# Nvidia Nano + Tensor RT + YOLOv3

主要資料來源：https://medium.com/@penolove15/nvidia-jetson-nano-%E7%94%A8-tensorrt-%E8%B7%91-yolov3-a87ebb0036a7

內容幾乎都一樣, 只是記錄自己實際跑起來的流程

----

裝完OS之後, 系統更新 & 裝相關套件

sudo apt-get update
sudo apt-get upgrade

# setup python3 related library
sudo apt-get install python3-dev python3-pip

# install jetson monitor tool (這個應該可有可無)
sudo -H pip3 install jetson-stats

# required by onnx
sudo apt-get install cmake protobuf-compiler libprotoc-dev

# required by scipy
sudo apt-get install libopenblas-dev gfortran

# others libs
sudo apt-get install libjpeg8-dev libxslt1-dev libfreetype6-dev

設定swap

sudo fallocate -l 4G /swapfile
sudo chmod 600 /swapfile
sudo mkswap /swapfile
sudo swapon /swapfile
sudo swapon -show

這個swap設定重開機後會消失, 所以要加到 /etc/fstab 裡

echo '/swapfile none swap sw 0 0' | sudo tee -a /etc/fstab

下載yolov3 tensorrt

git clone https://github.com/penolove/yolov3-tensorrt.git
cd yolov3-tensorrt

# download the weights
mkdir weights
wget -O weights/yolov3.weights https://pjreddie.com/media/files/yolov3.weights

# requirement for pycuda, please check your own cuda path
export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

# install required libs, `scipy`, `lxml` will take a long time to be installed
pip3 install cython
pip3 install -r requirements.txt

# convert darknet weight into onnx format
python3 yolov3_to_onnx.py

# max_workspace_size這個參數改不改都可以
# edit the following line in the onnx_to_tensorrt.py 
# builder.max_workspace_size = 1 << 28  # ori 30(1GB)->28(256MB)
# to reduce gpu mem usage from 1GB to 256MB

# this script already detect a example image
python3 onnx_to_tensorrt.py

# wrapper of the detector with eyewitness
python3 naive_detector.py --engine yolov3.engine

範例的naive_detector.py有內建偵測 dog.jpg, 結果如下:

# builder.max_workspace_size = 1 << 30
result1:
...Running inference on image demo/test_image.jpg...
=> time: 0.8821
...Running inference on image demo/test_image.jpg... done in 0:00:20.329890


# builder.max_workspace_size = 1 << 28  # ori 30(1GB)
result2:
...Running inference on image demo/test_image.jpg...
=> time: 0.9913
...Running inference on image demo/test_image.jpg... done in 0:00:12.393595

這份python只能跑YOLOv3, 不能跑tiny YOLOv3, 需要找另一份code來用