Running a Tensorflow Lite model on a Teensy microcontroller

Compiling

Basic instructions for building tflite from source can be found here

For Teensy these are the steps that I needed to take to compile. These intructions worked from the main branch of tflite-micro in July 2023. tflite doesn’t seem to have versioned releases so this may be different in the future.

Clone the tflite-micro repository in github
As in the instructions above run the python script to create a directory tree containing the sources necessary to build:
```
 python tensorflow/lite/micro/tools/project_generation/create_tflm_tree.py tmp/tflm-tree
```
The created tree can be copied into the src folder of a platformio project to build it.
You will need to update the default tool chain to a version which includes std::to_string by adding the following line to platformio.ini:
```
 platform_packages = toolchain-gccarmnoneeabi @ 1.90201.191206
```
You will then need to provide an updated math .a library file from here. This can be copied into /lib folder in the project and will need adding to the build flags in platofromio.ini:
```
 build_flags = -L ./lib
```
Add the following build flag to platformio.ini:
```
 build_flags = -D ARDUINOSTL_M_H
```

tflite/tensorflow/lite/micro/kernels/lstm_eval_common.cc

Remove the double cast from line 192:

 cell_state_info.quantized_cell_clip = static_cast<int16_t>(
     std::min(std::max(cell_clip / cell_state_scale, -32768.0), 32767.0));

tflite/third_party/gemmlowp/fixedpoint/fixedpoint.h

Add a double cast on line 519+:

     std::max(static_cast<double>(round(x * static_cast<double>(1ll << kFractionalBits))),
              min_bound),
     max_bound)));

tflite/tensorflow/lite/kernels/cppmath.h

Declare TfLiteRound directly instead of with macro expansion:

 // DECLARE_STD_GLOBAL_SWITCH1(TfLiteRound, round);
 template <class T>                                
 inline T TfLiteRound(const T x) {                    
     return round(x); 
 }

tflite/tensorflow/lite/micro/debug_log.cc

Import Arduino.h and change fprintf to Serial.println:
```
 #import <Arduino.h>
```
```
 Serial.println(s);
```
It should now be possible to compile the source without errors. I’ve done this and checked the resulting code into the xen_machinelearning repository.

This repository can be used as a library in other projects by adding it as a dependency in platformio.ini:
```
lib_deps = https://github.com/pigatron-industries/xen_machinelearning.git
```
and also the following build flag:
```
build_flags = -D ARDUINOSTL_M_H
```