kruglov-dmitry / tensorflow_lite_from_cpp_python

demo code to reproduce the issue of using XNNPACK delegate in tensorflow lite in c++ and python

What is this about

Minimum example to reproduce issue when the same network at the same input produce different result using different kind of tensorflow lite library - build with and without XNNPACK support:

cd <tensorflow_source_tree>/tensorflow/lite/

Case 1: build without xnnpack support

mkdir tf_build_no_xnnpack/
cd tf_build_no_xnnpack/
cmake -D TFLITE_ENABLE_XNNPACK=OFF ..
make -j 8

Case 2: build with xnnpack support

mkdir tf_build/
cd tf_build/
cmake ..
make -j 8

Cpp

1. Edit Makefile to specify include and lib paths to include for opencv and tensorflow:

• TF_SRC_PATH
• TF_LITE_BUILD_FOLDER
• OPENCV_INCLUDE_PATH
• OPENCV_LIB_PATH

NOTE: list of libs for linking is different for convinience in Makefile present commented version for list of libs without XNNPACK

2. build

make tf_load_sample

3. run

tf_load_sample

Python

1. create venv:

make venv

2. run python script

source VENV/bin/activate && cd python &&
python tf_load_sample.py

Network and conversion

Original keras network definition in python/network.py

for conversion python interface was used:

md.load_weights("FER.h5")
converter = tf.lite.TFLiteConverter.from_keras_model(md)
converter.experimental_new_converter = True
onverter.optimizations = [tf.lite.Optimize.DEFAULT]

converter.target_spec.supported_types = [tf.float32]
quantized_tflite_model = converter.convert()
open("FER.tflite", "wb").write(quantized_tflite_model)