Skip to content

Collection of STM32 projects making use of Tensorflow Lite Micro

License

Notifications You must be signed in to change notification settings

PhilippvK/stm32-tflm-demos

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

STM32 TensorFlow Lite Micro Demos

Collection of STM32 projects making use of Tensorflow Lite Micro. Examples inspired by the TinyML Book.

Toolchain:

CMake-based. See our template: https://github.com/alxhoff/STM3240G-EVAL-TensorFlow-Hello-World

Supported Examples

Supported Boards

Please consult the manufacturer's website for specs and datasheets.

Running

git clone https://github.com/PhilippvK/stm32-tflm-demos

You need to also pull the sub-repos where the actual examples are stored.

git submodule update --init --recursive

Then switch into the folder where the example you want to deploy is stored. Do not worry about the recursive argument, we have only linked our repos, which are essential in this case.

cd examples/hello_world

OR

cd examples/micro_speech

OR

cd examples/mnist

Then:

mkdir build
cd build
cmake -DTF_COMMIT=37c2bf5016fcbed261476386eced503e907cdc01 ..
make
make flash

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!

Compatibility

Due to recent updates of the TF Source tree which have not been implemented for the TFLM Compiler, you have to use Commit 37c2bf5016fcbed261476386eced503e907cdc01 or earlier when cloning the TF repositiory via CMake:

cd build
cmake -DTF_COMMIT=37c2bf5016fcbed261476386eced503e907cdc01 ..

If there are new changes for patched files in the TF Source Tree, they will probably not be merged with the modified file or patches might get overwritten. If something like this occours, please remove the build/_deps/tf-* directories and the CMakeCache.txt and try again!

Documentation