samhvw8 / daisykit

Toolkit for software engineers to Deploy AI Systems Yourself (DAISY). DaisyKit SDK is the core of models and algorithms, which can be used to develop wrappers and applications for different platforms: mobile, embedded or web browsers.

Daisykit

Daisykit is an AI toolkit for software engineers to Deploy AI Systems Yourself (DAISY). We develop this package with a focus on the ease of deployment. This repository contains:

• Daisykit SDK, the core of models and algorithms.
• Daisykit Python.

Website: https://daisykit.org/.

Demo Video: https://www.youtube.com/watch?v=zKP8sgGoFMc.

Daisykit.Development.Demo.21112021_360p.mp4

Our development plan for Daisykit. We are working to build the whole system gradually.

Environment Setup

For Ubuntu, we need build tools from build-essential package. For Windows, Visual Studio 2019 is recommended.

• Install OpenCV.

Ubuntu:

sudo apt install libopencv-dev

Windows:

Download and extract OpenCV from the official website, and add OpenCV_DIR to path.

• Install Vulkan development package.

Ubuntu:

sudo apt install -y libvulkan-dev vulkan-utils
sudo apt install mesa-vulkan-drivers # For Intel GPU support

• Download precompiled NCNN, extract it (version for your development computer).

Build and Run on PC

• Initialize / Update submodules

git submodule update --init

• Build

Ubuntu:

mkdir build
cd build
cmake .. -Dncnn_FIND_PATH="<path to ncnn lib>"
make

Windows:

mkdir build
cd build
cmake -G "Visual Studio 16 2019" -Dncnn_FIND_PATH="<path to ncnn lib>" ..
cmake --build . --config Release

• Run face detection example

Ubuntu:

./bin/demo_face_detector_graph

Windows:

./bin/Release/demo_face_detector_graph

Coding convention

Read coding convention and contribution guidelines here.

Build documentation

• Step 1: Install doxygen first.

• Step 2: Build the documentation:

cd docs
doxygen Doxyfile.in

• Step 3: Deploy html documentation from docs/_build/html.

• Step 4: Our lastest documentation is deployed at https://docs.daisykit.org.

Known issues and problems

1. Slow model inference - Low FPS

This issue can happen on development build. Add -DCMAKE_BUILD_TYPE=Debug to cmake command and build again. The FPS can be much better.

References

This toolkit is developed on top of other source code. Including

• Toolchains setup from ncnn.
• QR Scanner from ZXing-CPP.
• JSON support from nlohmann/json.
• Pretrained AI models from different sources: https://docs.daisykit.org/md_models.html.