Skip to content

Latest commit

 

History

History
260 lines (236 loc) · 7.93 KB

build-caffe-matlab.md

File metadata and controls

260 lines (236 loc) · 7.93 KB
Raw

Build caffe on Ubuntu

On Ubuntu 18.04, gcc/g++ 6.5 and Matlab 2018a is needed. install them first. Install gcc6/g++6

sudo apt-get install gcc-6 g++-6
sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-6 2 --slave /usr/bin/g++ g++ /usr/bin/g++-6
sudo update-alternatives --config gcc  # choose gcc 6
# make sure is gcc 6 and g++6
gcc -v
g++ -v

On Ubuntu 16.04 gcc/g++ 6.3 and Matlab 2018a is needed. you should build gcc/g++ 6.3 first, and then:

sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-6 2 --slave /usr/bin/g++ g++ /usr/bin/g++-6
sudo update-alternatives --config gcc  # choose gcc 6
# make sure is gcc 6 and g++6
gcc -v
g++ -v

If you will build caffe with CUDA 10, please update CMAKE to 3.14.3.

1. Install denpendencies

sudo apt-get update 
sudo apt-get install --yes cmake git build-essential
sudo apt-get install --yes libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
sudo apt-get install --yes --no-install-recommends libboost-all-dev
sudo apt-get install --yes libopenblas-dev liblapack-dev libatlas-base-dev
sudo apt-get install --yes libgflags-dev libgoogle-glog-dev liblmdb-dev
sudo apt-get install --yes python-pip
sudo apt-get install --yes graphviz

2. Clone caffe from github

git clone https://github.com/BVLC/caffe
cd caffe

3. Install pycaffe denpendencies

  • modify matplotlib>=1.3.1 in python/requirements.txt to matplotlib==2.2.4
  • modify ipython>=3.0.0 in python/requirements.txt to ipython==5.0.0
  • modify scikit-image>=0.9.3 to scikit-image==0.14.2
  • modify network>=1.8.1 to network==2.2

run command

pip install pydotplus
pip install -r python/requirements.txt

4. Build caffe

  • Modify cmake configuration

    • modify line 357 in cmake/Utils.cmake from: 
        message(FATAL_ERROR "Logic error. Need to update cmake script")
      to 
       # message(FATAL_ERROR "Logic error. Need to update cmake script")
    • add two lines to CMakeLists.txt, from: 
      ...
if(POLICY CMP0054)
  cmake_policy(SET CMP0054 NEW)
endif()
      to 
      ...
if(POLICY CMP0054)
  cmake_policy(SET CMP0054 NEW)
endif()
include_directories(include)
include_directories(${CUDA_TOOLKIT_ROOT_DIR}/include)  # CUDA include directory

  • Run command:

    mkdir cmake-build
cd cmake-build
cmake -D CMAKE_INSTALL_PREFIX=/opt/caffe \
-D CMAKE_BUILD_TYPE=Release \
-D OpenCV_DIR=/usr/share/OpenCV \
-D BUILD_matlab=ON \
-D Matlab_DIR=/usr/local/MATLAB/R2018a \
-D CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-9.0 ..

    CMAKE_INSTALL_PREFIX specify the install directory, caffe will be install to /opt/caffe. OpenCV_DIR specify the default OpenCV to use. BUILD_matlab set on to build matcaffe. Matlab_DIR specify the root directory of Matlab.

    cmake's output is like:

    ******************* Caffe Configuration Summary *******************
-- General:
...
-- 
--   BUILD_SHARED_LIBS :   ON
--   BUILD_python      :   ON
--   BUILD_matlab      :   ON
--   BUILD_docs        :   ON
--   CPU_ONLY          :   OFF
--   USE_OPENCV        :   ON
--   USE_LEVELDB       :   ON
--   USE_LMDB          :   ON
--   USE_NCCL          :   OFF
--   ALLOW_LMDB_NOLOCK :   OFF
--   USE_HDF5          :   ON
-- 
-- Dependencies:
--   BLAS              :   Yes (Atlas)
--   Boost             :   Yes (ver. 1.65)
--   glog              :   Yes
--   gflags            :   Yes
--   protobuf          :   Yes (ver. 3.0.0)
--   lmdb              :   Yes (ver. 0.9.21)
--   LevelDB           :   Yes (ver. 1.20)
--   Snappy            :   Yes (ver. ..)
--   OpenCV            :   Yes (ver. 3.2.0) # 3.2.0 or 2.4.9
--   CUDA              :   Yes (ver. 9.0)  # can be No, if you don't have cuda
-- 
-- NVIDIA CUDA:
--   Target GPU(s)     :   Auto
--   GPU arch(s)       :   sm_61
--   cuDNN             :   Yes (ver. 7.3.0)
-- 
-- Python:
--   Interpreter       :   /usr/bin/python2.7 (ver. 2.7.15)
--   Libraries         :   /usr/lib/x86_64-linux-gnu/libpython2.7.so (ver 2.7.15rc1)
--   NumPy             :   /home/creator/.local/lib/python2.7/site-packages/numpy/core/include (ver 1.16.2)
-- 
-- Matlab:
--   Matlab            :   Yes (/usr/local/MATLAB/R2018a/bin/mex, /usr/local/MATLAB/R2018a/bin/mexext
--   Octave            :   Yes (/usr/bin/mkoctfile)
--   Build mex using   :   Matlab
-- 
-- Documentaion:
--   Doxygen           :   /usr/bin/doxygen (1.8.13)
--   config_file       :   /home/creator/Downloads/caffe-master/.Doxyfile
-- 
-- Install:
--   Install path      :   /opt/caffe

    Matlab should be found.

  • Make:

    # replace 8 with your cpu kernel count
make all -j8
sudo make install
sudo chmod -R 777 /opt/caffe
# remove origin caffe
rm ~/.local/lib/python2.7/site-packages/caffe
# make symbolic to python libs
ln -s /opt/caffe/python/caffe ~/.local/lib/python2.7/site-packages/
# copy matlab directory to install directory
cp -r ../matlab/* /opt/caffe/matlab

  • Modify ~/.bashrc add

    export PATH=/opt/caffe/bin:$PATH  # (optional)
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/caffe/lib

    to the end of file ~/.bashrc

5. Test pycaffe

create caffe_test.py, write code:

import sys
import os

import caffe

if __name__ == '__main__':
	caffe.set_mode_cpu()
	print 'You have good luck!'

then:

python caffe_test.py

if there is no error message, congratulations, you made it!

Install code hints file for Pycharm

See: https://github.com/Mannix1994/PythonResources

6. Test matcaffe

Must run Matlab in a new Terminal.

matlab # run matlab in a terminal

Open Matlab and write test.m

PATH_TO_CAFFE_MATLAB='/opt/caffe/matlab';
addpath(genpath(PATH_TO_CAFFE_MATLAB));
caffe.set_mode_cpu();

Run test.m, there are some errors:

Invalid MEX-file '/opt/caffe/matlab/+caffe/private/caffe_.mexa64': /usr/lib/libgdal.so.20: symbol
TIFFReadRGBATileExt version LIBTIFF_4.0 not defined in file libtiff.so.5 with link time reference.

Error in caffe.set_mode_cpu (line 5)
caffe_('set_mode_cpu');

Error in test (line 3)
caffe.set_mode_cpu();

It looks like the lib LIBTIFF's problem. This is because Matlab will load the library of its self, we compile caffe and link system libtiff. so we should preload system libtiff.

Firstly, find where libtiff is:

locate libtiff.so

output:

/snap/gimp/113/usr/lib/x86_64-linux-gnu/libtiff.so.5
...
/snap/vlc/770/usr/lib/x86_64-linux-gnu/libtiff.so.5.2.4
/usr/lib/x86_64-linux-gnu/libtiff.so
/usr/lib/x86_64-linux-gnu/libtiff.so.5
/usr/lib/x86_64-linux-gnu/libtiff.so.5.3.0
/usr/local/MATLAB/R2018a/bin/glnxa64/libtiff.so.5
/usr/local/MATLAB/R2018a/bin/glnxa64/libtiff.so.5.0.5

We should choose the system libtiff /usr/lib/x86_64-linux-gnu/libtiff.so.5not the one in snap/ or the one in /usr/local/MATLAB/.
Open terminal and run:

# preload some libs
export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtiff.so.5
matlab  # run matlab from terminal

Then run test.m in matlab prompt, there will be no error. When you run matcaffe, if there are any problem about lib, you can solve it by the method mentioned above.
Some other libs need to be preload:

export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtiff.so.5:/usr/lib/x86_64-linux-gnu/libhdf5_cpp.so:/usr/lib/x86_64-linux-gnu/libhdf5_hl_cpp.so:/usr/lib/x86_64-linux-gnu/libhdf5_serial.so:/usr/lib/x86_64-linux-gnu/libhdf5_serial_fortran.so:/usr/lib/x86_64-linux-gnu/libhdf5_serial_hl.so:/usr/lib/x86_64-linux-gnu/hdf5/serial/libhdf5.so
matlab # run matlab from terminal

Good luck.