均值削减是数据预处理中常见的处理方式,按照之前在学习ufldl教程PCA的一章时,对于图像介绍了两种:第一种常用的方式叫做dimension_mean(个人命名),是依据输入数据的维度,每个维度内进行削减,这个也是常见的做法;第二种叫做per_image_mean,ufldl教程上说,在natural images上训练网络时;给每个像素(这里只每个dimension)计算一个独立的均值和方差是make little sense的;这是因为图像本身具有统计不变性,即在图像的一部分的统计特性和另一部分相同。作者最后建议,如果你训练你的算法在非natural images(如mnist,或者在白背景存在单个独立的物体),其他类型的规则化是值得考虑的。但是当在natural images上训练时,per_image_mean是一个合理的默认选择。
本文中在imagenet数据集上采用的是dimension_mean的方法。
一:程序开始
make_image_mean.sh文件调用代码:
[cpp] view plaincopy
- EXAMPLE=examples/imagenet
- DATA=data/ilsvrc12
- TOOLS=build/tools
- $TOOLS/compute_image_mean $EXAMPLE/ilsvrc12_train_lmdb \
- $DATA/imagenet_mean.binaryproto<strong>
- </strong>
二:make_image_mean.cpp函数分析
输入参数:lmdb文件 均值文件imagenet_mean.binaryproto
2.1 头文件分析
[cpp] view plaincopy
- #include<stdint.h>//定义了几种扩展的整数类型和宏
- #include<algorithm>//输出数组的内容、对数组进行排序、反转数组内容、复制数组内容等操作,
- #include<string>
- #include<utility>//utility头文件定义了一个pair类型,pair类型用于存储一对数据;它也提供一些常用的便利函数、或类、或模板。大小求值、值交换:min、max和swap。
- #include<vector>//可以自动扩展容量的数组
- #include"boost/scoped_ptr.hpp"
- #include"gflags/gflags.h"
- #include"glog/logging.h"
- #include"caffe/proto/caffe.pb.h"
- #include"caffe/util/db.hpp"//引入包装好的lmdb操作函数
- #include"caffe/util/io.hpp"//引入opencv中的图像操作函数
- usingnamespacecaffe; //引入caffe命名空间
- usingstd::max;//
- usingstd::pair;
- using boost::scoped_ptr;
2.2 gflags宏定义string变量
DEFINE_string(backend, "lmdb","The backend {leveldb, lmdb} containing theimages");
2.3 main函数分析
2.3.1 lmdb数据操作
[cpp] view plaincopy
- scoped_ptr<db::DB>db(db::GetDB(FLAGS_backend));
- db->Open(argv[1], db::READ);//只读的方式打开lmdb文件
- scoped_ptr<db::Cursor> cursor(db->NewCursor());
- //lmdb数据库的“光标”文件,一个光标保存一个从数据库根目录到数据库文件的路径;A cursorholds a path of (page pointer, key index) from the DB root to a position in theDB, plus other state.
2.3.4 声明中转对象变量
[cpp] view plaincopy
- BlobProtosum_blob;//声明blob变量;这个BlobProto在哪里定义的,没有找到;感觉应该在caffe.pb.h中定义的,因为db.cpp和io.cpp中没有找到
- int count = 0;
- // load first datum
- Datum datum;
- datum.ParseFromString(cursor->value());//这个cursor.value,感觉返回的应该是lmdb中存储的第一个键值对数据
2.3.5 给BlobProto类型变量赋值
每个blob对象,为一个4维的数组,分别为image_num*channels*height*width
[cpp] view plaincopy
- sum_blob.set_num(1);//设置图片的个数
- sum_blob.set_channels(datum.channels());
- sum_blob.set_height(datum.height());
- sum_blob.set_width(datum.width());
- constintdata_size = datum.channels() *datum.height() * datum.width();//每张图片的尺寸
- intsize_in_datum = std::max<int>(datum.data().size(),datum.float_data_size());
这个size()和float_data_size()有些不明白,图像数据正常应该是整形的数据(例如uint8_t),感觉这个size()应该对应的是整型数据的个数,例如一个50*50的彩色图片,最后应该是50*50*3=750个整型数来表示一幅50*50的图片;至于这个float_data_size()就不清楚了,感觉是某些图片数据使用float类型存储的,所以用float来统计数值的个数。开始感觉这个float的size应该是把int类型转换成float后,查看在float类型下的字节占用情况;但是由下面的代码来看,感觉这个size(),统计的是数据的个数也就是750,而不是占用的字节数。如果图像使用int类型存储的,那么float_data_size()=0;如果使用float类型存储的,那么datum.data.size=0。所以每次都要max操作
[cpp] view plaincopy
- for (inti= 0; i<size_in_datum; ++i) {
- sum_blob.add_data(0.);//设置初值为float型的0.0
- }
2.3.6利用循环和cursor读取lmdb中的数据
[cpp] view plaincopy
- while (cursor->valid()) {//如果cursor是有效的
- Datum datum;
- datum.ParseFromString(cursor->value());//解析cuisor.value返回的字符串值,到datum
- DecodeDatumNative(&datum);//感觉是把datum中字符串类型的值,变成相应的类型
- conststd::string& data =datum.data();//利用data来引用datum.data
- size_in_datum = std::max<int>(datum.data().size(),datum.float_data_size());
- CHECK_EQ(size_in_datum,data_size) <<"Incorrect data field size"<<size_in_datum;
- if (data.size() != 0) {//datum.data().size()!=0
- CHECK_EQ(data.size(),size_in_datum);//判断是否相等
- for (inti= 0; i<size_in_datum; ++i) {
- sum_blob.set_data(i, sum_blob.data(i) + (uint8_t)data[i]);//对应位置的像素值相加(uin8_t类型相加),相加的结果放在sum_blob中
- }
- } else{
- CHECK_EQ(datum.float_data_size(), size_in_datum);
- for (inti= 0; i<size_in_datum; ++i) {
- sum_blob.set_data(i, sum_blob.data(i) +
- static_cast<float>(datum.float_data(i)));//对应位置的像素值相加(float类型相加)
- }
- }
- ++count;
- if (count % 10000 == 0) {
- LOG(INFO) <<"Processed "<<count <<" files.";
- }
- cursor->Next();//光标下移(指针),指向下一个存储在lmdb中的数据
- }
2.3.7 求均值
[cpp] view plaincopy
- for (inti= 0; i<sum_blob.data_size(); ++i) {
- sum_blob.set_data(i, sum_blob.data(i) / count);
- }
2.3.8 存储到指定文件
[cpp] view plaincopy
- // Write to disk
- if (argc == 3) {
- LOG(INFO) <<"Write to "<<argv[2];
- WriteProtoToBinaryFile(sum_blob, argv[2]);
- }
2.3.9 计算每个channel的均值,这个貌似没有用到吧!
[cpp] view plaincopy
- constint channels = sum_blob.channels();
- constint dim = sum_blob.height() *sum_blob.width();
- std::vector<float>mean_values(channels,0.0);//容量为3的数组,初始值为0.0
- LOG(INFO) <<"Number of channels:"<< channels;
- for (intc = 0; c < channels; ++c) {
- for (inti= 0; i< dim; ++i) {
- mean_values[c] += sum_blob.data(dim * c + i);
- }
- LOG(INFO) <<"mean_value channel["<< c <<"]:"<<mean_values[c]/ dim;
- }
三,相关文件
compute_image_mean.cpp
[cpp] view plaincopy
- #include <stdint.h>
- #include <algorithm>
- #include <string>
- #include <utility>
- #include <vector>
- #include "boost/scoped_ptr.hpp"
- #include "gflags/gflags.h"
- #include "glog/logging.h"
- #include "caffe/proto/caffe.pb.h"
- #include "caffe/util/db.hpp"
- #include "caffe/util/io.hpp"
- using namespace caffe; // NOLINT(build/namespaces)
- using std::max;
- using std::pair;
- using boost::scoped_ptr;
- DEFINE_string(backend, "lmdb",
- "The backend {leveldb, lmdb} containing the images");
- int main(int argc, char** argv) {
- ::google::InitGoogleLogging(argv[0]);
- #ifndef GFLAGS_GFLAGS_H_
- namespace gflags = google;
- #endif
- gflags::SetUsageMessage("Compute the mean_image of a set of images given by"
- " a leveldb/lmdb\n"
- "Usage:\n"
- " compute_image_mean [FLAGS] INPUT_DB [OUTPUT_FILE]\n");
- gflags::ParseCommandLineFlags(&argc, &argv, true);
- if (argc < 2 || argc > 3) {
- gflags::ShowUsageWithFlagsRestrict(argv[0], "tools/compute_image_mean");
- return 1;
- }
- scoped_ptr<db::DB> db(db::GetDB(FLAGS_backend));
- db->Open(argv[1], db::READ);
- scoped_ptr<db::Cursor> cursor(db->NewCursor());
- BlobProto sum_blob;
- int count = 0;
- // load first datum
- Datum datum;
- datum.ParseFromString(cursor->value());
- if (DecodeDatumNative(&datum)) {
- LOG(INFO) << "Decoding Datum";
- }
- sum_blob.set_num(1);
- sum_blob.set_channels(datum.channels());
- sum_blob.set_height(datum.height());
- sum_blob.set_width(datum.width());
- const int data_size = datum.channels() * datum.height() * datum.width();
- int size_in_datum = std::max<int>(datum.data().size(),datum.float_data_size());
- for (int i = 0; i < size_in_datum; ++i) {
- sum_blob.add_data(0.);//设置初值为float型的0.0
- }
- LOG(INFO) << "Starting Iteration";
- while (cursor->valid()) {//如果cursor是有效的
- Datum datum;
- datum.ParseFromString(cursor->value());//解析cuisor.value返回的字符串值,到datum
- DecodeDatumNative(&datum);
- const std::string& data = datum.data();//利用data来引用datum.data
- size_in_datum = std::max<int>(datum.data().size(),datum.float_data_size());
- CHECK_EQ(size_in_datum, data_size) << "Incorrect data field size " <<size_in_datum;
- if (data.size() != 0) {
- CHECK_EQ(data.size(), size_in_datum);
- for (int i = 0; i < size_in_datum; ++i) {
- sum_blob.set_data(i, sum_blob.data(i) + (uint8_t)data[i]);
- }
- } else {
- CHECK_EQ(datum.float_data_size(), size_in_datum);
- for (int i = 0; i < size_in_datum; ++i) {
- sum_blob.set_data(i, sum_blob.data(i) +
- static_cast<float>(datum.float_data(i)));
- }
- }
- ++count;
- if (count % 10000 == 0) {
- LOG(INFO) << "Processed " << count << " files.";
- }
- cursor->Next();
- }
- if (count % 10000 != 0) {
- LOG(INFO) << "Processed " << count << " files.";
- }
- for (int i = 0; i < sum_blob.data_size(); ++i) {
- sum_blob.set_data(i, sum_blob.data(i) / count);
- }
- // Write to disk
- if (argc == 3) {
- LOG(INFO) << "Write to " << argv[2];
- WriteProtoToBinaryFile(sum_blob, argv[2]);
- }
- const int channels = sum_blob.channels();
- const int dim = sum_blob.height() * sum_blob.width();
- std::vector<float> mean_values(channels, 0.0);
- LOG(INFO) << "Number of channels: " << channels;
- for (int c = 0; c < channels; ++c) {
- for (int i = 0; i < dim; ++i) {
- mean_values[c] += sum_blob.data(dim * c + i);
- }
- LOG(INFO) << "mean_value channel [" << c << "]:" << mean_values[c] / dim;
- }
- return 0;
- }
四:以上代码注释为个人理解,如有遗漏,错误还望大家多多交流,指正,以便共同学习,进步!!
转载请标明出处:http://blog.csdn.net/whiteinblue/article/details/45540301