TensorRT Deformable Detr (Python)

TensorRT for Deformable Detr(use python)

Test enviroments

TensorRT 8.4.1.5
Ubuntu 18.04
mmdetection 2.24.1
polygraphy(TensorRT Tool)

A quick demo

1. Convert deformable detr model form pytorch to onnx

python3 onnx_deformable_detr.py --config ${mmdet_path}/configs/deformable_detr/deformable_detr_twostage_refine_r50_16x2_50e_coco.py --checkpoint ${mmdet_path}/work_dirs/deformable_detr_twostage_refine_r50_16x2_50e_coco_20210419_220613-9d28ab72.pth --output deformable_detr.onnx --h 750 --w 1333

2.Genrate TensorRT engine

polygraphy convert deformable_detr.onnx --workspace 8G -o deformable_detr_fp32.engine

3.Inference

python3 demo.py --engine deformable_detr_fp32.engine --thresh 0.3 --img ${your_picture_path} --output result.jpg

Inference performance(only inference time in GPU)

GPU	Model	Mode	Inference time
3090	deformable_detr_twostage_refine_r50_16x2_50e_coco	fp32	35ms
3090	deformable_detr_twostage_refine_r50_16x2_50e_coco	fp16	17ms