RT-DETR Deploy with ONNX and TensorRT

base on project https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rtdetr

not need NMS not need Anchors and soeasy post process

1.Clone project

git clone https://github.com/PaddlePaddle/PaddleDetection.git

2.Export paddlepaddle model

cd PaddleDetection/
python tools/export_model.py -c configs/rtdetr/rtdetr_hgnetv2_l_6x_coco.yml -o weights=https://bj.bcebos.com/v1/paddledet/models/rtdetr_hgnetv2_l_6x_coco.pdparams trt=True --output_dir=output_inference

note: need paddlepaddle python package >= 2.4.0

3.Convert paddlepaddle model to onnx model

paddle2onnx --model_dir=./output_inference/rtdetr_hgnetv2_l_6x_coco/ --model_filename model.pdmodel  --params_filename model.pdiparams --opset_version 16 --save_file rtdetr_hgnetv2_l_6x_coco.onnx

4.Simplify onnx model

onnxsim  rtdetr_hgnetv2_l_6x_coco.onnx rtdetr_hgnetv2_l_6x_coco-sim.onnx --overwrite-input-shape im_shape:1,2 image:1,3,640,640 scale_factor:1,2

5.ONNX infer with simplified model

python 1.onnx_run_original_onnx_post_process.py

and will generate a result.jpg

6.Modify onnx model to only one input

simplified model has three inputs (im_shape, image, scale_factor),but we want only one input of model. use script modify onnx model

python modify_onnx.py -i rtdetr_hgnetv2_l_6x_coco-sim.onnx -o rtdetr_hgnetv2_l_6x_coco-modify.onnx

7.ONNX infer with modified model

python 2.onnx_run_modify_onnx_no_post_process.py

8.Generate TensorRT engine

for supporting GatherND and GridSample op need TensorRT version >= 8.5

• GatherND TensorRT >= 8.4
• GridSample TensorRT >= 8.5

trtexec --onnx=rtdetr_hgnetv2_l_6x_coco-modify.onnx --saveEngine=rtdetr_hgnetv2_l_6x_coco-modify-fp32.engine