Challenges Loading YOLOv7 Detection Model on Xavier AGX with Jetpack 5.1.2

Deepstream SDK: 6.3
Jetpack: 5.1.2
Jetson Xavier AGX
I am currently facing challenges loading the YOLOv7 detection model on a Jetson Xavier AGX with Jetpack 5.1.2. The issues encountered are as follows:

  1. Low FPS with NVIDIA-AI-IOT Repository:

I utilized the YOLOv7 configuration files from the NVIDIA-AI-IOT repository here to convert them into an engine. However, when I integrate the engine into my application, the FPS drops significantly to 6.
GitHub - NVIDIA-AI-IOT/yolo_deepstream: yolo model qat and deploy with deepstream&tensorrt

  1. Memory Issues with Alternative Repository:

As an alternative, I tried another repository (GitHub - marcoslucianops/DeepStream-Yolo: NVIDIA DeepStream SDK 6.3 / 6.2 / 6.1.1 / 6.1 / 6.0.1 / 6.0 / 5.1 implementation for YOLO models) with different YOLOv7 configuration files. During the conversion process to an engine, I encountered memory-related errors, as shown below:

WARNING: [TRT]: DLA requests all profiles have the same min, max, and opt value. All DLA layers are falling back to GPU.
WARNING: [TRT]: Tactic Device request: 8435MB Available: 6341MB. Device memory is insufficient to use tactic.
WARNING: [TRT]: Skipping tactic 3 due to insufficient memory on requested size of 8435 detected for tactic 0x0000000000000004. Try decreasing the workspace size with IBuilderConfig::setMemoryPoolLimit().
...

I would greatly appreciate any insights, recommendations, or assistance you could provide to help overcome these issues and successfully load the YOLOv7 detection model on the Xavier AGX with Jetpack 5.1.2.

Thank you in advance for your help.

What FPS can you get if running the original sample? and which mode did you set for network-mode?

The “FPS” of the original sample is 20, and I initially used “network-model=0.” Later, I changed it to “network-mode=2,” but the issues were not resolved.

You may need to find where the latency is introduced in your program, the way to measure the latency of different components can be found here: DeepStream SDK FAQ - #12 by bcao

Sorry for the late reply, Is this still an DeepStream issue to support?

could you elaborate on your test steps? how did you generate yolov7 model? are you using the model from this link? how did you convert them into an engine? by DeepStream or trtexec?

thank you for answer i used now model yolov8 in jetson and he work correctly

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