All settings are ready in the project file. If the project file can be opened normally, it does not need to be modified. Please open the Deployment/MDK-ARM/deploy2.uvprojx in Keil. While ensuring proper hardware connectivity, build and download the project.
In case you want to replicate our work，please follow the setting below: For CubeMX Settings
1. Project Manager -> Advanced Settings: set "RCC","GPIO","CRC" in LL, keep "USART" in HAL.
2. Pinout -> System Core -> NVIC -> Code generation: only keep "Time base: System tick timer"
For Keil Compiling Settings
1.Option for Target->Target->Code generation：Use default compiler version 6 & Use Micro LIB;
2.Option for Target->C/C++(AC6)->Optimization:-Oz
3.Option for Target->C/C++(AC6)->Click on "One ELF Section per Function" & "Execute-only Code" & "Short enums/wchar"
Note: We did not use X-CUBE-AI due to its large memory overhead. In addition to the necessary/default files for the STM32 system, we also have several files that need to be included during compilation.
matrix.c / .h
aiRun.c / .h
parameter.c / .h
Our method extracts some features and inputs them into a three-layer FC.
- matrix.c / .h: Defines functions for matrix data format, calculating matrix multiplication, matrix addition, and Relu.
- parameter.c / .h: Stores FC parameters and hyperparameters when calculating features.
- aiRun.c / .h :
- Model_Init(): We deconvert our homemade float16 data to float32 and convert them into matrix data.
- aiRun(): We do feature extraction and matrix calculation here. weight_genenerate/to_float16.c file should be compiled and run on PC. This file manually sets the model's original float weights and defines a set of structures that compress two float16 data into a uint32 for the weights in parameter.c.