Caffe602是一个清晰而高效的深度学习258框架,其作者是博士毕业于UC Berkeley的贾扬清1.7K,目前在Google100工作。
Caffe52是纯粹的C++/CUDA架构,支持命令行、Python和MATLAB接口;可以在CPU和GPU163直接无缝切换:
Caffe::set_mode(Caffe::GPU);
Caffe的优势
- 上手快:模型与相应优化都是以文本形式而非代码形式给出。
Caffe给出了模型的定义、最优化设置以及预训练的权重,方便立即上手。 - 速度快:能够运行最棒的模型与海量的数据。
Caffe与cuDNN结合使用,测试AlexNet模型,在K40上处理每张图片只需要1.17ms. - 模块化:方便扩展到新的任务和设置上。
可以使用Caffe提供的各层类型来定义自己的模型。 - 开放性:公开的代码和参考模型用于再现。
- 社区好:可以通过BSD-2参与开发与讨论。
Caffe的网络定义
Caffe中的网络都是有向无环图的集合,可以直接定义:
name: "dummy-net"
layers {name: "data" …}
layers {name: "conv" …}
layers {name: "pool" …}
layers {name: "loss" …}
数据及其导数以blobs的形式在层间流动。
Caffe的各层定义
Caffe层的定义由2部分组成:层属性与层参数,例如
name:"conv1"
type:CONVOLUTION
bottom:"data"
top:"conv1"
convolution_param{
num_output:20
kernel_size:5
stride:1
weight_filler{
type: "xavier"
}
}
这段配置文件的前4行是层属性,定义了层名称、层类型以及层连接结构(输入blob和输出blob);而后半部分是各种层参数。
Blob
Blob是用以存储数据的4维数组,例如
- 对于数据:Number*Channel*Height*Width
- 对于卷积权重:Output*Input*Height*Width
- 对于卷积偏置:Output*1*1*1
训练网络
网络参数的定义也非常方便,可以随意设置相应参数。
甚至调用GPU运算只需要写一句话:
solver_mode:GPU
![](https://img-blog.csdn.net/20140930110655328?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQveWNoZW5nX3NqdHU=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
Caffe的安装与配置
Caffe需要预先安装一些依赖项,首先是CUDA驱动。不论是CentOS还是Ubuntu都预装了开源的nouveau显卡驱动(SUSE没有这种问题),如果不禁用,则CUDA驱动不能正确安装。以Ubuntu为例,介绍一下这里的处理方法,当然也有其他处理方法。
# sudo vi/etc/modprobe.d/blacklist.conf
# 增加一行 :blacklist nouveau
sudoapt-get --purge remove xserver-xorg-video-nouveau #把官方驱动彻底卸载:
sudoapt-get --purge remove nvidia-* #清除之前安装的任何NVIDIA驱动
sudo service lightdm stop #进命令行,关闭Xserver
sudo kill all Xorg
安装了CUDA之后,依次按照Caffe官网安装指南1.7K安装BLAS、OpenCV、Boost即可。
Caffe跑跑MNIST试试
在Caffe安装目录之下,首先获得MNIST数据集:
cd data/mnist
sh get_mnist.sh
生成mnist-train-leveldb/ 和 mnist-test-leveldb/,把数据转化成leveldb格式:
sh examples/mnist/create_mnist.sh
训练网络:
sh train_lenet.sh
![](https://img-blog.csdn.net/20140930110918372?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQveWNoZW5nX3NqdHU=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
让Caffe生成的数据集能在Theano上直接运行
不论使用何种框架进行CNNs训练,共有3种数据集:
- Training Set:用于训练网络
- Validation Set:用于训练时测试网络准确率
- Test Set:用于测试网络训练完成后的最终正确率
Caffe生成的数据分为2种格式:Lmdb和Leveldb
- 它们都是键/值对(Key/Value Pair)嵌入式数据库管理系统编程库。
- 虽然lmdb的内存消耗是leveldb的1.1倍,但是lmdb的速度比leveldb快10%至15%,更重要的是lmdb允许多种训练模型同时读取同一组数据集。
- 因此lmdb取代了leveldb成为Caffe默认的数据集生成格式。
Google Protocol Buffer的安装
Protocol Buffer是一种类似于XML的用于序列化数据的自动机制。
首先在Protocol Buffers的中下载最新版本:
https://developers.google.com/protocol-buffers/docs/downloads
解压后运行:
./configure
$ make
$ make check
$ make install
pip installprotobuf
添加动态链接库
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
Lmdb的安装
pip install lmdb
要parse(解析)一个protobuf类型数据,首先要告诉计算机你这个protobuf数据内部是什么格式(有哪些项,这些项各是什么数据类型的决定了占用多少字节,这些项可否重复,重复几次),安装protobuf这个module就可以用protobuf专用的语法来定义这些格式(这个是.proto文件)了,然后用protoc来编译这个.proto文件就可以生成你需要的目标文件。
想要定义自己的.proto文件请阅读:
https://developers.google.com/protocol-buffers/docs/proto?hl=zh-cn
编译.proto文件
protoc--proto_path=IMPORT_PATH --cpp_out=DST_DIR --java_out=DST_DIR--python_out=DST_DIR path/to/file.proto
--proto_path 也可以简写成-I 是.proto所在的路径
输出路径:
--cpp_out 要生成C++可用的头文件,分别是***.pb.h(包含申明类)***.pb.cc(包含可执行类),使用的时候只要include “***.pb.h”
--java_out 生成java可用的头文件
--python_out 生成python可用的头文件,**_pb2.py,使用的时候import**_pb2.py即可
最后一个参数就是你的.proto文件完整路径。
Caffe (CNN, deep learning) 介绍
Caffe -----------Convolution Architecture For Feature Embedding (Extraction)
- Caffe 是什么东东?
- CNN (Deep Learning) 工具箱
- C++ 语言架构
- CPU 和GPU 无缝交换
- Python 和matlab的封装
- 但是,Decaf只是CPU 版本。
-
为什么要用Caffe?
- 运算速度快。简单 友好的架构 用到的一些库:
- Google Logging library (Glog): 一个C++语言的应用级日志记录框架,提供了C++风格的流操作和各种助手宏.
- lebeldb(数据存储): 是一个google实现的非常高效的kv数据库,单进程操作。
- CBLAS library(CPU版本的矩阵操作)
- CUBLAS library (GPU 版本的矩阵操作)
-
Caffe 架构
- 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)
-
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)
Caffe深度学习之图像分类模型AlexNet解读
在imagenet上的图像分类challenge上Alex提出的alexnet网络结构模型赢得了2012届的冠军。要研究CNN类型DL网络模型在图像分类上的应用,就逃不开研究alexnet,这是CNN在图像分类上的经典模型(DL火起来之后)。
在DL开源实现caffe的model样例中,它也给出了alexnet的复现,具体网络配置文件如下train_val.prototxt
接下来本文将一步步对该网络配置结构中各个层进行详细的解读(训练阶段):
各种layer的operation更多解释可以参考 Caffe Layer Catalogue
从计算该模型的数据流过程中,该模型参数大概5kw+。
-
conv1阶段DFD(data flow diagram): ![](http://img.ptcms.csdn.net/article/201501/23/54c19b9b105a6.jpg)
-
conv2阶段DFD(data flow diagram): ![](http://img.ptcms.csdn.net/article/201501/23/54c19cfbc9bdf.jpg)
-
conv3阶段DFD(data flow diagram):
![](http://img.ptcms.csdn.net/article/201501/23/54c19dd3db20e.jpg)
-
conv4阶段DFD(data flow diagram):
-
conv5阶段DFD(data flow diagram):
![](http://img.ptcms.csdn.net/article/201501/23/54c19e34c8bf1.jpg)
-
fc6阶段DFD(data flow diagram):
![](http://img.ptcms.csdn.net/article/201501/23/54c19e4357d25.jpg)
-
fc7阶段DFD(data flow diagram):
![](http://img.ptcms.csdn.net/article/201501/23/54c19e5c37c9e.jpg)
-
fc8阶段DFD(data flow diagram):
caffe的输出中也有包含这块的内容日志,详情如下:
I0721 10:38:15.326920 4692 net.cpp:125] Top shape: 256 3 227 227 (39574272)
I0721 10:38:15.326971 4692 net.cpp:125] Top shape: 256 1 1 1 (256)
I0721 10:38:15.326982 4692 net.cpp:156] data does not need backward computation.
I0721 10:38:15.327003 4692 net.cpp:74] Creating Layer conv1
I0721 10:38:15.327011 4692 net.cpp:84] conv1 <- data
I0721 10:38:15.327033 4692 net.cpp:110] conv1 -> conv1
I0721 10:38:16.721956 4692 net.cpp:125] Top shape: 256 96 55 55 (74342400)
I0721 10:38:16.722030 4692 net.cpp:151] conv1 needs backward computation.
I0721 10:38:16.722059 4692 net.cpp:74] Creating Layer relu1
I0721 10:38:16.722070 4692 net.cpp:84] relu1 <- conv1
I0721 10:38:16.722082 4692 net.cpp:98] relu1 -> conv1 (in-place)
I0721 10:38:16.722096 4692 net.cpp:125] Top shape: 256 96 55 55 (74342400)
I0721 10:38:16.722105 4692 net.cpp:151] relu1 needs backward computation.
I0721 10:38:16.722116 4692 net.cpp:74] Creating Layer pool1
I0721 10:38:16.722125 4692 net.cpp:84] pool1 <- conv1
I0721 10:38:16.722133 4692 net.cpp:110] pool1 -> pool1
I0721 10:38:16.722167 4692 net.cpp:125] Top shape: 256 96 27 27 (17915904)
I0721 10:38:16.722187 4692 net.cpp:151] pool1 needs backward computation.
I0721 10:38:16.722205 4692 net.cpp:74] Creating Layer norm1
I0721 10:38:16.722221 4692 net.cpp:84] norm1 <- pool1
I0721 10:38:16.722234 4692 net.cpp:110] norm1 -> norm1
I0721 10:38:16.722251 4692 net.cpp:125] Top shape: 256 96 27 27 (17915904)
I0721 10:38:16.722260 4692 net.cpp:151] norm1 needs backward computation.
I0721 10:38:16.722272 4692 net.cpp:74] Creating Layer conv2
I0721 10:38:16.722280 4692 net.cpp:84] conv2 <- norm1
I0721 10:38:16.722290 4692 net.cpp:110] conv2 -> conv2
I0721 10:38:16.725225 4692 net.cpp:125] Top shape: 256 256 27 27 (47775744)
I0721 10:38:16.725242 4692 net.cpp:151] conv2 needs backward computation.
I0721 10:38:16.725253 4692 net.cpp:74] Creating Layer relu2
I0721 10:38:16.725261 4692 net.cpp:84] relu2 <- conv2
I0721 10:38:16.725270 4692 net.cpp:98] relu2 -> conv2 (in-place)
I0721 10:38:16.725280 4692 net.cpp:125] Top shape: 256 256 27 27 (47775744)
I0721 10:38:16.725288 4692 net.cpp:151] relu2 needs backward computation.
I0721 10:38:16.725298 4692 net.cpp:74] Creating Layer pool2
I0721 10:38:16.725307 4692 net.cpp:84] pool2 <- conv2
I0721 10:38:16.725317 4692 net.cpp:110] pool2 -> pool2
I0721 10:38:16.725329 4692 net.cpp:125] Top shape: 256 256 13 13 (11075584)
I0721 10:38:16.725338 4692 net.cpp:151] pool2 needs backward computation.
I0721 10:38:16.725358 4692 net.cpp:74] Creating Layer norm2
I0721 10:38:16.725368 4692 net.cpp:84] norm2 <- pool2
I0721 10:38:16.725378 4692 net.cpp:110] norm2 -> norm2
I0721 10:38:16.725389 4692 net.cpp:125] Top shape: 256 256 13 13 (11075584)
I0721 10:38:16.725399 4692 net.cpp:151] norm2 needs backward computation.
I0721 10:38:16.725409 4692 net.cpp:74] Creating Layer conv3
I0721 10:38:16.725419 4692 net.cpp:84] conv3 <- norm2
I0721 10:38:16.725427 4692 net.cpp:110] conv3 -> conv3
I0721 10:38:16.735193 4692 net.cpp:125] Top shape: 256 384 13 13 (16613376)
I0721 10:38:16.735213 4692 net.cpp:151] conv3 needs backward computation.
I0721 10:38:16.735224 4692 net.cpp:74] Creating Layer relu3
I0721 10:38:16.735234 4692 net.cpp:84] relu3 <- conv3
I0721 10:38:16.735242 4692 net.cpp:98] relu3 -> conv3 (in-place)
I0721 10:38:16.735250 4692 net.cpp:125] Top shape: 256 384 13 13 (16613376)
I0721 10:38:16.735258 4692 net.cpp:151] relu3 needs backward computation.
I0721 10:38:16.735302 4692 net.cpp:74] Creating Layer conv4
I0721 10:38:16.735312 4692 net.cpp:84] conv4 <- conv3
I0721 10:38:16.735321 4692 net.cpp:110] conv4 -> conv4
I0721 10:38:16.743952 4692 net.cpp:125] Top shape: 256 384 13 13 (16613376)
I0721 10:38:16.743988 4692 net.cpp:151] conv4 needs backward computation.
I0721 10:38:16.744000 4692 net.cpp:74] Creating Layer relu4
I0721 10:38:16.744010 4692 net.cpp:84] relu4 <- conv4
I0721 10:38:16.744020 4692 net.cpp:98] relu4 -> conv4 (in-place)
I0721 10:38:16.744030 4692 net.cpp:125] Top shape: 256 384 13 13 (16613376)
I0721 10:38:16.744038 4692 net.cpp:151] relu4 needs backward computation.
I0721 10:38:16.744050 4692 net.cpp:74] Creating Layer conv5
I0721 10:38:16.744057 4692 net.cpp:84] conv5 <- conv4
I0721 10:38:16.744067 4692 net.cpp:110] conv5 -> conv5
I0721 10:38:16.748935 4692 net.cpp:125] Top shape: 256 256 13 13 (11075584)
I0721 10:38:16.748955 4692 net.cpp:151] conv5 needs backward computation.
I0721 10:38:16.748965 4692 net.cpp:74] Creating Layer relu5
I0721 10:38:16.748975 4692 net.cpp:84] relu5 <- conv5
I0721 10:38:16.748983 4692 net.cpp:98] relu5 -> conv5 (in-place)
I0721 10:38:16.748998 4692 net.cpp:125] Top shape: 256 256 13 13 (11075584)
I0721 10:38:16.749011 4692 net.cpp:151] relu5 needs backward computation.
I0721 10:38:16.749022 4692 net.cpp:74] Creating Layer pool5
I0721 10:38:16.749030 4692 net.cpp:84] pool5 <- conv5
I0721 10:38:16.749039 4692 net.cpp:110] pool5 -> pool5
I0721 10:38:16.749050 4692 net.cpp:125] Top shape: 256 256 6 6 (2359296)
I0721 10:38:16.749058 4692 net.cpp:151] pool5 needs backward computation.
I0721 10:38:16.749074 4692 net.cpp:74] Creating Layer fc6
I0721 10:38:16.749083 4692 net.cpp:84] fc6 <- pool5
I0721 10:38:16.749091 4692 net.cpp:110] fc6 -> fc6
I0721 10:38:17.160079 4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.160148 4692 net.cpp:151] fc6 needs backward computation.
I0721 10:38:17.160166 4692 net.cpp:74] Creating Layer relu6
I0721 10:38:17.160177 4692 net.cpp:84] relu6 <- fc6
I0721 10:38:17.160190 4692 net.cpp:98] relu6 -> fc6 (in-place)
I0721 10:38:17.160202 4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.160212 4692 net.cpp:151] relu6 needs backward computation.
I0721 10:38:17.160222 4692 net.cpp:74] Creating Layer drop6
I0721 10:38:17.160230 4692 net.cpp:84] drop6 <- fc6
I0721 10:38:17.160238 4692 net.cpp:98] drop6 -> fc6 (in-place)
I0721 10:38:17.160258 4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.160265 4692 net.cpp:151] drop6 needs backward computation.
I0721 10:38:17.160277 4692 net.cpp:74] Creating Layer fc7
I0721 10:38:17.160286 4692 net.cpp:84] fc7 <- fc6
I0721 10:38:17.160295 4692 net.cpp:110] fc7 -> fc7
I0721 10:38:17.342094 4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.342157 4692 net.cpp:151] fc7 needs backward computation.
I0721 10:38:17.342175 4692 net.cpp:74] Creating Layer relu7
I0721 10:38:17.342185 4692 net.cpp:84] relu7 <- fc7
I0721 10:38:17.342198 4692 net.cpp:98] relu7 -> fc7 (in-place)
I0721 10:38:17.342208 4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.342217 4692 net.cpp:151] relu7 needs backward computation.
I0721 10:38:17.342228 4692 net.cpp:74] Creating Layer drop7
I0721 10:38:17.342236 4692 net.cpp:84] drop7 <- fc7
I0721 10:38:17.342245 4692 net.cpp:98] drop7 -> fc7 (in-place)
I0721 10:38:17.342254 4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.342262 4692 net.cpp:151] drop7 needs backward computation.
I0721 10:38:17.342274 4692 net.cpp:74] Creating Layer fc8
I0721 10:38:17.342283 4692 net.cpp:84] fc8 <- fc7
I0721 10:38:17.342291 4692 net.cpp:110] fc8 -> fc8
I0721 10:38:17.343199 4692 net.cpp:125] Top shape: 256 22 1 1 (5632)
I0721 10:38:17.343214 4692 net.cpp:151] fc8 needs backward computation.
I0721 10:38:17.343231 4692 net.cpp:74] Creating Layer loss
I0721 10:38:17.343240 4692 net.cpp:84] loss <- fc8
I0721 10:38:17.343250 4692 net.cpp:84] loss <- label
I0721 10:38:17.343264 4692 net.cpp:151] loss needs backward computation.
I0721 10:38:17.343305 4692 net.cpp:173] Collecting Learning Rate and Weight Decay.
I0721 10:38:17.343327 4692 net.cpp:166] Network initialization done.
I0721 10:38:17.343335 4692 net.cpp:167] Memory required for Data 1073760256
CIFAR-10在caffe上进行训练与学习
使用数据库:CIFAR-10
60000张 32X32 彩色图像 10类,50000张训练,10000张测试
![](http://img.ptcms.csdn.net/article/201501/23/54c19c99008d5.jpg)
准备
在终端运行以下指令:
cd $CAFFE_ROOT/data/cifar10
./get_cifar10.sh
cd $CAFFE_ROOT/examples/cifar10
./create_cifar10.sh
其中CAFFE_ROOT是caffe-master在你机子的地址
运行之后,将会在examples中出现数据库文件./cifar10-leveldb和数据库图像均值二进制文件./mean.binaryproto
![](http://img.ptcms.csdn.net/article/201501/23/54c19c890e291.jpg)
模型
该CNN由卷积层,POOLing层,非线性变换层,在顶端的局部对比归一化线性分类器组成。该模型的定义在CAFFE_ROOT/examples/cifar10 directory’s cifar10_quick_train.prototxt中,可以进行修改。其实后缀为prototxt很多都是用来修改配置的。
![](http://img.ptcms.csdn.net/article/201501/23/54c19c5adc1ff.jpg)
训练和测试
训练这个模型非常简单,当我们写好参数设置的文件cifar10_quick_solver.prototxt和定义的文件cifar10_quick_train.prototxt和cifar10_quick_test.prototxt后,运行train_quick.sh或者在终端输入下面的命令:
cd $CAFFE_ROOT/examples/cifar10
./train_quick.sh
即可,train_quick.sh是一个简单的脚本,会把执行的信息显示出来,培训的工具是train_net.bin,cifar10_quick_solver.prototxt作为参数。
然后出现类似以下的信息:这是搭建模型的相关信息
I0317 21:52:48.945710 2008298256 net.cpp:74] Creating Layer conv1
I0317 21:52:48.945716 2008298256 net.cpp:84] conv1 <- data
I0317 21:52:48.945725 2008298256 net.cpp:110] conv1 -> conv1
I0317 21:52:49.298691 2008298256 net.cpp:125] Top shape: 100 32 32 32 (3276800)
I0317 21:52:49.298719 2008298256 net.cpp:151] conv1 needs backward computation.
接着:
0317 21:52:49.309370 2008298256 net.cpp:166] Network initialization done.
I0317 21:52:49.309376 2008298256 net.cpp:167] Memory required for Data 23790808
I0317 21:52:49.309422 2008298256 solver.cpp:36] Solver scaffolding done.
I0317 21:52:49.309447 2008298256 solver.cpp:47] Solving CIFAR10_quick_train
之后,训练开始
I0317 21:53:12.179772 2008298256 solver.cpp:208] Iteration 100, lr = 0.001
I0317 21:53:12.185698 2008298256 solver.cpp:65] Iteration 100, loss = 1.73643
...
I0317 21:54:41.150030 2008298256 solver.cpp:87] Iteration 500, Testing net
I0317 21:54:47.129461 2008298256 solver.cpp:114] Test score #0: 0.5504
I0317 21:54:47.129500 2008298256 solver.cpp:114] Test score #1: 1.27805
其中每100次迭代次数显示一次训练时lr(learningrate),和loss(训练损失函数),每500次测试一次,输出score 0(准确率)和score 1(测试损失函数)
当5000次迭代之后,正确率约为75%,模型的参数存储在二进制protobuf格式在cifar10_quick_iter_5000
然后,这个模型就可以用来运行在新数据上了。
其他
另外,更改cifar*solver.prototxt文件可以使用CPU训练,
# solver mode: CPU or GPU
solver_mode: CPU
可以看看CPU和GPU训练的差别。
主要资料来源:caffe官网教程