tensorflow.sh

# source: https://gist.github.com/phatblat/1713458
# Save script's current directory
DIR=$(pwd)

set -e
set -u
set -x


echo "############################"
echo "# TensorFlow with GPU Support"
echo "############################"
echo "# "
echo "# MANUAL STEPS REQUIRED"
echo "# - make sure you source ~/.robotics_setup and ensure cuda and cudnn are on the LD_LIBRARY_PATH"
echo "# "
echo "# TF 1.8 special steps with CUDA 9.1!"
echo "# see https://github.com/tensorflow/tensorflow/issues/16526"
echo "# and https://github.com/tensorflow/tensorflow/issues/14573"
echo "# "
echo "# After tensorflow is cloned and the first failure, browse to the repository"
echo "# ~/src/tensorflow/configure.py"
echo "# "
echo "# and change:"
echo "# "
echo "#     _DEFAULT_CUDA_VERSION = '9.0'"
echo "# "
echo "#  to:"
echo "#     _DEFAULT_CUDA_VERSION = '9.1'"
echo "# "
echo "# "
echo "# Also ensure you've run nvidia.sh and cuda.sh to install those components"
echo "# "
echo "# It is important to note that as of tensorflow 1.5, "
echo "# the binary supports CUDA 9.0 but these scripts install 9.1,"
echo "# so building from source is recommended"
echo "# "
echo "# Install Steps From:"
echo "# https://www.tensorflow.org/get_started/os_setup#installing_from_sources"
echo "# https://alliseesolutions.wordpress.com/2016/09/08/install-gpu-tensorflow-from-sources-w-ubuntu-16-04-and-cuda-8-0-rc/"
echo "# Additional possible source:"
echo "# https://www.nvidia.com/object/gpu-accelerated-applications-tensorflow-installation.html"
echo "# https://github.com/tensorflow/tensorflow/issues/7449#issuecomment-282855210"
echo "#"
echo "# Options:"
echo "#"
echo "#   -p Enable/disable python bindings, options are ON, OFF"
echo "#   -b git branch to check out, such as master"
echo "#   -l git location to use, -b tensorflow means github.com/tensorflow/tensorflow,"
echo "#      while ahundt means github.com/ahundt/tensorflow"
echo "# "
echo "# Getting a permission denied error? Try this:"
echo "#     pip install --user --upgrade /tmp/tensorflow_pkg/tensorflow-*"
echo "# "
echo "# Getting a permission denied error? Try this:"
echo "#     sudo chmod o+r /usr/local/lib/python2.7/dist-packages/"
echo "# Another backup last step until I find a better workaround (you shouldn't use sudo pip, and you might hit weird problems later):"
echo "#     sudo pip install --upgrade /tmp/tensorflow_pkg/tensorflow-*"
echo "# "
echo "# Can't run? Make sure the right tensorflow package is on your PATH both in your shell and in python's path."
echo "#"
echo "# Build from source (default):"
echo "#    sh tensorflow.sh"
echo ""
echo "# Install with pip:"
echo "#    sh tensorflow.sh -p ON"
echo "#"
echo "#"



# Enable python bindings via cython by default
# set to "ON" to build python bindings and "OFF" to disable them
# Note: "ON" generates tons of warnings and the log size might prevent CI from succeeding.
PIP_INSTALL="ON"
location="tensorflow" # github.com/jrl-umi3218/Tasks # ongoing development happens here
#location="ahundt" # github.com/ahundt/Tasks # I have some patches here
#location="jorisv" # github.com/jorisv/Tasks # outdated original repository location

# modify if using a different branch
branch="master"

# Check if the user specified any command line options
# other than teh default and set the variable appropriately.
while getopts p:l:b: option
do
 case "${option}"
 in
 p) PIP_INSTALL="${OPTARG}";;
 l) location="${OPTARG}";;
 b) branch="${OPTARG}";;
 esac
done


# os specific setup
OS=`uname`
case $OS in
    'Linux')

        if [ "${PIP_INSTALL}" = "ON" ]
        then
            sh python.sh
        fi
        sudo apt-get update

        sudo apt-get install -y openjdk-8-jdk git python-dev python3-dev python-numpy python3-numpy build-essential python-pip python3-pip python-virtualenv swig python-wheel libcurl3-dev libcupti-dev
        sudo apt-get install -y openmpi-bin openmpi-common
    ;;
    *) ;;
    'Darwin')
        OS='Mac'
        brew install python python3 numpy
    ;;
esac


if [ "${PIP_INSTALL}" = "ON" ] ; then
	echo "# PIP INSTALL MANUAL STEPS: YOU MAY NEED TO RUN:"
	echo "# ./cuda_9.0_cudnn_7.0.sh"
	echo "# ./cudnn.sh"
    pip2 install tensorflow-gpu --user --upgrade
    pip3 install tensorflow-gpu --user --upgrade
else
    # install the bazel build system
    ./bazel.sh

    cd ~/src/
    if [ ! -d ~/src/tensorflow ]
    then
        git clone https://github.com/${location}/tensorflow
    fi

    cd ~/src/tensorflow
    git fetch --all
    git checkout r1.8

    echo "###########################################################################################################"
    echo "MANUAL STEPS YOU MAY NEED TO EDIT FOR YOUR SYSTEM"
    echo "Configuring, please use all defaults, except CUDA 8.0, cudnn 5, NO OPENCL and for GTX 1080 set compute capability to 5.2,6.1:"
    echo ""
    echo " To determine the compute capability for your gpu see https://developer.nvidia.com/cuda-gpus "
    echo ""
    echo "To see all the possible flags with the current master see:"
    echo "https://github.com/tensorflow/tensorflow/blob/master/configure.py"
    echo "be sure to switch to the appropriate branch if the options aren't working"
    echo ""
    echo "To find the CUDA_TOOLKIT_PATH try the following where 9-1 is the version (8-0 is another example):"
    echo "    dpkg-query -L cuda-cublas-9-1"
    echo "###########################################################################################################"
    export TF_NEED_CUDA=1
    export TF_CUDA_VERSION="9.1"
    export TF_CUDNN_VERSION=7
    export CUDA_TOOLKIT_PATH=/usr/local/cuda
    export TF_NEED_GCP=1
    export TF_NEED_HDFS=1
    export TF_NEED_OPENCL=0
    export TF_NEED_JEMALLOC=1
    export TF_ENABLE_XLA=1
    export TF_CUDA_CLANG=0
    # TensorRT is only for super high end GPUs with fp16 and int8:
    # https://developer.nvidia.com/tensorrt
    export TF_NEED_TENSORRT=0
    export TF_CUDA_COMPUTE_CAPABILITIES="5.2,6.1"
    # Open MPI settings, not working yet, see
    # https://github.com/tensorflow/tensorflow/issues/11903#issuecomment-332718012
    export TF_NEED_MPI=0
    # export MPI_HOME=/usr/
    # CC_OPT_FLAGS hack to remove when possible
    # export CC_OPT_FLAGS="-DOMPI_SKIP_MPICXX=1 -march=native"
	
	# workaround for https://github.com/tensorflow/tensorflow/issues/14573
	if [ ! -f /usr/local/cuda/lib64/stubs/libcuda.so.1 ] ; then
		sudo ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1
		sudo ln -s /usr/local/cuda/targets/x86_64-linux/lib/lib64 /usr/local/cuda/targets/x86_64-linux/lib
	fi

    # Note python3, python2, pip3 and pip2 are used explicily
    # because that ensures tensorflow is installed for both versions
    # of python, and there is no accidental install of the same version twice.

    # Python2 install, requires pip2
    if [ -x "$(command -v pip2)" ] ; then
        echo "python 2 install:"
        export PYTHON_BIN_PATH=`which python2`

        # answer yes to any config questions not covered by the above exports, run the configuration
        yes "" | $PYTHON_BIN_PATH configure.py

        # To be compatible with as wide a range of machines as possible, TensorFlow defaults to only using SSE4.1 SIMD instructions on x86 machines. Most modern PCs and Macs support more advanced instructions, so if you're building a binary that you'll only be running on your own machine, you can enable these by using --copt=-march=native in your bazel build command.
		#
        # For details on LD_LIBRARY_PATH see https://github.com/tensorflow/tensorflow/issues/15142#issuecomment-352773470
        bazel build --copt=-march=native -c opt --config=cuda //tensorflow/tools/pip_package:build_pip_package --action_env="LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"
        bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg --verbose_failures
        cd ~/
        python2 -m pip install --upgrade --user /tmp/tensorflow_pkg/tensorflow-*p27*
        python2 -c 'import tensorflow as tf; print(tf.__version__); sess = tf.InteractiveSession(); sess.close();'
    fi

    # Python 3 install, requires pip3
    if [ -x "$(command -v pip3)" ] ; then

        echo "python3 install:"
        cd ~/src/tensorflow
        export PYTHON_BIN_PATH=`which python3`

        # answer yes to any config questions not covered by the above exports, run the configuration
        yes "" | $PYTHON_BIN_PATH configure.py

        # For details on LD_LIBRARY_PATH see https://github.com/tensorflow/tensorflow/issues/15142#issuecomment-352773470
        bazel build --copt=-march=native -c opt --config=cuda //tensorflow/tools/pip_package:build_pip_package --action_env="LD_LIBRARY_PATH=${LD_LIBRARY_PATH}" --verbose_failures
        bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
        cd ~/
        python3 -m pip install --upgrade --user /tmp/tensorflow_pkg/tensorflow-*p3*
        python3 -c 'import tensorflow as tf; print(tf.__version__); sess = tf.InteractiveSession(); sess.close();'
    fi
fi

# TODO: BEGIN OLD NOTES TO REMOVE IF NO LONGER NEEDED
# At one point I had to make a symlink for mkl to be found properly:
# /home/ahundt/src/tensorflow/third_party/eigen3/mkl_include -> /home/ahundt/src/tensorflow/third_party/mkl/mklml_lnx_2018.0.20170425/include
# echo "# "
# echo "# Plus add the following to your .bashrc or .zshcrc:"
# echo "# export LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64\""
# echo "# export CUDA_HOME=/usr/local/cuda"
# TODO: END OLD NOTES TO REMOVE IF NO LONGER NEEDED


echo "###########################################################################################################"
echo "MANUAL STEPS with no spaces after tensorflow hit tab before hitting enter to fill in blanks with the following MANUAL line:"
echo "pip install /tmp/tensorflow_pkg/tensorflow # or: pip install /tmp/tensorflow_pkg/tensorflow-*"
echo "###########################################################################################################"
echo ""
echo "To test that TensorFlow installed correctly do the following with the python or python3 command depending on your version"
echo "but only from a directory that doesn't have the tensorflow source in it because those files might be picked up accidentally:'"
echo ""
echo "python2:"
echo "python2 -c 'import tensorflow as tf; print(tf.__version__); sess = tf.InteractiveSession(); sess.close();'"
echo ""
echo "python3:"
echo "python3 -c 'import tensorflow as tf; print(tf.__version__); sess = tf.InteractiveSession(); sess.close();'"
cd $DIR