installing Tensorflow on Ubuntu 16.04 with GPU support
It will take about two hours just before "$ bazel build ...".
Bazel building will take about one or two hours.
- BIOS setting: diable Secure Boot, which contradicts with CUDA
- CUDA Toolkit 8.0
pre-installation actions can be skipped (for Ubuntu 16.04, Nvidia Geforce GTX 1080 Ti)
select to download "deb(local)" file and install by Package Manager
(since runfile is difficult to install)
(proxy_server of aptitude and wget needs not be worried about)
And, install the patch from CUDA download page if exists: ex:
sudo dpkg -i /home/autosem/Downloads/cuda-repo-ubuntu1604-8-0-local-cublas-performance-update_8.0.61-1_amd64.deb
sudo apt-get update
sudo apt-get install cuda (<== maybe this line is not necessary)
post-installation actions
cuda environment variables is also to be contained in .bashrc
see below
cuda samples are copied in ~/cuda/
I didn't install driver: so driver is not verified
to test: try to compile the cuda samples copied as above
error for cudaDecodeGL for some library problem -> ignore
$ ~/cuda/NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/deviceQuery
check it shows GTX 1080 Ti and Result = PASS
$ ~/cuda/NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/bandwidthTest
check it shows Result = PASS
- Nvidia driver may have been installed during installing CUDA Toolkit
- cuDNN v5.1 Or v6.0 (cf. tried v7.0.1, with error during bazel building)
You have to sign up as an Nvidia developer
from https://developer.nvidia.com/cudnn
download "cuDNN vX.X Library for Linux" (for compatible CUDA version)
move and extract as eg. ~/cuda/cuDNNLib/cuda/
Or
download "cuDNN vX.X Runtime Library for Ubuntu 16.04 (Deb),
cuDNN vX.X Developer Library for Ubuntu 16.04 (Deb),
cuDNN vX.X Code Samples and User Guide for Ubuntu 16.04 (Deb)"
$ sudo dpkg -i ./Downloads/libcudnn6_6.0.21-1+cuda8.0_amd64.deb
$ sudo dpkg -i ./Downloads/libcudnn6-dev_6.0.21-1+cuda8.0_amd64.deb
$ sudo dpkg -i ./Downloads/libcudnn6-doc_6.0.21-1+cuda8.0_amd64.deb
libcudnn.so and etc. will be found in /usr/lib/x86_64-linux-gnu/
set environment variables as in .bashrc (or .profile)
export CUDA_HOME="/usr/local/cuda-8.0"
export LD_LIBRARY_PATH="$HOME/cuda/cuDNNLib/cuda/lib64:$HOME/cuda/cuDNNLib/cuda:${CUDA_HOME}/lib64"
(Or for the case of installing Runtime, Developer Library:
export LD_LIBRARY_PATH="/usr/lib/x86_64-linux-gnu:${CUDA_HOME}/lib64")
export PATH="/usr/local/cuda-8.0/bin:$HOME/bin:$HOME/.local/bin:$PATH"
"Add <installpath> to your build and link process
by adding -I<installpath> to your compile line
and -L<installpath> -lcudnn to your link line." => how ???
- $ sudo apt-get install libcupti-dev
- install Tensorflow from sources
prepare
install git, curl, bazel, pip3, numpy and etc. (python3 (for python3.5) is already installed in Ubuntu 16.04)
$ sudo apt-get install git
$ sudo apt-get install curl
$ 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
$ sudo apt-get install python3-numpy python3-dev python3-pip python3-wheel
download tensorflow directory
$ git clone https://github.com/tensorflow/tensorflow
configure:
$ cd ~/tensorflow
$ ./configure
select default except
select python3.5
enable CUDA: y
cuDNN lib version is 5.1.10 (Or 6.0.21)
cuDNN lib directory is /home/soh/cuda/cuDNNLib/cuda (Or /usr/lib/x86_64-linux-gnu)
compute capability is 6.1
build the pip package with bazel
for gcc 5, append --cxxopt...
$ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"
$ bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
install pip package
$ ls /tmp/tensorflow_pkg/
$ sudo pip3 install /tmp/tensorflow_pkg/tensorflow-1.2.0-cp35-cp35m-linux_x86_64.whl
test
$ cd
$ mkdir tf
$ cd tf
$ python3 /* to test 4-line program */
-------
history
-------
sudo dpkg -i cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb
sudo apt-get update
sudo apt-get install cuda
vi .bashrc /* insert as follow
export CUDA_HOME="/usr/local/cuda-8.0"
export LD_LIBRARY_PATH="$HOME/cuda/cuDNNLib/cuda/lib64:$HOME/cuda/cuDNNLib/cuda:${CUDA_HOME}/lib64"
export PATH="/usr/local/cuda-8.0/bin:$HOME/bin:$HOME/.local/bin:$PATH"
*/
source .bashrc
mkdir cuda
cd cuda
cuda-install-samples-8.0.sh .
cd NVIDIA_CUDA-8.0_Samples/
make
cd
~/cuda/NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/deviceQuery
~/cuda/NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/bandwidthTest
cd cuda
cp ../Downloads/cudnn-8.0-linux-x64-v5.1.tgz .
tar xvzf cudnn-8.0-linux-x64-v5.1.tgz
mkdir cuDNNLib
mv cuda cuDNNLib/
mv cudnn-8.0-linux-x64-v5.1.tgz ../Downloads/
sudo apt install git
sudo apt install curl
git clone https://github.com/tensorflow/tensorflow
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
sudo apt-get install python3-numpy python3-dev python3-pip python3-wheel
ls /* tensorflow repository is in directory tensorflow now */
cd tensorflow
./configure
gcc --version /* and then I found it is 5.0 or newer */
bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_
package --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
ls /tmp/tensorflow_pkg/
sudo pip3 install /tmp/tensorflow_pkg/tensorflow-1.2.0-cp35-cp35m-linux_x86_64.whl
cd
mkdir tf
cd tf
python3 /* to test 4-line program */