caffe训练模型并预测图片类型

参考：http://www.cnblogs.com/denny402/tag/caffe/（徐其华）

            http://blog.csdn.net/u010193446/article/details/53433631

目录： 1. 准备图像数据

             2.创建清单

             3.将图像数据转换为lmdb格式

             4.计算均值

             5.编写参数的配置文件及训练网络的配置文件

             6.训练模型

             7.用训练的模型预测图片的类别

1.  准备图像数据

     训练数据：将图片按类别进行分类，并标记，例如准备两类图片，一类为福码图片249张，用1表示此类，一类为其他各式的图片249张，用0表示此类

     测试数据：将图片按类别进行分类，并标记，例如准备两类图片，一类为福码图片61张，一类为其他各式的图片61张

    在主文件夹下创建work_liuyan_2017/readpractice文件夹，训练图片保存在trainimage，测试图片保存在testimage

2.  创建清单

     sudo vim work_liuyan_2017/readpractice create_list.sh

  

#/usr/bin/env sh
DATATRAIN=readpractice/
DATATEST=readpractice/
MY=readpractice
echo "create train.txt"
rm -f $MY/train.txt
for i in 0 1
do
find $DATATRAIN/trainimage -name $i*.png | cut -d '/' -f4-5 | sed "s/$/ $i/">>$MY/train.txt
done
echo "create test.txt"
rm -f $MY/test.txt
for i in 0 1
do
find $DATATEST/testimage -name $i*.png | cut -d '/' -f4-5 | sed "s/$/ $i/">>$MY/test.txt
done
echo "all done"

运行此脚本

cd work_liuyan_2017

sudo sh  readpractice/create_list.sh

若成功在work_liuyan_2017/readpractice下生成traintxt及test.txt，里面为图片的列表清单

3.将图像数据转换为lmdb格式

cd work_liuyan_2017

sudo vim readpractice/create_lmdb.sh

   

#!/usr/bin/env sh
MY=readpractice
echo "create train lmdb"
rm -rf $MY/img_train_lmdb
/home/liuyan/caffe-master/build/tools/convert_imageset \
--shuffle \
--resize_height=240 \
--resize_width=240 \
/home/liuyan/work_liuyan_2017/readpractice/trainimage/ \
$MY/train.txt \
$MY/img_train_lmdb

echo "create test lmdb"
rm -rf $MY/img_test_lmdb
/home/liuyan/caffe-master/build/tools/convert_imageset \
--shuffle \
--resize_height=240 \
--resize_width=240 \
/home/liuyan/work_liuyan_2017/readpractice/testimage/ \
$MY/test.txt \
$MY/img_test_lmdb

echo "all done"

cd  work_liuyan_2017

sudo sh   readpractice /create_lmdb.sh

生成 img_train_lmdb 和 img_test_lmdb文件

4. 计算均值

图片减去均值再训练，会提高训练速度和精度。因此，一般都会有这个操作。

caffe程序提供了一个计算均值的文件compute_image_mean.cpp，我们直接使用就可以了

sudo /home/liuyan/caffe-master/build/tools/compute_image_mean /home/liuyan/work_liuyan_2017/readpractice/img_train_lmdb /home/liuyan/work_liuyan_2017/readpractice/mean.binaryproto

5.编写参数的配置文件及训练网络的配置文件

编写参数的配置文件

cd worke_liuyan_2017

sudo vim readpractice/solver.prototxt

net: "/home/liuyan/work_liuyan_2017/readpractice/train_val.prototxt"
test_iter: 2
test_interval: 500
base_lr: 0.01
lr_policy: "step"
gamma: 0.1
stepsize: 100
display: 50
max_iter: 2000
momentum: 0.9
weight_decay: 0.0005
snapshot: 1000
snapshot_prefix: "/home/liuyan/work_liuyan_2017/readpractice/"
solver_mode: GPU

编写训练网络的配置文件

首先：sudo cp /home/liuyan/caffe-master/models/bvlc_reference_caffenet/train_val.prototxt readpractice

其次：修改部分train_val.prototxt

开始的

name: "CaffeNet"
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    mirror: true
    crop_size: 227
    mean_file: "readpractice/mean.binaryproto"   #路径 注意cd到哪里了
  }
# mean pixel / channel-wise mean instead of mean image
#  transform_param {
#    crop_size: 227
#    mean_value: 104
#    mean_value: 117
#    mean_value: 123
#    mirror: true
#  }
  data_param {
    source: "readpractice/img_train_lmdb"
    batch_size: 127
    backend: LMDB
  }
}
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    mirror: false
    crop_size: 227
    mean_file: "readpractice/mean.binaryproto"
  }
# mean pixel / channel-wise mean instead of mean image
#  transform_param {
#    crop_size: 227
#    mean_value: 104
#    mean_value: 117
#    mean_value: 123
#    mirror: false
#  }
  data_param {
    source: "readpractice/img_test_lmdb"
    batch_size: 61
    backend: LMDB
  }
}

最后的：

layer {
  name: "fc8"
  type: "InnerProduct"
  bottom: "fc7"
  top: "fc8"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 2                 #原来为1000，训练为n类，就写>=n,训练的为2类，我就写了2
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "accuracy"
  type: "Accuracy"
  bottom: "fc8"
  bottom: "label"
  top: "accuracy"
  include {
    phase: TEST
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "fc8"
  bottom: "label"
  top: "loss"
}

6.训练模型

cd work_liuyan_2017

开始训练 sudo /home/liuyan/caffe-master/build/tools/caffe train -solver  readpractice/solver.prototxt
模型会保存在

snapshot_prefix: "/home/liuyan/work_liuyan_2017/readpractice/"
这个路径中
训练结果： 

7.用训练的模型预测图片的类别

首先：

 sudo cp caffe-master/examples/myfile/deploy.prototxt work_liuyan_2017/readpractice/

修改deploy.prototxt最后部分

layer {
  name: "fc8"  
  type: "InnerProduct"
  bottom: "fc7"
  top: "fc8"
  inner_product_param {
    num_output: 2                  #输出类别为2类，保持一致
  }
}
layer {
  name: "prob"
  type: "Softmax"
  bottom: "fc8"
  top: "prob"
}

其次：准备一张将要预测的图片，命名为2.png

然后：

创建labels.txt:

在终端输入：

 sudo caffe-master/build/examples/cpp_classification/classification.bin \
> work_liuyan_2017/readpractice/deploy.prototxt \
> work_liuyan_2017/readpractice/_iter_1000.caffemodel \
> work_liuyan_2017/readpractice/mean.binaryproto \

>work_liuyan_2017/readpractice/labels.txt \

> work_liuyan_2017/readpractice/2.png

回车输出结果，最后的几行

不过图片数量和迭代次数都不够，只是试一试