mnist示例


* mnist数据集合

* 大型手写识别图片:600000个训练集合和100000个测试集合。

  • 每个图片归一化,数字居中,28*28

* 下载MNIST数据集

  • 下载数据集: data/mnist下使用脚本:get_mnist.sh
cd data/mnist/

./get_mnist.sh

tree

* mnsit数据格式

  • 训练集,图片(train-images-...)-像素值(0-255)
  • 训练集,标签(train-label-...)-标签值(0-9)
  • 测试集,图片(t10k-images-...)
  • 测试集,标签(t10k-labels-...)

* 格式转换

* 将二进制转换为levelDB或者LMDB

 ./examples/mnist/create_mnist.sh
 //需要从caffe的根目录执行,否则报错
 //build/examples/mnist/convert_mnist_data.bin: not found这样的错误

* Data Layer

  • 从lmdb中读出数据,数据层
layer {
  name: "mnist"  //名字
  type: "Data"    //数据类型(数据层)
  transform_param {
    scale: 0.00390625   //数据变化的数据缩放因子( 1 divided by 256)  
  }
  data_param {    //数据层参数
    source: "mnist_train_lmdb"  //LMDB的路径
    backend: LMDB   //数据格式
    batch_size: 64   //批处理的数目
  }
  top: "data"  //层的输出blob两个,data和label
  top: "label"
}
  • Blob 是Caffe处理和传输的真实数据的包装类,同时它还隐含提供了在CPU和GPU之间同步数据的能力。

  • caffe使用blob存储和交换数据。blob对不同数据提供了统一的内存接口;例如:一批图片,模型参数,优化过程的偏导数等。

  • Blob是Caffe的基本数据结构,具有CPU和GPU之间同步的能力,它是4维的数组(Num, Channels, Height, Width)。


* 卷积层:Convolution Layer

layer {
  name: "conv1"
  type: "Convolution"
  param { lr_mult: 1 }  //权值学习速率倍乘因子,1表示与全局参数一致
  param { lr_mult: 2 }  //bias学习速率倍乘因子,是全局参数的2倍
  convolution_param {    //卷积计算参数
    num_output: 20        //输出feature_map的数目:20
    kernel_size: 5        //卷积核额尺寸:5*5
    stride: 1             //卷积输出的间隔跳跃,1表示连续输出,无跳跃
    weight_filler {       //权值使用xavier填充器
      type: "xavier"
    }
    bias_filler {         //bias使用常数填充器,默认是0
      type: "constant"
    }
  }
  bottom: "data"
  top: "conv1"
}

* pooling层

layer {
  name: "pool1"
  type: "Pooling"
  pooling_param {
    kernel_size: 2   //使用最大值采样方法:2*2
    stride: 2        //下采样输出跳跃间隔2*2
    pool: MAX        //下窗口尺寸2*2

  }
  bottom: "conv1"
  top: "pool1"
}

* Fully Connected Layer

layer {
  name: "ip1"
  type: "InnerProduct"
  param { lr_mult: 1 }
  param { lr_mult: 2 }
  inner_product_param {
    num_output: 500      //输出元素个数为500
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
  bottom: "pool2"
  top: "ip1"
}

* ReLU Layer

layer {
  name: "relu1"
  type: "ReLU"
  bottom: "ip1"
  top: "ip1"
}

* innerproduct layer(内积层)

layer {
  name: "ip2"
  type: "InnerProduct"
  param { lr_mult: 1 }
  param { lr_mult: 2 }
  inner_product_param {
    num_output: 10
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
  bottom: "ip1"
  top: "ip2"
}

* accuracy层

layer {
  name: "accuracy"
  type: "Accuracy"
  bottom: "ip2"
  bottom: "label"
  top: "accuracy"
  include {
    phase: TEST
  }
}

* loss层

layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "ip2"
  bottom: "label"
}

2. 训练超参数

* 运行examples/mnist/lenet_train_test.prototxt

* lenet_solver.protext训练超参数

# The train/test net protocol buffer definition
net: "examples/mnist/lenet_train_test.prototxt"
# test_iter specifies how many forward passes the test should carry out.
# In the case of MNIST, we have test batch size 100 and 100 test iterations,
# covering the full 10,000 testing images.
test_iter: 100
# Carry out testing every 500 training iterations.
test_interval: 500
# The base learning rate, momentum and the weight decay of the network.
base_lr: 0.01
momentum: 0.9
weight_decay: 0.0005
# The learning rate policy
lr_policy: "inv"
gamma: 0.0001
power: 0.75
# Display every 100 iterations
display: 100
# The maximum number of iterations
max_iter: 10000
# snapshot intermediate results
snapshot: 5000
snapshot_prefix: "examples/mnist/lenet"
# solver mode: CPU or GPU
solver_mode: GPU

3. 训练日志

  • GLOG格式:日期-时间-进程号-源码文件-代码行号] 输出信息

4. 使用训练好的模型预测

./build/tools/caffe.bin test \  //表示只做预测,并进行参数更新
- model examples/mnist/lenet_train_test.portptxt \
- weights examples/mnist/lenet_iter_10000.caffemodel \
- iterations 100   //迭代次数

5. build/tools/caffe.bin

http://blog.csdn.net/langb2014/article/details/50458014

* 命令

  • train----训练或finetune模型(model),
  • test-----测试模型
  • device_query---显示gpu信息
  • time-----显示程序执行时间

* 参数

-solver -gpu -snapshot -weights -iteration -model -sighup_effect -sigint_effect

results matching ""

    No results matching ""