TensorFlow Lite Micro Hello World sample
Overview
This sample TensorFlow application replicates a sine wave and demonstrates the absolute basics of using TensorFlow Lite Micro.
The model included with the sample is trained to replicate a sine function and generates x values to print alongside the y values predicted by the model. The x values iterate from 0 to an approximation of 2π.
The sample also includes a full end-to-end workflow of training a model and converting it for use with TensorFlow Lite Micro for running inference on a microcontroller.
The sample comes in two flavors. One with TensorFlow Lite Micro reference kernels and one with CMSIS-NN optimized kernels.
Note
This README and sample have been modified from the TensorFlow Hello World sample for Zephyr.
Building and Running
The sample should work on most boards since it does not rely on any sensors.
Add the tflite-micro module to your West manifest and pull it:
west config manifest.project-filter -- +tflite-micro
west update
The reference kernel application can be built and executed on QEMU as follows:
west build -b qemu_x86 samples/modules/tflite-micro/hello_world
west build -t run
Exit QEMU by pressing CTRL+A x.
The CMSIS-NN kernel application can be built and executed on any Arm(R) Cortex(R)-M core based platform, for example based on Arm Corstone(TM)-300 software. A reference implementation of Corstone-300 can be downloaded either as a FPGA bitfile for the [MPS3 FPGA prototyping board](https://developer.arm.com/tools-and-software/development-boards/fpga-prototyping-boards/mps3), or as a [Fixed Virtual Platform](https://developer.arm.com/tools-and-software/open-source-software/arm-platforms-software/arm-ecosystem-fvps) that can be emulated on a host machine.
Assuming that the Corstone-300 FVP has been downloaded, installed and added to the PATH variable, then building and testing can be done with following commands.
`
$ west build -p auto -b mps3/an547 samples/modules/tflite-micro/hello_world/ -T sample.tensorflow.helloworld.cmsis_nn
$ FVP_Corstone_SSE-300_Ethos-U55 build/zephyr/zephyr.elf
`
Sample Output
...
x_value: 1.0995567*2^1, y_value: 1.6951603*2^-1
x_value: 1.2566366*2^1, y_value: 1.1527088*2^-1
x_value: 1.4137159*2^1, y_value: 1.1527088*2^-2
x_value: 1.5707957*2^1, y_value: -1.0849024*2^-6
x_value: 1.7278753*2^1, y_value: -1.0509993*2^-2
...
The modified sample prints 10 generated-x-and-predicted-y pairs. To see
the full period of the sine curve, increase the number of loops in main.c.
Modifying Sample for Your Own Project
It is recommended that you copy and modify one of the two TensorFlow
samples when creating your own TensorFlow project. To build with
TensorFlow, you must enable the below Kconfig options in your prj.conf:
CONFIG_CPP=y
CONFIG_REQUIRES_FULL_LIBC=y
CONFIG_TENSORFLOW_LITE_MICRO=y
Note that the CMSIS-NN kernel sample demonstrates how to use CMSIS-NN optimized kernels with TensorFlow Lite Micro, in that is sets below Kconfig option. Note also that this Kconfig option is only set for Arm Cortex-M cores, i.e. option CPU_CORTEX_M is set.
CONFIG_TENSORFLOW_LITE_MICRO_CMSIS_NN_KERNELS=y
Training
Follow the instructions in the train/ directory to train your
own model for use in the sample.