tflite-micro (TensorFlow Lite for Microcontrollers)¶

• Homepage: https://github.com/tensorflow/tflite-micro
• Manpage: https://www.tensorflow.org/lite/microcontrollers
• Changelog: https://github.com/tensorflow/tflite-micro/commits/main
• Repository: https://github.com/tensorflow/tflite-micro
• Package: master/make/pkgs/tflite-micro/
• Steward: -

Overview¶

TensorFlow Lite for Microcontrollers (TFLM) is Google's purpose-built runtime for machine learning
inference on resource-constrained devices — originally targeting bare-metal microcontrollers, and fully
capable of running on embedded Linux platforms like the FRITZ!Box.

The Freetz-EVO package cross-compiles TFLM into a static library (libtflm.a) for MIPS32, along
with 200+ C++ headers, making the full inference engine available to other packages at link time.
No Python runtime is needed on the device — the library is entirely self-contained.

Supported operations include the complete TFLM kernel set: convolutions (conv2d, depthwise conv,
transpose conv), pooling, LSTM, GRU, fully connected layers, softmax/sigmoid/tanh activations,
reshape, concatenation, dequantize, and more.

Installation¶

Enable the package in make menuconfig under:

Packages → T → tflite-micro [*]

An optional sub-option builds and installs a demo binary:

[*] Build hello_world demo binary

What Gets Installed¶

The library and headers are installed to the staging area (not the target filesystem root) so they
are available to other packages at build time. Only the optional tflm-hello-world binary is deployed
to the device filesystem.

Linking Against libtflm.a¶

To use TFLM from another Freetz package, add the staging include and lib paths and declare a select
dependency:

MY_PACKAGE_CXXFLAGS := -I$(TARGET_TOOLCHAIN_STAGING_DIR)/usr/include/tflite-micro
MY_PACKAGE_CXXFLAGS += -std=c++17 -fno-rtti -fno-exceptions
MY_PACKAGE_LDFLAGS  := -L$(TARGET_TOOLCHAIN_STAGING_DIR)/usr/lib -ltflm

# In your Config.in
select FREETZ_PACKAGE_TFLITE_MICRO

Build Pipeline¶

The build involves three stages:

1. Host tool installation — numpy and Pillow are installed into freetz's own Python 3
   (tools/path/python3) using python3 -m pip install. These are required by the TFLM
   source-tree generator script.

2. Source-tree generation — the official create_tflm_tree.py script is run on the host
   to produce a flat, single-directory source tree (tflm-tree/) containing all required .cc
   and .h files, including third-party dependencies (kissfft, flatbuffers, ruy, gemmlowp).

3. Cross-compilation — every .cc file in the flat tree is compiled to MIPS32 with
   $(TARGET_CC) / $(TARGET_CXX), then archived into libtflm.a with $(TARGET_AR).

Build Notes¶

• Python 3 host tools: freetz has its own Python 3.14 at tools/path/python3 that shadows any
  system Python during the build. numpy and Pillow are automatically installed into this interpreter
  by the .mk recipe (idempotent — subsequent builds skip re-installation if already present).

• array.h workaround: the TFLM generator omits a transitive dependency (array.h) needed by
  tensorflow/lite/micro/span.h; the recipe copies it from the TFLM source tree into the generated
  flat tree automatically.

• Kiss FFT include path: the flat source tree places kissfft under third_party/kissfft; the
  cross-compilation step adds -I$(TFLITE_MICRO_TREE_DIR)/third_party/kissfft for correct header
  resolution.

• Compiler flags: all files are built with -std=c++17 -fno-rtti -fno-exceptions -O2 — the same
  flags required by the upstream TFLM build system. Exceptions and RTTI are deliberately disabled to
  keep code size small and avoid uClibc compatibility issues.

• Static linking only: libtflm.a is a static archive; there is no shared library and no
  lib*.so is installed to the device.

Relevant Files¶