用faster-rcnn训练自己的数据集(VOC2007格式,python版)

用faster-rcnn训练自己的数据集(VOC2007格式,python版)

一. 配置caffe环境

ubunt16.04下caffe环境安装

二. 下载,编译及测试py-faster-rcnn源码

(一)下载源码

github链接

或者执行 git clone –recursive https://github.com/rbgirshick/py-faster-rcnn.git

注意加上–recursive关键字

(二)编译源码

编译过程中可能会出现缺失一些python模块,按提示安装

(1)编译Cython模块

cd $FRCN_ROOT/lib make

(2)修改Markfile配置

参考ubunt16.04下caffe环境安装
中修改Makefile.config

(3)编译python接口

cd $FRCN_ROOT/caffe-fast-rcnnmake -j8  多核编译,时间较长make pycaffe

(4)下载训练好的VGG16和ZF模型

cd $FRCN_ROOT./data/scripts/fetch_faster_rcnn_models.sh

时间太长的话可以考虑找网上别人分享的资源

(三)测试源码

cd $FRCN_ROOT./tool/demo.py

三. 使用faster-rcnn训练自己的数据集

(一)下载预训练参数及模型

cd $FRCN_ROOT./data/scripts/fetch_imagenet_models.sh./data/scripts/fetch_selective_search_data.sh

(二)制作数据集

制作数据集(VOC2007格式)

将制作好的VOC2007文件夹放置在data/VOCdevkit2007文件夹下,没有则新建VOCdevkit2007文件夹

(三)修改配置文件

(1)修改py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/stage1_fast_rcnn_train.pt和stage2_fast_rcnn_train.pt 两个文件

备注:3处修改及其附近的代码

name: "ZF"layer {  name: 'data'  type: 'Python' top: 'data'top: 'rois'top: 'labels' top: 'bbox_targets' top: 'bbox_inside_weights'top: 'bbox_outside_weights'python_param {  module: 'roi_data_layer.layer'  layer: 'RoIDataLayer'  param_str: "'num_classes': 2" #按训练集类别改，该值为类别数+1}}layer { name: "cls_score" type: "InnerProduct" bottom: "fc7" top: "cls_score" param { lr_mult: 1.0 } param { lr_mult: 2.0 }inner_product_param {    num_output: 2 #按训练集类别改，该值为类别数+1 weight_filler {   type: "gaussian"   std: 0.01 } bias_filler {    type: "constant"   value: 0  } }}layer {  name: "bbox_pred" type: "InnerProduct" bottom: "fc7"top: "bbox_pred" param { lr_mult: 1.0 } param { lr_mult: 2.0 } inner_product_param {   num_output: 8 #按训练集类别改，该值为（类别数+1）*4   weight_filler {     type: "gaussian"      std: 0.001    }    bias_filler {      type: "constant"      value: 0    }  }}

(2)修改py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/stage1_rpn_train.pt和stage2_rpn_train.pt 两个文件

备注:1处修改及其附近的代码

layer {  name: 'input-data'  type: 'Python'  top: 'data'  top: 'im_info'  top: 'gt_boxes'  python_param {    module: 'roi_data_layer.layer'    layer: 'RoIDataLayer'    param_str: "'num_classes': 2" #按训练集类别改，该值为类别数+1  }}

(3)修改py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/faster_rcnn_test.pt文件

备注:2处修改及其附近的代码

layer {  name: "cls_score"  type: "InnerProduct"  bottom: "fc7"  top: "cls_score"  param { lr_mult: 1.0 }  param { lr_mult: 2.0 }  inner_product_param {    num_output: 2 #按训练集类别改，该值为类别数+1    weight_filler {      type: "gaussian"      std: 0.01    }    bias_filler {      type: "constant"      value: 0    }  }}layer {  name: "bbox_pred"  type: "InnerProduct"  bottom: "fc7"  top: "bbox_pred"  param { lr_mult: 1.0 }  param { lr_mult: 2.0 }  inner_product_param {    num_output: 8 #按训练集类别改，该值为（类别数+1）*4    weight_filler {      type: "gaussian"      std: 0.001    }    bias_filler {      type: "constant"      value: 0    }  }}

(4)修改py-faster-rcnn/lib/datasets/pascal_voc.py

self._classes = ('__background__', # always index 0                         '你的标签1','你的标签2',你的标签3','你的标签4')注:如果只是在原始检测的20种类别:'aeroplane', 'bicycle', 'bird', 'boat','bottle', 'bus', 'car', 'cat', 'chair','cow', 'diningtable', 'dog', 'horse','motorbike', 'person', 'pottedplant','sheep', 'sofa', 'train', 'tvmonitor'中检测单一类别,可参考修改下面的代码:def _load_image_set_index(self):        """        Load the indexes listed in this dataset's image set file.        """        # Example path to image set file:        # self._devkit_path + /VOCdevkit2007/VOC2007/ImageSets/Main/val.txt        image_set_file = os.path.join(self._data_path, 'ImageSets', 'Main',                                      self._image_set + '.txt')        assert os.path.exists(image_set_file), \                'Path does not exist: {}'.format(image_set_file)        with open(image_set_file) as f:            image_index = [x.strip() for x in f.readlines()]注:如果需要在原始的20类别只检测车辆的话才需要修改这部分代码.        # only load index with cars obj        new_image_index = []        for index in image_index:            filename = os.path.join(self._data_path, 'Annotations', index + '.xml')            tree = ET.parse(filename)            objs = tree.findall('object')            num_objs = 0            for ix, obj in enumerate(objs):                curr_name = obj.find('name').text.lower().strip()                if curr_name == 'car':                    num_objs += 1                    break            if num_objs > 0:                new_image_index.append(index)        return new_image_indexdef _load_pascal_annotation(self, index):        """        Load image and bounding boxes info from XML file in the PASCAL VOC        format.        """        filename = os.path.join(self._data_path, 'Annotations', index + '.xml')        tree = ET.parse(filename)        objs = tree.findall('object')        if not self.config['use_diff']:            # Exclude the samples labeled as difficult            non_diff_objs = [                obj for obj in objs if int(obj.find('difficult').text) == 0]            # if len(non_diff_objs) != len(objs):            #     print 'Removed {} difficult objects'.format(            #         len(objs) - len(non_diff_objs))            objs = non_diff_objs注:如果需要在原始的20类别只检测车辆的话才需要修改这部分代码.        # change num objs , only read car        # num_objs = len(objs)        num_objs = 0        for ix, obj in enumerate(objs):            curr_name = obj.find('name').text.lower().strip()            if curr_name == 'car':                num_objs += 1        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" area for pascal is just the box area        seg_areas = np.zeros((num_objs), dtype=np.float32)#注:如果需要在原始的20类别只检测车辆的话才需要修改这部分代码# Load object bounding boxes into a data    frame.        tmp_ix = 0        for ix, obj in enumerate(objs):            bbox = obj.find('bndbox')            # Make pixel indexes 0-based            x1 = float(bbox.find('xmin').text) - 1            y1 = float(bbox.find('ymin').text) - 1            x2 = float(bbox.find('xmax').text) - 1            y2 = float(bbox.find('ymax').text) - 1            curr_name = obj.find('name').text.lower().strip()            if curr_name != 'car':                continue            cls = self._class_to_ind[curr_name]            boxes[tmp_ix, :] = [x1, y1, x2, y2]            gt_classes[tmp_ix] = cls            overlaps[tmp_ix, cls] = 1.0            seg_areas[tmp_ix] = (x2 - x1 + 1) * (y2 - y1 + 1)            tmp_ix += 1        overlaps = scipy.sparse.csr_matrix(overlaps)        return {'boxes' : boxes,                'gt_classes': gt_classes,                'gt_overlaps' : overlaps,                'flipped' : False,                'seg_areas' : seg_areas}

(4)py-faster-rcnn/lib/datasets/imdb.py修改

def append_flipped_images(self):        num_images = self.num_images        widths = [PIL.Image.open(self.image_path_at(i)).size[0]                  for i in xrange(num_images)]        for i in xrange(num_images):            boxes = self.roidb[i]['boxes'].copy()            oldx1 = boxes[:, 0].copy()            oldx2 = boxes[:, 2].copy()            boxes[:, 0] = widths[i] - oldx2 - 1            boxes[:, 2] = widths[i] - oldx1 - 1            for b in range(len(boxes)):                if boxes[b][2] < boxes[b][0]:                   boxes[b][0] = 0            assert (boxes[:, 2] >= boxes[:, 0]).all()

(5)py-faster-rcnn/tools/train_faster_rcnn_alt_opt.py修改迭代次数（建议修改）

max_iters=[8000,4000,8000,4000]建议:第一次训练使用较低的迭代次数,先确保能正常训练,如max_iters=[8,4,8,4]

训练分别为4个阶段（rpn第1阶段，fast rcnn第1阶段，rpn第2阶段，fast rcnn第2阶段）的迭代次数。可改成你希望的迭代次数。
如果改了这些数值，最好把py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt里对应的solver文件（有4个）也修改，stepsize小于上面修改的数值，stepsize的意义是经过stepsize次的迭代后降低一次学习率（非必要修改）。

(6)删除缓存文件(每次修改配置文件后训练都要做)

删除py-faster-rcnn文件夹下所有的.pyc文件及data文件夹下的cache文件夹,data/VOCdekit2007下的annotations_cache文件夹(最近一次成功训练的annotation和当前annotation一样的话这部分可以不删,否则可以正常训练,但是最后评价模型会出错)

(四)开始训练

cd $FRCN_ROOT./experiments/scripts/faster_rcnn_alt_opt.sh 0 ZF pascal_voc

成功训练后在py-faster-rcnn/output/faster_rcnn_alt_opt/voc_2007_trainval文件夹下
会有以final.caffemodel结尾的模型文件,一般为ZF_faster_rcnn_final.caffemodel

成功训练后会有一次模型性能的评估测试,成功的话会有MAP指标和平均MAP指标的输出,类似下文,
训练日志文件保存在experiments/logs文件夹下.

Evaluating detectionsWriting car VOC results fileVOC07 metric? YesAP for car = 0.0090Mean AP = 0.0090~~~~~~~~Results:0.0090.009~~~~~~~~--------------------------------------------------------------Results computed with the **unofficial** Python eval code.Results should be very close to the official MATLAB eval code.Recompute with `./tools/reval.py --matlab ...` for your paper.-- Thanks, The Management--------------------------------------------------------------real    1m43.822suser    1m25.764ssys 0m15.516s

(五)测试训练结果

(1)修改py-faster-rcnn\tools\demo.py

CLASSES = ('__background__',         '你的标签1','你的标签2',你的标签3','你的标签4')NETS = {'vgg16': ('VGG16',                  'VGG16_faster_rcnn_final.caffemodel'),        'zf': ('ZF',                  'ZF_faster_rcnn_final.caffemodel')}im_names = os.listdir(os.path.join(cfg.DATA_DIR, 'demo'))  

(2)放置模型及测试图片

将训练得到的py-faster-rcnn\output\faster_rcnn_alt_opt\***_trainval中ZF的final.caffemodel拷贝至py-faster-rcnn\data\faster_rcnn_models测试图片放在py-faster-rcnn\data\demo(与上面demo.py设置路径有关,可修改)

(3)进行测试

cd $FRCN_ROOT./tool/demo.py

四. 曾出现过的bug及当时的解决方法

(1) 训练时出现KeyError:’max_overlaps’ ,解决方法:删除data文件夹下的cache文件夹

(2) 训练结束后测试时出现类似

File "/home/hyzhan/py-faster-rcnn/tools/../lib/datasets/voc_eval.py", line 126, in voc_eval    R = [obj for obj in recs[imagename] if obj['name'] == classname]KeyError: '000002'

解决方法: 删除data/VOCdekit2007下的annotations_cache文件夹

(3) caffe-fast-rcnn编译时出现找不到nvcc命令的情况,解决方法:

export PATH=/usr/local/cuda-8.0/bin:$PATHexport LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64:$LD_LIBRARY_PATH

将cuda安装路径添加到环境变量中

(4) caffe-fast-rcnn编译时出现类似找不到opencv命令的情况,解决方法,添加环境变量:

export LD_LIBRARY_PATH=/home/hyzhan/software/opencv3/lib:$LD_LIBRARY_PATH

(5) 训练的时候执行”./experiments/scripts/faster_rcnn_alt_opt.sh 0 ZF pascal_voc”语句进行训练会出现找不到faster_rcnn_alt_opt.sh文件的情况,解决方法:重新手打命令

(6) 测试之前需要修改tool文件夹下的demo或者mydemo里面的class类别,不然会显示上次训练的类别