Chapter 3 Start Caffe with MNIST Demo

先从一个具体的例子来开始Caffe，以MNIST手写数据为例。 

1.下载数据

下载mnist到caffe-masterdatamnist文件夹。

THE MNIST DATABASE：Yann LeCun et al.

 train-images-idx3-ubyte.gz:  training set images (9912422 bytes)

 train-labels-idx1-ubyte.gz:  training set labels (28881 bytes)

 t10k-images-idx3-ubyte.gz:   test set images (1648877 bytes)

 t10k-labels-idx1-ubyte.gz:   test set labels (4542 bytes)

2.生成lmdb文件

使用convert_mnist_data project转换数据。

打开Caffe.sln，设置convert_mnist_data为启动项目，修改convert_mnist_data.cpp中代码。

在main函数中设置了转换数据的路径，具体的代码如下：

    //get mnist train and test lmdb data By Xiaopan Lyu=====================
    argc = 4;
    argv[0] = "lmdb";
    //convert train mnist data=============================================
    argv[1] = "../../data/mnist/train-images.idx3-ubyte";
    argv[2] = "../../data/mnist/train-labels.idx1-ubyte";
    argv[3] = "../../data/mnist/mnist_train_lmdb";

    //convert test mnist data=============================================
    argv[1] = "../../data/mnist/t10k-images.idx3-ubyte";
    argv[2] = "../../data/mnist/t10k-labels.idx1-ubyte";
    argv[3] = "../../data/mnist/mnist_test_lmdb";
    //======================================================================

这段代码在main函数中的位置如下：

int main(int argc, char** argv) {
#ifndef GFLAGS_GFLAGS_H_
    namespace gflags = google;
#endif

    FLAGS_alsologtostderr = 1;

    //get mnist train and test lmdb data By Xiaopan Lyu=====================
    argc = 4;
    argv[0] = "lmdb";
    //convert train mnist data=============================================
    argv[1] = "../../data/mnist/train-images.idx3-ubyte";
    argv[2] = "../../data/mnist/train-labels.idx1-ubyte";
    argv[3] = "../../data/mnist/mnist_train_lmdb";

    //convert test mnist data=============================================
    argv[1] = "../../data/mnist/t10k-images.idx3-ubyte";
    argv[2] = "../../data/mnist/t10k-labels.idx1-ubyte";
    argv[3] = "../../data/mnist/mnist_test_lmdb";
    //======================================================================

    gflags::SetUsageMessage("This script converts the MNIST dataset to
"
        "the lmdb/leveldb format used by Caffe to load data.
"
        "Usage:
"
        "    convert_mnist_data [FLAGS] input_image_file input_label_file "
        "output_db_file
"
        "The MNIST dataset could be downloaded at
"
        "    http://yann.lecun.com/exdb/mnist/
"
        "You should gunzip them after downloading,"
        "or directly use data/mnist/get_mnist.sh
");
    gflags::ParseCommandLineFlags(&argc, &argv, true);

    const string& db_backend = FLAGS_backend;

    if (argc != 4) {
        gflags::ShowUsageWithFlagsRestrict(argv[0],
            "examples/mnist/convert_mnist_data");
    }
    else {
        google::InitGoogleLogging(argv[0]);
        convert_dataset(argv[1], argv[2], argv[3], db_backend);
    }
    system("pause");
    return 0;
}

两次运行代码，分别得到train和test data。

get mnist_train_lmdb

get mnist_test_lmdb

Notes.

1）argv[0]、argv[1]、argv[2]、argv[3]分别表示的含义：[FLAGS]     input_image_file    input_label_file    output_db_file.

2）output_db_file设置中最后的一级文件夹不要事先自己建立好，代码中不支持覆盖，如果存在文件夹会报错。

3）这些路径的设置是在debug模式下，文件的层级是以当前的.cpp文件为基础的，与实际EXE文件有所不同。

3.配置网络为TRAIN模式

1）配置lenet_train_test.prototxt

caffe在mnist自带的是使用lenet的网络结。lenet网络的定义在examplesmnistlenet_train_test.prototxt文件中。

注意配置好改网络定义中的数据路径。如下所示，注意line 14&31，如果目录比较混淆，可以直接写成绝对路径。

   1:  name: "LeNet"
   2:  layer {
   3:    name: "mnist"
   4:    type: "Data"
   5:    top: "data"
   6:    top: "label"
   7:    include {
   8:      phase: TRAIN
   9:    }
  10:    transform_param {
  11:      scale: 0.00390625
  12:    }
  13:    data_param {
  14:      source: "E:/MyCode/DL/caffe-master/examples/mnist/mnist_train_lmdb"
  15:      batch_size: 64
  16:      backend: LMDB
  17:    }
  18:  }
  19:  layer {
  20:    name: "mnist"
  21:    type: "Data"
  22:    top: "data"
  23:    top: "label"
  24:    include {
  25:      phase: TEST
  26:    }
  27:    transform_param {
  28:      scale: 0.00390625
  29:    }
  30:    data_param {
  31:      source: "E:/MyCode/DL/caffe-master/examples/mnist/mnist_test_lmdb"
  32:      batch_size: 100
  33:      backend: LMDB
  34:    }
  35:  }

2）配置lenet_solver.prototxt

lenet_solver.prototxt中实际上是定义了一种解决方案。

注意line 2,23&25,这三行的数据需要修改，这里也是用了绝对路径。只使用CPU训练。

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

3）修改了source code为train模式

修改了caffe.cpp文件的相关内容。增加了line15到line21的代码，顺便说一句Google的gflags解析命令行参数甚是优雅。

   1:  int main(int argc, char** argv) {
   2:      // Print output to stderr (while still logging).
   3:      FLAGS_alsologtostderr = 1;
   4:      // Set version
   5:      gflags::SetVersionString(AS_STRING(CAFFE_VERSION));
   6:      // Usage message.
   7:      gflags::SetUsageMessage("command line brew
"
   8:          "usage: caffe <command> <args>

"
   9:          "commands:
"
  10:          "  train           train or finetune a model
"
  11:          "  test            score a model
"
  12:          "  device_query    show GPU diagnostic information
"
  13:          "  time            benchmark model execution time");
  14:      // Run tool or show usage.
  15:      //train lenet By XiaopanLyu====================================================
  16:      argc = 3;
  17:      argv[0] = "caffe";
  18:      argv[1] = "train";
  19:      argv[2] = "-solver=E:/MyCode/DL/caffe-master/examples/mnist/lenet_solver.prototxt";
  20:      //argv[1] = "solver=../examples/mnist/lenet_solver.prototxt";
  21:      //=============================================================================
  22:      caffe::GlobalInit(&argc, &argv);
  23:      if (argc == 2) {
  24:  #ifdef WITH_PYTHON_LAYER
  25:          try {
  26:  #endif
  27:              return GetBrewFunction(caffe::string(argv[1]))();
  28:              system("pause");
  29:  #ifdef WITH_PYTHON_LAYER
  30:          }
  31:          catch (bp::error_already_set) {
  32:              PyErr_Print();
  33:              return 1;
  34:          }
  35:  #endif
  36:      }
  37:      else {
  38:          gflags::ShowUsageWithFlagsRestrict(argv[0], "tools/caffe");
  39:      }
  40:  }

4.Training LeNet

运行caffe project，mnist demo就开始运行了，以下就是运行的过程和结果。

1）运行过程

2）运行结果

运行完成后，生成了4个文件。查看lenet_solver.prototxt可知，最大迭代次数为10000次，5000次保存一次快照结果。

5.配置网络为TEST模式

修改caffe.cpp文件，增加参数配置代码

   1:      //test lenet By XiaopanLyu====================================================
   2:      argc = 5;
   3:      argv[0] = "caffe";
   4:      argv[1] = "test";
   5:      argv[2] = "-model=E:/MyCode/DL/caffe-master/examples/mnist/lenet_train_test.prototxt";
   6:      argv[3] = "-weights=E:/MyCode/DL/caffe-master/examples/mnist/lenet_iter_10000.caffemodel";
   7:      argv[4] = "-iterations=100";
   8:      //=============================================================================
   9:      caffe::GlobalInit(&argc, &argv);

6.Testing LeNet

用LeNet的网络配置运行mnist测试数据集，几分钟的时间得到如下效果。

迭代100次，测试数据集的准确率为99.02%。

7.NOTES

在第3、5部分，配置网络的参数可以参考Caffe的官方辅导文档：http://caffe.berkeleyvision.org/tutorial/interfaces.html