tensorrt_yolov3

---

Description

Use nvidia tensorRT to speed up yolov3 inference

Requirements:

• CUDA>=10.0
• pytorch>=1.0
• pycuda
• opencv_python
• TensorRT 5.0.2.6/7

Installation Instructions:

CUDA:

CUDA is an NVIDIA GPU programming language, with installation instructions that can be found on NVIDIA's website at: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html

TensorRT:

TensorRT is a CUDA-based deep learning inference platform optimized for NVIDIA GPU's, used to run cone detections and keypoint detections on the MIT/DUT18D car. TensorRT installation instructions can also be found on NVIDIA's website at: https://docs.nvidia.com/deeplearning/sdk/tensorrt-install-guide/index.html

Yolo v3

official website: YOLO v3

Usage

1. Download darknet weights and model cfg

weightsfile only for COCO here,

Or, you could just type (if you're on Linux)

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

2. Convert the darknet weights to onnx

!!! only support python2

python darknet to onnx.py

3. Covert onnx to TensorRT engine file:

chage ONNX_FILE_PATH and ENGINE_FILE_PATH in onnx_to_tensorrt.py and run:

python3 onnx_to_tensorrt.py

4. Detect

python detect --engine_path <your engine file> --target_path <image>

optional arguments:

  -h, --help            show this help message and exit
  --cfg_path CFG_PATH
  --target_path TARGET_PATH
                        path to target image/video
  --engine_file ENGINE_FILE
                        path to tensortRT engine file
  --mode MODE           detect: image or video
  --camera_device CAMERA_DEVICE
                        code of camera device
  --vanilla_anchor      whether to use vanilla anchor boxes 
  --no_vanilla_anchor   

example:

python detect.py --engine_path yolov3-608.trt --target_path data/dog.jpg

5. On a Camera

python dectect.py --camera_device <code>  --mode video

example:

python dectect.py --camera_device 0  --mode video

6. for int8 model

this model inference with fp32

the int8 one is writing