TensorFlow 2.16.2 C++ library build

2. 手順
1) tensorflow 2.16.2 source tar.gz をダウンロードする。
https://github.com/tensorflow/tensorflow/releases から、 2.16.2 の tar.gz をダウンロード。
$ cd
$ mkdir -p local/tensorflow
$ cd local/tensorflow
ここで、tensorflow-2.16.2.tar.gz を展開する。

2) bazel でのビルド。
virtual env 上で行う。
$ cd
$ source kivy_env/bin/activate <-- 自分の環境を使う。

$cd ~/local/tensorflow/tensorflow-2.16.2
$ export TF_PYTHON_VERSION=3.10

i) config
$ ./configure
Please specify the location of python. [Default is /home/nishi/kivy_env/bin/python3]: Please input the desired Python library path to use. Default is [/home/nishi/colcon_ws/build/camera_calibration/src] /home/nishi/kivy_env/bin/python3/site-packages Do you wish to build TensorFlow with ROCm support? [y/N]: N Do you wish to build TensorFlow with CUDA support? [y/N]: N Do you want to use Clang to build TensorFlow? [Y/n]: Y Please specify the path to clang executable. [Default is /usr/bin/clang]: Please specify optimization flags to use during compilation when bazel option "--config=opt" is specified [Default is -Wno-sign-compare]: Would you like to interactively configure ./WORKSPACE for Android builds? [y/N]: N Preconfigured Bazel build configs. You can use any of the below by adding "--config=<>" to your build command. See .bazelrc for more details. --config=mkl # Build with MKL support. --config=mkl_aarch64 # Build with oneDNN and Compute Library for the Arm Architecture (ACL). --config=monolithic # Config for mostly static monolithic build. --config=numa # Build with NUMA support. --config=dynamic_kernels # (Experimental) Build kernels into separate shared objects. --config=v1 # Build with TensorFlow 1 API instead of TF 2 API. Preconfigured Bazel build configs to DISABLE default on features: --config=nogcp # Disable GCP support. --config=nonccl # Disable NVIDIA NCCL support. Configuration finished end

ii) start build
$ bazel build --verbose_failures --config=opt --copt=-march=native //tensorflow:libtensorflow_cc.so //tensorflow:libtensorflow_framework.so //tensorflow:install_headers

すべての出力は、下記ディレクトリーにできる。
/home/your-id/local/tensorflow/tensorflow-2.16.2/bazel-out/k8-opt/bin/tensorflow
$ls ~/local/tensorflow/tensorflow-2.16.2/bazel-out/k8-opt/bin/tensorflow c libtensorflow_cc.so.2.16.2 cc libtensorflow_cc.so.2.16.2-2.params compiler libtensorflow_framework.so core libtensorflow_framework.so.2 distribute libtensorflow_framework.so.2.16.2 dtensor libtensorflow_framework.so.2.16.2-2.params include lite install_headers.genrule_script.sh python libtensorflow_cc.so tools libtensorflow_cc.so.2

iii) あとは、必要なファイルをコピーするだけ。
copy shell を作った。
~/local/tensorflow/deploy-nishi2.sh
#!/bin/bash -eu # -*- coding: utf-8 -*- # Referece: # [copy_tensorflow_headers.sh](https://gist.github.com/saitodev/3cde48806a32272962899693700d9669) # $ sudo ./deploy-nishi2.sh VER=2.16.2 #DIST_DIR=/usr/local DIST_DIR=/home/nishi/usr/local HEADER_DIR=${DIST_DIR}/include/tensorflow-${VER} LIB_DIR=${DIST_DIR}/lib/tensorflow-${VER} SRC_DIR=./tensorflow-${VER} if [ ! -e $HEADER_DIR ]; then mkdir -p $HEADER_DIR fi if [ ! -e $LIB_DIR ]; then mkdir -p $LIB_DIR fi pushd ${SRC_DIR} pushd bazel-out/k8-opt/bin/tensorflow cp -af libtensorflow_cc.so ${LIB_DIR} cp -af libtensorflow_cc.so.2 ${LIB_DIR} cp -af libtensorflow_cc.so.2.16.2 ${LIB_DIR} cp -af libtensorflow_framework.so ${LIB_DIR} cp -af libtensorflow_framework.so.2 ${LIB_DIR} cp -af libtensorflow_framework.so.2.16.2 ${LIB_DIR} pushd include cp -arf * $HEADER_DIR popd popd # use system header /usr/include/google # $ sudo apt installe libprotobuf-dev # but version unmatch ? # もしかしたら、*.h,*.inc だけでなく protobuf.o もコピーして、リンカーでくっつけるべきか? pushd bazel-tensorflow-${VER}/external/com_google_protobuf/src find google -follow -type f -name "*.h" -exec cp --parents {} $HEADER_DIR \; find google -follow -type f -name "*.inc" -exec cp --parents {} $HEADER_DIR \; popd popd

$ chmod 755 deploy-nishi2.sh
$ sudo ./deploy-nishi2.sh
これで、一括コピーされる。

3. テストプログラムのビルドとRun
テストプログラムも、拝借させてもらった。
example.cpp
/* example.cpp /home/nishi/usr/local/src/cpp-nishi/nishi-many-test/tf-cpp/sample3/example.cpp https://toxweblog.toxbe.com/2017/11/10/tensorflow-cppapi-compile-by-g/ # export CPATH=$CPATH:/home/nishi/usr/local/include/tensorflow-2.16.2 # export LIBRARY_PATH=$LIBRARY_PATH:/home/nishi/usr/local/share/libtensorflow-cpu-2.16.2/lib 1) lib path export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/nishi/usr/local/lib/tensorflow-2.16.2 2) include path /home/nishi/usr/local/include/tensorflow-2.16.2 1. build $ g++ example.cpp -I/home/nishi/usr/local/include/tensorflow-2.16.2 -L/home/nishi/usr/local/lib/tensorflow-2.16.2 -std=c++17 -ltensorflow_cc -ltensorflow_framework -lprotobuf -o example -lprotobuf は、どこ? system か? libprotobuf-dev $ sudo apt install libprotobuf-dev libprotobuf.so libprotobuf.so.2 libprotobuf.so.version 2. run $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/nishi/usr/local/lib/tensorflow-2.16.2 $ ./example 2024-11-17 19:03:34.082568: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. 2024-11-17 19:03:34.083194: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:388] MLIR V1 optimization pass is not enabled 2024-11-17 19:03:34.113146: I example.cpp:61] 19 -3 outputs[0].matrix<float>():19 -3 */ // tensorflow/cc/example/example.cc #include <string> #include <iostream> #include <stdio.h> #include <unistd.h> #include "tensorflow/cc/client/client_session.h" #include "tensorflow/cc/ops/standard_ops.h" #include "tensorflow/core/framework/tensor.h" int main() { using namespace tensorflow; using namespace tensorflow::ops; Scope root = Scope::NewRootScope(); // Matrix A = [3 2; -1 0] auto A = Const(root, { {3.f, 2.f}, {-1.f, 0.f} }); // Vector b = [3 5] auto b = Const(root, { {3.f, 5.f} }); // v = Ab^T auto v = MatMul(root.WithOpName("v"), A, b, MatMul::TransposeB(true)); std::vector<Tensor> outputs; ClientSession session(root); // Run and fetch v TF_CHECK_OK(session.Run({v}, &outputs)); // Expect outputs[0] == [19; -3] LOG(INFO) << outputs[0].matrix<float>(); std::cout << " outputs[0].matrix<float>():" << outputs[0].matrix<float>() << std::endl; return 0; }
上部に、記載してあるので、それを参考にすれば、compile and run が出来る。
i) build
$ g++ example.cpp -I/home/nishi/usr/local/include/tensorflow-2.16.2 -L/home/nishi/usr/local/lib/tensorflow-2.16.2 -std=c++17 -ltensorflow_cc -ltensorflow_framework -lprotobuf -o example

ii) run
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/nishi/usr/local/lib/tensorflow-2.16.2
$ ./example
2024-11-17 19:03:34.082568: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. 2024-11-17 19:03:34.083194: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:388] MLIR V1 optimization pass is not enabled 2024-11-17 19:03:34.113146: I example.cpp:61] 19 -3 outputs[0].matrix<float>():19 -3

3.1 C++ load model のサンプル。
はてさて、 C++ で、どうやるのかさっぱりわからん。
とりあえず、google であさって、load model を作ってみた。
load_model.cpp
/* load_model.cpp /home/nishi/usr/local/src/cpp-nishi/nishi-many-test/tf-cpp/sample3/load_model.cpp https://stackoverflow.com/questions/61552420/tensorflow-2-0-c-load-pre-trained-model https://www.tensorflow.org/guide/saved_model?hl=ja#c_%E3%81%AB%E3%82%88%E3%82%8B_savedmodel_%E3%81%AE%E8%AA%AD%E3%81%BF%E8%BE%BC%E3%81%BF https://gist.github.com/OneRaynyDay/c79346890dda095aecc6e9249a9ff3e1 # export CPATH=$CPATH:/home/nishi/usr/local/include/tensorflow-2.16.2 # export LIBRARY_PATH=$LIBRARY_PATH:/home/nishi/usr/local/share/libtensorflow-cpu-2.16.2/lib 1) lib path export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/nishi/usr/local/lib/tensorflow-2.16.2 2) include path /home/nishi/usr/local/include/tensorflow-2.16.2 1. build $ g++ load_model.cpp -I/home/nishi/usr/local/include/tensorflow-2.16.2 -L/home/nishi/usr/local/lib/tensorflow-2.16.2 -std=c++17 -ltensorflow_cc -ltensorflow_framework [-lprotobuf] -o load_model -lprotobuf は、不要みたい 2. run $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/nishi/usr/local/lib/tensorflow-2.16.2 $ ./load_model */ // tensorflow/cc/example/example.cc #include <string> #include <iostream> #include <stdio.h> #include <unistd.h> #include "tensorflow/cc/client/client_session.h" #include "tensorflow/cc/ops/standard_ops.h" #include <tensorflow/cc/saved_model/loader.h> #include <tensorflow/cc/saved_model/tag_constants.h> #include <tensorflow/core/public/session_options.h> #include <tensorflow/core/framework/tensor.h> int main() { //using namespace tensorflow; //using namespace tensorflow::ops; tensorflow::SavedModelBundle model; tensorflow::Status status = tensorflow::LoadSavedModel( tensorflow::SessionOptions(), tensorflow::RunOptions(), //"path/to/model/folder", "/home/nishi/Documents/VisualStudio-TF/opp_with_lstm/Models/test_opp/a.model", {tensorflow::kSavedModelTagServe}, &model); std::cout << " exit" << std::endl; return 0; }

注) "/home/nishi/Documents/VisualStudio-TF/opp_with_lstm/Models/test_opp/a.model" は、
tosa-no-onchan/opp_with_lstm/tree/main/Models/test_opp の a.model.zip を解凍したmodel

1 build.
$ g++ load_model.cpp -I/home/nishi/usr/local/include/tensorflow-2.16.2 -L/home/nishi/usr/local/lib/tensorflow-2.16.2 -std=c++17 -ltensorflow_cc -ltensorflow_framework -o load_model

2. run
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/nishi/usr/local/lib/tensorflow-2.16.2
$ ./load_model
一応、OK だった。
すごい。...

4. tensorflow オリジナル examples の label_image が、参考になるので、試してみる。
download した、 tensorflow-2.16.2.tar.gz の中に入っているが、
tensorflow/examples/label_image をみると、
build は、
$ bazel build tensorflow/examples/label_image/...
でできるみたい。
$ cd
$ source kivy_env/bin/activate
$ cd ~/local/tensorflow/tensorflow-2.16.2
$ export TF_PYTHON_VERSION=3.10
$ bazel build tensorflow/examples/label_image/...

runは、
$ bazel-bin/tensorflow/examples/label_image/label_image

tensorflow/examples/label_image/data/inception_v3_2016_08_28_frozen.pb が必要みたい。
$ curl -L "https://storage.googleapis.com/download.tensorflow.org/models/inception_v3_2016_08_28_frozen.pb.tar.gz" | tar -C tensorflow/examples/label_image/data -xz
でダウンロードできるようだが、virtual_env から抜けて curl を実行しないといけないみたい。

再度実行してみる。
$ bazel-bin/tensorflow/examples/label_image/label_image 2024-11-19 15:21:42.219751: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. 2024-11-19 15:21:43.897880: I tensorflow/examples/label_image/main.cc:255] military uniform (653): 0.834306 2024-11-19 15:21:43.897910: I tensorflow/examples/label_image/main.cc:255] mortarboard (668): 0.0218695 2024-11-19 15:21:43.897918: I tensorflow/examples/label_image/main.cc:255] academic gown (401): 0.0103581 2024-11-19 15:21:43.897924: I tensorflow/examples/label_image/main.cc:255] pickelhaube (716): 0.00800817 2024-11-19 15:21:43.897939: I tensorflow/examples/label_image/main.cc:255] bulletproof vest (466): 0.0053509
OK じゃ。

試しに、上記 label_image を、/home/nishi/Documents/VisualStudio-CPP/tf2_cpp/label_image にコピーして、g++ build and run させてみた。

$ g++ main.cc -I/home/nishi/usr/local/include/tensorflow-2.16.2 -L/home/nishi/usr/local/lib/tensorflow-2.16.2 -std=c++17 -ltensorflow_cc -ltensorflow_framework -o predict

$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/nishi/usr/local/lib/tensorflow-2.16.2
$ ./predict
OK でした。
これで、先に作った C++ ライブラリーと、ヘッダーファイルが問題ないことが確認できたぞね。

あとは、main.cc を参考に、やり方を学ぶだけじゃ。