Run rbgirshick/py-faster-rcnn on ubuntu with gpu

1. Setup

git clone --recursive https://github.com/rbgirshick/py-faster-rcnn.git
cd py-faster-rcnn/caffe-fast-rcnn

# edit Makefile.config
# set CPU_ONLY or cuda path
# set USE_OPENCV
# set OPENCV_VERSION
# set PYTHON_INCLUDE (python2.7, python3.5, or anaconda, prefer to python 2.7)
# set PYTHON_LIB
# set WITH_PYTHON_LAYER

make all -j8
cd python && pip install -r requirements.txt
make pycaffe
export PYTHONPATH=$PYTHONPATH:`pwd`
# test
python -c 'import caffe'

cd ../../lib
make 
python -c 'import fast_rcnn'

1.1 Problem 1

Since caffe in py-faster-rcnn is not the latest version, so when using cudnn, some errors may occur. Replace cudnn-related files in caffe-fast-rcnn file. More details, see [here].(http://blog.csdn.net/u011070171/article/details/52399317)

1.2 Problem 2

While pip install -R requirements.txt, install gfortran first. gfortran is needed for scipy.

sudo apt-get install libgfortran3 build-essential gfortran python-all-dev libatlas-base-dev 

2. Run a demo

cd $FRCN_ROOT
./tools/demo.py

3. Train VGG16 network on VOC 2007

./tools/train_net.py --gpu 0 --solver models/pascal_voc/VGG16/fast_rcnn/solver.prototxt --weights data/imagenet_models/VGG16.v2.caffemodel

to be continued.