt/nvidia/deepstream/deepstream-4.0/sources/objectDetector_SSD

Here is an implementation from community for your reference:
https://github.com/thatbrguy/Deep-Stream-ONNX

Thanks.

Thanks for you reply.
I am read the code from your given links.
But I found my custom network is not for the popular detection or segmentation.
So I am curious about how to write the ''custom_bbox_parser ‘’ accordingly.
Except the popular network like YOLO or SSD,can self-defined model run with the DeepStream ?

Hi,

Please refer to this page:
https://towardsdatascience.com/how-to-deploy-onnx-models-on-nvidia-jetson-nano-using-deepstream-b2872b99a031

Thanks.

Thanks.
But the link is unreachable, even with a VPN.

Hi,

Could you try it again?
I can open it without issue.

The tutorial title is “How to deploy ONNX models on NVIDIA Jetson Nano usingDeepStream”, published by Bharath Raj.

Thanks.

Hi, I am trying the https://github.com/thatbrguy/Deep-Stream-ONNX
but there is error in the end.
The path needs modification and the folder name needs update.

Is there any video tutorial?

I am working on a tracking device using Nvidia nano.
I have onnx from Azure and just want to import it to deep stream. Then output the x,y coordinates from my webcam.

Thanks