yolov9

YOLOv9 Export - Support Detection and Segmentation Models

Description

This implementation use two TensorRT Plugins YoloNMS and ROIAling.
YoloNMS is a custom TRT Plugin. For now we must build this Plugin.
ROIAling is available on TensorRT since 8.5

Export Model - End2End for Detection and Segmentation

1. Clone Original YOLOv9, Patch Export Files and Install the requirements

git clone https://github.com/levipereira/deepstream-yolov9.git
git clone https://github.com/WongKinYiu/yolov9.git

## Patch Export Files
cp deepstream-yolov9/yolov9/export_onnxtrt.py yolov9/
cp deepstream-yolov9/yolov9/models/experimental_trt.py yolov9/models/

## Start docker pytorch container
mkdir yolov9/runs
docker run --gpus all  \
 -it \
 --rm \
 --name yolov9-onnx-trt \
 --net host  \
 --ipc=host \
 -v $(pwd)/yolov9:/yolov9 \
 -v $(pwd)/runs:/yolov9/runs \
 -w /yolov9 \
 nvcr.io/nvidia/pytorch:23.02-py3

## Install Pre-Reqs 
apt-get update 
apt-get install -y zip htop screen libgl1-mesa-glx libfreetype6-dev

pip3 install seaborn \
thop \
markdown-it-py==2.2.0 \
onnx-simplifier==0.4.35 \
onnxsim==0.4.35 \
onnxruntime==1.16.3                
pip install onnx_graphsurgeon --extra-index-url https://pypi.ngc.nvidia.com 
pip install pillow==9.5.0 --no-cache-dir --force-reinstall

2. Download model

skip this step if you are using custom models

cd /yolov9
mkdir -p /yolov9/runs/models
cd /yolov9/runs/models
wget https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-c-seg-converted.pt
wget https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-c-converted.pt

3. Export Models

The export_onnxtrt.py automatically detect Detection model and Segmentation Model

Available Arguments:

• --weights: Path to the model.pt file.
• --imgsz: Image size (height, width). Default 640
• --device: Device for execution (cpu, cuda device). Default cpu
• --class-agnostic: Enable Agnostic NMS (single class). Default False
• --topk-all: Topk to keep for all classes. Default 100
• --iou-thres: IoU threshold. Default 0.45
• --conf-thres: Confidence threshold. Default 0.25
• --mask-resolution: Mask pooled output. Default 160
• --pooler-scale: Multiplicative factor used to translate the ROI coordinates. Default 0.25
• --sampling-ratio: Number of sampling points in the interpolation. Default 0

python3 export_onnxtrt.py --weights runs/models/yolov9-c-converted.pt

python3 export_onnxtrt.py --weights runs/models/yolov9-c-seg-converted.pt

Note: Only yolov9-c-seg-converted.pt has an issue after reparametrization. It was configured as a DetectionModel instead of a SegmentationModel, gelan-c-seg.pt does not have this issue.
To workaround this, you can force the export to recognize it as a segmentation model by using the environment variable MODEL_DET=0.

If you don't use the environment variable, you'll encounter the following error:

ONNX TRT: export failure ❌ 0.9s: 'tuple' object has no attribute 'permute'

To workaround this issue, use the following command:

MODEL_DET=0 python3 export_onnxtrt.py --weights runs/models/yolov9-c-seg-converted.pt

4. Copy ONNX Models

exit from docker to host.

Copy the generated ONNX model file to the deepstream-yolov9/models dir.

cp yolov9/run/models/*.onnx deepstream-yolov9/models/

Acknowledgements

https://github.com/hiennguyen9874/yolov7-seg