PyTorch-Onnx-Tensorrt

Test yolov3-trt on jetson nano

Requirements

1. Python 3
2. OpenCV
3. PyTorch
4. Onnx 1.4.1
5. Tensorrt

Downloading YoloV3 Configs and Weights

mkdir cfg
cd cfg 
原链接失效百度网盘地址
链接：https://pan.baidu.com/s/1CtWGfJQ2PWl6kLoMsH7Yhg 
提取码：gm4v

mkdir weights
cd weights
wget https://pjreddie.com/media/files/yolov3.weights

Editing Config File

Inorder to Run the model in Pytorch or creating Onnx / Tensorrt File for different Input image Sizes ( 416, 608, 960 etc), you need to edit the Batch Size and Input image size in the config file - net info section.

batch=1
width=416
height=416

Generating the Onnx File

python3 create_onnx.py --reso 416

Generating the Tensorrt File

python3 onnx_to_tensorrt.py --model yolov3-416
 

Creating the Tensorrt engine takes some time. So have some patience.

Test the YOLOv3 TensorRT engine with the "dog.jpg" image.(jetson nano run 2.55 FPS,jetson xavier nx run 11.59 FPS)

python3 trt_yolov3.py --model yolov3-416
                        --image --filename ${HOME}/Pictures/dog.jpg

Run the "trt_yolov3.py" demo program.(jetson nano run 3.18 FPS,jetson xavier nx run 13.01FPS)

python3 trt_yolov3.py --usb --vid 0 --width 1280 --height 720   (or 640x480)

evaluating mAP of the optimized YOLOv3 engine (jetson nano coco map@IOU=0.5 → 61.6%)

python3 eval_yolov3.py --model yolov3-416 

相关链接

• https://github.com/jkjung-avt/tensorrt_demos#yolov3
• https://github.com/Rapternmn/PyTorch-Onnx-Tensorrt