Deep Learning for POWER

IBM Deep Learning packages now include:

   - caffe-bvlc (upstream Berkeley Vision and Learning Center)
   - caffe-nv (NVIDIA fork)
   - digits (NVIDIA DIGITS)
   - theano
   - torch
   - power-mldl (meta-package to install all above)

Pre-requisites now include cuDNN v5.1 .deb packages:

   1. Ubuntu 14.04
   2. CUDA 7.5
   3. cuDNN v5.1 for CUDA 7.5 .debs (https://developer.nvidia.com/cudnn)
      - cuDNN v5.1 Runtime Library for Power8 (Deb)
      - cuDNN v5.1 Developer Library for Power8 (Deb)
      - cuDNN v5.1 Code Samples and User Guide Power8 (Deb)
   4. "trusty-backports universe" and "trusty multiverse" repos
      $ sudo apt-get install software-properties-common
      $ sudo apt-add-repository "deb http://ports.ubuntu.com/ubuntu-ports/ trusty-backports universe"
      $ sudo apt-add-repository "deb http://ports.ubuntu.com/ubuntu-ports/ trusty multiverse"
      $ sudo apt-get update

Pre-requisites (for DIGITS) include a liblmdb0 backport. Installer may
suggest to omit Deep Learning packages rather than install the backport.
Rejecting that suggestion should cause the installer to offer to install
the liblmdb0 backport.

UPGRADING:

The easiest method to upgrade from the previous release is to uninstall
the old packages first then install the cuDNN .debs and new packages:

    $ sudo apt-get purge libopenblas cudnn-req caffe torch theano
    $ sudo dpkg -i libcudnn*.deb
    $ sudo apt-get install power-mldl

NOTES:

Each package provides an environment setup script. We recommend updating
shell rc file (e.g. `.bashrc`) to source setup scripts:

        source /opt/DL/<framework>/bin/<framework>-activate

Each package provides a test script to run some self-tests:

        $ <framework>-test

digits-activate script will create $HOME/.digits for cfg and log files.

Caffe packages install to separate directories (/opt/DL/caffe-{bvlc|nv}).
By default caffe-nv is presented as /opt/DL/caffe; can be changed with:

        $ sudo update-alternatives --config caffe

Caffe packages include scripts to copy samples to a local directory:

        $ caffe-install-samples <somedir>