原文链接 www.cnblogs.com/zhiyishou/p/5651321.html
具体代码见https://github.com/zhiyishou/py-faster-rcnn
![[转载]使用ImageNet在faster-rcnn上训练自己的分类网络_第1张图片](http://img.e-com-net.com/image/info10/5bc7471fd2f640b6b9a0e90ba02ca69c.jpg)
这是我对cup, glasses训练的识别
faster-rcnn在fast-rcnn的基础上加了rpn来将整个训练都置于GPU内，以用来提高效率，这里我们将使用ImageNet的数据集来在faster-rcnn上来训练自己的分类器。从ImageNet上可下载到很多类别的Image与bounding box annotation来进行训练（每一个类别下的annotation都少于等于image的个数，所以我们从annotation来建立索引）。
在lib/dataset/factory.py中提供了coco与voc的数据集获取方法，而我们要做的就是在这里加上我们自己的ImageNet获取方法，我们先来建立ImageNet数据获取主文件。coco与pascal_voc的获取都是继承于父类imdb，所以我们可根据pascal_voc的获取方法来做模板修改完成我们的ImageNet类。
创建ImageNet类
由于在faster-rcnn里使用rpn来代替了selective_search，所以我们可以在使用时直接略过有关selective_search的方法，根据pascal_voc类做模板，我们需要留下的方法有：
__init__//初始化image_path_at//根据数据集列表的index来取图片绝对地址image_path_from_index//配合上面_load_image_set_index//获取数据集列表_gt_roidb//获取ground-truth数据rpn_roidb//获取region proposal数据_load_rpn_roidb//根据gt_roidb生成rpn_roidb数据并合成_load_psacal_annotation//加载annotation文件并对bounding box进行数据整理
__init__:
def __init__(self, image_set): imdb.__init__(self,'imagenet')self._image_set = image_setself._data_path = os.path.join(cfg.DATA_DIR,"imagenet")#类别与对应的wnid，可以修改成自己要训练的类别self._class_wnids = {'cup':'n03147509','glasses':'n04272054'}#类别，修改类别时同时要修改这里self._classes = ('__background__',self._class_wnids['cup'],self._class_wnids['glasses'])self._class_to_ind = dict(zip(self.classes, xrange(self.num_classes)))#bounding box annotation 文件的目录self._xml_path = os.path.join(self._data_path,"Annotations")self._image_ext ='.JPEG'#我们使用xml文件名来做数据集的索引# the xml file name and each one corresponding to image file nameself._image_index =self._load_xml_filenames()self._salt = str(uuid.uuid4())self._comp_id ='comp4'self.config = {'cleanup':True,'use_salt':True,'use_diff':False,'matlab_eval':False,'rpn_file': None,'min_size':2} assert os.path.exists(self._data_path), \'Path does not exist: {}'.format(self._data_path)
image_path_at
defimage_path_at(self, i):#使用index来从xml_filenames取到filename，生成绝对路径returnself.image_path_from_image_filename(self._image_index[i])
image_path_from_image_filename(类似pascal_voc中的image_path_from_index)
defimage_path_from_image_filename(self, image_filename):image_path = os.path.join(self._data_path,'Images', image_filename + self._image_ext)assertos.path.exists(image_path), \'Path does not exist: {}'.format(image_path)returnimage_path
_load_xml_filenames(类似pascal_voc中的_load_image_set_index)
def_load_xml_filenames(self):#从Annotations文件夹中拿取到bounding box annotation文件名#用来做数据集的索引xml_folder_path = os.path.join(self._data_path,"Annotations")assertos.path.exists(xml_folder_path), \'Path does not exist: {}'.format(xml_folder_path)fordirpath, dirnames, filenamesinos.walk(xml_folder_path): xml_filenames = [xml_filename.split(".")[0]forxml_filenameinfilenames]returnxml_filenames
gt_roidb
defgt_roidb(self):#Ground-Truth 数据缓存cache_file = os.path.join(self.cache_path, self.name +'_gt_roidb.pkl')ifos.path.exists(cache_file):withopen(cache_file,'rb')asfid: roidb = cPickle.load(fid)print'{} gt roidb loaded from {}'.format(self.name, cache_file)returnroidb#从xml中获取Ground-Truth数据gt_roidb = [self._load_imagenet_annotation(xml_filename)forxml_filenameinself._image_index]withopen(cache_file,'wb')asfid: cPickle.dump(gt_roidb, fid, cPickle.HIGHEST_PROTOCOL)print'wrote gt roidb to {}'.format(cache_file)returngt_roidb
rpn_roidb
defrpn_roidb(self):#根据gt_roidb生成rpn_roidb，并进行合并gt_roidb = self.gt_roidb() rpn_roidb = self._load_rpn_roidb(gt_roidb) roidb = imdb.merge_roidbs(gt_roidb, rpn_roidb)returnroidb
_load_rpn_roidb
def_load_rpn_roidb(self, gt_roidb):filename = self.config['rpn_file']print'loading {}'.format(filename)assertos.path.exists(filename), \'rpn data not found at: {}'.format(filename)withopen(filename,'rb')asf: box_list = cPickle.load(f)returnself.create_roidb_from_box_list(box_list, gt_roidb)
_load_imagenet_annotation(类似于pascal_voc中的_load_pascal_annotation)
def_load_imagenet_annotation(self, xml_filename):#从annotation的xml文件中拿取bounding box数据filepath = os.path.join(self._data_path,'Annotations', xml_filename +'.xml')#这里使用了ap，是我写的一个annotation parser，在后面贴出代码#它会返回这个xml文件的wnid, 图像文件名，以及里面包含的注解物体wnid, image_name, objects = ap.parse(filepath) num_objs = len(objects) boxes = np.zeros((num_objs,4), dtype=np.uint16) gt_classes = np.zeros((num_objs), dtype=np.int32) overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32) seg_areas = np.zeros((num_objs), dtype=np.float32)# Load object bounding boxes into a data frame.forix, objinenumerate(objects): box = obj["box"] x1 = box['xmin'] y1 = box['ymin'] x2 = box['xmax'] y2 = box['ymax']# 如果这个bounding box并不是我们想要学习的类别，那则跳过# go next if the wnid not exist in declared classestry: cls = self._class_to_ind[obj["wnid"]]exceptKeyError:print"wnid %s isn't show in given"%obj["wnid"]continueboxes[ix, :] = [x1, y1, x2, y2] gt_classes[ix] = cls overlaps[ix, cls] =1.0seg_areas[ix] = (x2 - x1 +1) * (y2 - y1 +1) overlaps = scipy.sparse.csr_matrix(overlaps)return{'boxes': boxes,'gt_classes': gt_classes,'gt_overlaps': overlaps,'flipped':False,'seg_areas': seg_areas}
annotation_parser.py文件
importosimportxml.dom.minidomdefgetText(node):returnnode.firstChild.nodeValuedefgetWnid(node):returngetText(node.getElementsByTagName("name")[0])defgetImageName(node):returngetText(node.getElementsByTagName("filename")[0])defgetObjects(node):objects = []forobjinnode.getElementsByTagName("object"): objects.append({"wnid": getText(obj.getElementsByTagName("name")[0]),"box":{"xmin": int(getText(obj.getElementsByTagName("xmin")[0])),"ymin": int(getText(obj.getElementsByTagName("ymin")[0])),"xmax": int(getText(obj.getElementsByTagName("xmax")[0])),"ymax": int(getText(obj.getElementsByTagName("ymax")[0])), } })returnobjectsdefparse(filepath):dom = xml.dom.minidom.parse(filepath) root = dom.documentElement image_name = getImageName(root) wnid = getWnid(root) objects = getObjects(root)returnwnid, image_name, objects
则对数据结构的要求是：
|---data|---imagenet|---Annotations|---n03147509|---n03147509_*.xml|---...|---n04272054|---n04272054_*.xml|---...|---Images|---n03147508_*.JPEG|---...|---n04272054_*.JPEG|---...
同时我在github上也提供了draw方法，可以用来将bounding box画于Image文件上，用来甄别该annotation的正确性
训练
这样，我们的ImageNet类则是生成好了，下面我们则可以训练我们的数据，但是在开始之前，还有一件事情，那就是修改prototxt中的与类别数目有关的值，我将models/pascal_voc拷贝到了models/imagenet进行修改，比如我想要训练ZF，如果使用的是train_faster_rcnn_alt_opt.py，则需要修改models/imagenet/ZF/faster_rcnn_alt_opt/下的所有pt文件里的内容，用如下的法则去替换：
//num为类别的个数input-data->num_classes = numclass_score->num_output = numbbox_pred->num_output = num*4
我这里使用train_faster_rcnn_alt_opt.py进行的训练，这样的话则需要把添加的models/imagenet作为可选项
//pt_type 则是添加的选择项，默认使用psacal_voc的models./tools/train_faster_rcnn_alt_opt.py--gpu 0 \--net_name ZF \--weights data/imagenet_models/ZF.v2.caffemodel[optional] \--imdb imagenet \--cfg experiments/cfgs/faster_rcnn_alt_opt.yml \--pt_type imagenet
识别
这里我们则需要使用刚训练出来的模型进行识别
#就像demo.py一样，但是使用训练的models，我创建了tools/classify.py来单独识别prototxt = os.path.join(cfg.ROOT_DIR, 'models/imagenet', NETS[args.demo_net][0], 'faster_rcnn_alt_opt', 'faster_rcnn_test.pt')caffemodel = os.path.join(cfg.ROOT_DIR, 'output/faster_rcnn_alt_opt/imagenet/'+ NETS[args.demo_net][0] +'_faster_rcnn_final.caffemodel')
同样，在识别前我们要对识别方法里的Classes进行修改，修改成你自己训练的类别后
执行
./tools/classify.py--net zf
则可对data/demo下的图片文件使用训练的zf网络进行识别
Have fun