Deep Learning software installation guide on fresh Ubuntu
Guide to building and installing CUDA, CuDNN, OpenCV, FFMPEG, Theano, Tensorflow, Keras, Lasagne, Torch and Caffe. It also includes common issues faced and recommended libraries and versions.
Table of Contents
- Image processing and Computer Vision
- Install GPU Libraries
Recently I assembled a machine with 4 GPU (Titan X), a clone of NVIDIA DevBox. There are few other blog posts which describe the hardware guide, so I will not go into the same detail. Please refer this. The actual list of parts I bought can be found at PCPart Picker.
However this blog post is about the software guide. Although, most of the researchers or grad students like me will have their own custom requirement I thought I should share my software installation guide, for someone who might be new and would benefit from some of these. I have had to do this process couple of times now, and I have learned from my mistakes.
NOTE 1: Docker : If getting maximum performance is not a requirement, I would suggest the reader to download Docker Images for Deep Learning packages freely available online, for e.g. Time required for all these standalone installation is justified only when these are going to be used over and again, for e.g in a research lab.
NOTE 2: Redundancy : Some of the commands can be combined to be done in a single line, especially installing libs from apt-get. But I prefer to run them one at a time, for more control and feedback on installation process.
NOTE 3: Python 2 : While most of the scientific community is moving on with Python 3, it seems Deep Learning communmmunty is happy with Python 2. Thus all instllations except where mentioned pertain to Python 2.7.
NOTE 4: Anaconda : Why not Anaconda ? I have done some previous installations with Anaconda, but it becomes messy with smaller libs, and also compiling OpenCV. More importantly, I do not need all the utils of Anaconda, and on this machine I would like to minimize the libs installed.
Upgrade the Ubuntu
Always, upgrade the ubuntu for security and sanity purposes !
sudo apt-get update# Fetches the list of available updates
sudo apt-get upgrade# Strictly upgrades the current packages
sudo apt-get dist-upgrade# Installs updates (new ones)
Partition and mount the harddrives
Since my configuration has two HDD of 3 TB, I have configured them as RAID 1, this provides data redundancy. You do not want to lose weeks worth of training due to HDD crash ! If you do not want RAID 1, then skip this step.
sudo apt-get install -y mdadm# Install mdadm, tool to manage RAID 1
sudo mdadm --assemble --scan# check for existing raids ## found existing !! If not prepare, refer tutorial here http://askubuntu.com/questions/526747/setting-up-raid-1-on-14-04-with-an-existing-drive
df -aTh# shows list of all mounts
sudo mount /dev/md0 /media/hdd/# (Manual) mount existing
blkid# shows uuid for drives to add to fstab Add the following line to /etc/fstab
the RAID 1 mount of two hdd
UUID=06ad59d9-3176-4c16-95e9-77356cc572d7 /media/hdd ext2 defaults 0 1
sudo mount -a# (Permanent) mount using fstab
Install Essentials, Extras, Git, Zsh
sudo apt-get install -y build-essential
sudo apt-get install -y ubuntu-restricted-extras
sudo apt-get install -y vim
sudo apt-get install -y git
sudo apt-get install -y git-core
- `sudo apt-get install -y checkinstall # allows for easy install/uninstall of packages from source
sudo apt-get install -y zsh
sudo apt-get install -y tmux
sudo apt-get install -y CMake
sudo apt-get install -y libopenblas-dev
- `sudo apt-get install -y libhdf5-dev
sudo apt-get install -y gcc-4.8# because CUDA (7.0 & 7.5) works will less than 4.9.0
- make soft link for gcc in /usr/bin
sudo apt-get install -y g++-4.8# because CUDA (7.0 & 7.5) works will less than 4.9.0
- make soft link for g++ in /usr/bin
sudo apt-get install -y apache2
- sudo /etc/init.d/apache2 start ` # start the Apache Server
Python and Libs
python get-pip.py# install pip
sudo apt-get install python-dev# pythonLibs
sudo apt-get install libblas-dev liblapack-dev libatlas-base-dev gfortran
sudo pip install cython git+https://github.com/scipy/scipy# Installs Cython and Scipy both (Cython is requirement for scipy)
sudo pip install -U scikit-learn# Requires numpy and scipy
- Matplotlib better to install using APT, lot of system lib dependency
sudo apt-get install -y matplotlib
sudo pip install numpy
sudo pip install markupsafe
sudo pip install h5py
sudo pip install nltk
sudo pip install nose
- Install leveldb # for efficient use with Caffe, and probably other libs
- git clone email@example.com:google/leveldb.git
- sudo mv out-shared/libleveldb.* /usr/local/lib/
- sudo cp -R include/leveldb /usr/local/include
- sudo ldconfig
Python3 and Ipython (Jupyter)
sudo apt-get install python3-pip# to install jupyter for python3, it needs pip3 and does not work using pip
sudo pip install jupyter
sudo pip3 install jupyter# I don’t know why it requires sepearate installation, especially when not done using Anaconda !
Install NVIDIA Drivers
I would highly advice against installation using apt-get as it always installs one version older, and with NVIDIA every generation of driver brings performance boosts.
sudo service lightdm stop# stop the X server before running the installation NOTE! Could not sign kernels, make a note of this. It might lead to problems with some libraries later on
- Download AMD 64 bit Linux drivers from NVIDIA website (manual download and transfer and execute the provided script or binary)
sudo ./cuda_7.0.28_linux.run --override# for cuda 7.0
sudo ln -s /usr/bin/g++-4.8 /usr/local/cuda/bin/g++# this might be omitted if a symbolic link has been made from /usr/bin/g++ like mentioned above
sudo ln -s /usr/bin/gcc-4.8 /usr/local/cuda/bin/gcc# this might be omitted if a symbolic link has been made from /usr/bin/gcc like mentioned above
CuDNN improves the performance of deep learning libraries. Unfornately, not all of them work with latest version. Google’s Tensorflow requires v2, Theano & Caffe works with v3, whereas latest version is 4. To get the library you will have to register at NVIDIA CuDNN
- Download the archive folder
Although not required, I find it useful to rather install both the version of CuDNN (2 and 3). Put the libs of v2 in CUDA 6.5 or CUDA 7.0 (for Tensorflow), and v3 in CUDA 7.5.
Image processing and Computer Vision
This is to install OpenCV 3.1 on Ubuntu 15.10. Thanks to the people at BVLC/Caffe, I did not have to tweak the settings, and pretty much their provided instructions worked.
sudo apt-get -y install libopencv-dev build-essential cmake git libgtk2.0-dev pkg-config python-dev python-numpy libdc1394-22 libdc1394-22-dev libjpeg-dev libpng12-dev libtiff5-dev libjasper-dev libavcodec-dev libavformat-dev libswscale-dev libxine2-dev libgstreamer0.10-dev libgstreamer-plugins-base0.10-dev libv4l-dev libtbb-dev libqt4-dev libfaac-dev libmp3lame-dev libopencore-amrnb-dev libopencore-amrwb-dev libtheora-dev libvorbis-dev libxvidcore-dev x264 v4l-utils unzip
Download OpenCV from http://opencv.org/downloads.html and unpack. Enter the unpacked directory. Execute:
mkdir build && cd build/# If you have Oh-my-zsh then just
cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local -D WITH_TBB=ON -D WITH_V4L=ON -D WITH_QT=ON -D WITH_OPENGL=ON ..
make -j12 # j 12 makes 12 parallel processes using all the cores to speed up the process, in a single core, the build might take couple of hours.
sudo checkinstall # to create the deb package, and install it.
sudo apt-get install -y ffmpeg
sudo apt-get install -y imagemagick
Install GPU Libraries
sudo pip install Theano
sudo apt-get install -y python-pycuda# also installs the dependencies, but it is best to have own installation of nvidia-cuda, to make sure the version is proper and to maintain multiple version installation
sudo pip install --upgrade https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow-0.6.0-cp27-none-linux_x86_64.whl
Tensorflow (as of Jan 15, 2016), works on Cuda 7.0 and CuDNN v2 ), thus change /usr/local/cuda softlink
sudo pip install keras# Could it get any easier !! Thanks people developer of Keras !
sudo pip install https://github.com/Lasagne/Lasagne/archive/master.zip# Preferred way to install Lasagne !
curl -s https://raw.githubusercontent.com/torch/ezinstall/master/install-deps | bash
git clone https://github.com/torch/distro.git ~/torch --recursive
cd ~/torch; ./install.sh
Aditionally to load caffe models in torch,
luarocks install loadcaffe
- Switch back to GCC 5 and G++ 5, because the prebuilt libraries of Ubuntu 15.10 are built on gcc 5, and thus compiling caffe in 4.8 will
not link them. However, caffe will refuse to compile with gcc 5, with error from following file
/usr/local/cuda/include/host_config.hJust comment out the line 115, which checks the version of GCC
- Use the CMake instead of default make (checkout https://github.com/BVLC/caffe/pull/1667 )
sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
sudo apt-get install --no-install-recommends libboost-all-dev
sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
for req in $(cat python/requirements.txt); do pip install $req; done
Linking Error in Leveldb e.g
Linking CXX executable caffe ../lib/libcaffe.so.1.0.0-rc3: undefined reference to `leveldb::Status::ToString[abi:cxx11]() const' ../lib/libcaffe.so.1.0.0-rc3: undefined reference to `leveldb::DB::Open(leveldb::Options const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, leveldb::DB**)' collect2: error: ld returned 1 exit status tools/CMakeFiles/caffe.bin.dir/build.make:122: recipe for target 'tools/caffe' failed make: *** [tools/caffe] Error 1
Solution, recompile Leveldb with GCC 5
sudo apt-get install libsnappy-dev
wget https://leveldb.googlecode.com/files/leveldb-1.9.0.tar.gz\ntar -xzf leveldb-1.9.0.tar.gz\ncd leveldb-1.9.0make
sudo mv libleveldb.* /usr/local/lib
sudo cp -R leveldb /usr/local/include
Linking Error in Protobuf e.g
Linking CXX executable caffe CMakeFiles/caffe.bin.dir/caffe.cpp.o: In function `std::string* google::MakeCheckOpString<cudaError, cudaError>(cudaError const&, cudaError const&, char const*)': caffe.cpp:(.text._ZN6google17MakeCheckOpStringI9cudaErrorS1_EEPSsRKT_RKT0_PKc[_ZN6google17MakeCheckOpStringI9cudaErrorS1_EEPSsRKT_RKT0_PKc]+0x43): undefined reference to google::base::CheckOpMessageBuilder::NewString()' CMakeFiles/caffe.bin.dir/caffe.cpp.o: In function `std::string* google::MakeCheckOpString<unsigned long, int>(unsigned long const&, int const&, char const*)': caffe.cpp:(.text._ZN6google17MakeCheckOpStringImiEEPSsRKT_RKT0_PKc[_ZN6google17MakeCheckOpStringImiEEPSsRKT_RKT0_PKc]+0x43): undefined reference to `google::base::CheckOpMessageBuilder::NewString()' CMakeFiles/caffe.bin.dir/caffe.cpp.o: In function `main': caffe.cpp:(.text.startup+0x3e): undefined reference to `google::SetVersionString(std::string const&)' caffe.cpp:(.text.startup+0x6e): undefined reference to `google::SetUsageMessage(std::string const&)' ../lib/libcaffe.so.1.0.0-rc3: undefined reference to `google::protobuf::internal::WireFormatLite::WriteStringMaybeAliased(int, std::string const&, google::protobuf::io::CodedOutputStream*)
Solution, change your current compiler to GCC Use
gcc --versionto make sure the correct version, and
which gccto check the softlinks to actual gcc
In Ubuntu 15.10 when using GCC 5.2, compiling might fail as Protobuf was compiled with wrong flags. Add the following to Makefile
CXXFLAGS += -D_GLIBCXX_USE_CXX11_ABI=0
Issues with Google Lib Protobuf (even after the ABI CXX11 fix)
sudo mkdir libprotobuf_
mv libprotobuf* libprotobuf_
apt-get source libprotobuf-dev
sudo make install
- mkdir build
- cd build
- cmake .. # Make kept giving me issues, but CMake worked.
- make all
- make test
- make runtest
- make pycaffe # To be able to import caffe
To be updated with instructions for Nervana
Remove unnecessary Ubuntu folders
When working with only terminal via ssh, you do not need some of these folders //
rm -rf ~/Desktop
rm -rf ~/Public# Is not persistent, after restart Ubuntu recreates this directory
rm -rf ~/Pictures
rm -rf ~/Music
rm -rf ~/Videos
rm -rf ~/Downloads
rm -rf ~/Templates
rm -rf ~/Documents
rm -rf ~/examples.desktop
These are my personal customization which I do on every machine I use. You may find them useful. They increase visual appeal and efficiency of working on Linux Machine.
chsh -s which zsh# Change the shell to ZSH (which zsh should be inside backticks, but Github Markdown messes up the formatting)
sudo reboot now# Requires restart
git clone firstname.lastname@example.org:iamaaditya/dotfiles.git
~/dotfiles/install.sh# Run the commands to make the shortcuts
git clone https://github.com/VundleVim/Vundle.vim.git ~/.vim/bundle/Vundle.vim# Install bundle
Solarized Color for vim
Powerline & associated fonts
pip install powerline-status# Powerline
wget https://github.com/Lokaltog/powerline/raw/develop/font/PowerlineSymbols.otf https://github.com/Lokaltog/powerline/raw/develop/font/10-powerline-symbols.conf
mkdir -p ~/.fonts/ && mv PowerlineSymbols.otf ~/.fonts/
fc-cache -vf ~/.fonts
mkdir -p ~/.config/fontconfig/conf.d/ && mv 10-powerline-symbols.conf ~/.config/fontconfig/conf.d/Note - if some aspects of powerline does not show up, check http://askubuntu.com/questions/283908/how-can-i-install-and-use-powerline-plugin
If you are curious on what this customization looks like, take a look at this screenshot. Here it shows Tmux session with multiple windows, and a multiple panes in the same window. Switching windows and panes in Tmux is very easy. Copy information across windows, sessions and servers is really at the tip of the finger. Also Solarized colors make it tolerable to use the screen for hours.