Debian Installation

Caffe packages are available for several Debian versions, as shown in thefollowing chart:

  1. Your Distro     |  CPU_ONLY  |  CUDA  | Codename
  2. ----------------+------------+--------+-------------------
  3. Debian/oldstable|     ✘      |   ✘    | Jessie (8.0)
  4. Debian/stable   |     ✔      |   ✔    | Stretch (9.0)
  5. Debian/testing  |     ✔      |   ✔    | Buster
  6. Debian/unstable |     ✔      |   ✔    | Buster

Last update: 2017-07-08

Binary installation with APT

Apart from the installation methods based on source, Debian users can installpre-compiled Caffe packages from the official archive with APT.

Make sure that your /etc/apt/sources.list contains contrib and non-freesections if you want to install the CUDA version, for instance:

  1. deb http://ftp2.cn.debian.org/debian sid main contrib non-free

Then we update APT cache and directly install Caffe. Note, the cpu version andthe cuda version cannot coexist.

  1. $ sudo apt update
  2. $ sudo apt install [ caffe-cpu | caffe-cuda ]
  3. $ caffe                                              # command line interface working
  4. $ python3 -c 'import caffe; print(caffe.__path__)'   # python3 interface working

These Caffe packages should work for you out of box. However, the CUDA versionmay break if your NVIDIA driver and CUDA toolkit are not installed with APT.

Customizing caffe packages

Some users may need to customize the Caffe package. The way to customizethe package is beyond this guide. Here is only a brief guide of producingthe customized .deb packages.

Make sure that there is a dec-src source in your /etc/apt/sources.list,for instance:

  1. deb http://ftp2.cn.debian.org/debian sid main contrib non-free
  2. deb-src http://ftp2.cn.debian.org/debian sid main contrib non-free

Then we build caffe deb files with the following commands:

  1. $ sudo apt update
  2. $ sudo apt install build-essential debhelper devscripts  # standard package building tools
  3. $ sudo apt build-dep [ caffe-cpu | caffe-cuda ]          # the most elegant way to pull caffe build dependencies
  4. $ apt source [ caffe-cpu | caffe-cuda ]                  # download the source tarball and extract
  5. $ cd caffe-XXXX
  6. [ ... optional, customizing caffe code/build ... ]
  7. $ dch --local "Modified XXX"                             # bump package version and write changelog
  8. $ debuild -B -j4                                         # build caffe with 4 parallel jobs (similar to make -j4)
  9. [ ... building ...]
  10. $ debc                                                   # optional, if you want to check the package contents
  11. $ sudo debi                                              # optional, install the generated packages
  12. $ ls ../                                                 # optional, you will see the resulting packages

It is a BUG if the package failed to build without any change.The changelog will be installed at e.g. /usr/share/doc/caffe-cpu/changelog.Debian.gz.

Source installation

Source installation under Debian/unstable and Debian/testing is similar to that of Ubuntu, buthere is a more elegant way to pull caffe build dependencies:

  1. $ sudo apt build-dep [ caffe-cpu | caffe-cuda ]

Note, this requires a deb-src entry in your /etc/apt/sources.list.

Compiler Combinations

Some users may find their favorate compiler doesn’t work with CUDA.

  1. CXX compiler |  CUDA 7.5  |  CUDA 8.0  |  CUDA 9.0  |
  2. -------------+------------+------------+------------+
  3. GCC-8        |     ?      |     ?      |     ?      |
  4. GCC-7        |     ?      |     ?      |     ?      |
  5. GCC-6        |     ✘      |     ✘      |     ✔      |
  6. GCC-5        |     ✔ [1]  |     ✔      |     ✔      |
  7. -------------+------------+------------+------------+
  8. CLANG-4.0    |     ?      |     ?      |     ?      |
  9. CLANG-3.9    |     ✘      |     ✘      |     ✔      |
  10. CLANG-3.8    |     ?      |     ✔      |     ✔      |

[1] CUDA 7.5 ‘s host_config.h must be patched before working with GCC-5.

[2] CUDA 9.0: https://devblogs.nvidia.com/parallelforall/cuda-9-features-revealed/

BTW, please forget the GCC-4.X series, since its libstdc++ ABI is not compatible with GCC-5’s.You may encounter failure linking GCC-4.X object files against GCC-5 libraries.(See https://wiki.debian.org/GCC5 )

Notes

  • Consider re-compiling OpenBLAS locally with optimization flags for sake ofperformance. This is highly recommended for any kind of production use, includingacademic research.

  • If you are installing caffe-cuda, APT will automatically pull some of theCUDA packages and the nvidia driver packages. Please be careful if you havemanually installed or hacked nvidia driver or CUDA toolkit or any otherrelated stuff, because in this case APT may fail.

  • Additionally, a manpage (man caffe) and a bash complementation script(caffe <TAB><TAB>, caffe train <TAB><TAB>) are provided.Both of the two files are still not merged into caffe master.

  • The python interface is Python 3 version: python3-caffe-{cpu,cuda}.No plan to support python2.

  • If you encountered any problem related to the packaging system (e.g. failed to install caffe-*),please report bug to Debian via Debian’s bug tracking system. See https://www.debian.org/Bugs/ .Patches and suggestions are also welcome.

FAQ

  • where is caffe-cudnn?

CUDNN library seems not redistributable currently. If you really want thecaffe-cudnn deb packages, the workaround is to install cudnn by yourself,and hack the packaging scripts, then build your customized package.

  • I installed the CPU version. How can I switch to the CUDA version?

sudo apt install caffe-cuda, apt’s dependency resolver is smart enough to deal with this.

  • Where are the examples, the models and other documentation stuff?
  1. $ sudo apt install caffe-doc
  2. $ dpkg -L caffe-doc