SSD框架训练自己的数据集
SSD demo中详细介绍了如何在VOC数据集上使用SSD进行物体检测的训练和验证。
本文介绍如何使用SSD实现对自己数据集的训练和验证过程，内容包括：
1 数据集的标注
2 数据集的转换
3 使用SSD如何训练
4 使用SSD如何测试
1 数据集的标注
数据的标注使用BBox-Label-Tool工具，该工具使用python实现，使用简单方便。修改后的工具支持多label的标签标注。
该工具生成的标签格式是:
object_number
className x1min y1min x1max y1max
classname x2min y2min x2max y2max
...
1.1 labelTool工具的使用说明
BBox-Label-Tool工具实现较简单，原始的git版本使用起来有一些小问题，进行了简单的修改，修改后的版本
#------------------------------------------------------------------------------- # Name: Object bounding box label tool # Purpose: Label object bboxes for ImageNet Detection data # Author: Qiushi # Created: 06/06/2014 # #------------------------------------------------------------------------------- from __future__ import division from Tkinter import * import tkMessageBox from PIL import Image, ImageTk import os import glob import random # colors for the bboxes COLORS = ['red', 'blue', 'yellow', 'pink', 'cyan', 'green', 'black'] # image sizes for the examples SIZE = 256, 256 classLabels=['mat', 'door', 'sofa', 'chair', 'table', 'bed', 'ashcan', 'shoe'] class LabelTool(): def __init__(self, master): # set up the main frame self.parent = master self.parent.title("LabelTool") self.frame = Frame(self.parent) self.frame.pack(fill=BOTH, expand=1) self.parent.resizable(width = False, height = False) # initialize global state self.imageDir = '' self.imageList= [] self.egDir = '' self.egList = [] self.outDir = '' self.cur = 0 self.total = 0 self.category = 0 self.imagename = '' self.labelfilename = '' self.tkimg = None # initialize mouse state self.STATE = {} self.STATE['click'] = 0 self.STATE['x'], self.STATE['y'] = 0, 0 # reference to bbox self.bboxIdList = [] self.bboxId = None self.bboxList = [] self.hl = None self.vl = None self.currentClass = '' # ----------------- GUI stuff --------------------- # dir entry & load self.label = Label(self.frame, text = "Image Dir:") self.label.grid(row = 0, column = 0, sticky = E) self.entry = Entry(self.frame) self.entry.grid(row = 0, column = 1, sticky = W+E) self.ldBtn = Button(self.frame, text = "Load", command = self.loadDir) self.ldBtn.grid(row = 0, column = 2, sticky = W+E) # main panel for labeling self.mainPanel = Canvas(self.frame, cursor='tcross') self.mainPanel.bind("<Button-1>", self.mouseClick) self.mainPanel.bind("<Motion>", self.mouseMove) self.parent.bind("<Escape>", self.cancelBBox) # press <Espace> to cancel current bbox self.parent.bind("s", self.cancelBBox) self.parent.bind("a", self.prevImage) # press 'a' to go backforward self.parent.bind("d", self.nextImage) # press 'd' to go forward self.mainPanel.grid(row = 1, column = 1, rowspan = 4, sticky = W+N) # showing bbox info & delete bbox self.lb1 = Label(self.frame, text = 'Bounding boxes:') self.lb1.grid(row = 1, column = 2, sticky = W+N) self.listbox = Listbox(self.frame, width = 22, height = 12) self.listbox.grid(row = 2, column = 2, sticky = N) self.btnDel = Button(self.frame, text = 'Delete', command = self.delBBox) self.btnDel.grid(row = 3, column = 2, sticky = W+E+N) self.btnClear = Button(self.frame, text = 'ClearAll', command = self.clearBBox) self.btnClear.grid(row = 4, column = 2, sticky = W+E+N) #select class type self.classPanel = Frame(self.frame) self.classPanel.grid(row = 5, column = 1, columnspan = 10, sticky = W+E) label = Label(self.classPanel, text = 'class:') label.grid(row = 5, column = 1, sticky = W+N) self.classbox = Listbox(self.classPanel, width = 4, height = 2) self.classbox.grid(row = 5,column = 2) for each in range(len(classLabels)): function = 'select' + classLabels[each] print classLabels[each] btnMat = Button(self.classPanel, text = classLabels[each], command = getattr(self, function)) btnMat.grid(row = 5, column = each + 3) # control panel for image navigation self.ctrPanel = Frame(self.frame) self.ctrPanel.grid(row = 6, column = 1, columnspan = 2, sticky = W+E) self.prevBtn = Button(self.ctrPanel, text='<< Prev', width = 10, command = self.prevImage) self.prevBtn.pack(side = LEFT, padx = 5, pady = 3) self.nextBtn = Button(self.ctrPanel, text='Next >>', width = 10, command = self.nextImage) self.nextBtn.pack(side = LEFT, padx = 5, pady = 3) self.progLabel = Label(self.ctrPanel, text = "Progress: / ") self.progLabel.pack(side = LEFT, padx = 5) self.tmpLabel = Label(self.ctrPanel, text = "Go to Image No.") self.tmpLabel.pack(side = LEFT, padx = 5) self.idxEntry = Entry(self.ctrPanel, width = 5) self.idxEntry.pack(side = LEFT) self.goBtn = Button(self.ctrPanel, text = 'Go', command = self.gotoImage) self.goBtn.pack(side = LEFT) # example pannel for illustration self.egPanel = Frame(self.frame, border = 10) self.egPanel.grid(row = 1, column = 0, rowspan = 5, sticky = N) self.tmpLabel2 = Label(self.egPanel, text = "Examples:") self.tmpLabel2.pack(side = TOP, pady = 5) self.egLabels = [] for i in range(3): self.egLabels.append(Label(self.egPanel)) self.egLabels[-1].pack(side = TOP) # display mouse position self.disp = Label(self.ctrPanel, text='') self.disp.pack(side = RIGHT) self.frame.columnconfigure(1, weight = 1) self.frame.rowconfigure(10, weight = 1) # for debugging ## self.setImage() ## self.loadDir() def loadDir(self, dbg = False): if not dbg: s = self.entry.get() self.parent.focus() self.category = int(s) else: s = r'D:\workspace\python\labelGUI' ## if not os.path.isdir(s): ## tkMessageBox.showerror("Error!", message = "The specified dir doesn't exist!") ## return # get image list self.imageDir = os.path.join(r'./Images', '%d' %(self.category)) self.imageList = glob.glob(os.path.join(self.imageDir, '*.jpg')) if len(self.imageList) == 0: print 'No .JPEG images found in the specified dir!' return # set up output dir self.outDir = os.path.join(r'./Labels', '%d' %(self.category)) if not os.path.exists(self.outDir): os.mkdir(self.outDir) labeledPicList = glob.glob(os.path.join(self.outDir, '*.txt')) for label in labeledPicList: data = open(label, 'r') if '0\n' == data.read(): data.close() continue data.close() picture = label.replace('Labels', 'Images').replace('.txt', '.jpg') if picture in self.imageList: self.imageList.remove(picture) # default to the 1st image in the collection self.cur = 1 self.total = len(self.imageList) self.loadImage() print '%d images loaded from %s' %(self.total, s) def loadImage(self): # load image imagepath = self.imageList[self.cur - 1] self.img = Image.open(imagepath) self.imgSize = self.img.size self.tkimg = ImageTk.PhotoImage(self.img) self.mainPanel.config(width = max(self.tkimg.width(), 400), height = max(self.tkimg.height(), 400)) self.mainPanel.create_image(0, 0, image = self.tkimg, anchor=NW) self.progLabel.config(text = "%04d/%04d" %(self.cur, self.total)) # load labels self.clearBBox() self.imagename = os.path.split(imagepath)[-1].split('.')[0] labelname = self.imagename + '.txt' self.labelfilename = os.path.join(self.outDir, labelname) bbox_cnt = 0 if os.path.exists(self.labelfilename): with open(self.labelfilename) as f: for (i, line) in enumerate(f): if i == 0: bbox_cnt = int(line.strip()) continue tmp = [int(t.strip()) for t in line.split()] ## print tmp self.bboxList.append(tuple(tmp)) tmpId = self.mainPanel.create_rectangle(tmp[0], tmp[1], \ tmp[2], tmp[3], \ width = 2, \ outline = COLORS[(len(self.bboxList)-1) % len(COLORS)]) self.bboxIdList.append(tmpId) self.listbox.insert(END, '(%d, %d) -> (%d, %d)' %(tmp[0], tmp[1], tmp[2], tmp[3])) self.listbox.itemconfig(len(self.bboxIdList) - 1, fg = COLORS[(len(self.bboxIdList) - 1) % len(COLORS)]) def saveImage(self): with open(self.labelfilename, 'w') as f: f.write('%d\n' %len(self.bboxList)) for bbox in self.bboxList: f.write(' '.join(map(str, bbox)) + '\n') print 'Image No. %d saved' %(self.cur) def mouseClick(self, event): if self.STATE['click'] == 0: self.STATE['x'], self.STATE['y'] = event.x, event.y #self.STATE['x'], self.STATE['y'] = self.imgSize[0], self.imgSize[1] else: x1, x2 = min(self.STATE['x'], event.x), max(self.STATE['x'], event.x) y1, y2 = min(self.STATE['y'], event.y), max(self.STATE['y'], event.y) if x2 > self.imgSize[0]: x2 = self.imgSize[0] if y2 > self.imgSize[1]: y2 = self.imgSize[1] self.bboxList.append((self.currentClass, x1, y1, x2, y2)) self.bboxIdList.append(self.bboxId) self.bboxId = None self.listbox.insert(END, '(%d, %d) -> (%d, %d)' %(x1, y1, x2, y2)) self.listbox.itemconfig(len(self.bboxIdList) - 1, fg = COLORS[(len(self.bboxIdList) - 1) % len(COLORS)]) self.STATE['click'] = 1 - self.STATE['click'] def mouseMove(self, event): self.disp.config(text = 'x: %d, y: %d' %(event.x, event.y)) if self.tkimg: if self.hl: self.mainPanel.delete(self.hl) self.hl = self.mainPanel.create_line(0, event.y, self.tkimg.width(), event.y, width = 2) if self.vl: self.mainPanel.delete(self.vl) self.vl = self.mainPanel.create_line(event.x, 0, event.x, self.tkimg.height(), width = 2) if 1 == self.STATE['click']: if self.bboxId: self.mainPanel.delete(self.bboxId) self.bboxId = self.mainPanel.create_rectangle(self.STATE['x'], self.STATE['y'], \ event.x, event.y, \ width = 2, \ outline = COLORS[len(self.bboxList) % len(COLORS)]) def cancelBBox(self, event): if 1 == self.STATE['click']: if self.bboxId: self.mainPanel.delete(self.bboxId) self.bboxId = None self.STATE['click'] = 0 def delBBox(self): sel = self.listbox.curselection() if len(sel) != 1 : return idx = int(sel[0]) self.mainPanel.delete(self.bboxIdList[idx]) self.bboxIdList.pop(idx) self.bboxList.pop(idx) self.listbox.delete(idx) def clearBBox(self): for idx in range(len(self.bboxIdList)): self.mainPanel.delete(self.bboxIdList[idx]) self.listbox.delete(0, len(self.bboxList)) self.bboxIdList = [] self.bboxList = [] def selectmat(self): self.currentClass = 'mat' self.classbox.delete(0,END) self.classbox.insert(0, 'mat') self.classbox.itemconfig(0,fg = COLORS[0]) def selectdoor(self): self.currentClass = 'door' self.classbox.delete(0,END) self.classbox.insert(0, 'door') self.classbox.itemconfig(0,fg = COLORS[0]) def selectsofa(self): self.currentClass = 'sofa' self.classbox.delete(0,END) self.classbox.insert(0, 'sofa') self.classbox.itemconfig(0,fg = COLORS[0]) def selectchair(self): self.currentClass = 'chair' self.classbox.delete(0,END) self.classbox.insert(0, 'chair') self.classbox.itemconfig(0,fg = COLORS[0]) def selecttable(self): self.currentClass = 'table' self.classbox.delete(0,END) self.classbox.insert(0, 'table') self.classbox.itemconfig(0,fg = COLORS[0]) def selectbed(self): self.currentClass = 'bed' self.classbox.delete(0,END) self.classbox.insert(0, 'bed') self.classbox.itemconfig(0,fg = COLORS[0]) def selectashcan(self): self.currentClass = 'ashcan' self.classbox.delete(0,END) self.classbox.insert(0, 'ashcan') self.classbox.itemconfig(0,fg = COLORS[0]) def selectshoe(self): self.currentClass = 'shoe' self.classbox.delete(0,END) self.classbox.insert(0, 'shoe') self.classbox.itemconfig(0,fg = COLORS[0]) def prevImage(self, event = None): self.saveImage() if self.cur > 1: self.cur -= 1 self.loadImage() def nextImage(self, event = None): self.saveImage() if self.cur < self.total: self.cur += 1 self.loadImage() def gotoImage(self): idx = int(self.idxEntry.get()) if 1 <= idx and idx <= self.total: self.saveImage() self.cur = idx self.loadImage() ## def setImage(self, imagepath = r'test2.png'): ## self.img = Image.open(imagepath) ## self.tkimg = ImageTk.PhotoImage(self.img) ## self.mainPanel.config(width = self.tkimg.width()) ## self.mainPanel.config(height = self.tkimg.height()) ## self.mainPanel.create_image(0, 0, image = self.tkimg, anchor=NW) if __name__ == '__main__': root = Tk() tool = LabelTool(root) root.mainloop()
使用方法：
(1) 在BBox-Label-Tool/Images目录下创建保存图片的目录， 目录以数字命名(BBox-Label-Tool/Images/1), 然后将待标注的图片copy到1这个目录下;
(2) 在BBox-Label-Tool目录下执行命令 python main.py
(3) 在工具界面上, Image Dir 框中输入需要标记的目录名(比如 1), 然后点击load按钮, 工具自动将Images/1目录下的图片加载进来;
需要说明一下, 如果目录中的图片已经标注过,点击load时不会被重新加载进来.
(4) 该工具支持多类别标注, 画bounding boxs框标定之前,需要先选定类别,然后再画框.
(5) 一张图片标注完后, 点击Next>>按钮, 标注下一张图片, 图片label成功后,会在BBox-Label-Tool/Labels对应的目录下生成与图片文件名对应的label文件.
2 数据集的转换
caffe训练使用LMDB格式的数据，ssd框架中提供了voc数据格式转换成LMDB格式的脚本。
所以实践中先将BBox-Label-Tool标注的数据转换成voc数据格式，然后再转换成LMDB格式。
2.1 voc数据格式
(1)Annotations中保存的是xml格式的label信息
<?xml version="1.0" ?> <annotation> <folder>VOC2007</folder> <filename>1.jpg</filename> <source> <database>My Database</database> <annotation>VOC2007</annotation> <image>flickr</image> <flickrid>NULL</flickrid> </source> <owner> <flickrid>NULL</flickrid> <name>idaneel</name> </owner> <size> <width>320</width> <height>240</height> <depth>3</depth> </size> <segmented>0</segmented> <object> <name>door</name> <pose>Unspecified</pose> <truncated>0</truncated> <difficult>0</difficult> <bndbox> <xmin>109</xmin> <ymin>3</ymin> <xmax>199</xmax> <ymax>204</ymax> </bndbox> </object> </annotation>
(2)ImageSet目录下的Main目录里存放的是用于表示训练的图片集和测试的图片集
(3)JPEGImages目录下存放所有图片集
(4)label目录下保存的是BBox-Label-Tool工具标注好的bounding box坐标文件，
该目录下的文件就是待转换的label标签文件。
2.2 Label转换成VOC数据格式
BBox-Label-Tool工具标注好的bounding box坐标文件转换成VOC数据格式的形式.
具体的转换过程包括了两个步骤：
（1）将BBox-Label-Tool下的txt格式保存的bounding box信息转换成VOC数据格式下以xml方式表示；
（2）生成用于训练的数据集和用于测试的数据集。
用python实现了上述两个步骤的换转。
createXml.py 完成txt到xml的转换； 执行脚本./createXml.py
#!/usr/bin/env python import os import sys import cv2 from itertools import islice from xml.dom.minidom import Document labels='label' imgpath='JPEGImages/' xmlpath_new='Annotations/' foldername='VOC2007' def insertObject(doc, datas): obj = doc.createElement('object') name = doc.createElement('name') name.appendChild(doc.createTextNode(datas[0])) obj.appendChild(name) pose = doc.createElement('pose') pose.appendChild(doc.createTextNode('Unspecified')) obj.appendChild(pose) truncated = doc.createElement('truncated') truncated.appendChild(doc.createTextNode(str(0))) obj.appendChild(truncated) difficult = doc.createElement('difficult') difficult.appendChild(doc.createTextNode(str(0))) obj.appendChild(difficult) bndbox = doc.createElement('bndbox') xmin = doc.createElement('xmin') xmin.appendChild(doc.createTextNode(str(datas[1]))) bndbox.appendChild(xmin) ymin = doc.createElement('ymin') ymin.appendChild(doc.createTextNode(str(datas[2]))) bndbox.appendChild(ymin) xmax = doc.createElement('xmax') xmax.appendChild(doc.createTextNode(str(datas[3]))) bndbox.appendChild(xmax) ymax = doc.createElement('ymax') if '\r' == str(datas[4])[-1] or '\n' == str(datas[4])[-1]: data = str(datas[4])[0:-1] else: data = str(datas[4]) ymax.appendChild(doc.createTextNode(data)) bndbox.appendChild(ymax) obj.appendChild(bndbox) return obj def create(): for walk in os.walk(labels): for each in walk[2]: fidin=open(walk[0] + '/'+ each,'r') objIndex = 0 for data in islice(fidin, 1, None): objIndex += 1 data=data.strip('\n') datas = data.split(' ') if 5 != len(datas): print 'bounding box information error' continue pictureName = each.replace('.txt', '.jpg') imageFile = imgpath + pictureName img = cv2.imread(imageFile) imgSize = img.shape if 1 == objIndex: xmlName = each.replace('.txt', '.xml') f = open(xmlpath_new + xmlName, "w") doc = Document() annotation = doc.createElement('annotation') doc.appendChild(annotation) folder = doc.createElement('folder') folder.appendChild(doc.createTextNode(foldername)) annotation.appendChild(folder) filename = doc.createElement('filename') filename.appendChild(doc.createTextNode(pictureName)) annotation.appendChild(filename) source = doc.createElement('source') database = doc.createElement('database') database.appendChild(doc.createTextNode('My Database')) source.appendChild(database) source_annotation = doc.createElement('annotation') source_annotation.appendChild(doc.createTextNode(foldername)) source.appendChild(source_annotation) image = doc.createElement('image') image.appendChild(doc.createTextNode('flickr')) source.appendChild(image) flickrid = doc.createElement('flickrid') flickrid.appendChild(doc.createTextNode('NULL')) source.appendChild(flickrid) annotation.appendChild(source) owner = doc.createElement('owner') flickrid = doc.createElement('flickrid') flickrid.appendChild(doc.createTextNode('NULL')) owner.appendChild(flickrid) name = doc.createElement('name') name.appendChild(doc.createTextNode('idaneel')) owner.appendChild(name) annotation.appendChild(owner) size = doc.createElement('size') width = doc.createElement('width') width.appendChild(doc.createTextNode(str(imgSize[1]))) size.appendChild(width) height = doc.createElement('height') height.appendChild(doc.createTextNode(str(imgSize[0]))) size.appendChild(height) depth = doc.createElement('depth') depth.appendChild(doc.createTextNode(str(imgSize[2]))) size.appendChild(depth) annotation.appendChild(size) segmented = doc.createElement('segmented') segmented.appendChild(doc.createTextNode(str(0))) annotation.appendChild(segmented) annotation.appendChild(insertObject(doc, datas)) else: annotation.appendChild(insertObject(doc, datas)) try: f.write(doc.toprettyxml(indent = ' ')) f.close() fidin.close() except: pass if __name__ == '__main__': create()
createTest.py 生成训练集和测试集标识文件； 执行脚本
./createTest.py %startID% %endID% %testNumber%
#!/usr/bin/env python import os import sys import random try: start = int(sys.argv[1]) end = int(sys.argv[2]) test = int(sys.argv[3]) allNum = end-start+1 except: print 'Please input picture range' print './createTest.py 1 1500 500' os._exit(0) b_list = range(start,end) blist_webId = random.sample(b_list, test) blist_webId = sorted(blist_webId) allFile = [] testFile = open('ImageSets/Main/test.txt', 'w') trainFile = open('ImageSets/Main/trainval.txt', 'w') for i in range(allNum): allFile.append(i+1) for test in blist_webId: allFile.remove(test) testFile.write(str(test) + '\n') for train in allFile: trainFile.write(str(train) + '\n') testFile.close() trainFile.close()
说明： 由于BBox-Label-Tool实现相对简单，该工具每次只能对一个类别进行打标签，所以转换脚本
每一次也是对一个类别进行数据的转换，这个问题后续需要优化改进。
优化后的BBox-Label-Tool工具，支持多类别标定，生成的label文件中增加了类别名称信息。
使用时修改classLabels，改写成自己的类别, 修改后的工具代码参见1.1中的main.py
2.3 VOC数据转换成LMDB数据
SSD提供了VOC数据到LMDB数据的转换脚本 data/VOC0712/create_list.sh 和 ./data/VOC0712/create_data.sh，这两个脚本是完全针对VOC0712目录下的数据进行的转换。
实现中为了不破坏VOC0712目录下的数据内容，针对我们自己的数据集，修改了上面这两个脚本，
将脚本中涉及到VOC0712的信息替换成我们自己的目录信息。
在处理我们的数据集时，将VOC0712替换成indoor。
具体的步骤如下：
(1) 在 $HOME/data/VOCdevkit目录下创建indoor目录，该目录中存放自己转换完成的VOC数据集；
(2) $CAFFE_ROOT/examples目录下创建indoor目录；
(3) $CAFFE_ROOT/data目录下创建indoor目录，同时将data/VOC0712下的create_list.sh,create_data.sh,labelmap_voc.prototxt
这三个文件copy到indoor目录下，分别重命名为create_list_indoor.sh,create_data_indoor.sh, labelmap_indoor.prototxt
(4)对上面新生成的两个create文件进行修改，主要修改是将VOC0712相关的信息替换成indoor
修改后的这两个文件分别为：
#!/bin/bash root_dir=$HOME/data/VOCdevkit/ sub_dir=ImageSets/Main bash_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" for dataset in trainval test do dst_file=$bash_dir/$dataset.txt if [ -f $dst_file ] then rm -f $dst_file fi for name in indoor do if [[ $dataset == "test" && $name == "VOC2012" ]] then continue fi echo "Create list for $name $dataset..." dataset_file=$root_dir/$name/$sub_dir/$dataset.txt img_file=$bash_dir/$dataset"_img.txt" cp $dataset_file $img_file sed -i "s/^/$name\/JPEGImages\//g" $img_file sed -i "s/$/.jpg/g" $img_file label_file=$bash_dir/$dataset"_label.txt" cp $dataset_file $label_file sed -i "s/^/$name\/Annotations\//g" $label_file sed -i "s/$/.xml/g" $label_file paste -d' ' $img_file $label_file >> $dst_file rm -f $label_file rm -f $img_file done # Generate image name and size infomation. if [ $dataset == "test" ] then $bash_dir/../../build/tools/get_image_size $root_dir $dst_file $bash_dir/$dataset"_name_size.txt" fi # Shuffle trainval file. if [ $dataset == "trainval" ] then rand_file=$dst_file.random cat $dst_file | perl -MList::Util=shuffle -e 'print shuffle(<STDIN>);' > $rand_file mv $rand_file $dst_file fi done
cur_dir=$(cd $( dirname ${BASH_SOURCE[0]} ) && pwd ) root_dir=$cur_dir/../.. cd $root_dir redo=1 data_root_dir="$HOME/data/VOCdevkit" dataset_name="indoor" mapfile="$root_dir/data/$dataset_name/labelmap_indoor.prototxt" anno_type="detection" db="lmdb" min_dim=0 max_dim=0 width=0 height=0 extra_cmd="--encode-type=jpg --encoded" if [ $redo ] then extra_cmd="$extra_cmd --redo" fi for subset in test trainval do python $root_dir/scripts/create_annoset.py --anno-type=$anno_type --label-map-file=$mapfile --min-dim=$min_dim --max-dim=$max_dim --resize-width=$width --resize-height=$height --check-label $extra_cmd $data_root_dir $root_dir/data/$dataset_name/$subset.txt $data_root_dir/$dataset_name/$db/$dataset_name"_"$subset"_"$db examples/$dataset_name done
(5)修改labelmap_indoor.prototxt，将该文件中的类别修改成和自己的数据集相匹配，注意需要保留一个label 0 , background类别
item {
name: "none_of_the_above"
label: 0
display_name: "background"
}
item {
name: "door"
label: 1
display_name: "door"
}
完成上面步骤的修改后，可以开始LMDB数据数据的制作，在$CAFFE_ROOT目录下分别运行：
./data/indoor/create_list_indoor.sh
./data/indoor/create_data_indoor.sh
命令执行完毕后，可以在$CAFFE_ROOT/indoor目录下查看转换完成的LMDB数据数据。
3 使用SSD进行自己数据集的训练
训练时使用ssd demo中提供的预训练好的VGGnet model : VGG_ILSVRC_16_layers_fc_reduced.caffemodel
将该模型保存到$CAFFE_ROOT/models/VGGNet下。
将ssd_pascal.py copy一份 ssd_pascal_indoor.py文件， 根据自己的数据集修改ssd_pascal_indoor.py
主要修改点：
（1）train_data和test_data修改成指向自己的数据集LMDB
train_data = "examples/indoor/indoor_trainval_lmdb"
test_data = "examples/indoor/indoor_test_lmdb"
（2） num_test_image该变量修改成自己数据集中测试数据的数量
（3）num_classes 该变量修改成自己数据集中 标签类别数量数 + 1
针对我的数据集，ssd_pascal_indoor.py的内容为：
from __future__ import print_function import caffe from caffe.model_libs import * from google.protobuf import text_format import math import os import shutil import stat import subprocess import sys # Add extra layers on top of a "base" network (e.g. VGGNet or Inception). def AddExtraLayers(net, use_batchnorm=True): use_relu = True # Add additional convolutional layers. from_layer = net.keys()[-1] # TODO(weiliu89): Construct the name using the last layer to avoid duplication. out_layer = "conv6_1" ConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 256, 1, 0, 1) from_layer = out_layer out_layer = "conv6_2" ConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 512, 3, 1, 2) for i in xrange(7, 9): from_layer = out_layer out_layer = "conv{}_1".format(i) ConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 128, 1, 0, 1) from_layer = out_layer out_layer = "conv{}_2".format(i) ConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 256, 3, 1, 2) # Add global pooling layer. name = net.keys()[-1] net.pool6 = L.Pooling(net[name], pool=P.Pooling.AVE, global_pooling=True) return net ### Modify the following parameters accordingly ### # The directory which contains the caffe code. # We assume you are running the script at the CAFFE_ROOT. caffe_root = os.getcwd() # Set true if you want to start training right after generating all files. run_soon = True # Set true if you want to load from most recently saved snapshot. # Otherwise, we will load from the pretrain_model defined below. resume_training = True # If true, Remove old model files. remove_old_models = False # The database file for training data. Created by data/VOC0712/create_data.sh train_data = "examples/indoor/indoor_trainval_lmdb" # The database file for testing data. Created by data/VOC0712/create_data.sh test_data = "examples/indoor/indoor_test_lmdb" # Specify the batch sampler. resize_width = 300 resize_height = 300 resize = "{}x{}".format(resize_width, resize_height) batch_sampler = [ { 'sampler': { }, 'max_trials': 1, 'max_sample': 1, }, { 'sampler': { 'min_scale': 0.3, 'max_scale': 1.0, 'min_aspect_ratio': 0.5, 'max_aspect_ratio': 2.0, }, 'sample_constraint': { 'min_jaccard_overlap': 0.1, }, 'max_trials': 50, 'max_sample': 1, }, { 'sampler': { 'min_scale': 0.3, 'max_scale': 1.0, 'min_aspect_ratio': 0.5, 'max_aspect_ratio': 2.0, }, 'sample_constraint': { 'min_jaccard_overlap': 0.3, }, 'max_trials': 50, 'max_sample': 1, }, { 'sampler': { 'min_scale': 0.3, 'max_scale': 1.0, 'min_aspect_ratio': 0.5, 'max_aspect_ratio': 2.0, }, 'sample_constraint': { 'min_jaccard_overlap': 0.5, }, 'max_trials': 50, 'max_sample': 1, }, { 'sampler': { 'min_scale': 0.3, 'max_scale': 1.0, 'min_aspect_ratio': 0.5, 'max_aspect_ratio': 2.0, }, 'sample_constraint': { 'min_jaccard_overlap': 0.7, }, 'max_trials': 50, 'max_sample': 1, }, { 'sampler': { 'min_scale': 0.3, 'max_scale': 1.0, 'min_aspect_ratio': 0.5, 'max_aspect_ratio': 2.0, }, 'sample_constraint': { 'min_jaccard_overlap': 0.9, }, 'max_trials': 50, 'max_sample': 1, }, { 'sampler': { 'min_scale': 0.3, 'max_scale': 1.0, 'min_aspect_ratio': 0.5, 'max_aspect_ratio': 2.0, }, 'sample_constraint': { 'max_jaccard_overlap': 1.0, }, 'max_trials': 50, 'max_sample': 1, }, ] train_transform_param = { 'mirror': True, 'mean_value': [104, 117, 123], 'resize_param': { 'prob': 1, 'resize_mode': P.Resize.WARP, 'height': resize_height, 'width': resize_width, 'interp_mode': [ P.Resize.LINEAR, P.Resize.AREA, P.Resize.NEAREST, P.Resize.CUBIC, P.Resize.LANCZOS4, ], }, 'emit_constraint': { 'emit_type': caffe_pb2.EmitConstraint.CENTER, } } test_transform_param = { 'mean_value': [104, 117, 123], 'resize_param': { 'prob': 1, 'resize_mode': P.Resize.WARP, 'height': resize_height, 'width': resize_width, 'interp_mode': [P.Resize.LINEAR], }, } # If true, use batch norm for all newly added layers. # Currently only the non batch norm version has been tested. use_batchnorm = False # Use different initial learning rate. if use_batchnorm: base_lr = 0.0004 else: # A learning rate for batch_size = 1, num_gpus = 1. base_lr = 0.00004 # Modify the job name if you want. job_name = "SSD_{}".format(resize) # The name of the model. Modify it if you want. model_name = "VGG_VOC0712_{}".format(job_name) # Directory which stores the model .prototxt file. save_dir = "models/VGGNet/VOC0712/{}".format(job_name) # Directory which stores the snapshot of models. snapshot_dir = "models/VGGNet/VOC0712/{}".format(job_name) # Directory which stores the job script and log file. job_dir = "jobs/VGGNet/VOC0712/{}".format(job_name) # Directory which stores the detection results. output_result_dir = "{}/data/VOCdevkit/results/VOC2007/{}/Main".format(os.environ['HOME'], job_name) # model definition files. train_net_file = "{}/train.prototxt".format(save_dir) test_net_file = "{}/test.prototxt".format(save_dir) deploy_net_file = "{}/deploy.prototxt".format(save_dir) solver_file = "{}/solver.prototxt".format(save_dir) # snapshot prefix. snapshot_prefix = "{}/{}".format(snapshot_dir, model_name) # job script path. job_file = "{}/{}.sh".format(job_dir, model_name) # Stores the test image names and sizes. Created by data/VOC0712/create_list.sh name_size_file = "data/indoor/test_name_size.txt" # The pretrained model. We use the Fully convolutional reduced (atrous) VGGNet. pretrain_model = "models/VGGNet/VGG_ILSVRC_16_layers_fc_reduced.caffemodel" # Stores LabelMapItem. label_map_file = "data/indoor/labelmap_indoor.prototxt" # MultiBoxLoss parameters. num_classes = 2 share_location = True background_label_id=0 train_on_diff_gt = True normalization_mode = P.Loss.VALID code_type = P.PriorBox.CENTER_SIZE neg_pos_ratio = 3. loc_weight = (neg_pos_ratio + 1.) / 4. multibox_loss_param = { 'loc_loss_type': P.MultiBoxLoss.SMOOTH_L1, 'conf_loss_type': P.MultiBoxLoss.SOFTMAX, 'loc_weight': loc_weight, 'num_classes': num_classes, 'share_location': share_location, 'match_type': P.MultiBoxLoss.PER_PREDICTION, 'overlap_threshold': 0.5, 'use_prior_for_matching': True, 'background_label_id': background_label_id, 'use_difficult_gt': train_on_diff_gt, 'do_neg_mining': True, 'neg_pos_ratio': neg_pos_ratio, 'neg_overlap': 0.5, 'code_type': code_type, } loss_param = { 'normalization': normalization_mode, } # parameters for generating priors. # minimum dimension of input image min_dim = 300 # conv4_3 ==> 38 x 38 # fc7 ==> 19 x 19 # conv6_2 ==> 10 x 10 # conv7_2 ==> 5 x 5 # conv8_2 ==> 3 x 3 # pool6 ==> 1 x 1 mbox_source_layers = ['conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'pool6'] # in percent % min_ratio = 20 max_ratio = 95 step = int(math.floor((max_ratio - min_ratio) / (len(mbox_source_layers) - 2))) min_sizes = [] max_sizes = [] for ratio in xrange(min_ratio, max_ratio + 1, step): min_sizes.append(min_dim * ratio / 100.) max_sizes.append(min_dim * (ratio + step) / 100.) min_sizes = [min_dim * 10 / 100.] + min_sizes max_sizes = [[]] + max_sizes aspect_ratios = [[2], [2, 3], [2, 3], [2, 3], [2, 3], [2, 3]] # L2 normalize conv4_3. normalizations = [20, -1, -1, -1, -1, -1] # variance used to encode/decode prior bboxes. if code_type == P.PriorBox.CENTER_SIZE: prior_variance = [0.1, 0.1, 0.2, 0.2] else: prior_variance = [0.1] flip = True clip = True # Solver parameters. # Defining which GPUs to use. gpus = "0" gpulist = gpus.split(",") num_gpus = len(gpulist) # Divide the mini-batch to different GPUs. batch_size = 4 accum_batch_size = 32 iter_size = accum_batch_size / batch_size solver_mode = P.Solver.CPU device_id = 0 batch_size_per_device = batch_size if num_gpus > 0: batch_size_per_device = int(math.ceil(float(batch_size) / num_gpus)) iter_size = int(math.ceil(float(accum_batch_size) / (batch_size_per_device * num_gpus))) solver_mode = P.Solver.GPU device_id = int(gpulist[0]) if normalization_mode == P.Loss.NONE: base_lr /= batch_size_per_device elif normalization_mode == P.Loss.VALID: base_lr *= 25. / loc_weight elif normalization_mode == P.Loss.FULL: # Roughly there are 2000 prior bboxes per image. # TODO(weiliu89): Estimate the exact # of priors. base_lr *= 2000. # Which layers to freeze (no backward) during training. freeze_layers = ['conv1_1', 'conv1_2', 'conv2_1', 'conv2_2'] # Evaluate on whole test set. num_test_image = 800 test_batch_size = 1 test_iter = num_test_image / test_batch_size solver_param = { # Train parameters 'base_lr': base_lr, 'weight_decay': 0.0005, 'lr_policy': "step", 'stepsize': 40000, 'gamma': 0.1, 'momentum': 0.9, 'iter_size': iter_size, 'max_iter': 60000, 'snapshot': 40000, 'display': 10, 'average_loss': 10, 'type': "SGD", 'solver_mode': solver_mode, 'device_id': device_id, 'debug_info': False, 'snapshot_after_train': True, # Test parameters 'test_iter': [test_iter], 'test_interval': 10000, 'eval_type': "detection", 'ap_version': "11point", 'test_initialization': False, } # parameters for generating detection output. det_out_param = { 'num_classes': num_classes, 'share_location': share_location, 'background_label_id': background_label_id, 'nms_param': {'nms_threshold': 0.45, 'top_k': 400}, 'save_output_param': { 'output_directory': output_result_dir, 'output_name_prefix': "comp4_det_test_", 'output_format': "VOC", 'label_map_file': label_map_file, 'name_size_file': name_size_file, 'num_test_image': num_test_image, }, 'keep_top_k': 200, 'confidence_threshold': 0.01, 'code_type': code_type, } # parameters for evaluating detection results. det_eval_param = { 'num_classes': num_classes, 'background_label_id': background_label_id, 'overlap_threshold': 0.5, 'evaluate_difficult_gt': False, 'name_size_file': name_size_file, } ### Hopefully you don't need to change the following ### # Check file. check_if_exist(train_data) check_if_exist(test_data) check_if_exist(label_map_file) check_if_exist(pretrain_model) make_if_not_exist(save_dir) make_if_not_exist(job_dir) make_if_not_exist(snapshot_dir) # Create train net. net = caffe.NetSpec() net.data, net.label = CreateAnnotatedDataLayer(train_data, batch_size=batch_size_per_device, train=True, output_label=True, label_map_file=label_map_file, transform_param=train_transform_param, batch_sampler=batch_sampler) VGGNetBody(net, from_layer='data', fully_conv=True, reduced=True, dilated=True, dropout=False, freeze_layers=freeze_layers) AddExtraLayers(net, use_batchnorm) mbox_layers = CreateMultiBoxHead(net, data_layer='data', from_layers=mbox_source_layers, use_batchnorm=use_batchnorm, min_sizes=min_sizes, max_sizes=max_sizes, aspect_ratios=aspect_ratios, normalizations=normalizations, num_classes=num_classes, share_location=share_location, flip=flip, clip=clip, prior_variance=prior_variance, kernel_size=3, pad=1) # Create the MultiBoxLossLayer. name = "mbox_loss" mbox_layers.append(net.label) net[name] = L.MultiBoxLoss(*mbox_layers, multibox_loss_param=multibox_loss_param, loss_param=loss_param, include=dict(phase=caffe_pb2.Phase.Value('TRAIN')), propagate_down=[True, True, False, False]) with open(train_net_file, 'w') as f: print('name: "{}_train"'.format(model_name), file=f) print(net.to_proto(), file=f) shutil.copy(train_net_file, job_dir) # Create test net. net = caffe.NetSpec() net.data, net.label = CreateAnnotatedDataLayer(test_data, batch_size=test_batch_size, train=False, output_label=True, label_map_file=label_map_file, transform_param=test_transform_param) VGGNetBody(net, from_layer='data', fully_conv=True, reduced=True, dilated=True, dropout=False, freeze_layers=freeze_layers) AddExtraLayers(net, use_batchnorm) mbox_layers = CreateMultiBoxHead(net, data_layer='data', from_layers=mbox_source_layers, use_batchnorm=use_batchnorm, min_sizes=min_sizes, max_sizes=max_sizes, aspect_ratios=aspect_ratios, normalizations=normalizations, num_classes=num_classes, share_location=share_location, flip=flip, clip=clip, prior_variance=prior_variance, kernel_size=3, pad=1) conf_name = "mbox_conf" if multibox_loss_param["conf_loss_type"] == P.MultiBoxLoss.SOFTMAX: reshape_name = "{}_reshape".format(conf_name) net[reshape_name] = L.Reshape(net[conf_name], shape=dict(dim=[0, -1, num_classes])) softmax_name = "{}_softmax".format(conf_name) net[softmax_name] = L.Softmax(net[reshape_name], axis=2) flatten_name = "{}_flatten".format(conf_name) net[flatten_name] = L.Flatten(net[softmax_name], axis=1) mbox_layers[1] = net[flatten_name] elif multibox_loss_param["conf_loss_type"] == P.MultiBoxLoss.LOGISTIC: sigmoid_name = "{}_sigmoid".format(conf_name) net[sigmoid_name] = L.Sigmoid(net[conf_name]) mbox_layers[1] = net[sigmoid_name] net.detection_out = L.DetectionOutput(*mbox_layers, detection_output_param=det_out_param, include=dict(phase=caffe_pb2.Phase.Value('TEST'))) net.detection_eval = L.DetectionEvaluate(net.detection_out, net.label, detection_evaluate_param=det_eval_param, include=dict(phase=caffe_pb2.Phase.Value('TEST'))) with open(test_net_file, 'w') as f: print('name: "{}_test"'.format(model_name), file=f) print(net.to_proto(), file=f) shutil.copy(test_net_file, job_dir) # Create deploy net. # Remove the first and last layer from test net. deploy_net = net with open(deploy_net_file, 'w') as f: net_param = deploy_net.to_proto() # Remove the first (AnnotatedData) and last (DetectionEvaluate) layer from test net. del net_param.layer[0] del net_param.layer[-1] net_param.name = '{}_deploy'.format(model_name) net_param.input.extend(['data']) net_param.input_shape.extend([ caffe_pb2.BlobShape(dim=[1, 3, resize_height, resize_width])]) print(net_param, file=f) shutil.copy(deploy_net_file, job_dir) # Create solver. solver = caffe_pb2.SolverParameter( train_net=train_net_file, test_net=[test_net_file], snapshot_prefix=snapshot_prefix, **solver_param) with open(solver_file, 'w') as f: print(solver, file=f) shutil.copy(solver_file, job_dir) max_iter = 0 # Find most recent snapshot. for file in os.listdir(snapshot_dir): if file.endswith(".solverstate"): basename = os.path.splitext(file)[0] iter = int(basename.split("{}_iter_".format(model_name))[1]) if iter > max_iter: max_iter = iter train_src_param = '--weights="{}" \\\n'.format(pretrain_model) if resume_training: if max_iter > 0: train_src_param = '--snapshot="{}_iter_{}.solverstate" \\\n'.format(snapshot_prefix, max_iter) if remove_old_models: # Remove any snapshots smaller than max_iter. for file in os.listdir(snapshot_dir): if file.endswith(".solverstate"): basename = os.path.splitext(file)[0] iter = int(basename.split("{}_iter_".format(model_name))[1]) if max_iter > iter: os.remove("{}/{}".format(snapshot_dir, file)) if file.endswith(".caffemodel"): basename = os.path.splitext(file)[0] iter = int(basename.split("{}_iter_".format(model_name))[1]) if max_iter > iter: os.remove("{}/{}".format(snapshot_dir, file)) # Create job file. with open(job_file, 'w') as f: f.write('cd {}\n'.format(caffe_root)) f.write('./build/tools/caffe train \\\n') f.write('--solver="{}" \\\n'.format(solver_file)) f.write(train_src_param) if solver_param['solver_mode'] == P.Solver.GPU: f.write('--gpu {} 2>&1 | tee {}/{}.log\n'.format(gpus, job_dir, model_name)) else: f.write('2>&1 | tee {}/{}.log\n'.format(job_dir, model_name)) # Copy the python script to job_dir. py_file = os.path.abspath(__file__) shutil.copy(py_file, job_dir) # Run the job. os.chmod(job_file, stat.S_IRWXU) if run_soon: subprocess.call(job_file, shell=True)
训练命令：
python examples/ssd/ssd_pascal_indoor.py4 测试
SSD框架中提供了测试代码，有C++版本和python版本
4.1 c++版本
编译完SSD后，C++版本的的可执行文件存放目录： .build_release/examples/ssd/ssd_detect.bin
测试命令 ./.build_release/examples/ssd/ssd_detect.binmodels/VGGNet/indoor/deploy.prototxt models/VGGNet/indoor/VGG_VOC0712_SSD_300x300_iter_60000.caffemodel pictures.txt
其中pictures.txt中保存的是待测试图片的list
4.2 python版本
python 版本的测试过程参见examples/detection.ipynb
参考：
1 将数据集做成VOC2007格式用于Faster-RCNN训练
2 SSD的配置安装与测试
