课程频道导航
课程分类导航
当前位置：首页 > 操作系统 > Windows

Windows caffe 跑mnist实例

　    一. 装完caffe当然要来跑跑自带的demo，在examples文件夹下。

先来试试用于手写数字识别的mnist，在 examples/mnist/ 下有需要的代码文件，但是没有图像库。

mnist库有50000个训练样本，10000个测试样本，都是手写数字图像。

　　caffe支持的数据格式为：lmdb　　leveldb

　　imdb比leveldb大，但是速度更快，且允许多种训练模型同时读取同一数据集。

　　默认情况，examples里支持的是imdb文件，不过你可以修改为leveldb，后面详解。

　　mnist数据集建议网上搜索下载，网盘有很多，注意将文件夹放到examplesmnist目录下，且最好命名为图中格式，

否则可能无法读取文件需手动配置。

　　笔者之前下的数据集命名的下划线是连接线就会报错无法读取文件，所以注意文件夹名字！

 windows下最好选择leveldb文件，linux则随意了。下好了leveldb文件就不用再使用convert_imageset函数了，省去了转换图片格式和计算均值的步骤。

　　二. 训练mnist模型

　　mnist的网络训练模型文件为： lenet_train_test.prototxt

name: <span>"</span><span>lenet</span><span>"</span><span> layer {
  name: </span><span>"</span><span>mnist</span><span>"</span><span>   type: </span><span>"</span><span>data</span><span>"</span><span>   top: </span><span>"</span><span>data</span><span>"</span><span>   top: </span><span>"</span><span>label</span><span>"</span><span>   include {
    phase: train
  }
  transform_param {
    scale: </span><span>0.00390625</span><span>   }
  data_param {
    source: </span><span>"</span><span>examples/mnist/mnist_train_leveldb</span><span>"</span><span>     batch_size: </span><span>64</span><span>     backend: leveldb
  }
}
layer {
  name: </span><span>"</span><span>mnist</span><span>"</span><span>   type: </span><span>"</span><span>data</span><span>"</span><span>   top: </span><span>"</span><span>data</span><span>"</span><span>   top: </span><span>"</span><span>label</span><span>"</span><span>   include {
    phase: test
  }
  transform_param {
    scale: </span><span>0.00390625</span><span>   }
  data_param {
    source: </span><span>"</span><span>examples/mnist/mnist_test_leveldb</span><span>"</span><span>     batch_size: </span><span>100</span><span>     backend: leveldb
  }
}
layer {
  name: </span><span>"</span><span>conv1</span><span>"</span><span>   type: </span><span>"</span><span>convolution</span><span>"</span><span>   bottom: </span><span>"</span><span>data</span><span>"</span><span>   top: </span><span>"</span><span>conv1</span><span>"</span><span>   param {
    lr_mult: </span><span>1</span><span>   }
  param {
    lr_mult: </span><span>2</span><span>   }
  convolution_param {
    num_output: </span><span>20</span><span>     kernel_size: </span><span>5</span><span>     stride: </span><span>1</span><span>     weight_filler {
      type: </span><span>"</span><span>xavier</span><span>"</span><span>     }
    bias_filler {
      type: </span><span>"</span><span>constant</span><span>"</span><span>     }
  }
}
layer {
  name: </span><span>"</span><span>pool1</span><span>"</span><span>   type: </span><span>"</span><span>pooling</span><span>"</span><span>   bottom: </span><span>"</span><span>conv1</span><span>"</span><span>   top: </span><span>"</span><span>pool1</span><span>"</span><span>   pooling_param {
    pool: max
    kernel_size: </span><span>2</span><span>     stride: </span><span>2</span><span>   }
}
layer {
  name: </span><span>"</span><span>conv2</span><span>"</span><span>   type: </span><span>"</span><span>convolution</span><span>"</span><span>   bottom: </span><span>"</span><span>pool1</span><span>"</span><span>   top: </span><span>"</span><span>conv2</span><span>"</span><span>   param {
    lr_mult: </span><span>1</span><span>   }
  param {
    lr_mult: </span><span>2</span><span>   }
  convolution_param {
    num_output: </span><span>50</span><span>     kernel_size: </span><span>5</span><span>     stride: </span><span>1</span><span>     weight_filler {
      type: </span><span>"</span><span>xavier</span><span>"</span><span>     }
    bias_filler {
      type: </span><span>"</span><span>constant</span><span>"</span><span>     }
  }
}
layer {
  name: </span><span>"</span><span>pool2</span><span>"</span><span>   type: </span><span>"</span><span>pooling</span><span>"</span><span>   bottom: </span><span>"</span><span>conv2</span><span>"</span><span>   top: </span><span>"</span><span>pool2</span><span>"</span><span>   pooling_param {
    pool: max
    kernel_size: </span><span>2</span><span>     stride: </span><span>2</span><span>   }
}
layer {
  name: </span><span>"</span><span>ip1</span><span>"</span><span>   type: </span><span>"</span><span>innerproduct</span><span>"</span><span>   bottom: </span><span>"</span><span>pool2</span><span>"</span><span>   top: </span><span>"</span><span>ip1</span><span>"</span><span>   param {
    lr_mult: </span><span>1</span><span>   }
  param {
    lr_mult: </span><span>2</span><span>   }
  inner_product_param {
    num_output: </span><span>500</span><span>     weight_filler {
      type: </span><span>"</span><span>xavier</span><span>"</span><span>     }
    bias_filler {
      type: </span><span>"</span><span>constant</span><span>"</span><span>     }
  }
}
layer {
  name: </span><span>"</span><span>relu1</span><span>"</span><span>   type: </span><span>"</span><span>relu</span><span>"</span><span>   bottom: </span><span>"</span><span>ip1</span><span>"</span><span>   top: </span><span>"</span><span>ip1</span><span>"</span><span> }
layer {
  name: </span><span>"</span><span>ip2</span><span>"</span><span>   type: </span><span>"</span><span>innerproduct</span><span>"</span><span>   bottom: </span><span>"</span><span>ip1</span><span>"</span><span>   top: </span><span>"</span><span>ip2</span><span>"</span><span>   param {
    lr_mult: </span><span>1</span><span>   }
  param {
    lr_mult: </span><span>2</span><span>   }
  inner_product_param {
    num_output: </span><span>10</span><span>     weight_filler {
      type: </span><span>"</span><span>xavier</span><span>"</span><span>     }
    bias_filler {
      type: </span><span>"</span><span>constant</span><span>"</span><span>     }
  }
}
layer {
  name: </span><span>"</span><span>accuracy</span><span>"</span><span>   type: </span><span>"</span><span>accuracy</span><span>"</span><span>   bottom: </span><span>"</span><span>ip2</span><span>"</span><span>   bottom: </span><span>"</span><span>label</span><span>"</span><span>   top: </span><span>"</span><span>accuracy</span><span>"</span><span>   include {
    phase: test
  }
}
layer {
  name: </span><span>"</span><span>loss</span><span>"</span><span>   type: </span><span>"</span><span>softmaxwithloss</span><span>"</span><span>   bottom: </span><span>"</span><span>ip2</span><span>"</span><span>   bottom: </span><span>"</span><span>label</span><span>"</span><span>   top: </span><span>"</span><span>loss</span><span>"</span><span> }</span>

一般修改两个data层的 “source”文件路径就行，上面的例子中，我已经改了，改为mnist的训练集和测试集文件夹路径。再就是注意“backend: leveldb”,默认的backend应该是imdb要修改！

　　网络模型 lenet_train_test.prototxt修改后再修改 lenet_solver.prototxt

该文件主要是一些学习参数和策略：

            <span> 1</span> # the train/<span>test net protocol buffer definition</span><span> 2</span> net: <span>"</span><span>examples/mnist/lenet_train_test.prototxt</span><span>"</span><span> 3</span> # test_iter specifies how many forward passes the test should carry <span>out</span><span>.</span><span> 4</span> # in the <span>case</span> of mnist, we have test batch size <span>100</span> and <span>100</span><span> test iterations,</span><span> 5</span> # covering the full <span>10</span>,<span>000</span><span> testing images.</span><span> 6</span> test_iter: <span>100</span><span> 7</span> # carry <span>out</span> testing every <span>500</span><span> training iterations.</span><span> 8</span> test_interval: <span>500</span><span> 9</span> # the <span>base</span><span> learning rate, momentum and the weight decay of the network.</span><span>10</span> base_lr: <span>0.01</span><span>11</span> momentum: <span>0.9</span><span>12</span> weight_decay: <span>0.0005</span><span>13</span><span># the learning rate policy</span><span>14</span> lr_policy: <span>"</span><span>inv</span><span>"</span><span>15</span> gamma: <span>0.0001</span><span>16</span> power: <span>0.75</span><span>17</span> # display every <span>100</span><span> iterations</span><span>18</span> display: <span>100</span><span>19</span><span># the maximum number of iterations</span><span>20</span> max_iter: <span>10000</span><span>21</span><span># snapshot intermediate results</span><span>22</span> snapshot: <span>5000</span><span>23</span> snapshot_prefix: <span>"</span><span>examples/mnist/lenet</span><span>"</span><span>24</span><span># solver mode: cpu or gpu</span><span>25</span> solver_mode: cpu

带#的注释可以不管，能理解最好：

　　第二行的 net:　　路径需改为自己的网络模型xx_train_test.prototxt路径。其他的学习率 base_lr，lr_policy等不建议修改；max_iter最大迭代次数可以稍微改小，display显示间隔也可以随意修改~最后一行，我是只有cpu模式所以设为cpu，如果可以用gpu加速可设为gpu！

　　到这基本设置就结束了，然后就是写命令执行测试程序了：

我选择写了批处理.bat文件执行，也可以直接在cmd环境输命令执行。

　　新建mnist_train.bat，内容如下：

cd ../../<span>"</span><span>build/x64/debug/caffe.exe</span><span>"</span> train --solver=examples/mnist/<span>lenet_solver.prototxt 
pause </span>

根据自己的情况修改第二行的路径位置，windows应该都是在build/x64目录下，有的博客写的/bin/目录其实是linux的并不适用于windows环境。还要注意使用斜线“/”，不要使用“”无法识别，python代码多为后者要修改！

我的环境只有debug目录，如果你有realease目录，使用realease目录。

 运行.bat成功后，会开始训练，训练结束界面如下：

　　最后几行可以看到accuracy的准确率可以达到99%，也是相当准确了！

提示，caffe文件夹内会生成.caffemodel文件

使用caffemodel文件开始测试：

　　三.测试数据

　　由于测试数据集也是直接下载好了的leveldb文件，所以省了不少步骤

　　直接新建mnist_test.bat文件，类似训练mnist模型一样，对该模型进行数据测试。　　

cd ../../<span>"</span><span>build/x64/debug/caffe.exe</span><span>"</span> test --model=examples/mnist/lenet_train_test.prototxt -weights=examples/mnist/<span>lenet_iter_10000.caffemodel
pause </span>

　　类似mnits_train.bat，修改文件路径名，test表示用于测试，model指向自己的网络模型文件，最后添加权值文件.caffemodel进行测试。

　　运行mnist_test.bat后，成功界面如下：

　　最后一行还是有98%的准确率还是很不错的，说明模型生成的还不错。

总结：其实还遇到了不少零零碎碎的问题，大多都可以百度解决，主要是记得修改对自己的文件路径目录，windows下一定要使用leveldb数据文件，.prototxt也记得修改，然后就是等待模型跑完看结果了，看到高准确率还是很开心的~

　　四. 使用该模型

　　模型训练好了，数据也只是测试了，那么我们要使用该模型判断一张图片是数字几该如何做呢？

这个时候需要生成 classification.exe，然后执行相应的.bat命令来预测图片的分类结果。

　　mnist分类使用可以参考http://www.cnblogs.com/yixuan-xu/p/5862657.html

　　发现opencv可以加载caffe 框架模型，准备再写一篇博客进行实践介绍~

http://docs.opencv.org/3.1.0/d5/de7/tutorial_dnn_googlenet.html


【说明】：本文章由站长整理发布，文章内容不代表本站观点，如文中有侵权行为，请与本站客服联系(QQ:254677821)！

相关教程推荐

其他课程推荐