Compiler.md

TFLite Micro Compiler

The TFLM Compiler to build offline models wich can be run without the TFLM Interpreter can be used in the examples.

Further Information

See official repo by cpetig or fork by tum-ei-eda.

Info: Due to some fixes, the submodule in the examples currently used the following extended version of the compiler:

https://github.com/PhilippvK/tflite_micro_compiler

Usage

Detailed Documantation can be found in the external/README.md of each example. F.e. here

Here are the very short summary of the most important steps:

1. Run git submodule update --init --recursive
2. Make sure that the dependencies are loaded via running: download_dependencies.sh
3. Set SET(ENABLE_TFLM_COMPILER ON) after FIND_PACKAGE(TFLite REQUIRED) to enable the feature in the CMakeLists.txt
4. Convert your TFLite model via make convert FILE=model.tflite
5. Try out the generate model under Src/offline_model.cc

Problems

Unfortunately from the start of October 20, due to the changes in listed below, the compiler produces Code which intruduces namespace-issues into the compile process. Until there is a fix available for the compiler, you might have to manually adapt the generated offline model file like explained later.

Related commits:

• https://github.com/tensorflow/tensorflow/commit/c95c69473f32e89b8d4b75f822f459afb97d0860
• https://github.com/tensorflow/tensorflow/commit/b13e49836a92241d653e96225552bd8df3bfd92e and similar

Examples fixes in offline_model.cc:

• Remove replace the prefix tflite::ops::micro:: with ``tflite::` for the following registration methods which have been moved to a flat namespace:
  • Register_CONV_2D
  • Register_QUANTIZE
  • Register_SHAPE
  • Register_SOFTMAX
  • Register_SVDF
  • Register_FULLY_CONNECTED
• Add #include "tensorflow/lite/micro/kernels/fully_connected.h" at the beginning at the file if you need to use the fully_connected kernel.