折腾faster-rcnn（三）--训练篇

标签（空格分隔）： Linux 深度学习

---

前面，我们已经成功跑起来faster-rcnn，并且自己动手训练了PascalVOC2007图像集，但是笔者接下来继续沿着官方文件跑PascalVOC2012，训练到第二阶段的某个点，程序就开始报错退出。查看logs发现，找不到图像的Annotation文件。遂放弃，转而生成自己的图像集。

图像收集

这里笔者需要先完成毕业设计，所以图像集设计的种类不多，也就三类。毕竟PascalVOC一开始也只有四类呢！看了这么多经验帖，总结需要注意一下几个问题：文件名不要大写，后缀只能是.jpg或.JPEG，图像不要长得太过“瘦长”。这里为了重新命名收集来的文件，笔者自己写了一段Python脚本，变量命名注释极不规范，而且并不能识别某些后缀是jpg但并不是jpg的“假图”，大家看看就好，勿喷。

# -*-coding:utf-8-*-
import os
from PIL import Image
import argparse

'''
This scrpit was designed for rename these jpeg images downloaded from different websites,we can use it to rename
as such format: 0000001.jpg
'''


def readimage(FileDir):
    imageset = []
    for image in os.listdir(FileDir):
        if os.path.isfile(os.path.join(FileDir, image)):
            imageset.append(image)
    return imageset


def RenameFiles(FileDir, count=1):
    extlist = ['.jpg', '.jpeg']
    imageset = readimage(FileDir)
    for image in imageset:
        extension = os.path.splitext(image)[1].lower()
        if FileDir == 'kiss':
            clsnum = 0
        if FileDir == 'hug':
            clsnum = 1
        if FileDir == 'hands':
            clsnum = 2
        if extension in extlist:
            oldname = os.path.join(FileDir, image)
            newimage = '0000' + str(count)
            count = count + 1
            newimagename = str(clsnum) + newimage[-5:] + '.jpg'

            print newimagename
            newname = os.path.join(FileDir, newimagename)
            os.rename(oldname, newname)
        else:
            print "Exist non-jpeg files: " + image
            break


def parse_args():
    """Parse input arguments."""

    parser = argparse.ArgumentParser(description='Rename demo')
    parser.add_argument('-set', dest='dataset', help='choose dataset')
    parser.add_argument('-count', dest='countnum', help='start number is set to 0',
                        default=0, type=int)
    args = parser.parse_args()

    return args


if __name__ == "__main__":
    args = parse_args()
    if args.countnum == 0:
        print "please change countnum"
    else:
        RenameFiles(args.dataset, args.countnum)

图像标定

这里采用的是labelImg的标定工具，发动同学和学姐学弟一起帮我标定2333~~~

图像分类

这里我们需要产生和PascalVOC一样格式的txt文件，这里笔者事先研究过了，产生是有点规律的，具体见我的上一篇博客：折腾faster-rcnn（二）--数据篇
然后是几个集合的比例问题：
一般情况下trainval占总数据集的50%，test占总数据集的50%；train占trainval的50%，val占trainval的50%。上面所占百分比可根据自己的数据集修改，如果数据集比较少，test和val可少一些。这个在生成以上四个文件的脚本里面可以调比例，暂时就用默认比例试试，跑跑。生成这几个文件的脚本如下：

# -*-coding:utf-8-*-
import os
import random

import FileOperate as fo

'''
Generate txt files in directory VOC2012/ImageSet
'''

def GenerateSet(abspath, imgsetdir):
    imgset = []
    imgfiles = fo.readfile(os.path.join(abspath, imgsetdir))
    for imgfile in imgfiles:
        imgset.append(os.path.splitext(imgfile)[0])
    rate = [0.5, 0.5, 0.5, 0.5]
    clsset = ['kiss', 'hug', 'hands']
    txtset = ['trainval.txt', 'train.txt', 'val.txt', 'test.txt']
    trainvalset = []
    testset = []
    trainvalposset = set(random.sample(xrange(len(imgset)), int(len(imgset) * rate[0])))
    testposset = set(xrange(len(imgset))) - trainvalposset
    for pos in trainvalposset:
        trainvalset.append(imgset[pos])
    for pos in testposset:
        testset.append(imgset[pos])
    trainset = []
    valset = []
    trainposset = set(random.sample(xrange(len(trainvalset)), int(len(trainvalset) * rate[1])))
    valposset = set(xrange(len(trainvalset))) - trainposset
    for pos in trainposset:
        trainset.append(trainvalset[pos])
    for pos in valposset:
        valset.append(trainvalset[pos])
    fo.createtext(txtset[0], trainvalset, 2)
    fo.createtext(txtset[1], trainset, 2)
    fo.createtext(txtset[2], valset, 2)
    fo.createtext(txtset[3], testset, 2)
    for trainvalimg in trainvalset:
        imgname = []
        imgname.append(trainvalimg)
        clsname = trainvalimg.split("_")[0]
        if clsset[0] == clsname:
            fo.createtext(clsset[0] + "_" + txtset[0], imgname, 1)
            fo.createtext(clsset[1] + "_" + txtset[0], imgname, -1)
            fo.createtext(clsset[2] + "_" + txtset[0], imgname, -1)
        elif clsset[1] == clsname:
            fo.createtext(clsset[0] + "_" + txtset[0], imgname, -1)
            fo.createtext(clsset[1] + "_" + txtset[0], imgname, 1)
            fo.createtext(clsset[2] + "_" + txtset[0], imgname, -1)
        else:
            fo.createtext(clsset[0] + "_" + txtset[0], imgname, -1)
            fo.createtext(clsset[1] + "_" + txtset[0], imgname, -1)
            fo.createtext(clsset[2] + "_" + txtset[0], imgname, 1)
    for trainimg in trainset:
        imgname = []
        imgname.append(trainimg)
        clsname = trainimg.split("_")[0]
        if clsset[0] == clsname:
            fo.createtext(clsset[0] + "_" + txtset[1], imgname, 1)
            fo.createtext(clsset[1] + "_" + txtset[1], imgname, -1)
            fo.createtext(clsset[2] + "_" + txtset[1], imgname, -1)
        elif clsset[1] == clsname:
            fo.createtext(clsset[0] + "_" + txtset[1], imgname, -1)
            fo.createtext(clsset[1] + "_" + txtset[1], imgname, 1)
            fo.createtext(clsset[2] + "_" + txtset[1], imgname, -1)
        else:
            fo.createtext(clsset[0] + "_" + txtset[1], imgname, -1)
            fo.createtext(clsset[1] + "_" + txtset[1], imgname, -1)
            fo.createtext(clsset[2] + "_" + txtset[1], imgname, 1)
    for valimg in valset:
        imgname = []
        imgname.append(valimg)
        clsname = valimg.split("_")[0]
        if clsset[0] == clsname:
            fo.createtext(clsset[0] + "_" + txtset[2], imgname, 1)
            fo.createtext(clsset[1] + "_" + txtset[2], imgname, -1)
            fo.createtext(clsset[2] + "_" + txtset[2], imgname, -1)
        elif clsset[1] == clsname:
            fo.createtext(clsset[0] + "_" + txtset[2], imgname, -1)
            fo.createtext(clsset[1] + "_" + txtset[2], imgname, 1)
            fo.createtext(clsset[2] + "_" + txtset[2], imgname, -1)
        else:
            fo.createtext(clsset[0] + "_" + txtset[2], imgname, -1)
            fo.createtext(clsset[1] + "_" + txtset[2], imgname, -1)
            fo.createtext(clsset[2] + "_" + txtset[2], imgname, 1)
    for testimg in testset:
        imgname = []
        imgname.append(testimg)
        fo.createtext(clsset[0] + "_" + txtset[3], imgname, 0)
        fo.createtext(clsset[1] + "_" + txtset[3], imgname, 0)
        fo.createtext(clsset[2] + "_" + txtset[3], imgname, 0)


if __name__ == "__main__":
    abspath = os.path.abspath('.')
    imgsetdir = 'JPEGImages'
    GenerateSet(abspath, imgsetdir)

建立图像库

将以上几步搞好后，我们需要仿照PSACALVOC2007的结构建立自己的图像库，目录结构如下图所示

VOCdevkit2007
    |-results
        |-VOC2007
            |-Main #空目录，用来存放test集结果
    |-VOC2007
        |-Annoations #标定文件xml
        |-ImageSets #txt文件存放至此
        |-JPEGImages #jpg图像存放至此

修改模型

这一步肯定是需要的，因为PascalVOC有20类，所以我们需要根据自己的需求稍微修改神经网络参数。所有参数都在/models下，分为coco和pascal_voc格式，每种格式又有两到三种网络模型，每种模型还有两种训练方式Alternative training(alt-opt)和Approximate joint training(end-to-end)。总之，由于各种原因，本文以Alternative training+ZF model为例。笔者的需求是四类，还有一类在这里叫__background__类，一共五类。
要修改的文件：
1.py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/stage1_rpn_train.pt

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': 5" #按训练集类别改，该值为类别数+1  
  }  
}  

2.py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/stage1_fast_rcnn_train.pt

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': 5" #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: 5 #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: 20 #按训练集类别改，该值为（类别数+1）*4，四个顶点坐标  
    weight_filler {  
      type: "gaussian"  
      std: 0.001  
    }  
    bias_filler {  
      type: "constant"  
      value: 0  
    }  
  }  
}  

3.py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/stage2_rpn_train.pt

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': 5" #按训练集类别改，该值为类别数+1  
  }  
}  

4.py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/stage2_fast_rcnn_train.pt

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': 5" #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: 5 #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: 20 #按训练集类别改，该值为（类别数+1）*4,四个顶点坐标  
    weight_filler {  
      type: "gaussian"  
      std: 0.001  
    }  
    bias_filler {  
      type: "constant"  
      value: 0  
    }  
  }  
}  

5.py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/faster_rcnn_test.pt

layer {  
  name: "cls_score"  
  type: "InnerProduct"  
  bottom: "fc7"  
  top: "cls_score"  
  inner_product_param {  
    num_output: 5 #按训练集类别改，该值为类别数+1  
  }  
}  

layer {  
  name: "bbox_pred"  
  type: "InnerProduct"  
  bottom: "fc7"  
  top: "bbox_pred"  
  inner_product_param {  
    num_output: 20 #按训练集类别改，该值为（类别数+1）*4  
  }  
} 

---

注意，注意，这里文件夹进行了切换，并不属于模型的参数而是属于支持库的改动：
6.py-faster-rcnn/lib/datasets/pascal_voc.py

class pascal_voc(imdb):
    def __init__(self, image_set, year, devkit_path=None):
        imdb.__init__(self, 'voc_' + year + '_' + image_set)
        self._year = year
        self._image_set = image_set
        self._devkit_path = self._get_default_path() if devkit_path is None \
            else devkit_path
        self._data_path = os.path.join(self._devkit_path, 'VOC' + self._year)
        self._classes = ('__background__',  # always index 0
                         'kiss', 'hug', 'hands','person')# 有几个类别此处就写几个，注意第一个（索引0位置）永远是__background__类

7.py-faster-rcnn/lib/datasets/imdb.py

   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
            assert (boxes[:, 2] >= boxes[:, 0]).all()
            entry = {'boxes': boxes,
                     'gt_overlaps': self.roidb[i]['gt_overlaps'],
                     'gt_classes': self.roidb[i]['gt_classes'],
                     'flipped': True}
            self.roidb.append(entry)
        self._image_index = self._image_index * 2

8.根目录执行./experiments/scripts/faster_rcnn_alt_opt.sh 0 ZF pascal_voc,训练和测试一气呵成，并且还有日志记录。如果只需要训练，那么执行（无日志记录）：

./tools/train_faster_rcnn_alt_opt.py --gpu 0 --net_name ZF --weights data/imagenet_models/ZF.v2.caffemodel --imdb voc_2007_trainval --cfg experiments/cfgs/faster_rcnn_alt_opt.yml

Q&A：
1.如果标签含有大写字母，那么修改py-faster-rcnn/lib/datasets/pascal_voc.py第209行：

cls = self._class_to_ind[obj.find('name').text.lower().strip()]  

2.出现错误：

TypeError: 'NoneType' object has no attribute '__getitem__'

这是因为某些图像根本不是jpeg图像，只是后缀被改动了，这一下被学姐坑惨了。建议大家训练前统统用图转工具将图像转换一下格式，以绝后患。