Step 0: System Check. (Usually the staff of the company will do it for you, so you can jump to Step 1.)

First, check whether the Ubuntu system is installed in the solid state drive, if installed in the mechanical disk, please reinstall the system. After that, turn off the Secure Boot entry in the system BIOS.

Solve the problem that Sunflower cannot connect after restart:

sudo apt-get update
sudo apt-get upgrade
sudo apt-get install lightdm

Step 1: NVIDIA Graph Device Driver Install

Change the system source to Aliyun source: Click the Software & Updates icon, select the Download from item in the interface as http://mirrors.aliyun.com/ubuntu, click Close, and click Reload.

Click the Software & Updates icon again, select Additional Drivers, and click to download the 450, 460, and 470 NVIDIA graphics card drivers. After the installation is successful, restart the host. After restarting, click the Setting icon to see if the Graphics item has changed to the NVIDIA driver, and type in the terminal:

nvidia-smi

Check whether the graphics card usage information can be viewed. If all are satisfied, the graphics card driver is installed successfully.

Step 2: Anaconda Install

Download the latest installation package from the Anaconda official website and run the command in the terminal:

bash {Anconda Package Name}.sh

Select yes in the whole process, restart the terminal, if the line header appears (base), the installation is successful.

Step 3: Pycharm Install

Download the latest version from the JetBrain official website, copy it to the directory where your home is located, and run the command in the terminal to decompress it:

sudo tar zxvf {Pycharm Package Name}.tar.gz
cd {Unzipped dictory name}/bin
sh ./pycharm.sh

Remember to log in to pycharm with your student account to activate the product.

Step 4: CUDA and cuDNN Install（Please ensure that Step 1 is completed properly.）

Download the CUDA-11.0 version from the CUDA archive on the NVIDIA official website and install it according to the official requirements.

Download the latest version of cuDNN corresponding to CUDA-11.0 from the NVIDIA official website, and enter:

tar zxvf {cuDNN Package Name}.tar.gz
sudo cp cuda/include/cudnn* /usr/local/cuda-11.0/include/
sudo cp cuda/lib64/libcudnn* /usr/local/cuda-11.0/lib64/
sudo chmod a+r /usr/local/cuda-11.0/include/cudnn*
sudo chmod a+r /usr/local/cuda-11.0/lib64/libcudnn*

Add environment variables, terminal input:

sudo gedit ~/.bashrc

Add text in the blanks of the document:

# CUDA config
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-11.0/lib64
export PATH=$PATH:/usr/local/cuda-11.0/bin
export CUDA_HOME=$CUDA_HOME:/usr/local/cuda-11.0

Step 5: Git Install

Terminal input：

sudo apt install git

Step 6: Replace the source of conda

Terminal input：

conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
conda config --set show_channel_urls yes

Step 7: Replace the source of pip

Create a .pip folder in the home directory, and create a pip.conf file in the folder with the following contents:

[global]
timeout = 6000
index-url = https://pypi.tuna.tsinghua.edu.cn/simple
trusted-host = pypi.tuna.tsinghua.edu.cn

Step 1: Source code clone

You can directly copy the PCIA-local folder from the code folder under this repo as the source code folder.

Or clone from my other repo (not yet open to clone):

git clone https://github.com/JoshuaWenHIT/PCIA-local.git

Step 2: Operating environment configuration

Relevant library installation:

pip install -r requirements.txt

Deep learning framework installation (recommended to use the version I specified):

conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 cudnn=7.6.5

Push stream library installation (recommended to use the version I specified):

conda install ffmpeg==4.3.1

Step 3: DCNv2 compilation (can be omitted, try to run this step when demo.py runs incorrectly)

Terminal input:

cd DCNv2
export CUDA_HOME=/usr/local/cuda-11.0
python setup.py build develop

If there are no problems with the previous steps, this step will naturally succeed. If it fails, check the steps from the beginning for any inconsistencies.

The demo.py file uses test2.mp4 as the test video by default, which can be run directly:

python demo.py

The test result will generate the test2_res.mp4 result video in the ./video folder. If the whole calculation process is normal and the result video is normal, the demo test will pass.

It is worth mentioning that the demo test is mainly used to test whether the deployment of the counting network is successful, and does not involve tasks such as communication, streaming, and cloud uploading in the PCIA product part. This part of the test is mainly completed in the main test.

The main test can be divided into three steps: 1) code authorization; 2) code startup; 3) joint debugging.

Step 1: Code authorization

This part involves the code encryption part in ./lib/model.py. The content of the encryption part will be explained in detail in the "How to encrypt PCIA?" section. Here, only the generation of the license.lic file used for running is explained.

In the ./code folder, the encode folder represents the encryption function file. Before generating the license.lic file, you need to know the mac address of the deployment host:

ifconfig

Take the above figure as an example, three sets of parameters are obtained respectively, among which enp0s31f6 is the model of the network card in use, and the model will vary according to different devices.

Please pay attention to your device network card model! ! ! And modify it in line87 of ./lib/model.py in PCIA-local folder! ! !

The mac address is represented after ether: b0:6e:bf:60:2a:24, after obtaining the mac address, enter it in the terminal (in the encode root directory):

python CreateLicense.py $your_macAddress-$month-$day # e.g. b06ebf602a24-05-01

Among them, $your_macAddress fills in the mac address of the deployed device, and $month-$day fills in the authorization expiration date of the license.lic file (if you want permanent authorization, you can fill in 13-01, which is a month that does not exist in normal months). After running, you can obtain the license.lic file in the encode root directory, and copy the file to the PCIA-local root directory to complete the authorization.

Step 2: Program start

After completing the authorization, modify it in main.py: 1) modify the mac address of line18; 2) modify the camera address of line20. When done, type in the terminal:

python main.py

If it runs normally, you can observe the following output:

已授权!
生成模型...
加载模型 ./weights/pcia_v7_b

模型加载完毕!
等待接收服务器指令......

Step 3: Joint debugging

It is necessary to contact the company personnel for cooperative testing, and the company staff will use the tablet to control the start and end of the algorithm, which requires practical operation drills.

It is necessary to ensure that there are legally named videos under ./video/output. It is recommended to perform joint debugging in step 3 of the main test! ! !

Terminal input:

python upload.py

If the terminal shows that the upload is successful, it means that the upload test has passed.

Code encryption mainly encrypts the .py file into a .so library file through the Cython method, and runs it on the terminal. Copy the .py file to be encrypted to the process folder under encode/encryptionforcode, and run the terminal:

python encryptcode.py build_ext --inplace

You can obtain the encrypted .so file in the process folder and replace it with the .py file in the ./lib of PCIA-local to complete the code encryption.

Model encryption mainly encrypts the .pth file into a weight file without a suffix name through the password hash check method. The encrypted file already exists in the ./weights folder in PCIA-local. Copy the .pth file to be encrypted to the encode/encryptionformodel/encryption/temp folder, Modify line15 in encode/encryptionformodel/encryption/main.py to customize the input weight file name and output file path, Finally the terminal runs:

python main.py # Note that the main.py file path here is encode/encryptionformodel/encryption/main.py, please do not confuse it with main.py under PCIA-local

You can obtain the encrypted model weight file in the temp folder, and replace the .pth file in ./weights of PCIA-local to complete the model encryption.