### Caffe

时间:2021-07-11 12:08:08

Caffe学习。

#@author:       gr
#@date: 2015-08-30
#@email: forgerui@gmail.com

1. Install

详细可以见官方文档博客1博客2

1.1 Prerequisites

  • CUDA is required for GPU mode.

    library version 7.0 and the latest driver version are recommended, but 6.* is fine too

    5.5, and 5.0 are compatible but considered legacy
  • BLAS via ATLAS, MKL, or OpenBLAS.
  • Boost >= 1.55
  • OpenCV >= 2.4 including 3.0
  • protobuf, glog, gflags
  • IO libraries hdf5, leveldb, snappy, lmdb

Caffe requires BLAS as the backend of its matrix and vector computations. There are several implementations of this library. The choice is yours:

  • ATLAS: free, open source, and so the default for Caffe.
  • Intel MKL: commercial and optimized for Intel CPUs, with a free trial and student licenses.

    Install MKL.

    Set BLAS := mkl in Makefile.config
  • OpenBLAS: free and open source; this optimized and parallel BLAS could require more effort to install, although it might offer a speedup.

    Install OpenBLAS

    Set BLAS := open in Makefile.config

我们这里使用atlas。

1.2 Compilation

  1. 拷贝配置文件

     cp Makefile.config.example Makefile.config
  2. 在Makefile.config文件中第73行LIBRARY_DIRS加上atlas库所在的位置,我的在/usr/lib64/atlas/,修改后:

     LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib64/atlas/
  3. Makefile文件中在236行将boost_thread修改为boost_thread-mt,修改后:

     LIBRARIES += boost_thread-mt stdc++
  4. 编译

     make all -j 20            #多核编译,根据机子情况选定
  5. 编译matlab

    修改Makefile.config,MATLAB_DIR中加入matlab在机器中的位置:

     MATLAB_DIR := /usr/local/MATLAB/MATLAB_Production_Server/R2013a

    编译:

     make matcaffe -j 20
  6. 编译python

    修改Makefile.config,将PYTHON_INCLUDE, PYTHON_LIB修改为你机子正确的配置。

     PYTHON_INCLUDE := /usr/local/include/python2.7 \
    /usr/lib/python2.7/site-packages/numpy/core/include/numpy/ PYTHON_LIB := /usr/local/lib

    编译:

     make pycaffe -j 20

    注意:如果遇到如下问题,

     /usr/bin/ld: /usr/local/lib/libpython2.7.a(abstract.o): relocation R_X86_64_32 against `a local symbol' can not be used when making a shared object; recompile with -fPIC
    /usr/local/lib/libpython2.7.a: could not read symbols: Bad value
    collect2: ld returned 1 exit status

    可以下载python,加上--enable-shared-fPIC选项重新编译安装,命令如下:

     ./configure --prefix=/usr/local/  --enable-shared CFLAGS=-fPIC
    make
    make install

2. Usage

2.1 caffe中的例子

可以参见博客

2.1.1 mnist

mnist的网络框架在文件examples/mnist/lenet.prototxt中。分别运行如下命令,即可实现mnist:

sh data/mnist/get_mnist.sh
sh examples/mnist/create_mnist.sh
sh examples/mnist/train_lenet.sh

最后运行的结果,可以看到accuracy = 0.9907

I0830 21:56:59.506049 12371 solver.cpp:326] Iteration 10000, loss = 0.00290909
I0830 21:56:59.506080 12371 solver.cpp:346] Iteration 10000, Testing net (#0)
I0830 21:57:00.983238 12371 solver.cpp:414] Test net output #0: accuracy = 0.9907
I0830 21:57:00.983290 12371 solver.cpp:414] Test net output #1: loss = 0.0304467 (* 1 = 0.0304467 loss)
I0830 21:57:00.983304 12371 solver.cpp:331] Optimization Done.
I0830 21:57:00.983314 12371 caffe.cpp:214] Optimization Done.
2.1.2 cifair
sh data/cifar10/get_cifar10.sh
sh examples/cifar10/create_cifar10.sh
sh examples/cifar10/train_quick.sh

2.2 caffe 框架学习

2.2.1 框架

caffe的框架如下:

### Caffe

  1. 预处理图像的leveldb构建

    输入:一批图像和label (2和3)

    输出:leveldb (4)

    指令里包含如下信息:

    conver_imageset (构建leveldb的可运行程序)

    train/ (此目录放处理的jpg或者其他格式的图像)

    label.txt (图像文件名及其label信息)

    输出的leveldb文件夹的名字

    CPU/GPU (指定是在cpu上还是在gpu上运行code)

  2. CNN网络配置文件

    Imagenet_solver.prototxt (包含全局参数的配置的文件)

    Imagenet.prototxt (包含训练网络的配置的文件)

    Imagenet_val.prototxt (包含测试网络的配置文件)

2.2.2 Caffe层次

**Blob: **基础的数据结构,是用来保存学习到的参数以及网络传输过程中产生数据的类。

**Layer: **是网络的基本单元,由此派生出了各种层类。修改这部分的人主要是研究特征表达方向的。

**Net: **是网络的搭建,将Layer所派生出层类组合成网络。

**Solver: **是Net的求解,修改这部分人主要会是研究DL求解方向的。

2.3 RCNN

Training your own R-CNN detector on PASCAL VOC

!!! tvmonitor : 0.6483 0.6614
~~~~~~~~~~~~~~~~~~~~
Results:
0.6428
0.6963
0.5016
0.4191
0.3191
0.6251
0.7087
0.6036
0.3266
0.5852
0.4627
0.5616
0.6037
0.6684
0.5414
0.3157
0.5285
0.4889
0.5772
0.6483 0.5412 ~~~~~~~~~~~~~~~~~~~~ test_results = 1x20 struct array with fields: recall
prec
ap
ap_auc

Reference

1. http://caffe.berkeleyvision.org/installation.html

2. http://www.rthpc.com/plus/view.php?aid=351

3. http://www.cnblogs.com/platero/p/3993877.html

4. http://www.csdn.net/article/2015-01-22/2823663