Skip to content

Latest commit

 

History

History
216 lines (124 loc) · 9.77 KB

RT-AK Quick Start EN.md

File metadata and controls

216 lines (124 loc) · 9.77 KB
Raw

OpenSource AI Kit: One Click to Deploy AI Models on RT-Thread

0. RT-AK Intro

RT-AK is an AI Kit developed by the RT-Thread Team, committed to help developers one-click deploy the AI models, allowing developers to develop business code on top of a unified API and ultimately optimize performance on the target platform, making it easier and simple to develop end-side AI applications.

With the RT-AK support, taking one line of command of python aitools.py --model xxx... can deploy the AI model to the RT-Thread embedded system:

Commands List: How to run aitools.py

GitHub Repository: https://github.com/RT-Thread/RT-AK

We are striving to minimize the difficulties and thresholds for embedded AI deployment!

1. Preparation

  • Windows 10
  • Python >= 3.7

Required Documents:

Index Documents Example
1 Hardware &BSP ART-PI BSP
2 Neural network model ./rt_ai_tools/Model/keras_mnist.h5
3 STM32 AI Plug-in X-CUBE-AI download & decompress, will be described below
4 RT-AK RT-AK code is cloned locally

1.1 X-CUBE-AI Intro

X-CUBE-AI is a part of STM32Cube Expansion Package in STM32Cube.AI ecosystem and extending STM32CubeMX capabilities with automatic conversion of pre-trained Neural Network and integration of generated optimized library into the user's project.

image-20210525153037964

X-CUBE-AI

image-20210525153057003

Download

1.2 X-CUBE-AI Decompress

Downloaded File Folder

Includes:

  • stm32ai-windows-5.2.0.zip has stored the X-CUBE-AI model conversion software: stm32ai. This file is required.
  • STMxxx.pack is a static library file of STM32Cube.AI , it doesn't need to decompress, and located in ./RT-AK/rt_ai_tools/platforms/stm32/X-CUBE-AI.5.2.0

Unzip stm32ai-windows-5.2.0.zip file

toD:\Program Files (x86)\stm32ai-windows-5.2.0 , so you can get a windows folder under this path.

Keep in mind about this path. HIGH ATTENTION!

STM32: X-CUBE-AI unzip path

1.3 ART-PI BSP

  1. Open RT-Thread Studio

  2. Create ART-PI BSP

Create ART-PI Project in RT-Thread Studio

2. Execute

Step1 Run rt_ai_tools/aitools.py

Code will automatically use the STM32Cube.AI model conversion tool to obtain a BSP with AI integrated. YES. Aha Moment!

Internal detailed processes please refer to the source code or the readme documentation under the plugin_stm32 warehouse.

Run Command

Enter edge-ai/RTAK/tools path, run aitools.py

image-20210528112918056

# Run Command
python aitools.py --project=<your_project_path> --model=<your_model_path> --platform=stm32 --ext_tools=<your_x-cube-ai_path> --clear

# Sample
python aitools.py --project="D:\RT-ThreadStudio\workspace\test" --model="./Models/keras_mnist.h5" --platform=stm32 --ext_tools="D:\Program Files (x86)\stm32ai-windows-5.2.0\windows" --clear

Works done so far, some more additional instructions to supplement are shown following

image-20210414174056329

Other Parameters for Run Command

# Specify the name of the conversion model,--model_name defaults to network
python aitools.py --project=<your_project_path> --model=<your_model_path>  --model_name=<model_name>  --platform=stm32 --ext_tools=<your_x-cube-ai_path>

# Save files generated during running the stm32ai thread,-- clear defaults to empty

# If exited, the work folder generated during running 'stm32ai', i.e. '--stm_out, will be deleted

python aitools.py --project=<your_project_path> --model=<your_model_path> --platform=stm32 --ext_tools=<your_x-cube-ai_path>

# Specify to save the run log, --log defaults to empty

python aitools.py --project=<your_project_path> --model=<your_model_path> --log=./log.log --platform=stm32 --ext_tools=<your_x-cube-ai_path>

# Specify the name of the saved folder,--stm_out defaults to the time of day, such as './20210223'

python aitools.py --project=<your_project_path> --model=<your_model_path> --platform=stm32 --ext_tools=<your_x-cube-ai_path> --stm_out <new_dir>

# Specify the generated c-model name ,--c_model_name defaults to network

python aitools.py --project=<your_project_path> --model=<your_model_path> --platform=stm32 --ext_tools=<your_x-cube-ai_path> --c_model_name=<new_model_name>

Run Parameters Specs

  • Main functions parameters
Parameter Description
--log ./log.log The log stores the path, which defaults to empty, and if so, the log file is saved, e.g. ./log.log
--project OS+BSP project engineering folder, empty, user-specified
--model Neural network model file path, defaults to ./Models/keras_mnist.h5
--model_name New model name after neural network model conversion, defaults to network
--rt_ai_lib RT-AK Lib provided by RT-Thread defaults to ../rt_ai_lib
--platform Specified hardware platform information, currently supported: stm32, k210, defaults to example
  • STM32 platform plug-in parameters
Parameter Description
--ext_tools X-CUBE-AI storage path, model conversion tool, stm32ai executable software inside, requiring user-specified
--cube_ai X-CUBE-AI runs the required static library, defaults to./platforms/stm32/X-CUBE-AI.5.2.0
--rt_ai_example Storert_ai_<model_name>_model.c sample file, defaults to./platforms/stm32/docs
--stm_out The intermediate folder path that results from the stm32ai thread processing is named by default for the timestamp of the day
--workspace Temporary workspace generated during running stm32ai, defaults to ./stm32ai_ws
--val_data Defaults to empty, allows users to customize the test dataset even when using an internally self-generated random dataset.
--compress Represents the global compressibility factor that will be applied, applied only to the full connection layer, with the option of "1|4|8", defaults to 1
--batches Indicates how many random data samples have been generated, with the default of 10
--mode "analyze|validate" mode (optional) plus "generate" mode (must have), 1 indicates selected, select one from {'001', '011', '101', '111'}, defaults to 001
--network The model name of the template file in Documents, defaults to mnist.
--enable_rt_lib Open macro definitions in project/rtconfgi.h and default to RT_AI_USE_CUBE
--clear If need to delete the intermediate folder stm_out that generated in stm32ai, defaults to False

Step2 Compile and Burn

This tutorial shows only 'RT-Thread Studio' compilation method, but RT-AK also supports:

  • Keil
  • Scons based on RT-Thread Env
  • Scons not based on RT-Thread Env

RT-Thread Studio Compile and Burn

In RT-Thread Studio, find the project engineering and right-click

Update packages and refresh project,

Then compile（）

Finally, burn and display.

编译成功

Now, you're successfully get a new ART-Pi BSP that integrates AI and RT-Thread.

AI-related application development can be easily processed on RT-Thread.

Sample Application Code

You may reach to Here for the sample code to run model reasoning mnist_app.c:

  1. Download and decompress and place it under the <BSP>/applications path

image-20210409192729527

  1. Select project engineering in RT-Thread Studio and right-click to refresh

image-20210409193019579

  1. Compile and burn.

  2. Enter the command: mnsit_app

image-20210409192638719

3. GitHub Address

https://github.com/EdgeAIWithRTT/Project3-Mnist_Cube_RTT/tree/master/Mnist_RTT