4.caffe：train_val.prototxt、 solver.prototxt 、 deploy.prototxt（ 创建模型与编写配置文件）

一，train_val.prototxt

name: "CIFAR10_quick"
layer {
  name: "cifar"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    # mirror: true
    # mean_file: "examples/cifar10/mean.binaryproto"uu
    mean_file: "myself/00b/00bmean.binaryproto" 
  }
  data_param {
    # source: "examples/cifar10/cifar10_train_lmdb"
    source: "myself/00b/00b_train_lmdb"
    batch_size: 50
    backend: LMDB
  }
}
layer {
  name: "cifar"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    # mean_file: "examples/cifar10/mean.binaryproto"
    mean_file: "myself/00b/00bmean.binaryproto"
  }
  data_param {
    # source: "examples/cifar10/cifar10_test_lmdb"
    source: "myself/00b/00b_val_lmdb"
    batch_size: 50
    backend: LMDB
  }
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 32
    # pad: 1
    kernel_size: 4
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.0001
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "pool1"
  top: "pool1"
}
layer {
  name: "conv2"
  type: "Convolution"
  bottom: "pool1"
  top: "conv2"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 32
    # pad: 2
    kernel_size: 4
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "relu2"
  type: "ReLU"
  bottom: "conv2"
  top: "conv2"
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2"
  top: "pool2"
  pooling_param {
    pool: AVE
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv3"
  type: "Convolution"
  bottom: "pool2"
  top: "conv3"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 32
    # pad: 2
    kernel_size: 4
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "relu3"
  type: "ReLU"
  bottom: "conv3"
  top: "conv3"
}
layer {
  name: "pool3"
  type: "Pooling"
  bottom: "conv3"
  top: "pool3"
  pooling_param {
    pool: AVE
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv4"
  type: "Convolution"
  bottom: "pool3"
  top: "conv4"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 32
    # pad: 2
    kernel_size: 4
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "relu4"
  type: "ReLU"
  bottom: "conv4"
  top: "conv4"
}
layer {
  name: "pool4"
  type: "Pooling"
  bottom: "conv4"
  top: "pool4"
  pooling_param {
    pool: AVE
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "pool4"
  top: "ip1"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 200
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "ip2"
  type: "InnerProduct"
  bottom: "ip1"
  top: "ip2"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 3
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "accuracy"
  type: "Accuracy"
  bottom: "ip2"
  bottom: "label"
  top: "accuracy"
  include {
    phase: TEST
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "ip2"
  bottom: "label"
  top: "loss"
}

二，solver.prototxt

# reduce the learning rate after 8 epochs (4000 iters) by a factor of 10

# The train/test net protocol buffer definition
net: "myself/00b/train_val.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: 10
# Carry out testing every 500 training iterations.
test_interval: 70
# The base learning rate, momentum and the weight decay of the network.
base_lr: 0.001
momentum: 0.9
weight_decay: 0.004
# The learning rate policy
lr_policy: "fixed"
# lr_policy: "step"
gamma: 0.1
stepsize: 100
# Display every 100 iterations
display: 10
# The maximum number of iterations
max_iter: 2000
# snapshot intermediate results
# snapshot: 3000
# snapshot_format: HDF5
 snapshot_prefix: "myself/00b/00b"
# solver mode: CPU or GPU
solver_mode: CPU

三，deploy.prototxt

name: "CIFAR10_quick"
layer {
  name: "data"
  type: "Input"
  top: "data"
  input_param { shape: { dim: 1 dim: 3 dim: 101 dim: 101 } }
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  convolution_param {
    num_output: 32
    kernel_size: 4
    stride: 1
  }
}
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "conv1"
  top: "conv1"
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv2"
  type: "Convolution"
  bottom: "pool1"
  top: "conv2"
  convolution_param {
    num_output: 32
    kernel_size: 4
    stride: 1
  }
}
layer {
  name: "relu2"
  type: "ReLU"
  bottom: "conv2"
  top: "conv2"
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2"
  top: "pool2"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv3"
  type: "Convolution"
  bottom: "pool2"
  top: "conv3"
  convolution_param {
    num_output: 32
    kernel_size: 4
    stride: 1
  }
}
layer {
  name: "relu3"
  type: "ReLU"
  bottom: "conv3"
  top: "conv3"
}
layer {
  name: "pool3"
  type: "Pooling"
  bottom: "conv3"
  top: "pool3"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv4"
  type: "Convolution"
  bottom: "pool3"
  top: "conv4"
  convolution_param {
    num_output: 32
    kernel_size: 4
    stride: 1
  }
}
layer {
  name: "relu4"
  type: "ReLU"
  bottom: "conv4"
  top: "conv4"
}
layer {
  name: "pool4"
  type: "Pooling"
  bottom: "conv4"
  top: "pool4"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "pool4"
  top: "ip1"
  inner_product_param {
    num_output: 200
  }
}
layer {
  name: "ip2"
  type: "InnerProduct"
  bottom: "ip1"
  top: "ip2"
  inner_product_param {
    num_output: 3
  }
}
layer {
  #name: "loss"

  name: "prob"
  type: "Softmax" 
  bottom: "ip2"
  top: "prob"

  #top: "loss"
}

参考一：

模型就用程序自带的caffenet模型，位置在 models/bvlc_reference_caffenet/文件夹下, 将需要的两个配置文件，复制到myfile文件夹内

# sudo cp models/bvlc_reference_caffenet/solver.prototxt examples/myfile/
# sudo cp models/bvlc_reference_caffenet/train_val.prototxt examples/myfile/

修改train_val.protxt，只需要修改两个阶段的data层就可以了，其它可以不用管。





复制代码


name: "CaffeNet"
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    mirror: true
    crop_size: 227
    mean_file: "examples/myfile/mean.binaryproto"
  }
  data_param {
    source: "examples/myfile/img_train_lmdb"
    batch_size: 256
    backend: LMDB
  }
}
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    mirror: false
    crop_size: 227
    mean_file: "examples/myfile/mean.binaryproto"
  }
  data_param {
    source: "examples/myfile/img_test_lmdb"
    batch_size: 50
    backend: LMDB
  }
}




 


实际上就是修改两个data layer的mean_file和source这两个地方，其它都没有变化 。


修改其中的solver.prototxt




# sudo vi examples/myfile/solver.prototxt






复制代码


net: "examples/myfile/train_val.prototxt"
test_iter: 2
test_interval: 50
base_lr: 0.001
lr_policy: "step"
gamma: 0.1
stepsize: 100
display: 20
max_iter: 500
momentum: 0.9
weight_decay: 0.005
solver_mode: GPU


复制代码




100个测试数据，batch_size为50，因此test_iter设置为2，就能全cover了。在训练过程中，调整学习率，逐步变小。




参考二：


前面做好了lmdb和均值文件，下面以Googlenet为例修改网络并训练模型。


 


我们将caffe-master\models下的bvlc_googlenet文件夹复制到caffe-master\examples\imagenet下。（因为我们的lmdb和均值都在这里，放一起方便些）


打开train_val.txt，修改：


1.修改data层：


 






[plain] view plain copy print?在CODE上查看代码片派生到我的代码片




    

  1. layer {  
    2. 

  2.   name: "data"  
    3. 

  3.   type: "Data"  
    4. 

  4.   top: "data"  
    5. 

  5.   top: "label"  
    6. 

  6.   include {  
    7. 

  7.     phase: TRAIN  
    8. 

  8.   }  
    9. 

  9.   transform_param {  
    10. 

  10.     mirror: true  
    11. 

  11.     crop_size: 224  
    12. 

  12.     mean_file: "examples/imagenet/mydata_mean.binaryproto" #均值文件  
    13. 

  13.     #mean_value: 104 #这些注释掉  
    14. 

  14.     #mean_value: 117  
    15. 

  15.     #mean_value: 123  
    16. 

  16.   }  
    17. 

  17.   data_param {  
    18. 

  18.     source: "examples/imagenet/mydata_train_lmdb" #训练集的lmdb  
    19. 

  19.     batch_size: 32 #根据GPU修改  
    20. 

  20.     backend: LMDB  
    21. 

  21.   }  
    22. 

  22. }  
    23. 






 






[plain] view plain copy print?在CODE上查看代码片派生到我的代码片




    

  1. layer {  
    2. 

  2.   name: "data"  
    3. 

  3.   type: "Data"  
    4. 

  4.   top: "data"  
    5. 

  5.   top: "label"  
    6. 

  6.   include {  
    7. 

  7.     phase: TEST  
    8. 

  8.   }  
    9. 

  9.   transform_param {  
    10. 

  10.     mirror: false  
    11. 

  11.     crop_size: 224  
    12. 

  12.     mean_file: "examples/imagenet/mydata_mean.binaryproto" #均值文件  
    13. 

  13.     #mean_value: 104  
    14. 

  14.     #mean_value: 117  
    15. 

  15.     #mean_value: 123  
    16. 

  16.   }  
    17. 

  17.   data_param {  
    18. 

  18.     source: "examples/imagenet/mydata_val_lmdb" #验证集lmdb  
    19. 

  19.     batch_size: 50 #和solver中的test_iter相乘约等于验证集大小  
    20. 

  20.     backend: LMDB  
    21. 

  21.   }  
    22. 

  22. }  
    23. 






 


2.修改输出：


由于Googlenet有三个输出，所以改三个地方，其他网络一般只有一个输出，则改一个地方即可。


如果是微调，那么输出层的层名也要修改。（参数根据层名来初始化，由于输出改了，该层参数就不对应了，因此要改名）






[plain]







layer {
  name: "loss1/classifier"
  type: "InnerProduct"
  bottom: "loss1/fc"
  top: "loss1/classifier"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1000 #改成你的数据集类别数
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}






[plain]






layer {
  name: "loss2/classifier"
  type: "InnerProduct"
  bottom: "loss2/fc"
  top: "loss2/classifier"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1000 #改成你的数据集类别数
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}






[plain]






layer {
  name: "loss3/classifier"
  type: "InnerProduct"
  bottom: "pool5/7x7_s1"
  top: "loss3/classifier"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1000 #改成你的数据集类别数
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}


3.打开deploy.prototxt，修改：






[plain]







layer {
  name: "loss3/classifier"
  type: "InnerProduct"
  bottom: "pool5/7x7_s1"
  top: "loss3/classifier"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1000 #改成你的数据集类别数
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}


如果是微调，该层层名和train_val.prototxt修改一致。


接着，打开solver，修改：






[plain]







net: "examples/imagenet/bvlc_googlenet/train_val.prototxt" #路径不要错
test_iter: 1000 #前面已说明该值
test_interval: 4000 #迭代多少次测试一次
test_initialization: false
display: 40
average_loss: 40
base_lr: 0.01
lr_policy: "step"
stepsize: 320000 #迭代多少次改变一次学习率
gamma: 0.96
max_iter: 10000000 #迭代次数
momentum: 0.9
weight_decay: 0.0002
snapshot: 40000
snapshot_prefix: "examples/imagenet/bvlc_googlenet" #生成的caffemodel保存在imagenet下，形如bvlc_googlenet_iter_***.caffemodel
solver_mode: GPU



这时，我们回到caffe-master\examples\imagenet下，打开train_caffenet.sh，修改：


(如果是微调，在脚本里加入-weights **/**/**.caffemodel即可，即用来微调的caffemodel路径)




[plain]





#!/usr/bin/env sh

./build/tools/caffe train \
    -solver examples/imagenet/bvlc_googlenet/solver.prototxt -gpu 0



（如果有多个GPU，可自行选择） 然后，在caffe-master下执行改脚本即可开始训练:$caffe-master ./examples/imagenet/train_caffenet.sh


训练得到的caffemodel就可以用来做图像分类了，此时，需要（1）得到的labels.txt，（2）得到的mydata_mean.binaryproto，（3）得到的caffemodel以及已经修改过的deploy.prototxt，共四个文件，具体过程看：http://blog.csdn.net/sinat_30071459/article/details/50974695


参考三：


*_train_test.prototxt,*_deploy.prototxt,*_slover.prototxt文件编写时注意


1、*_train_test.prototxt文件


这是训练与测试网络配置文件


（1）在数据层中 参数include{


                                 phase：TRAIN/TEST


                             }


TRAIN与TEST不能有“...”否则会报错，还好提示信息里，会提示哪一行出现了问题，如下图：




数字8就代表配置文件的第8行出现了错误


（2）卷积层和全连接层相似：卷积层（Convolution），全连接层（InnerProduct，容易翻译成内积层）相似处有两个【1】：都有两个param{lr_mult：1


                                           decay_mult：1                            


                               }


                             param{lr_mult： 2


                                        decay_mult: 0            


                              }


【2】：convolution_param{}与inner_product_param{}里面的参数相似，甚至相同


今天有事，明天再续！


续上！


（3）平均值文件*_mean.binaryproto要放在transform_param{}里，训练与测试数据集放在data_param{}里


2.*_deploy.prototxt文件


【1】*_deploy.prototxt文件的构造和*_train_test.prototxt文件的构造稍有不同首先没有test网络中的test模块，只有训练模块


【2】数据层的写法和原来也有不同，更加简洁：


input: "data" input_dim: 1 input_dim: 3 input_dim: 32 input_dim: 32


注意红色部分，那是数据层的名字，没有这个的话，第一卷积层无法找到数据，我一开始没有加这句就报错。下面的四个参数有点类似batch_size（1,3,32,32）里四个参数


【3】卷积层和全连接层中weight_filler{}与bias_filler{}两个参数不用再填写，应为这两个参数的值，由已经训练好的模型*.caffemodel文件提供


【4】输出层的变化（1）没有了test模块测试精度（2）输出层


*_train_test.prototxt文件：


layer{   name: "loss"   type: "SoftmaxWithLoss"#注意此处与下面的不同   bottom: "ip2"   bottom: "label"#注意标签项在下面没有了，因为下面的预测属于哪个标签，因此不能提供标签   top: "loss" }


*_deploy.prototxt文件：


layer {   name: "prob"   type: "Softmax"   bottom: "ip2"   top: "prob" }


***注意在两个文件中输出层的类型都发生了变化一个是SoftmaxWithLoss，另一个是Softmax。另外为了方便区分训练与应用输出，训练是输出时是loss，应用时是prob。


3、*_slover.prototxt


net: "test.prototxt" #训练网络的配置文件 test_iter: 100 #test_iter 指明在测试阶段有多上个前向过程（也就是有多少图片）被执行。 在MNIST例子里，在网络配置文件里已经设置test网络的batch size=100，这里test_iter 设置为100，那在测试阶段共有100*100=10000 图片被处理 test_interval: 500 #每500次训练迭代后，执行一次test base_lr: 0.01 #学习率初始化为0.01 momentum：0.9 #u=0.9 weight_decay:0.0005 # lr_policy: "inv" gamma: 0.0001 power: 0.75 #以上三个参数都和降低学习率有关，详细的学习策略和计算公式见下面 // The learning rate decay policy. The currently implemented learning rate   


// policies are as follows:   


//    - fixed: always return base_lr.   


//    - step: return base_lr * gamma ^ (floor(iter / step))   


//    - exp: return base_lr * gamma ^ iter 


//    - inv: return base_lr * (1 + gamma * iter) ^ (- power)   


//    - multistep: similar to step but it allows non uniform steps defined by   


//      stepvalue   


//    - poly: the effective learning rate follows a polynomial decay, to be   


//      zero by the max_iter. return base_lr (1 - iter/max_iter) ^ (power)   


//    - sigmoid: the effective learning rate follows a sigmod decay  


 //      return base_lr ( 1/(1 + exp(-gamma * (iter - stepsize))))   


// where base_lr, max_iter, gamma, step, stepvalue and power are defined   


// in the solver parameter protocol buffer, and iter is the current iteration. display:100 #每100次迭代，显示结果 snapshot: 5000 #每5000次迭代，保存一次快照 snapshot_prefix: "path_prefix" #快照保存前缀：更准确的说是快照保存路径+前缀，应为文件名后的名字是固定的 solver_mode:GPU #选择解算器是用cpu还是gpu


批处理文件编写：


F:/caffe/caffe-windows-master/bin/caffe.exe train --solver=C:/Users/Administrator/Desktop/caffe_test/cifar-10/cifar10_slover_prototxt --gpu=all pause


参考四：


 


06


 


原


 


将train_val.prototxt 转换成deploy.prototxt


 




                         


                                                                                                                                              




 




1.删除输入数据（如：type:data...inckude{phase: TRAIN}），然后添加一个数据维度描述。


 






[plain] view plain copy print?在CODE上查看代码片派生到我的代码片




    

  1. input: "data"   
    2. 

  2. input_dim: 1   
    3. 

  3. input_dim: 3   
    4. 

  4. input_dim: 224   
    5. 

  5. input_dim: 224  
    6. 

  6. force_backward: true  
    7. 






input: "data" 
input_dim: 1 
input_dim: 3 
input_dim: 224 
input_dim: 224
force_backward: true


 




 


2.移除最后的“loss” 和“accuracy” 层，加入“prob”层。


 






[plain] view plain copy print?在CODE上查看代码片派生到我的代码片




    

  1. layers {  
    2. 

  2.   name: "prob"  
    3. 

  3.   type: SOFTMAX  
    4. 

  4.   bottom: "fc8"  
    5. 

  5.   top: "prob"  
    6. 

  6. }  
    7. 







layers {
  name: "prob"
  type: SOFTMAX
  bottom: "fc8"
  top: "prob"
}

如果train_val文件中还有其他的预处理层，就稍微复杂点。如下，在'data'层，在‘data’层和‘conv1’层(with bottom:”data”  / top:”conv1″). 插入一个层来计算输入数据的均值。

 


 






[plain] view plain copy print?在CODE上查看代码片派生到我的代码片




    

  1. layer {  
    2. 

  2. name: “mean”  
    3. 

  3. type: “Convolution”  
    4. 

  4. <strong>bottom: “data”  
    5. 

  5. top: “data”</strong>  
    6. 

  6. param {  
    7. 

  7. lr_mult: 0  
    8. 

  8. decay_mult: 0  
    9. 

  9. }  
    10. 

  10.   
    11. 

  11. …}  
    12. 






在deploy.prototxt文件中，“mean” 层必须保留，只是容器改变，相应的‘conv1’也要改变 ( bottom:”mean”/ top:”conv1″ )。






[plain]




    

  1. layer {  
    2. 

  2. name: “mean”  
    3. 

  3. type: “Convolution”  
    4. 

  4. <strong>bottom: “data”  
    5. 

  5. top: “mean“</strong>  
    6. 

  6. param {  
    7. 

  7. lr_mult: 0  
    8. 

  8. decay_mult: 0  
    9. 

  9. }  
    10. 

  10.   
    11. 

  11. …}  
    12. 





 







	
posted on 
2016-08-17 11:30 
WP的烂笔头 
阅读(9726) 
评论(0) 
编辑 
收藏 
举报












    
    

    
    

    

        

    

        

            
                
                
            
        

    

        

            
                
            
        

    

    

    











	
		

			

			
                
                    

			

                    

	





	Powered by: 博客园
	Copyright © 2024 WP的烂笔头

Powered by .NET 9.0 on Kubernetes