Skip to content

caffe2.md

Latest commit

 

History

History
44 lines (29 loc) · 1.74 KB

caffe2.md

File metadata and controls

44 lines (29 loc) · 1.74 KB
Raw

Run detectron for caffe2 in pytorch

    export PYTHONPATH=$PYTHONPATH:`pwd`

libcaffe2_detectron_ops_gpu.so is in /usr/local/lib/python2.7/dist-packages/torch, so need to modify /media/data/vesor_detectron/detectron/utils/env.py to add the following line (line71)

    prefixes.append('/usr/local/lib/python2.7/dist-packages/torch')

NOTE don't add /usr/local/lib/python2.7/dist-packages/torch to PYTHONPATH, it will crash when loading libtorch.so

Set envirotment var

caffe2 installed in /usr/local/caffe2 So we need to tell python the include path by setting PYTHONPATH. Note "YOURPATH=/xxx" is not the same as "export YOURPATH=/xxx", it won't affect the "env"

    env | grep YOURPATH

Note the pip / pip2 / pip3

You need to "pip install future" for caffe2, but I don't know why it installed in pip3. The "pip -V" show it is pip2, not pip3. Don't know what's wrong. So I have to use "pip2 install future"

Build with CPU feature

If you compile caffe2 using default configs, it will complain features like avx, avx2 and fma when running. facebookarchive/caffe2#1789

Change the CMakeLists.txt: option(USE_NATIVE_ARCH "Use -march=native" OFF) Set to "ON"

Specify your own OpenCV path

modify cmake/Dependencies.cmake, comment out the find opencv stuff, and add following:

    set(OPENCV_PATH /usr/local/opencv320)
    caffe2_include_directories(${OPENCV_PATH}/include)
    list(APPEND Caffe2_DEPENDENCY_LIBS ${OPENCV_PATH}/lib/libopencv_core.so)
    list(APPEND Caffe2_DEPENDENCY_LIBS ${OPENCV_PATH}/lib/libopencv_highgui.so)
    list(APPEND Caffe2_DEPENDENCY_LIBS ${OPENCV_PATH}/lib/libopencv_imgproc.so)
    list(APPEND Caffe2_DEPENDENCY_LIBS ${OPENCV_PATH}/lib/libopencv_imgcodecs.so)