Please provide complete information as applicable to your setup.
• Hardware Platform (Jetson / GPU) jetson • DeepStream Version 6.3 • JetPack Version (valid for Jetson only) 5.1.3 • TensorRT Version 8.4
Question 1: I used the uploaded export_yolov8.py to convert the PT model to onnx, resulting in an onnx model with only one output header. Then, I used the uploaded config.inpfer_primary_YOLOV8.txt and deepstream configuration file to convert the engine model. I found a problem where the engine converted from config.inpfer_primary_YOLOV8.txt with network type=0 crashes when running, while the engine converted from config.inpfer_primary_YOLOV8.txt with network type=2 can run normally. Why is this?
Question 2: Where can I find out the meaning of each parameter in config.inpfer_primary_YOLOV8.txt? I couldn’t find any introduction to the network type parameter in the introduction file of the deepstrea-yolo folder. upload.tar.gz (2.3 KB)
I promise it’s a detection model,but use the latest deepstream-yolo convert file(export_yolov8.py), the onnx model only have one output header,should I use the convert file that not the latest version?
The postprocessing I mentioned in this topic is the postprocessing which parse the model’s output and calculate the bboxes from the model’s output. The postprocessing should be aligned to the model you use.
Okay, I understand what you mean. How should I configure the config_infer_primary_yoloV8.txt file for post-processing if the output header of the model is one?
i use the deepstream-yolo proj to generate onnx,and use the
deepstream-app -c config_primary_yolov8.txt
tool to generate the engine format that can run detection analysis on jetson device.i use the above upload file to generate onnx and engine but when I run it’s acount a segmentation error.
you must generate a correct onnx model
here is how I export the onnx with python:
from ultralytics import YOLO
model = YOLO("yolov8s.yaml")
path = model.export(format="onnx", imgsz=(640, 640)) # return path to exported model
yolov8’s output format is different with yolov7, so you must rewrite the cuda function “NvDsInferParseYoloCuda” to adapt the yolov8’s post processing. Without that you can not got correct result with this repo
You’ve already configured your postprocessing function “NvDsInferParseYolo” with “network-type=0” in the configuration file. This postprocessing function caused the crash. So you need to debug your postprocessing function “NvDsInferParseYolo”, it may not match to your model.
ok, thank you so…much for your reply ! I don’t know if I can understand it this way. Network type parameters equal to 0, 1, and 2 do not necessarily correspond exactly to the model type. Previously, I saw that network type=0 corresponds to detection, and equal to 2 represents classification. In fact, the detection model can also be set to network type=2, right？
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