Skip to main content

Development Environment

1. RKNN Development Environment​

Image

On the PC side, the main tasks are model training and model conversion. You can choose:

  • Windows system
  • Ubuntu on a Windows virtual machine
  • Docker Linux system
  • Cloud servers, etc.

The PC side requires the installation of:

  • Common software and libraries (e.g., PyCharm, Python, cross-compilers)
  • Deep learning frameworks (PyTorch, TensorFlow, PaddlePaddle)
  • It is recommended to use a virtual environment (Python virtual environment/Anaconda/Miniconda)

Board environment:

  • System: Debian
  • Pre-installed components: RKNN drivers and other related components
  • Common software: Python, CMake, Make, GCC, OpenCV, etc.

Testing environment:

  • PC side: WSL2 (used with PyCharm)
  • Board system: Debian12

2. RKNN Development Process​

Process

Main steps:

  1. Model Training

  2. Model Conversion

    • Convert the model to RKNN format

  3. Model Evaluation

    • Use RKNN-Toolkit2 for quantization and performance analysis
    • Refer to the RKNPU User Guide

  4. On-Board Inference

3.1. Anaconda Installation​

Installation Steps:

wget https://repo.anaconda.com/archive/Anaconda3-2023.07-2-Linux-x86_64.sh
bash Anaconda3-2023.07-2-Linux-x86_64.sh
source ~/.bashrc

Common Commands:

conda create -n env_name python=3.8   # Create environment
conda activate env_name               # Activate environment
conda deactivate                      # Deactivate environment
conda config --set auto_activate_base false  # Disable auto-activation

Mirror Configuration:

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

3.2. RKNN-Toolkit2 Installation​

Installation Steps:

conda create -n toolkit2_1.6 python=3.8
conda activate toolkit2_1.6
git clone https://github.com/airockchip/rknn-toolkit2
pip3 install -r packages/requirements_cp38-1.6.0.txt
pip3 install packages/rknn_toolkit2-1.6.0+81f21f4d-cp38-cp38-linux_x86_64.whl

Verify Installation:

from rknn.api import RKNN
rknn = RKNN()  # Success if no errors occur

3.3. Jupyter Notebook Installation​

Installation Method:

conda install jupyter notebook   # Using conda
# Or
pip3 install jupyter -i https://pypi.tuna.tsinghua.edu.cn/simple

3.4. Deep Learning Framework Installation​

PaddlePaddle Installation:

# CPU Version
conda install paddlepaddle==2.5.1 --channel https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/Paddle/

# GPU Version (requires CUDA)
conda install paddlepaddle-gpu==2.5.1 cudatoolkit=11.2 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/Paddle/

TensorFlow Installation:

# Create virtual environment
python3 -m venv .tensorflow_venv
source .tensorflow_venv/bin/activate
pip3 install tensorflow

PyTorch Installation:

conda create -n pytorch python=3.8
conda activate pytorch
pip3 install torch torchvision torchaudio
  1. CUDA Toolkit Archive
  2. cuDNN Archive
  3. Anaconda Mirror Configuration
  4. PaddlePaddle Conda Installation Guide
  5. TensorFlow Installation Guide