使用ImageNet在faster-rcnn上训练本身的分类网络

具体代码见https://github.com/zhiyishou/py-faster-rcnn

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，以用来提升效率，这里咱们将使用ImageNet的数据集来在faster-rcnn上来训练本身的分类器。从ImageNet上可下载到不少类别的Image与bounding box annotation来进行训练（每个类别下的annotation都少于等于image的个数，因此咱们从annotation来创建索引）。

在lib/dataset/factory.py中提供了coco与voc的数据集获取方法，而咱们要作的就是在这里加上咱们本身的ImageNet获取方法，咱们先来创建ImageNet数据获取主文件。coco与pascal_voc的获取都是继承于父类imdb，因此咱们可根据pascal_voc的获取方法来作模板修改完成咱们的ImageNet类。

建立ImageNet类

因为在faster-rcnn里使用rpn来代替了selective_search，因此咱们能够在使用时直接略过有关selective_search的方法，根据pascal_voc类作模板，咱们须要留下的方法有：

__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_set
        self._data_path = os.path.join(cfg.DATA_DIR, "imagenet")
        #类别与对应的wnid，能够修改为本身要训练的类别
        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'
        #咱们使用xml文件名来作数据集的索引
        # the xml file name and each one corresponding to image file name
        self._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

def image_path_at(self, i):
        #使用index来从xml_filenames取到filename，生成绝对路径
        return self.image_path_from_image_filename(self._image_index[i])

image_path_from_image_filename(相似pascal_voc中的image_path_from_index)

def image_path_from_image_filename(self, image_filename):
        image_path = os.path.join(self._data_path, 'Images',
                                  image_filename + self._image_ext)
        assert os.path.exists(image_path), \
                'Path does not exist: {}'.format(image_path)
        return image_path

_load_xml_filenames(相似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
        #用来作数据集的索引
        xml_folder_path = os.path.join(self._data_path, "Annotations")
        assert os.path.exists(xml_folder_path), \
            'Path does not exist: {}'.format(xml_folder_path)

        for dirpath, dirnames, filenames in os.walk(xml_folder_path):
                xml_filenames = [xml_filename.split(".")[0] for xml_filename in filenames]

        return xml_filenames

gt_roidb

def gt_roidb(self):
        #Ground-Truth 数据缓存
        cache_file = os.path.join(self.cache_path, self.name + '_gt_roidb.pkl')
        if os.path.exists(cache_file):
            with open(cache_file, 'rb') as fid:
                roidb = cPickle.load(fid)
            print '{} gt roidb loaded from {}'.format(self.name, cache_file)
            return roidb

        #从xml中获取Ground-Truth数据
        gt_roidb = [self._load_imagenet_annotation(xml_filename)
                    for xml_filename in self._image_index]
        with open(cache_file, 'wb') as fid:
            cPickle.dump(gt_roidb, fid, cPickle.HIGHEST_PROTOCOL)
        print 'wrote gt roidb to {}'.format(cache_file)

        return gt_roidb

rpn_roidb

def rpn_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)

        return roidb

_load_rpn_roidb

def _load_rpn_roidb(self, gt_roidb):
        filename = self.config['rpn_file']
        print 'loading {}'.format(filename)
        assert os.path.exists(filename), \
               'rpn data not found at: {}'.format(filename)
        with open(filename, 'rb') as f:
            box_list = cPickle.load(f)
        return self.create_roidb_from_box_list(box_list, gt_roidb)

_load_imagenet_annotation(相似于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.
        for ix, obj in enumerate(objects):
            box = obj["box"]
            x1 = box['xmin']
            y1 = box['ymin']
            x2 = box['xmax']
            y2 = box['ymax']
            # 若是这个bounding box并非咱们想要学习的类别，那则跳过
            # go next if the wnid not exist in declared classes
            try:
                cls = self._class_to_ind[obj["wnid"]]
            except KeyError:
                print "wnid %s isn't show in given"%obj["wnid"]
                continue
            boxes[ix, :] = [x1, y1, x2, y2]
            gt_classes[ix] = cls
            overlaps[ix, cls] = 1.0
            seg_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

import os
import xml.dom.minidom

def getText(node):
    return node.firstChild.nodeValue

def getWnid(node):
    return getText(node.getElementsByTagName("name")[0])

def getImageName(node):
    return getText(node.getElementsByTagName("filename")[0])

def getObjects(node):
    objects = []
    for obj in node.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])),
            }
        })
    return objects

def parse(filepath):
    dom = xml.dom.minidom.parse(filepath)
    root = dom.documentElement
    image_name = getImageName(root)
    wnid = getWnid(root)
    objects = getObjects(root)
    
    return wnid, 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，能够用来将bounding box画于Image文件上，用来甄别该annotation的正确性

训练

这样，咱们的ImageNet类则是生成好了，下面咱们则能够训练咱们的数据，可是在开始以前，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，好比我想要训练ZF，若是使用的是train_faster_rcnn_alt_opt.py，则须要修改models/imagenet/ZF/faster_rcnn_alt_opt/下的全部pt文件里的内容，用以下的法则去替换：

//num为类别的个数
input-data->num_classes = num
class_score->num_output = num
bbox_pred->num_output   = num*4

我这里使用train_faster_rcnn_alt_opt.py进行的训练，这样的话则须要把添加的models/imagenet做为可选项

//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

这里咱们则须要使用刚训练出来的模型进行识别

#就像demo.py同样，可是使用训练的models，我建立了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')

一样，在识别前咱们要对识别方法里的Classes进行修改，修改为你本身训练的类别后

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