pva-faster-rcnn训练过程介绍

   pva-faster-rcnn是在py-faster-rcnn的基础之上，把cnn部分修改为pvanet网络。 PVANET是目前目标检测算法比较好的实现，也是Region Proposal系列方法的一个新实现，目前达到了real-time的单张图像50ms，压缩后的模型达到了real-time的单张图像30ms。

论文地址：https://www.arxiv.org/pdf/1608.08021v3.pdf
github项目:https://github.com/sanghoon/pva-faster-rcnn

pva-faster-rcnn的搭建
1.获取项目

git clone --recursive https://github.com/sanghoon/pva-faster-rcnn.git

2.编译建立Cython模块
安装python依赖

sudo pip install Cythonsudo pip install easydictcd pva-faster-rcnn/lib

此处需要修改lib下的setup.py第135行
GPU 计算能力查看地址'nvcc': ['-arch=sm_35',
https://developer.nvidia.com/cuda-gpus

比如我的是TESLA M40，计算能力为5.2，则修改为

'nvcc': ['-arch=sm_52',

执行

make all -j16

进行编译
3.编译建立Caffe和pycaffe

cd pva-faster-rcnn/caffe-fast-rcnncp Makefile.config.example Makefile.configgedit Makefile.config

修改Makefile.config文件:
需要修改的行如下

USE_CUDNN := 1
OPENCV_VERSION := 3
#CUDA directory contains bin/ and lib/ directories that we need.
CUDA_DIR := /usr/local/cuda
CUDA_DIR := /usr/local/cuda-8.0
# Whatever else you find you need goes here.
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial/
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-Linux-gnu/hdf5/serial/
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial/
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-Linux-gnu/hdf5/serial/
#Uncomment to support layers written in Python (will link against Python libs)
WITH_PYTHON_LAYER := 1
#Uncomment to use pkg-config to specify OpenCV library paths.
# (Usually not necessary -- OpenCV libraries are normally installed in one of the above $LIBRARY_DIRS.)
USE_PKG_CONFIG := 1

Makefile修改

cd ~/pva-faster-rcnn/caffe-faster-rcnn
gedit Makefile

原有

NVCCFLAGS += -ccbin=$(CXX) -Xcompiler -fPIC $(COMMON_FLAGS)

修改为

NVCCFLAGS += -D_FORCE_INLINES -ccbin=$(CXX) -Xcompiler -fPIC $(COMMON_FLAGS)

更新caffe-fast-rcnn的caffe部分，因为此项目caffe不是最新涉及到cudnn的计算会报错。

# Make sure to clone with --recursive
cd ~/pva-faster-rcnn/caffe-fast-rcnn
git remote add -f caffe https://github.com/BVLC/caffe.git
git merge -X theirs caffe/master
gedit include/caffe/layers/python_layer.hpp
# Remove
self_.attr("phase") = static_cast<int>(this->phase_);
cd ~/pva-faster-rcnn/caffe-faster-rcnn/python
for req in $(cat requirements.txt); do sudo -H pip install $req --upgrade; done

编译安装

mkdir build
cd build
cmake ..
make -j16 all
make -j16 pycaffe
make -j16 install

Caffe路径设置

gedit ~/.bashrc
export CAFFE_ROOT=~/pva-faster-rcnn/caffe-fast-rcnn
export PYTHONPATH=~/pva-faster-rcnn/caffe-fast-rcnn/python:$PYTHONPATH
source ~/.bashrc

检查安装

$ python
>>> import caffe
>>> caffe.version
sudo -H pip install easydict
sudo apt-get install python-gi-cairo python-tk

4.下载预训练模型
采用pva-faster-rcnn/models下的的shell脚本进行下载预训练模型，打开文件将pva9.1_pretrained_no_fc6.caffemodel放入$pva-faster-rcnn/models/pvanet/这个目录下

5.voc2007数据集下载
打开终端（任何目录）输入：

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar

解压（严格按照此顺序）：

tar xvf VOCtrainval_06-Nov-2007.tar
tar xvf VOCtest_06-Nov-2007.tar
tar xvf VOCdevkit_08-Jun-2007.tar

将生成的VOCdevkit文件夹更名为VOCdevkit2007移动至$pva-faster-rcnn/data/目录下面

6.训练

./tools/train_net.py --gpu 0 --solver models/pvanet/example_train/solver.prototxt --weights models/pvanet/pva9.1_pretrained_no_fc6.caffemodel --iters 100000 --cfg models/pvanet/cfgs/train.yml

7、测试

./tools/test_net.py --gpu 0 --def models/pvanet/example_train/test.pt --net models/pvanet/pvanet_frcnn_iter_100000.caffemodel --cfg models/pvanet/cfgs/submit_1019.yml