This demo implements a simple handwriting recognizer, which can transcribe digits from 0 to 9 into characters. It uses a CNN architecture and can display the percent of confidence the CNN delivers for each of the 10 possible outputs.
Moreover, the input can be taken either from the touchscreen, enabling us to draw the digits ourselves, or from images showing digits loaded from the memory card.
You may also be interested in our wrapper repository stm32-tflm-demos for further details.
First configure your board and features like CMSIS-NN, Benchmarking,... as explained here.
Then you have to choose whether you want to use the board's touchscreen to draw and recognize digits real-time, or load pre-recorded samples from the SD card. You make this choice by setting FAKE_TOUCH in CMakeLists.txt to either ON or OFF.
SET(FAKE_TOUCH OFF) # Use on board touchscreen
SET(FAKE_TOUCH ON) # Use samples from the SD card
By default, we are using the board's built-in touchscreen. For actual samples and the required file structure, see the media folder.
Finally, simply run:
mkdir build
cd build
cmake -DTF_COMMIT=37c2bf5016fcbed261476386eced503e907cdc01 ..
make
make flash
For more details about build instruction, see the main README.
Other make targets we provide include make debug and make convert (to convert the *.tflite file into tfite_micro_compiler sources).
Fetching the dependencies (Tensorflow, STM32 Firmware) for the first time can take a very long time. You can either take a break or do the following to speed-up the process when using multiple examples:
export SOME_DIRECTORY=<CHOOSE_SOME_PATH>
git clone --recursive https://github.com/tensorflow/tensorflow.git $SOME_DIRECTORY/tensorflow
git clone https://github.com/mborgerding/kissfft.git $SOME_DIRECTORY/kissfft
git clone https://github.com/STMicroelectronics/STM32CubeF4 $SOME_DIRECTORY/STM32CubeF4
git clone https://github.com/STMicroelectronics/STM32CubeF7 $SOME_DIRECTORY/STM32CubeF7
git clone https://github.com/ARM-software/CMSIS_5.git $SOME_DIRECTORY/CMSIS_5
$SOME_DIRECTORY/tensorflow/tensorflow/lite/tools/make/download_dependencies.sh
After doing this a single time, just replace cmake .. with the following command in the future:
cmake -DTF_SRC=$SOME_DIRECTORY/tensorflow -DSTM32Cube_DIR=$SOME_DIRECTORY/STM32CubeF7 -DARM_CMSIS_SRC=$SOME_DIRECTORY/CMSIS_5 -DKISSFFT_SRC=$SOME_DIRECTORY/kissfft ..
Tipp: Swap out STM32CubeF7 with STM32CubeF4 when building for the F413ZH board!
It is mandatory to delete CMakeCache.txt after changing the STM32 board. Sometimes, you need to delete the _deps folder under build before running cmake .. again.