Caffe是一个清晰而高效的深度学习框架,其作者是博士毕业于UC Berkeley的 贾扬清,目前在Google工作。
Caffe是纯粹的C++/CUDA架构,支持命令行、Python和MATLAB接口;可以在CPU和GPU直接无缝切换:
1
|
Caffe::set_mode(Caffe::GPU);
|
Caffe中的网络都是有向无环图的集合,可以直接定义:
1
2
3
4
5
|
name:
"dummy-net"
layers {<span><span>name: <span>
"data"
…<
/span
><
/span
><
/span
>}
layers {<span><span>name: <span>
"conv"
…<
/span
><
/span
><
/span
>}
layers {<span><span>name: <span>
"pool"
…<
/span
><
/span
><
/span
>}
layers {<span><span>name: <span>
"loss"
…<
/span
><
/span
><
/span
>}
|
1
2
3
4
5
6
7
8
9
10
11
12
|
name:
"conv1"
type
:CONVOLUTION
bottom:
"data"
top
:
"conv1"
convolution_param{
num_output:<span>20
kernel_size:5
stride:1
weight_filler{
type
:
"<span style="
color:
#c0504d;">xavier</span>"
}
}
|
Blob
Blob是用以存储数据的4维数组,例如
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:
1
|
solver_mode:GPU
|
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。
生成mnist-train-leveldb/ 和 mnist-test-leveldb/,把数据转化成leveldb格式:
训练网络:
1
2
3
4
5
6
|
# 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官网安装指南安装BLAS、OpenCV、Boost即可。
在Caffe安装目录之下,首先获得MNIST数据集:
1
2
|
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:
1
2
|
cd
examples
/lenet
sh create_mnist.sh
|
训练网络:
1
|
sh train_lenet.sh
|
不论使用何种框架进行CNNs训练,共有3种数据集:
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:
1
2
3
4
5
|
.
/configure
$
make
$
make
check
$
make
install
pip installprotobuf
|
1
|
export
LD_LIBRARY_PATH=
/usr/local/lib
:$LD_LIBRARY_PATH
|
1
|
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
1
|
protoc--proto_path=IMPORT_PATH --cpp_out=DST_DIR --java_out=DST_DIR--python_out=DST_DIR path
/to/file
.proto
|
1
2
3
4
5
6
|
--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 -----------Convolution Architecture For Feature Embedding (Extraction)
为什么要用Caffe?
Caffe 架构
CNN网络配置文件
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+。
conv3阶段DFD(data flow diagram):
conv4阶段DFD(data flow diagram):
fc6阶段DFD(data flow diagram):
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
|
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
60000张 32X32 彩色图像 10类,50000张训练,10000张测试
在终端运行以下指令:
1
2
3
4
|
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
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。
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:
1
2
|
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
1
2
3
4
5
|
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:
1
2
3
4
|
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
1
2
3
4
5
6
|
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。
另外,更改cifar*solver.prototxt文件可以使用CPU训练,
1
2
|
# solver mode: CPU or GPU
solver_mode: CPU
|
可以看看CPU和GPU训练的差别。