tensorflow_cc
This repository makes possible the usage of the TensorFlow C++ API from the outside of the TensorFlow source code folders and without the use of the Bazel build system.
This repository contains two CMake projects. The tensorflow_cc project downloads, builds and installs the TensorFlow C++ API into the operating system and the example project demonstrates its simple usage.
Docker
If you wish to start using this project right away, fetch a prebuilt image on Docker Hub!
Running the image on CPU:
docker run -it floopcz/tensorflow_cc:ubuntu-shared /bin/bash
If you also want to utilize your NVIDIA GPU, install NVIDIA Docker and run:
docker run --runtime=nvidia -it floopcz/tensorflow_cc:ubuntu-shared-cuda /bin/bash
The list of available images:
Image name | Description |
---|---|
floopcz/tensorflow_cc:ubuntu-static |
Ubuntu + static build of tensorflow_cc |
floopcz/tensorflow_cc:ubuntu-shared |
Ubuntu + shared build of tensorflow_cc |
floopcz/tensorflow_cc:ubuntu-shared-cuda |
Ubuntu + shared build of tensorflow_cc + NVIDIA CUDA |
floopcz/tensorflow_cc:archlinux-shared |
Arch Linux + shared build of tensorflow_cc |
To build one of the images yourself, e.g. ubuntu-shared
, run:
docker build -t floopcz/tensorflow_cc:ubuntu-shared -f Dockerfiles/ubuntu-shared .
Installation
1) Install requirements
Ubuntu 18.04:
sudo apt-get install build-essential curl git cmake unzip autoconf autogen automake libtool mlocate \
zlib1g-dev g++-7 python python3-numpy python3-dev python3-pip python3-wheel wget
sudo updatedb
If you require GPU support on Ubuntu, please also install Bazel, NVIDIA CUDA Toolkit (>=9.2), NVIDIA drivers, cuDNN, and cuda-command-line-tools
package. The tensorflow build script will automatically detect CUDA if it is installed in /opt/cuda
or /usr/local/cuda
directories.
Arch Linux:
sudo pacman -S base-devel cmake git unzip mlocate python python-numpy wget
sudo updatedb
For GPU support on Arch, also install the following:
sudo pacman -S gcc7 bazel cuda cudnn nvidia
Warning: Newer versions of TensorFlow sometimes fail to build with the latest version of Bazel. You may wish to install an older version of Bazel (e.g., 0.16.1).
2) Clone this repository
git clone https://github.com/FloopCZ/tensorflow_cc.git
cd tensorflow_cc
3) Build and install the library
There are two possible ways to build the TensorFlow C++ library:
- As a static library (default):
- Faster to build.
- Provides only basic functionality, just enough for inferring using an existing network (see contrib/makefile).
- No GPU support.
- As a shared library:
- Requires Bazel.
- Slower to build.
- Provides the full TensorFlow C++ API.
- GPU support.
cd tensorflow_cc
mkdir build && cd build
# for static library only:
cmake ..
# for shared library only (requires Bazel):
# cmake -DTENSORFLOW_STATIC=OFF -DTENSORFLOW_SHARED=ON ..
make && sudo make install
Warning: Optimizations for Intel CPU generation >=ivybridge
are enabled by default. If you have a
processor that is older than ivybridge
generation, you may wish to run export CC_OPT_FLAGS="-march=native"
before the build. This command provides the best possible optimizations for your current CPU generation, but
it may cause the built library to be incompatible with older generations.
4) (Optional) Free disk space
# cleanup bazel build directory
rm -rf ~/.cache
# remove the build folder
cd .. && rm -rf build
Usage
1) Write your C++ code:
// example.cpp
#include <tensorflow/core/platform/env.h>
#include <tensorflow/core/public/session.h>
#include <iostream>
using namespace std;
using namespace tensorflow;
int main()
{
Session* session;
Status status = NewSession(SessionOptions(), &session);
if (!status.ok()) {
cout << status.ToString() << "\n";
return 1;
}
cout << "Session successfully created.\n";
}
2) Link TensorflowCC to your program using CMake
# CMakeLists.txt
find_package(TensorflowCC REQUIRED)
add_executable(example example.cpp)
# Link the static Tensorflow library.
target_link_libraries(example TensorflowCC::Static)
# Altenatively, link the shared Tensorflow library.
# target_link_libraries(example TensorflowCC::Shared)
# For shared library setting, you may also link cuda if it is available.
# find_package(CUDA)
# if(CUDA_FOUND)
# target_link_libraries(example ${CUDA_LIBRARIES})
# endif()
3) Build and run your program
mkdir build && cd build
cmake .. && make
./example
If you are still unsure, consult the Dockerfiles for Ubuntu and Arch Linux.
Compile Tensorflow C++ with GPU support on RCI cluster
Elapsed time: 13096.862s for using 12 cores.
1) Ask for a machine with 2x CPU, GPU and with 50G memory and 100G disk space (can be reduced, but that is verified it is working)
srun -v -p gpu --gres=gpu:1 --mem 50G --tmp 100G --pty bash -i
2) Go into the local scratch storage (/data/temporary [1])
export SCRATCH_DIR=/data/temporary
mkdir $SCRATCH_DIR
3) Create a directory for Bazel's cache
mkdir bazel_cache
4) Git clone my version of FloopCZ/tensorflow_cc [2] you find here
git clone https://github.com/jan-rudolf/tensorflow_cc
5) Make a symlink for Bazel's cache
ln -fs $SCRATCH_DIR/bazel_cache $HOME/.cache/bazel
6) Set up Bazel's enviroment variable with the custome cache directory
export TEST_TMPDIR=$SCRATCH_DIR/bazel_cache
7) Add all dependant modules
ml cuDNN/7.5.0.56-fosscuda-2019a # this also includes GCC, CUDA, libtool etc.
ml Autotools/20180311-GCCcore-8.2.0
ml wheel/0.31.1-fosscuda-2019a-Python-3.7.2
ml NCCL/2.4.2-fosscuda-2019a # mozna neni potreba
ml Bazel/0.20.0-GCCcore-8.2.0
ml CMake/3.13.3-GCCcore-8.2.0
8) Create a build directory
cd tensorflow_cc/tensorflow_cc/
mkdir build && cd build
9) Create makefile for shared library (default allow CUDA is ON)
cmake -DTENSORFLOW_STATIC=OFF -DTENSORFLOW_SHARED=ON ..
10) Run compilation (takes few hours)
make
11) After compilation, copy header files and object files where you need
make install
References: [1] https://login.rci.cvut.cz/wiki/storage [2] https://github.com/FloopCZ/tensorflow_cc