Reference doc¶

[1] : Guide tensorflow lite
[2] : Convertion format NN flat buffer
[3] : Inférence avec Tensorflow Lite
[4] : Tensorflow pour ARM64 linux

[ 96%] Building CXX object CMakeFiles/tensorflowlite_c.dir/c_api_experimental.cc.o
/home/grand/install/xilinx/SDK/2018.3/gnu/aarch64/lin/aarch64-linux/bin/aarch64-linux-gnu-g++ -DPTHREADPOOL_NO_DEPRECATED_API=1 -Dtensorflowlite_c_EXPORTS -I/home/grand/install/tf_lite/tensorflow_src -I/home/grand/install/tf_lite/tensorflow_src/tensorflow/lite/schema -I/home/grand/install/tf_lite/tflite_build/eigen -I/home/grand/install/tf_lite/tflite_build/neon2sse -I/home/grand/install/tf_lite/tflite_build/abseil-cpp -I/home/grand/install/tf_lite/tflite_build/farmhash/src -I/home/grand/install/tf_lite/tflite_build/flatbuffers/include -I/home/grand/install/tf_lite/tflite_build/gemmlowp/public -I/home/grand/install/tf_lite/tflite_build/gemmlowp -I/home/grand/install/tf_lite/tflite_build/ruy -I/home/grand/install/tf_lite/tflite_build/xnnpack/include -I/home/grand/install/tf_lite/tflite_build/pthreadpool-source/include -funsafe-math-optimizations -std=c++11 -fPIC -DTFLITE_BUILD_WITH_XNNPACK_DELEGATE -DTFL_STATIC_LIBRARY_BUILD -pthread -MD -MT CMakeFiles/tensorflowlite_c.dir/c_api_experimental.cc.o -MF CMakeFiles/tensorflowlite_c.dir/c_api_experimental.cc.o.d -o CMakeFiles/tensorflowlite_c.dir/c_api_experimental.cc.o -c /home/grand/install/tf_lite/tensorflow_src/tensorflow/lite/c/c_api_experimental.cc
[100%] Linking CXX shared library libtensorflowlite_c.so
/usr/local/lib/python3.6/dist-packages/cmake/data/bin/cmake -E cmake_link_script CMakeFiles/tensorflowlite_c.dir/link.txt --verbose=1
/home/grand/install/xilinx/SDK/2018.3/gnu/aarch64/lin/aarch64-linux/bin/aarch64-linux-gnu-g++ -fPIC -funsafe-math-optimizations -Wl,--version-script,/home/grand/install/tf_lite/tensorflow_src/tensorflow/lite/c/version_script.lds -shared -Wl,-soname,libtensorflowlite_c.so -o libtensorflowlite_c.so CMakeFiles/tensorflowlite_c.dir/common.c.o CMakeFiles/tensorflowlite_c.dir/c_api.cc.o CMakeFiles/tensorflowlite_c.dir/c_api_experimental.cc.o  tensorflow-lite/libtensorflow-lite.a _deps/abseil-cpp-build/absl/flags/libabsl_flags.a _deps/abseil-cpp-build/absl/flags/libabsl_flags_internal.a _deps/abseil-cpp-build/absl/flags/libabsl_flags_marshalling.a _deps/abseil-cpp-build/absl/flags/libabsl_flags_reflection.a _deps/abseil-cpp-build/absl/flags/libabsl_flags_config.a _deps/abseil-cpp-build/absl/flags/libabsl_flags_program_name.a _deps/abseil-cpp-build/absl/flags/libabsl_flags_private_handle_accessor.a _deps/abseil-cpp-build/absl/flags/libabsl_flags_commandlineflag.a _deps/abseil-cpp-build/absl/flags/libabsl_flags_commandlineflag_internal.a _deps/abseil-cpp-build/absl/container/libabsl_raw_hash_set.a _deps/abseil-cpp-build/absl/container/libabsl_hashtablez_sampler.a _deps/abseil-cpp-build/absl/base/libabsl_exponential_biased.a _deps/abseil-cpp-build/absl/hash/libabsl_hash.a _deps/abseil-cpp-build/absl/hash/libabsl_city.a _deps/abseil-cpp-build/absl/hash/libabsl_wyhash.a _deps/abseil-cpp-build/absl/status/libabsl_status.a _deps/abseil-cpp-build/absl/strings/libabsl_cord.a _deps/abseil-cpp-build/absl/types/libabsl_bad_optional_access.a _deps/abseil-cpp-build/absl/strings/libabsl_str_format_internal.a _deps/abseil-cpp-build/absl/synchronization/libabsl_synchronization.a _deps/abseil-cpp-build/absl/debugging/libabsl_stacktrace.a _deps/abseil-cpp-build/absl/debugging/libabsl_symbolize.a _deps/abseil-cpp-build/absl/debugging/libabsl_debugging_internal.a _deps/abseil-cpp-build/absl/debugging/libabsl_demangle_internal.a _deps/abseil-cpp-build/absl/synchronization/libabsl_graphcycles_internal.a _deps/abseil-cpp-build/absl/base/libabsl_malloc_internal.a _deps/abseil-cpp-build/absl/time/libabsl_time.a _deps/abseil-cpp-build/absl/strings/libabsl_strings.a _deps/abseil-cpp-build/absl/strings/libabsl_strings_internal.a _deps/abseil-cpp-build/absl/base/libabsl_throw_delegate.a _deps/abseil-cpp-build/absl/base/libabsl_base.a _deps/abseil-cpp-build/absl/base/libabsl_spinlock_wait.a -lrt _deps/abseil-cpp-build/absl/numeric/libabsl_int128.a _deps/abseil-cpp-build/absl/time/libabsl_civil_time.a _deps/abseil-cpp-build/absl/time/libabsl_time_zone.a _deps/abseil-cpp-build/absl/types/libabsl_bad_variant_access.a _deps/abseil-cpp-build/absl/base/libabsl_raw_logging_internal.a _deps/abseil-cpp-build/absl/base/libabsl_log_severity.a _deps/farmhash-build/libfarmhash.a _deps/fft2d-build/libfft2d_fftsg2d.a _deps/fft2d-build/libfft2d_fftsg.a _deps/flatbuffers-build/libflatbuffers.a _deps/ruy-build/libruy.a -ldl _deps/xnnpack-build/libXNNPACK.a /home/grand/install/xilinx/SDK/2018.3/gnu/aarch64/lin/aarch64-linux/aarch64-linux-gnu/libc/usr/lib/libm.so cpuinfo/libcpuinfo.a clog/libclog.a -pthread pthreadpool/libpthreadpool.a -lpthread 
make[2]: Leaving directory '/home/grand/install/tf_lite/tflite_build'
[100%] Built target tensorflowlite_c
make[1]: Leaving directory '/home/grand/install/tf_lite/tflite_build'
/usr/local/lib/python3.6/dist-packages/cmake/data/bin/cmake -E cmake_progress_start /home/grand/install/tf_lite/tflite_build/CMakeFiles 0

grand@xilinx2:~/install/tf_lite/tflite_build$ ll

avec cette page https://www.tensorflow.org/lite/guide/build_cmake_arm?hl=fr#build_for_aarch64_arm64

. ~/install/xilinx/SDK/2018.3/settings64.sh
 1732  ARMCC_PREFIX=/home/grand/install/xilinx/SDK/2018.3/gnu/aarch64/lin/aarch64-linux/bin/aarch64-linux-gnu-
 1733  ARMCC_FLAGS="-funsafe-math-optimizations" 
 1736  /usr/local/bin/cmake -DCMAKE_C_COMPILER=${ARMCC_PREFIX}gcc -DCMAKE_CXX_COMPILER=${ARMCC_PREFIX}g++ -DCMAKE_C_FLAGS="${ARMCC_FLAGS}" -DCMAKE_CXX_FLAGS="${ARMCC_FLAGS}" -DCMAKE_VERBOSE_MAKEFILE:BOOL=ON -DCMAKE_SYSTEM_NAME=Linux -DCMAKE_SYSTEM_PROCESSOR=aarch64  ../tensorflow_src/tensorflow/lite/c
 1737  /usr/local/bin/cmake --build . -j 2

3) conversion modele Keras

tflite_convert --keras_model_file=trigger_icrc_80.keras --output_file=trigger_icrc_80.tflite

Tensorflow Lite sur CPU ARM

1) Compilation de la bibliothèque tensorflow lite pour ARM xilinx avec cmake

pour clone le package tensotflow lite

https://www.tensorflow.org/lite/guide/build_cmake?hl=fr#step_1_install_cmake_tool

Compilation sur ARM

https://www.tensorflow.org/lite/guide/build_cmake_arm?hl=fr#build_for_aarch64_arm64

mon adaption à Xilinx avec un cmake plus récent que celui de Xilinx installé ici /usr/local/bin/cmake

source ~/install/xilinx/SDK/2018.3/settings64.sh
1732 ARMCC_PREFIX=/home/grand/install/xilinx/SDK/2018.3/gnu/aarch64/lin/aarch64-linux/bin/aarch64-linux-gnu-
1733 ARMCC_FLAGS="-funsafe-math-optimizations"
1736 /usr/local/bin/cmake -DCMAKE_C_COMPILER=${ARMCC_PREFIX}gcc -DCMAKE_CXX_COMPILER=${ARMCC_PREFIX}g++ -DCMAKE_C_FLAGS="${ARMCC_FLAGS}" -DCMAKE_CXX_FLAGS="${ARMCC_FLAGS}" -DCMAKE_VERBOSE_MAKEFILE:BOOL=ON -DCMAKE_SYSTEM_NAME=Linux -DCMAKE_SYSTEM_PROCESSOR=aarch64 ../tensorflow_src/tensorflow/lite/c
1737 /usr/local/bin/cmake --build . -j 2

2) Code d'inférence en C

L'inférence de réseau de neurone est juste la réponse du réseau à une entrée, à ce moment l’entraînement du réseau est terminé.

Démarche général:

https://www.tensorflow.org/lite/guide/inference?hl=fr

La description du réseau est dans un format spécifique à tensorflow lite, une "moulinette" de conversion existe entre tensorflow et keras.

Pour GRAND nous avons utilisé Keras et c'est

https://www.tensorflow.org/lite/convert/index?hl=fr

tflite_convert --keras_model_file=trigger_icrc_80.keras --output_file=trigger_icrc_80.tflite

Ensuite il faut faire le code d'inférence de C , la doc donne cette exemple

https://www.tensorflow.org/lite/guide/inference?hl=fr#using_c_api_in_objective-c_code

voici mon adapation pour le projet GRAND

https://github.com/grand-mother/grand-daq-master/blob/815e9b8a44be24ed2fb50541b455753a53a508a9/dudaq_1/proto_tf/main.c#L522C4-L522C4

et ici le "makefile" qui compile et lie avec la lib tensorflow lite (libtensorflowlite_c.so) avec l'outil python SCons

https://github.com/grand-mother/grand-daq-master/blob/dev_tensorflow/dudaq_1/proto_tf/SConstruct.py