Faster-RCNN+ZF用自己的数据集训练模型(Python版本)

柔光的暖阳◎ 2022-09-29 12:58 18阅读 0赞

**原文链接：**http://blog.csdn.net/sinat\_30071459/article/details/51332084

说明：本博文假设你已经做好了自己的数据集，该数据集格式和VOC2007相同。下面是训练前的一些修改。

（做数据集的过程可以看[http://blog.csdn.net/sinat\_30071459/article/details/50723212][http_blog.csdn.net_sinat_30071459_article_details_50723212]）

Faster-RCNN源码下载地址：

Matlab版本：[https://github.com/ShaoqingRen/faster\_rcnn][https_github.com_ShaoqingRen_faster_rcnn]

[Python][]版本:[https://github.com/rbgirshick/py-faster-rcnn][https_github.com_rbgirshick_py-faster-rcnn]

本文用到的是**[Python][]**版本，在**Linux**下运行。

Matlab版本的训练过程：[http://blog.csdn.net/sinat\_30071459/article/details/50546891][http_blog.csdn.net_sinat_30071459_article_details_50546891]

**准备工作：**

1.配置caffe

这个不多说，网上教程很多。

2.其他的注意事项

这里说的挺详细了，认真看看吧。地址：[https://github.com/rbgirshick/py-faster-rcnn][https_github.com_rbgirshick_py-faster-rcnn]（主要内容如下）

下面大概翻译一下上面网址的内容吧。

（1）安装cython, `python-opencv`,`easydict`

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  pip install cython
2.  pip install easydict
3.  apt-get install python-opencv

![save_snippets.png][]

pip install cython
    pip install easydict
    apt-get install python-opencv

（2）下载py-faster-rcnn

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  \# Make sure to clone with –recursive
2.  git clone –recursive https://github.com/rbgirshick/py-faster-rcnn.git

![save_snippets_01.png][]

# Make sure to clone with --recursive
    git clone --recursive https://github.com/rbgirshick/py-faster-rcnn.git

如图：

![20160506201811276][]  
（3）进入py-faster-rcnn/lib

执行make

如图：

![20160506195644498][]

（4）进入py-faster-rcnn\\caffe-fast-rcnn

执行 cp Makefile.config.example Makefile.config

然后，配置Makefile.config文件，可参考我的配置：[Makefile.config文件][Makefile.config]

配置好Makefile.config文件后，执行：

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  make -j8 && make pycaffe

![save_snippets_01.png][]

make -j8 && make pycaffe

如图：

![20160506200505806][]

（5）下载VOC2007数据集

提供一个百度云地址：[http://pan.baidu.com/s/1mhMKKw4][http_pan.baidu.com_s_1mhMKKw4]

解压，然后，将该数据集放在py-faster-rcnn\\data下，用你的数据集替换VOC2007数据集。（替换Annotations，ImageSets和JPEGImages）

（用你的Annotations，ImagesSets和JPEGImages替换py-faster-rcnn\\data\\VOCdevkit2007\\VOC2007中对应文件夹）

（6）下载ImageNet数据集下预训练得到的模型参数（用来初始化）

提供一个百度云地址：[http://pan.baidu.com/s/1hsxx8OW][http_pan.baidu.com_s_1hsxx8OW]

解压，然后将该文件放在py-faster-rcnn\\data下

下面是训练前的一些修改。

### 1.py-faster-rcnn/models/pascal\_voc/ZF/faster\_rcnn\_alt\_opt/stage1\_fast\_rcnn\_train.pt修改 ###

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  layer \{
2.  name: ‘data’
3.  type: ‘Python’
4.  top: ‘data’
5.  top: ‘rois’
6.  top: ‘labels’
7.  top: ‘bbox\_targets’
8.  top: ‘bbox\_inside\_weights’
9.  top: ‘bbox\_outside\_weights’
10. python\_param \{
11. module: ‘roi\_data\_layer.layer’
12. layer: ‘RoIDataLayer’
13. param\_str: “‘num\_classes’: 16” \#按训练集类别改，该值为类别数+1
14. \}
15. \}

![save_snippets.png][]

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': 16" #按训练集类别改，该值为类别数+1
      }
    }

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  layer \{
2.  name: “cls\_score”
3.  type: “InnerProduct”
4.  bottom: “fc7”
5.  top: “cls\_score”
6.  param \{ lr\_mult: 1.0 \}
7.  param \{ lr\_mult: 2.0 \}
8.  inner\_product\_param \{
9.  num\_output: 16 \#按训练集类别改，该值为类别数+1
10. weight\_filler \{
11. type: “gaussian”
12. std: 0.01
13. \}
14. bias\_filler \{
15. type: “constant”
16. value: 0
17. \}
18. \}
19. \}

![save_snippets.png][]

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: 16 #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
        }
      }
    }

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  layer \{
2.  name: “bbox\_pred”
3.  type: “InnerProduct”
4.  bottom: “fc7”
5.  top: “bbox\_pred”
6.  param \{ lr\_mult: 1.0 \}
7.  param \{ lr\_mult: 2.0 \}
8.  inner\_product\_param \{
9.  num\_output: 64 \#按训练集类别改，该值为（类别数+1）\*4
10. weight\_filler \{
11. type: “gaussian”
12. std: 0.001
13. \}
14. bias\_filler \{
15. type: “constant”
16. value: 0
17. \}
18. \}
19. \}

![save_snippets.png][]

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: 64 #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.py-faster-rcnn/models/pascal\_voc/ZF/faster\_rcnn\_alt\_opt/stage1\_rpn\_train.pt修改 ###

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  layer \{
2.  name: ‘input-data’
3.  type: ‘Python’
4.  top: ‘data’
5.  top: ‘im\_info’
6.  top: ‘gt\_boxes’
7.  python\_param \{
8.  module: ‘roi\_data\_layer.layer’
9.  layer: ‘RoIDataLayer’
10. param\_str: “‘num\_classes’: 16” \#按训练集类别改，该值为类别数+1
11. \}
12. \}

![save_snippets.png][]

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': 16" #按训练集类别改，该值为类别数+1
      }
    }

### 3.py-faster-rcnn/models/pascal\_voc/ZF/faster\_rcnn\_alt\_opt/stage2\_fast\_rcnn\_train.pt修改 ###

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

![save_snippets.png][]

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

![save_snippets.png][]

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

![save_snippets.png][]

### 4.py-faster-rcnn/models/pascal\_voc/ZF/faster\_rcnn\_alt\_opt/stage2\_rpn\_train.pt修改 ###

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

![save_snippets.png][]

### 5.py-faster-rcnn/models/pascal\_voc/ZF/faster\_rcnn\_alt\_opt/faster\_rcnn\_test.pt修改 ###

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  layer \{
2.  name: “cls\_score”
3.  type: “InnerProduct”
4.  bottom: “fc7”
5.  top: “cls\_score”
6.  inner\_product\_param \{
7.  num\_output: 16 \#按训练集类别改，该值为类别数+1
8.  \}
9.  \}

![save_snippets.png][]

layer {
      name: "cls_score"
      type: "InnerProduct"
      bottom: "fc7"
      top: "cls_score"
      inner_product_param {
        num_output: 16 #按训练集类别改，该值为类别数+1
      }
    }

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  layer \{
2.  name: “bbox\_pred”
3.  type: “InnerProduct”
4.  bottom: “fc7”
5.  top: “bbox\_pred”
6.  inner\_product\_param \{
7.  num\_output: 64 \#按训练集类别改，该值为（类别数+1）\*4
8.  \}
9.  \}

![save_snippets.png][]

layer {
      name: "bbox_pred"
      type: "InnerProduct"
      bottom: "fc7"
      top: "bbox_pred"
      inner_product_param {
        num_output: 64 #按训练集类别改，该值为（类别数+1）*4
      }
    }

### 6.py-faster-rcnn/lib/datasets/pascal\_voc.py修改 ###

（1）

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  class pascal\_voc(imdb):
2.  def \_\_init\_\_(self, image\_set, year, devkit\_path=None):
3.  imdb.\_\_init\_\_(self, ‘voc\_’ + year + ‘\_’ + image\_set)
4.  self.\_year = year
5.  self.\_image\_set = image\_set
6.  self.\_devkit\_path = self.\_get\_default\_path() if devkit\_path is None \\
7.  else devkit\_path
8.  self.\_data\_path = os.path.join(self.\_devkit\_path, ‘VOC’ + self.\_year)
9.  self.\_classes = (‘\_\_background\_\_’, \# always index 0
10. ’你的标签1’,’你的标签2’,你的标签3’,’你的标签4’
11. )

![save_snippets.png][]

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
                             '你的标签1','你的标签2',你的标签3','你的标签4'
                          )

上面要改的地方是

**修改训练集文件夹：**

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  self.\_data\_path = os.path.join(self.\_devkit\_path, ‘VOC’+self.\_year)

![save_snippets_01.png][]

self._data_path = os.path.join(self._devkit_path, 'VOC'+self._year)

用你的数据集直接替换原来VOC2007内的Annotations，ImageSets和JPEGImages即可,以免出现各种错误。

**修改标签：**

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  self.\_classes = (‘\_\_background\_\_’, \# always index 0
2.  ’你的标签1’,’你的标签2’,’你的标签3’,’你的标签4’
3.  )

![save_snippets.png][]

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'
                          )

修改成你的数据集的标签就行。

（2）

**\[html\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  cls = self.\_class\_to\_ind\[obj.find(‘name’).text.lower().strip()\]

![save_snippets.png][]

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

这里把标签转成小写，如果你的标签含有大写字母，可能会出现KeyError的错误，所以建议标签用小写字母。

（去掉lower应该也行）

建议训练的标签还是用小写的字母，如果最终需要用大写字母或中文显示标签，可参考：

[http://blog.csdn.net/sinat\_30071459/article/details/51694037][http_blog.csdn.net_sinat_30071459_article_details_51694037]

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

该文件的append\_flipped\_images(self)函数修改为：

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  def append\_flipped\_images(self):
2.  num\_images = self.num\_images
3.  widths = \[PIL.Image.open(self.image\_path\_at(i)).size\[0\]
4.  for i in xrange(num\_images)\]
5.  for i in xrange(num\_images):
6.  boxes = self.roidb\[i\]\[‘boxes’\].copy()
7.  oldx1 = boxes\[:, 0\].copy()
8.  oldx2 = boxes\[:, 2\].copy()
9.  boxes\[:, 0\] = widths\[i\] - oldx2 - 1
10. print boxes\[:, 0\]
11. boxes\[:, 2\] = widths\[i\] - oldx1 - 1
12. print boxes\[:, 0\]
13. assert (boxes\[:, 2\] >= boxes\[:, 0\]).all()
14. entry = \{‘boxes’ : boxes,
15. ‘gt\_overlaps’ : self.roidb\[i\]\[‘gt\_overlaps’\],
16. ‘gt\_classes’ : self.roidb\[i\]\[‘gt\_classes’\],
17. ‘flipped’ : True\}
18. self.roidb.append(entry)
19. self.\_image\_index = self.\_image\_index \* 2

![save_snippets.png][]

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
                print boxes[:, 0]
                boxes[:, 2] = widths[i] - oldx1 - 1
                print boxes[:, 0]
                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

这里assert (boxes\[:, 2\] >= boxes\[:, 0\]).all()可能出现AssertionError，具体解决办法参考：

[http://blog.csdn.net/xzzppp/article/details/52036794][http_blog.csdn.net_xzzppp_article_details_52036794]

**！！！为防止与之前的模型搞混,训练前把output文件夹删除（或改个其他名），还要把py-faster-rcnn/data/cache中的文件和**

**py-faster-rcnn/data/VOCdevkit2007/annotations\_cache中的文件删除（如果有的话）。**

**至于学习率等之类的设置，可在py-faster-rcnn/models/pascal\_voc/ZF/faster\_rcnn\_alt\_opt中的solve文件设置，迭代次数可在py-faster-rcnn\\tools的train\_faster\_rcnn\_alt\_opt.py中修改：**

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  max\_iters = \[80000, 40000, 80000, 40000\]

![save_snippets.png][]

max_iters = [80000, 40000, 80000, 40000]

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。

### 8.开始训练 ###

进入py-faster-rcnn，执行：

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  ./experiments/scripts/faster\_rcnn\_alt\_opt.sh 0 ZF pascal\_voc

![save_snippets.png][]

./experiments/scripts/faster_rcnn_alt_opt.sh 0 ZF pascal_voc

这样，就开始训练了。

![20160506201954200][]

### 9.测试 ###

将训练得到的py-faster-rcnn\\output\\faster\_rcnn\_alt\_opt\\\*\*\*\_trainval中ZF的caffemodel拷贝至py-faster-rcnn\\data\\faster\_rcnn\_models（如果没有这个文件夹，就新建一个），然后，修改：

py-faster-rcnn\\tools\\demo.py,主要修改：

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  CLASSES = (‘\_\_background\_\_’,
2.  ’你的标签1’, ’你的标签2’, ’你的标签3’, ’你的标签4’)

![save_snippets_01.png][]

CLASSES = ('__background__',
               '你的标签1', '你的标签2', '你的标签3', '你的标签4')

改成你的数据集标签；

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  NETS = \{‘vgg16’: (‘VGG16’,
2.  ‘VGG16\_faster\_rcnn\_final.caffemodel’),
3.  ‘zf’: (‘ZF’,
4.  ‘ZF\_faster\_rcnn\_final.caffemodel’)\}

![save_snippets.png][]

NETS = {'vgg16': ('VGG16',
                      'VGG16_faster_rcnn_final.caffemodel'),
            'zf': ('ZF',
                      'ZF_faster_rcnn_final.caffemodel')}

上面ZF的caffemodel改成你的caffemodel。

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  im\_names = \[‘1559.jpg’,’1564.jpg’\]

![save_snippets_01.png][]

im_names = ['1559.jpg','1564.jpg']

改成你的测试图片。(测试图片放在py-faster-rcnn\\data\\demo中)

### 10.结果 ###

在py-faster-rcnn下，

执行：

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  ./tools/demo.py –net zf

![save_snippets_01.png][]

./tools/demo.py --net zf

或者将默认的模型改为zf：

**\[html\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  parser.add\_argument(‘–net’, dest=‘demo\_net’, help=‘Network to use \[vgg16\]’,
2.  choices=NETS.keys(), default=‘vgg16’)

![save_snippets_01.png][]

parser.add_argument('--net', dest='demo_net', help='Network to use [vgg16]',
                            choices=NETS.keys(), default='vgg16')

修改：

**\[html\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  default=‘zf’

![save_snippets.png][]

default='zf'

执行：

**\[plain\]** view plain copy

print ? [![在CODE上查看代码片][CODE]][CODE_CODE] [![派生到我的代码片][ico_fork.svg]][ico_fork.svg 1]

1.  ./tools/demo.py

![save_snippets.png][]

./tools/demo.py

![20160508201052360][]

[http_blog.csdn.net_sinat_30071459_article_details_50723212]: http://blog.csdn.net/sinat_30071459/article/details/50723212
[https_github.com_ShaoqingRen_faster_rcnn]: https://github.com/ShaoqingRen/faster_rcnn
[Python]: http://lib.csdn.net/base/11
[https_github.com_rbgirshick_py-faster-rcnn]: https://github.com/rbgirshick/py-faster-rcnn
[http_blog.csdn.net_sinat_30071459_article_details_50546891]: http://blog.csdn.net/sinat_30071459/article/details/50546891
[CODE]: https://code.csdn.net/assets/CODE_ico.png
[CODE_CODE]: https://code.csdn.net/snippets/1674959
[ico_fork.svg]: https://code.csdn.net/assets/ico_fork.svg
[ico_fork.svg 1]: https://code.csdn.net/snippets/1674959/fork
[save_snippets.png]: /images/20220708/0de96e6039dd4046aa89bbfdffc56613.png
[save_snippets_01.png]: /images/20220708/f16f4e93a19c4e3a8b96f4914e7de1ef.png
[20160506201811276]: /images/20220708/52d710c40fe042d2825a40cf4923abee.png
[20160506195644498]: /images/20220708/5e0a79008f2b41a4bb6295554f5b7f55.png
[Makefile.config]: http://download.csdn.net/detail/sinat_30071459/9512429
[20160506200505806]: /images/20220708/2c1563d0d3ee4b41bcfd6461e6341696.png
[http_pan.baidu.com_s_1mhMKKw4]: http://pan.baidu.com/s/1mhMKKw4
[http_pan.baidu.com_s_1hsxx8OW]: http://pan.baidu.com/s/1hsxx8OW
[http_blog.csdn.net_sinat_30071459_article_details_51694037]: http://blog.csdn.net/sinat_30071459/article/details/51694037
[http_blog.csdn.net_xzzppp_article_details_52036794]: http://blog.csdn.net/xzzppp/article/details/52036794
[20160506201954200]: /images/20220708/74a7e587eb964b88bb4f284f14de98ff.png
[20160508201052360]: /images/20220708/88dc479fd27c4b7d94a82b7da770f450.png