Caffe

This is Zhuo's fork of Caffe, which fix build(compilation, linking, dependencies) problems. It's not hacking, but better maintainance.

Caffe is the best, really ONNX for model conversion for mobile deployment!

Installation & Dependencies

Linux Only tested under ubuntu 16.04

1. install conda & python & pip

• miniconda / anaconda are both OK.
• use python3 since python2 will die from 2020. On default python version set to 3 in CMakeLists.txt
• better management for different python minor versions
• don't need apt install python packages, and also don't recommend
• if which pip returns /usr/local/bin/pip, do this:

sudo pip uninstall pip
conda install pip

2. install cmake

• manually install cmake, instead of apt install cmake since this only gives cmake 3.5.1, which is too old and buggy.
• 3.14 or later is recomended
• manually install cmake binary from official website, or conda install cmake
• cmake were born for cross-platform make, however, Caffe's cmake script is not cross-platform. Let's fix it.

3. install boost

• boost if heavy, but caffe use boost. I'll try to replace boost with C++11 gradually
• ubuntu16.04's apt provided boost-python is only for python2. Python2 will retire in 2020. Hence don't use apt to install boost-python if you're using python3.
• just conda install it: conda install boost and conda install py-boost. For python3.7, it provides boost 1.67
• cmake's finding boost scripts were buggy. Being fixed but may still not work for other not tested versions.
• modify cmake/Custom.cmake to specifying boost include directory implicitly
• on MacOSX, either brew installed boost or conda installed boost are not OK with Caffe + Python3.

4. HDF5: turned off by default

• I've never been using hdf5 format for both training and inference
• the cmake scripts were wrongly finding HDF5 dependency whatever you turn on or off it. Fixed.

5. leveldb: turned off by default

• thus we don't need snappy on default

6. lmdb: We need it. It's kept for training purpose thou many modern boys only use tf/pytorch/mxnet. sudo apt install liblmdb-dev

7. opencv: we need it. For load image and preprocess. tested version: manually compiled opencv 4.1.1 / 3.4.5's C/C++ lib, 4.1.1's python lib.

Please compile it like this:

cmake .. \
    -DCMAKE_INSTALL_PREFIX=/usr/local/opencv-4.1 \
    -DINSTALL_PYTHON_EXAMPLES=ON \
    -DOPENCV_PYTHON3_VERSION=3.7 \
    -DPYTHON3_EXECUTABLE=/home/zz/soft/miniconda3/bin/python \
    -DPYTHON3_INCLUDE_DIR=/home/zz/soft/miniconda3/include/python3.7m \
    -DPYTHON3_LIBRARY=/home/zz/soft/miniconda3/lib/libpython3.7m.so \
    -DBUILD_opencv_python3=ON \
    -DBUILD_opencv_python2=OFF \
    -DPYTHON_DEFAULT_EXECUTABLE=/home/zz/soft/miniconda3/bin/python \
    -DHAVE_opencv_python3=ON \
    -DBUILD_TIFF=ON

make -j8
sudo make install

cd python_loader
python setup.py develop

8. cuda/cudnn: we need it for NVidia GPU speedup.

9. protobuf: we need it for model loading and saving. tested version: sudo apt install libprotobuf-dev protobuf-compiler

10. gflags: I'll remove it gradually.

11. glog: I'll remove it gradually. Will be replaced by spdlog Now include/spdlog folder using spdlog v1.x code.

12. blas: use OpenBLAS by default. sudo apt install libopenblas-dev

13. Python packages

• numpy: pip install numpy
• skimage: pip install scikit-image
• jupyter notebook: conda install jupyter notebook
• protobuf: must be equal to the version used in cmake (for C/C++). e.g. pip install protobuf==2.6.1.

License and Citation

Caffe is released under the BSD 2-Clause license. The BAIR/BVLC reference models are released for unrestricted use.

Please cite Caffe in your publications if it helps your research:

@article{jia2014caffe,
  Author = {Jia, Yangqing and Shelhamer, Evan and Donahue, Jeff and Karayev, Sergey and Long, Jonathan and Girshick, Ross and Guadarrama, Sergio and Darrell, Trevor},
  Journal = {arXiv preprint arXiv:1408.5093},
  Title = {Caffe: Convolutional Architecture for Fast Feature Embedding},
  Year = {2014}
}