在读取ssd-caffe代码的时候有一些问题,真的需要您的帮助。
原生caffe只支持分类,数据读取层常用的是读取LMDB数据库和读取图像进行训练。
为了支持多个标签和输入注释框的输入,我决定使用ssd-caffe,它在原生caffe中增加了一个AnnotatedDataLayer层。这个新添加的层可以支持多个标签和注释框,但它有局限性。原因是它读取的数据类型仍然是lmdb。
我们现在需要随机读取数据集的数据,但根据查询结果,lmdb是B+树结构,只能通过迭代器依次读取,所以我们想把lmdb改为直接读取图片。但是,原生caffe的直接读取图片不支持多标签和注释框。如何修改caffe的image_data_layers来支持注释框的输入(能否按照AnnotatedDataLayer的方法来解决这个问题)?
注意事项
修改了ssd-caffe源码。https:/github.comeric612MobileNet-YOLO。
新添加的注释框的文件路径。MobileNet-YOLOsrccaffelayersannotated_data_layer.cpp。
原生caffe文件路径,可直接读取图片。MobileNet-YOLOsrccaffelayersimage_data_layer.cpp。
数据层 提供了从硬盘中异步读取随机数据的可能性(它使用2个线程:一个线程读取数据,另一个线程将数据传送给神经网络)。你的顶部blob是由数据和标签组成的。不幸的是,标签是一维的。为了解决这个问题,可以将我们的lmdb数据库以特殊的顺序组织起来。然后当我们读取数据时,在将数据交付给神经网络之前,我们会对数据进行转换,使其适应我们的问题。下面我用一个例子来说明。首先我会写一个LMDB数据库,里面有10张不同的图片(是同一张图片,但我们假设它们是不同的),10个随机的边界框和10个随机的标签,每个标签的维度是3。
注注意:要重现以下代码,您必须安装caffe。如果你只编译了 caffe 文件夹,那么就在 root_caffeexamplesnew_folder。,把代码放在那里,然后编译做 使.
#include <caffe/caffe.hpp>
#include "caffe/proto/caffe.pb.h"
#include "caffe/util/db.hpp"
#include "boost/scoped_ptr.hpp"
#include <opencv2/imgcodecs.hpp>
#include <iostream>
#include <stdlib.h>
using namespace caffe;
using boost::scoped_ptr;
std::vector<float> generate_random_boxes(const int max_num_bbx){
std::vector<float> bbx(4*max_num_bbx);
for(int i = 0; i < max_num_bbx; i++){
float scale = 500*static_cast <float> (rand()) / static_cast <float> (RAND_MAX);
float x1 = static_cast <float> (rand()) / static_cast <float> (RAND_MAX);
float y1 = static_cast <float> (rand()) / static_cast <float> (RAND_MAX);
float x2 = x1 + static_cast <float> (rand()) / static_cast <float> (RAND_MAX);
float y2 = x1 + static_cast <float> (rand()) / static_cast <float> (RAND_MAX);
bbx[i*4] = scale*x1;
bbx[i*4 + 1] = scale*y1;
bbx[i*4 + 2] = scale*x2;
bbx[i*4 + 3] = scale*y2;
}
return bbx;
}
std::vector<float> generate_random_labels(const int dim_label, const int max_num_bbx){
std::vector<float> labels(dim_label*max_num_bbx);
for(int i = 0; i < max_num_bbx; i++){
for(int j = 0; j < dim_label; j++){
labels[dim_label*i + j] = static_cast <float> (rand()) / static_cast <float> (RAND_MAX);
}
}
return labels;
}
int main(){
const std::string root_path = "/path/for/test/";
const std::string path_lmdb = root_path + "lmdb";
std::string rm_lmdb = std::string("rm -rf ") + path_lmdb.c_str();
system(rm_lmdb.c_str());
scoped_ptr<db::DB> db(db::GetDB("lmdb"));
db->Open(path_lmdb, db::NEW);
scoped_ptr<db::Transaction> txn(db->NewTransaction());
int n = 10;
int max_num_bbx = 7;
int dim_label = 3;
cv::Mat aux_img = cv::imread(root_path + "image.jpg");
int rows = aux_img.rows;
int cols = aux_img.cols;
std::vector<cv::Mat> vec_img(n);
std::vector< std::vector<float> > vec_bbx(n);
std::vector< std::vector<float> > vec_label(n);
for(int i = 0; i < n; i++){
vec_img[i] = aux_img.clone();
vec_bbx[i] = generate_random_boxes(max_num_bbx);
vec_label[i] = generate_random_labels(dim_label, max_num_bbx);
}
for(int i = 0; i< n; i++){
int sz = 3*rows*cols + 4*max_num_bbx + dim_label*max_num_bbx;
Datum datum;
datum.set_label(0); //no used
datum.set_channels(1);
datum.set_height(1);
datum.set_width(sz);
google::protobuf::RepeatedField<float>* datumFloatData = datum.mutable_float_data();
//store images
cv::Mat img = vec_img[i];
for(int d = 0; d < 3; d++){ //BGR
for(int r = 0; r < rows; r++){
for(int c = 0; c < cols; c++){
cv::Vec3b pixel = img.at<cv::Vec3b>(r, c);
datumFloatData->Add(float(pixel[d]));
}
}
}
//store bounding-boxes
std::vector<float>& bbx = vec_bbx[i];
for(int j = 0; j < 4*max_num_bbx; j++)
datumFloatData->Add(bbx[j]);
//store labels
std::vector<float>& label = vec_label[i];
for(int j = 0; j < dim_label*max_num_bbx; j++)
datumFloatData->Add(label[j]);
//store lmdb
std::string key_str = caffe::format_int(i);
std::string out;
CHECK(datum.SerializeToString(&out));
txn->Put(key_str, out);
txn->Commit();
txn.reset(db->NewTransaction());
std::cout<<"save data: "<<i<<std::endl;
}
return 0;
}
然后在文件夹 "路径测试" 我们将有一个名为 lmdb 包含我们的数据库。现在我们必须读取数据,并按照所需的顺序组织数据。为此,我将使用一个 切片 层,它允许将一个输入的底部数据划分为多个顶部。因此,由一批图像、边界框和标签组成的输入数据将被划分为5个顶部blobs:img_b、img_g、img_r、bbx、标签。
#include <caffe/caffe.hpp>
#include <opencv2/imgcodecs.hpp>
#include <opencv2/core.hpp>
#include <opencv2/imgcodecs.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include "boost/scoped_ptr.hpp"
#include <iostream>
#include <stdio.h>
#include <stdlib.h>
using namespace caffe;
using boost::scoped_ptr;
int main(){
const std::string root_path = "/path/for/test/";
const std::string path_lmdb = root_path + "lmdb";
//parameters used to store lmdb data base
int n = 10;
int max_num_bbx = 7;
int dim_label = 3;
cv::Mat aux_img = cv::imread(root_path + "image.jpg");
int rows = aux_img.rows;
int cols = aux_img.cols;
//here we build the network input
NetParameter net_param;
LayerParameter* db_layer_param = net_param.add_layer();
db_layer_param->set_name("data");
db_layer_param->set_type("Data");
DataParameter* db_data_param = db_layer_param->mutable_data_param();
db_data_param->set_batch_size(2);
db_data_param->set_prefetch(3);
db_data_param->set_source(path_lmdb);
db_data_param->set_backend(DataParameter_DB_LMDB);
db_layer_param->add_top("data");
LayerParameter* slice_layer_param = net_param.add_layer();
slice_layer_param->set_name("slice");
slice_layer_param->set_type("Slice");
slice_layer_param->mutable_slice_param()->set_axis(3);//starting B
slice_layer_param->mutable_slice_param()->add_slice_point(rows*cols);//starting G
slice_layer_param->mutable_slice_param()->add_slice_point(2*rows*cols);//starting R
slice_layer_param->mutable_slice_param()->add_slice_point(3*rows*cols);//starting bbx
slice_layer_param->mutable_slice_param()->add_slice_point(3*rows*cols + 4*max_num_bbx);//starting labels
slice_layer_param->add_bottom("data");
slice_layer_param->add_top("img_b");
slice_layer_param->add_top("img_g");
slice_layer_param->add_top("img_r");
slice_layer_param->add_top("bbx");
slice_layer_param->add_top("labels");
//NOTE: you must add the additional layers of your model
/*
.
.
.
.
*/
//here we store and load the model
//NOTE:In this example is not necessary to store the model in prototxt file
const std::string net_file = root_path + "model.prototxt";
Net<float> net(net_param);
WriteProtoToTextFile(net_param,net_file);
//here we make forward in order to read our data
net.Forward();
/*Note that in this example we read 2 images, but then we will only show the first*/
//read first image
boost::shared_ptr< Blob< float > > img_b = net.blob_by_name("img_b");
boost::shared_ptr< Blob< float > > img_g = net.blob_by_name("img_g");
boost::shared_ptr< Blob< float > > img_r = net.blob_by_name("img_r");
cv::Mat img(rows,cols,CV_8UC3);
for(int r = 0; r < rows; r++){
for(int c = 0; c < cols; c++){
img.at<cv::Vec3b>(r,c)[0] = (uchar) img_b->cpu_data()[r*cols + c];
img.at<cv::Vec3b>(r,c)[1] = (uchar) img_g->cpu_data()[r*cols + c];
img.at<cv::Vec3b>(r,c)[2] = (uchar) img_r->cpu_data()[r*cols + c];
}
}
//read bounding boxes
boost::shared_ptr< Blob< float > > bbx = net.blob_by_name("bbx");
for(int i = 0; i < max_num_bbx; i++){
float x1 = bbx->cpu_data()[4*i];
float y1 = bbx->cpu_data()[4*i + 1];
float x2 = bbx->cpu_data()[4*i + 2];
float y2 = bbx->cpu_data()[4*i + 3];
cv::Point pt1(y1, x1);
cv::Point pt2(y2, x2);
cv::rectangle(img, pt1, pt2, cv::Scalar(0, 255, 0));
}
//read labels
boost::shared_ptr< Blob< float > > labels = net.blob_by_name("labels");
std::cout<<"labels: "<<std::endl;
for(int i = 0; i < max_num_bbx; i++){
for(int j = 0; j < dim_label; j++){
std::cout<<labels->cpu_data()[i*dim_label + j]<<" ";
}
std::cout<<std::endl;
}
cv::imshow("img", img);
cv::waitKey(0);
return 0;
}
生成的输出结果如下。
用WriteProtoToTextFile(net_param, net_file)生成的prototxt由以下的文件生成: 数据 和 切片层 是这样的。
layer {
name: "data"
type: "Data"
top: "data"
data_param {
source: "/path/for/test/lmdb"
batch_size: 2
backend: LMDB
prefetch: 3
}
}
layer {
name: "slice"
type: "Slice"
bottom: "data"
top: "img_b"
top: "img_g"
top: "img_r"
top: "bbx"
top: "labels"
slice_param {
slice_point: 344000
slice_point: 688000
slice_point: 1032000
slice_point: 1032028
axis: 3
}
}
之后 切片层,您可能需要添加额外的 重塑图层 以适应后续各层的数据。