从零开始理解caffe网络的参数

LeNet网络介绍

LeNet-5模型，大神Yann LeCun的作品，用以手写字体的识别，在邮编识别上得到应用，论文《Gradient-Based Learning Applied
to Document Recognition》堪称CNN开山之作。

LeNet网络详解

网络名称

name: "LeNet"         # 网络（NET）名称为LeNet

mnist层-train

layer {               # 定义一个层  name: "mnist"       # 层名字mnist  type: "Data"        # 层的类型，数据层  top: "data"         # 层的输出blob有两个；data和label  top: "label"        # blob是数据库中用来存储二进制文件的字段类型  include {    phase: TRAIN      # 该层仅在训练阶段有效  }  transform_param {    scale: 0.00390625 # 数据变换使用的数据缩放因子  }  data_param {        # 数据层参数     source: "A:/Caffe/caffe/examples/mnist/mnist-train-leveldb"   # 数据路径    batch_size: 64     # 批量数目，一次读取64张图    backend: LEVELDB   # 数据格式为LEVELDB  }}

mnist层-test

layer {                # 同是名为mnist且输出为data和label的数据层，只是定义的参数只在检测阶段有效  name: "mnist"  type: "Data"  top: "data"  top: "label"  include {    phase: TEST  }  transform_param {    scale: 0.00390625  }  data_param {    source: "A:/Caffe/caffe/examples/mnist/mnist-test-leveldb"    batch_size: 1    backend: LEVELDB  }}

卷积层1

# lr_mult: 学习率的系数，最终的学习率是这个数乘以solver.prototxt配置文件中的base_lr。# 如果有两个lr_mult, 则第一个表示权值的学习率，第二个表示偏置项的学习率。一般偏置项的学习率是权值学习率的两倍。layer {                      # 定义了一个卷积层conv1  name: "conv1"  type: "Convolution"  bottom: "data"             # 输入blob为data  top: "conv1"               # 输出blob为conv1  param {    lr_mult: 1               # 全局学习速率倍乘因子，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数量为20    kernel_size: 5           # 卷积核尺寸，5X5    stride: 1                # 卷积输出跳跃间隔，1表示连续输出，无跳跃    weight_filler {          # 权值使用xavier填充器      type: "xavier"    }    bias_filler {      type: "constant"       # bias使用常数填充器，默认是0    }  }}

下采样层pool1

layer {                      # 下采样层pol1，输入conv1，输出pool1  name: "pool1"  type: "Pooling"  bottom: "conv1"  top: "pool1"  pooling_param {    pool: MAX                # 最大值下采样法    kernel_size: 2           # 下采样窗口尺寸2X2    stride: 2                # 下采样输出跳跃间隔2X2  }}

卷积层2

layer {  name: "conv2"  type: "Convolution"  bottom: "pool1"  top: "conv2"  param {    lr_mult: 1  }  param {    lr_mult: 2  }  convolution_param {    num_output: 50    kernel_size: 5    stride: 1    weight_filler {      type: "xavier"    }    bias_filler {      type: "constant"    }  }}

下采样层pool2

layer {  name: "pool2"  type: "Pooling"  bottom: "conv2"  top: "pool2"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}

全连接层1

layer {                        # 全连接层  name: "ip1"  type: "InnerProduct"  bottom: "pool2"  top: "ip1"  param {    lr_mult: 1  }  param {    lr_mult: 2  }  inner_product_param {    num_output: 500            # 该层输出参数 500X500    weight_filler {      type: "xavier"    }    bias_filler {      type: "constant"    }  }}

采用ReLU的非线性层

layer {                         # 非线性层，用ReLU方法  name: "relu1"  type: "ReLU"  bottom: "ip1"  top: "ip1"}

全连接层2

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

accuracy层

layer {                        # 分类准确率层  name: "accuracy"  type: "Accuracy"  bottom: "ip2"  bottom: "label"  top: "accuracy"  include {    phase: TEST  }}

损失层

layer {                      # 损失层  name: "loss"  type: "SoftmaxWithLoss" # 采用softmax回归  bottom: "ip2"  bottom: "label"  top: "loss"}

网络结构图

参考资料：《深度学习-21天实战Caffe》