Compile Tensorflow C++ without Bazel
Updates:
- The configurations to compile for tensorflow 1.8.0 is described in this post. If you want tensorflow to work nicely with OpenCV, follow this post with only one change described in that post.
In this post, I will give detailed instructions on how to compile the official C++ Tensorflow project label_image with gcc instead of bazel.
The reason why I write this blog is because officially, to compile a C++ Tensorflow project, you have to integrate it in the source tree of tensorflow, create a BUILD file and compile it with bazel. For some reason, if you want to create a C++ Tensorflow project in your favorite C++ IDEs and build it with Makefile or CMake, you will need to do some extra work to allow gcc to be able to compile successfully C++ Tensorflow codes. The detailed instructions are in the second part, you can skip the first part (Create a Ubuntu docker image) if you want to do directly on your machine instead of on a docker image.
I. Create a Ubuntu docker image
In the first part, I will create a docker image with latest version of tensorflow on the latest stable Ubuntu:
- Ubuntu 17.10
- gcc 7.2 (comes with Ubuntu 17.10)
- tensorflow 1.4.0
The only reason why I use Docker is to create an independent environment to test the latest version of tensorfow on the latest version of Ubuntu.
I assume that you know the basic of Docker, here I create a ubuntu 17.10 image with this Dockerfile:
FROM ubuntu:17.10
# Install.
RUN \
sed -i 's/# \(.*multiverse$\)/\1/g' /etc/apt/sources.list && \
apt-get update && \
apt-get -y upgrade && \
apt-get install -y build-essential && \
apt-get install -y software-properties-common && \
apt-get install -y byobu curl git htop man unzip vim wget && \
apt-get update && \
apt-get install python-dev python-pip python-setuptools python-sphinx python-yaml python-h5py python3-pip python-numpy python-scipy python-nose && \
rm -rf /var/lib/apt/lists/*
Then use the following command to create the image:
docker build -t ubuntu:17.10 .
From now on, we will work on the interactive shell of docker image, to go into the shell:
docker run -it ubuntu:17.10 /bin/sh
II. Steps to make a standalone C++ Tensorflow
1. Compile tensorflow
Install bazel to compile tensorflow:
sudo apt-get install openjdk-8-jdk
echo "deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8" | sudo tee /etc/apt/sources.list.d/bazel.list
curl https://bazel.build/bazel-release.pub.gpg | sudo apt-key add -
sudo apt-get update && sudo apt-get install bazel
sudo apt-get upgrade bazel
Clone tensorflow repo:
git clone https://github.com/tensorflow/tensorflow.git
Compile tensorflow for c++ (default parameters, python3)
cd tensorflow
./configure
bazel build //tensorflow:libtensorflow_cc.so
2. Install additional libraries
You need to install cmake to compile others libraries, here I install cmake 3.9.6:
wget https://cmake.org/files/v3.9/cmake-3.9.6.tar.gz
tar -xzvf cmake-3.9.6.tar.gz
cd cmake-3.9.6
./configure
make
sudo make install
Install protobuf from sources:
(Note that the version v3.4.0 goes with tensorflow 1.4.0. If you have a wrong version of protobuf, when compile the label_images program in the final step of this post you will see error:
./tensorflow/core/framework/tensor.pb.h:12:2: error: #error This file was generated by a newer version of protoc which is
Then go look at ./include/third_party/tensorflow/core/framework/tensor.pb.h (which is created after step 3) and go to the line of error to get the right version of protobuf)
cd /home
sudo apt-get install autoconf automake libtool
git clone https://github.com/google/protobuf.git
cd protobuf
git checkout v3.4.0
./autogen.sh
./configure
make
make check
sudo make install
Install Eigen 3.3.4 from sources
cd /home
wget http://bitbucket.org/eigen/eigen/get/3.3.4.tar.bz2
tar -xvjf 3.3.4.tar.bz2
cd eigen-folder
mkdir build
cd build
cmake ..
make
sudo make install
3. Create and compile a standalone tensorflow C++ project
Create a standalone folder to test compilation of C++ tensorflow with gcc
cd /home
mkdir standalone
To compile tensorflow with gcc, it has to get all the header files required to compile. Create an include folder and copy header files to that folder:
export TENSORFLOW_DIR=/home/tensorflow
cd /home/standalone
mkdir -p include/third_party
cp -r $TENSORFLOW_DIR/tensorflow include/third_party/
cp -r $TENSORFLOW_DIR/bazel-genfiles/tensorflow include/third_party/
cp -r $TENSORFLOW_DIR/third_party/eigen3 include/third_party/
cp -r /home/protobuf/src/google include/third_party/
cp -r /home/eigen-folder/. include/third_party/eigen3/
cp -r include/third_party/eigen3/Eigen include/third_party/
Clone google nsync:
cd /home/standalone/include
git clone https://github.com/google/nsync.git
We need also 2 libraries libtensorflow_cc.so and libtensorflow_framework.so. Copy those libraries to /usr/local/lib:
cd /home/standalone
mkdir lib
cp -r $TENSORFLOW_DIR/bazel-bin/tensorflow/libtensorflow_cc.so /home/standalone/lib
cp -r $TENSORFLOW_DIR/bazel-bin/tensorflow/libtensorflow_framework.so /home/standalone/lib
Refresh shared library cache:
sudo ldconfig
Copy C++ example label_image under tensorflow source tree to standalone folder:
cd /home/standalone
cp $TENSORFLOW_DIR/tensorflow/examples/label_image/main.cc .
cp -r $TENSORFLOW_DIR/tensorflow/examples/label_image/data .
Get inception model
cd /home/standalone/data
wget https://storage.googleapis.com/download.tensorflow.org/models/inception_v3_2016_08_28_frozen.pb.tar.gz
tar -xzvf inception_v3_2016_08_28_frozen.pb.tar.gz
Create Makefile in standalone folder with content:
CC = g++
CFLAGS = -std=c++11 -g -Wall -D_DEBUG -Wshadow -Wno-sign-compare -w
DIR = /home/standalone/
INC = -I/usr/local/include/eigen3
INC += -I./include/third_party
INC += -I./include
INC += -I./include/nsync/public/
LDFLAGS = -lprotobuf -pthread -lpthread
LDFLAGS += -L$(DIR)/lib/ -Wl,-R$(DIR)/lib/ '-Wl,-R$$ORIGIN'
LDFLAGS += -ltensorflow_cc -ltensorflow_framework
all: main
main:
$(CC) $(CFLAGS) -o main main.cc $(INC) $(LDFLAGS)
run:
./main --image=./data/grace_hopper.jpg --graph=./data/inception_v3_2016_08_28_frozen.pb --labels=./data/imagenet_slim_labels.txt
clean:
rm -f main
Now do:
make
make run
Normally, it wil output:
./main --image=./data/grace_hopper.jpg --graph=./data/inception_v3_2016_08_28_frozen.pb --labels=./data/imagenet_slim_labels.txt
2017-11-19 13:10:23.989200: I tensorflow/core/platform/cpu_feature_guard.cc:137]
2017-11-19 13:10:25.205406: I main.cc:250] military uniform (653): 0.834306
2017-11-19 13:10:25.205491: I main.cc:250] mortarboard (668): 0.0218692
2017-11-19 13:10:25.205520: I main.cc:250] academic gown (401): 0.0103579
2017-11-19 13:10:25.205544: I main.cc:250] pickelhaube (716): 0.00800814
2017-11-19 13:10:25.205563: I main.cc:250] bulletproof vest (466): 0.00535088
For those who use docker image, now you can quit the interactive shell of docker and commit it so that it keeps all the things that you just did:
# On the interactive shell of docker do
exit
# On the local shell do the following to get docker container's id
docker ps -l
# Then commit
docker commit <container-id> <image-name:tag>
I hope that this blog helps. If you have any question, feel free to contact me to: phan at tuanphuc dot com.