Improve inference performances yolov5

Autonomous Machines Jetson & Embedded Systems Jetson Nano
,
1

Hello everyone. I am trying to deploy a little software on the Jetson Nano 2gb to do object detection in real time. What I am currently using is the yolov5n model (compiled to tensorrt) to do inference, but for processing a frame it needs ~140ms. I am a newbie and my knowledges are very limited, I read something about DeepStream and how it could improve performances, but I have no idea if and how it can be implemented with my current software (written in Python). Is there something I can do to improve performances, using DeepStream or something else?

This is how I load and use my model:
model = torch.hub.load( ‘yolov5’, ‘custom’, path=path_to_model.engine, force_reload=True, source=‘local’)
predictions = model(opencv_image)

I’ll be very grateful for any help you can give me with this

3

Hi,

Have you tried it with TensorRT? Since the API you used is PyTorch.

Thanks.

4

I used TensorRT, the problem was related to how I handled the frame capture from the camera, using multiple threads improved performances as desired. Thanks for your reply though

5

Hi,

Do you run it with the below API?
If yes, you might use PyTorch built-in TensorRT rather than pure TensorRT.

model = torch.hub.load( ‘yolov5’, ‘custom’, path=path_to_model.engine, force_reload=True, source=‘local’)
predictions = model(opencv_image)

If you want to accelerate a camera pipeline, it’s recommended to try our Deepstream SDK.
Below is a sample for YOLOv5 from the community for your reference:

Thanks.

1 Like
7

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