TensorFlow with support for GPU and python on NVIDIA GTX 950 in Debian Stretch

Why is there a need for this recipe at all; why doesn't TensorFlow and CUDA just work together?

There are two underlying problems:

1. TensorFlow is not packaged by Debian, and the versions of TensorFlow provided by google assume that the CUDA libraries (in particular the cuDNN library and the header file for the cupti library) are located in non-standard locations. Yes, non-standard, not just where Debian happens to put libraries and include files. Since TensorFlow is free software, we can fix it.
2. The deep-learning part of CUDA, cuDNN, is proprietary software, and is outdated in that it requires outdated versions of gcc. And since it is proprietary, the free software community cannot fix it. Instead, we have to install obsolete versions of gcc, however, different versions of gcc can be installed simultaneously, and we don't have to make the obsolete version (4.9) the system default, so other things relying on gcc will unaffected.

We will first build TensorFlow with support for

• cuDNN (nvidias deep-learning library)
• cudaToolKit ()

Then we will build a python-package which uses our custom build of TensorFlow.

1. Install the Debian base system, and add contrib non-free to /etc/apt/sources.list
2. Install the nvidia kernel modules and the nvidia-cuda-toolkit from debian non-free
3. Install nvidias deeplearning library cuDNN. You have to get these packages directly from nvidia here. You will have to register as a nvidia developer to get access to these packages. The registration is free. The packages needed are available at "Download cuDNN v7.0.5 (Dec 5, 2017), for CUDA 8.0":
   • cuDNN v7.0.5 Runtime Library for Ubuntu16.04 (Deb)
   • cuDNN v7.0.5 Developer Library for Ubuntu16.04 (Deb)
   • after downloading them, install them with dpkg:

sudo dpkg -i libcudnn7-dev_7.0.5.15-1+cuda8.0_amd64.deb libcudnn7_7.0.5.15-1+cuda8.0_amd64.deb

• Optional: Download and install the documentation as well: libcudnn7-doc_7.0.5.15-1+cuda8.0_amd64.deb
  1. Install git, curl and some python development libraries python-numpy python-dev python-pip python-wheel python-setuptools and openjdk-8-jdk (required by bazel)

sudo apt-get install git curl python-numpy python-dev python-pip python-wheel python-setuptools openjdk-8-jdk

1. Get the source code for tensorflow: git clone https://github.com/tensorflow/tensorflow
2. Install bazel (from googles repository, I don't know why bazel is not packaged by debian yet)
   • 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
3. cd tensorflow
4. git checkout r1.5
5. Install obsolete version of gcc, gcc 4.9 (improved version of https://github.com/dasGringuen/debian9-install)
   • ./configure
   • say Yes to CUDA
   • make a directory /usr/local/cuda
   • make symlink from /usr/lib/x86_64-linux-gnu to /usr/local/coda/lib64

:/usr/local/cuda# ls -l
total 0
lrwxrwxrwx 1 root staff 25 feb  9 23:55 lib64 -> /usr/lib/x86_64-linux-gnu

• Say that you have version 7.0.5 of libcudnn when asked about it.
  1. Compile
• export LC_ALL=en_US.utf8
• export LANG=en_US.utf8
• make a symlink from

/usr/local/cuda# ls -l include
lrwxrwxrwx 1 root staff 12 feb 10 00:16 include -> /usr/include

• create a dir /usr/local/cuda/nvvm
• create a dir /usr/local/cuda/extras/CUPTI
• create a symlink

:/usr/local/cuda/extras/CUPTI# ln -s /usr/lib/x86_64-linux-gnu include

• compile. For low RAM requirement include --local_resources 2048,.5,1.0

bazel build --local_resources 2048,.5,1.0 --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

16. Build and install a wheel

• build the wheel

bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

• install it too

sudo pip3 install /tmp/tensorflow_pkg/tensorflow-1.5.0-cp35-cp35m-linux_x86_64.whl

• sudo pip install /tmp/tensorflow_pkg/tensorflow-1.5.0-cp27-cp27mu-linux_x86_64.whl
  1. Install DeepBach and its dependencies:
• git clone https://github.com/Ghadjeres/DeepBach
• Musescore

sudo apt-get install musescore

• Install stuff needed by the python module music21
• matplotlib scipy

sudo apt-get install python-matplotlib python-scipy

• sudo pip install music21
• Inform python about where the music21 binary is located by running this code in the interactive python environment. Start that environment by just typing python in a terminal

$ python

Once inside, issue the following:

import music21
us=music21.environment.UserSettings()
us['musicxmlPath']='/usr/bin/musescore'
us['warnings'] = 0
exit()

The last line:

us['warnings'] = 0

Is optional, and only relevant if you run python 2. If you run python 3, it is pointless (but also harmless).

What does this do? It creates a file .music21rc with among other things has this element:

<preference name="musicxmlPath" value="/usr/bin/musescore" />

Back to the dependencies of DeepBach, we are done with musescore, next is support for Hierarchical Data Format library, version 5, aka HDF5, packaged in Debian as python-h5py:

sudo apt-get install python-h5py

• keras, not available in Debian.

    sudo pip install keras
    

• make, available in Debian:

sudo apt-get install python-mako

• nose, available in Debian:

sudo apt-get install python-nose

• libprotobuf-c1 (not sure if needed)

sudo apt-get install libprotobuf-c1

• python-setuptools python-tqdm python-werkzeug python-sqlite python-wheel python-lzma (xz) python-appdirs python-click python-flask python-packaging

sudo apt-get install python-setuptools python-tqdm python-werkzeug python-sqlite python-wheel python-lzma (xz) python-appdirs python-click python-flask python-packaging

• ignore gpuarray (old stuff, only required if you use theano, it seems), mkl (hey, we want GPU computing, right), tensorflow-gpu (we have compiled our own version of tensorflow which supports gpu)
• Put deepBachMuseScore.qml file in your MuseScore2/Plugins directory:
  • According to `man musescore`, the per user plugins directory for musescore2 is ~/.local/share/data/MuseScore/MuseScore2/ and deepBachMuseScore.qml resides in the root directory of DeepBach, so

cp ~/DeepBach/deepBachMuseScore.qml .local/share/data/MuseScore/MuseScore2/plugins/

• Fix the server, the model_name is "deepbach", not "skip_new". Change in DeepBach/plugin_flask_server.py:

model_name = 'deepbach'
# model_name = 'skip_new'

• Start the server

cd /home/hans/DeepBach/DeepBach
export FLASK_APP=/home/hans/DeepBach/plugin_flask_server.py
python -m flask run --host=0.0.0.0

• To access the server from another computer, put the deepBachMuseScore.qml in your $HOME/.local/share/data/MuseScore/MuseScore2/plugins/ on the client, and change adress of the server, eg:

            /* text: "http://localhost:5000/" */
            text: "http://foo.bar.com:5000/"

Python 3

sudo apt-get install python3-pip python3-setuptools python3-matplotlib python3-scipy python3-numpy python3-h5py python3-mako python3-nose libprotobuf-c1 python3-tqdm python3-werkzeug python3-wheel python3-xopen python3-appdirs python3-click python3-flask python3-packaging
sudo pip3 install keras music21 tensorflow
python3

import music21
us=music21.environment.UserSettings()
us['musicxmlPath']='/usr/bin/musescore'
exit()

cd /home/hans/DeepBach/DeepBach
export FLASK_APP=/home/hans/DeepBach/plugin_flask_server.py
python3 -m flask run --host=0.0.0.0

[14:45:11]hans@Haleno:~>sudo pip3 install tensorflow
Collecting tensorflow
  Using cached tensorflow-1.6.0-cp35-cp35m-manylinux1_x86_64.whl
Collecting termcolor>=1.1.0 (from tensorflow)
  Downloading termcolor-1.1.0.tar.gz
Requirement already satisfied: wheel>=0.26 in /usr/lib/python3/dist-packages (from tensorflow)
Collecting protobuf>=3.4.0 (from tensorflow)
  Using cached protobuf-3.5.2-cp35-cp35m-manylinux1_x86_64.whl
Collecting numpy>=1.13.3 (from tensorflow)
  Downloading numpy-1.14.2-cp35-cp35m-manylinux1_x86_64.whl (12.1MB)
Collecting gast>=0.2.0 (from tensorflow)
  Downloading gast-0.2.0.tar.gz
Collecting absl-py>=0.1.6 (from tensorflow)
  Using cached absl-py-0.1.11.tar.gz
Collecting grpcio>=1.8.6 (from tensorflow)
  Using cached grpcio-1.10.0-cp35-cp35m-manylinux1_x86_64.whl
Collecting tensorboard<1.7.0,>=1.6.0 (from tensorflow)
  Using cached tensorboard-1.6.0-py3-none-any.whl
Requirement already satisfied: six>=1.10.0 in /usr/lib/python3/dist-packages (from tensorflow)
Collecting astor>=0.6.0 (from tensorflow)
  Using cached astor-0.6.2-py2.py3-none-any.whl
Requirement already satisfied: setuptools in /usr/lib/python3/dist-packages (from protobuf>=3.4.0->tensorflow)
Collecting bleach==1.5.0 (from tensorboard<1.7.0,>=1.6.0->tensorflow)
  Downloading bleach-1.5.0-py2.py3-none-any.whl
Requirement already satisfied: werkzeug>=0.11.10 in /usr/lib/python3/dist-packages (from tensorboard<1.7.0,>=1.6.0->tensorflow)
Collecting markdown>=2.6.8 (from tensorboard<1.7.0,>=1.6.0->tensorflow)
  Downloading Markdown-2.6.11-py2.py3-none-any.whl (78kB)
Collecting html5lib==0.9999999 (from tensorboard<1.7.0,>=1.6.0->tensorflow)
  Downloading html5lib-0.9999999.tar.gz (889kB)
Building wheels for collected packages: termcolor, gast, absl-py, html5lib
  Running setup.py bdist_wheel for termcolor ... done
  Stored in directory: /root/.cache/pip/wheels/de/f7/bf/1bcac7bf30549e6a4957382e2ecab04c88e513117207067b03
  Running setup.py bdist_wheel for gast ... done
  Stored in directory: /root/.cache/pip/wheels/8e/fa/d6/77dd17d18ea23fd7b860e02623d27c1be451521af40dd4a13e
  Running setup.py bdist_wheel for absl-py ... done
  Stored in directory: /root/.cache/pip/wheels/3c/0f/0a/6c94612a8c26070755559045612ca3645fea91c11f2148363e
  Running setup.py bdist_wheel for html5lib ... done
  Stored in directory: /root/.cache/pip/wheels/6f/85/6c/56b8e1292c6214c4eb73b9dda50f53e8e977bf65989373c962
Successfully built termcolor gast absl-py html5lib
Installing collected packages: termcolor, protobuf, numpy, gast, absl-py, grpcio, html5lib, bleach, markdown, tensorboard, astor, tensorflow
  Found existing installation: numpy 1.12.1
    Not uninstalling numpy at /usr/lib/python3/dist-packages, outside environment /usr
Successfully installed absl-py-0.1.11 astor-0.6.2 bleach-1.5.0 gast-0.2.0 grpcio-1.10.0 html5lib-0.9999999 markdown-2.6.11 numpy-1.14.2 protobuf-3.5.2 tensorboard-1.6.0 tensorflow-1.6.0 termcolor-1.1.0

comments powered by Disqus

Back to the index

Blog roll

R-bloggers, Debian Weekly

---

Last modified: mars 25, 2018