深度学习与人脸识别系列（3）__利用caffe训练深度学习模型

作者：wjmishuai

出处：http://blog.csdn.net/wjmishuai/article/details/50854162

一：下载训练集和测试集

训练集的下载地址:http://www.cbsr.ia.ac.cn/english/CASIA-WebFace-Database.html

测试集的下载地址: http://vis-www.cs.umass.edu/lfw/

二:利用linux脚本生成训练集

#!/bin/bash


#-----------------------------------------------------------------------------------------


#生成训练集和验证集
#made by 郭开


#-----------------------------------------------------------------------------------------


#保存图片的路径  
PATH=/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa
echo "start..."
#遍历文件夹
for name in $PATH/webface_img/*;do
        var=0
#遍历文件夹中的图片
        for file_name in $name/*;do  
            var=$(($var+1));  
            str=$file_name  
            #保存图片路径
            str=${str#*img}  
            #保存图片lable  
            lable=${str#*/}  
            lable=${lable%%/*}  
            if [ "$var" = "1" ] || [ "$var" = "3" ] || [ "$var" = "5" ]; then  
            #验证集  
            echo $str" "$lable>>/home/gk/val.txt  
            else  
            #测试集  
            echo $str" "$lable>>/home/gk/train.txt  
            fi  
done  
done  
echo "Done."

生成的文件如下图:

三：将图片转换成LMDB格式

#!/usr/bin/env sh  


#-----------------------------------------------------------------------------------------  
  
  
# 将图片批量生成lmbd格式的数据存储
#made by 郭开  
  
  
#-----------------------------------------------------------------------------------------  
  
#保存生成的lmdb的目录  
EXAMPLE=/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa/lmdb_webface
  
#train.txt和val.txt所在的目录
DATA=/home/gk
  
#转换图片的工具所在的目录  
TOOLS=/home/gk/caffe-master/build/tools  
  
#图片所在的目录
TRAIN_DATA_ROOT=/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa/webface_img

VAL_DATA_ROOT=/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa/webface_img
  
# 设置 RESIZE=true 可以把图片resize成想要的尺寸。
RESIZE=true
if $RESIZE; then
  RESIZE_HEIGHT=227
  RESIZE_WIDTH=227
else
  RESIZE_HEIGHT=0
  RESIZE_WIDTH=0
fi


if [ ! -d "$TRAIN_DATA_ROOT" ]; then
  echo "Error: TRAIN_DATA_ROOT is not a path to a directory: $TRAIN_DATA_ROOT"
  echo "Set the TRAIN_DATA_ROOT variable in create_imagenet.sh to the path" \
       "where the ImageNet training data is stored."
  exit 1
fi


if [ ! -d "$VAL_DATA_ROOT" ]; then
  echo "Error: VAL_DATA_ROOT is not a path to a directory: $VAL_DATA_ROOT"
  echo "Set the VAL_DATA_ROOT variable in create_imagenet.sh to the path" \
       "where the ImageNet validation data is stored."
  exit 1
fi


echo "Creating train lmdb..."


GLOG_logtostderr=1 $TOOLS/convert_imageset \
    --resize_height=$RESIZE_HEIGHT \
    --resize_width=$RESIZE_WIDTH \
    --shuffle \
    $TRAIN_DATA_ROOT \
    $DATA/train.txt \
    $EXAMPLE/train_lmdb


echo "Creating val lmdb..."


GLOG_logtostderr=1 $TOOLS/convert_imageset \
    --resize_height=$RESIZE_HEIGHT \
    --resize_width=$RESIZE_WIDTH \
    --shuffle \
    $VAL_DATA_ROOT \
    $DATA/val.txt \
    $EXAMPLE/val_lmdb


echo "Done."

四：计算图片的mean或者设置scale：0.00390625

#!/usr/bin/env sh  


#-----------------------------------------------------------------------------------------    
    
    
# 计算图像均值
#made by 郭开    
    
    
#-----------------------------------------------------------------------------------------    
#lmdb格式的文件所在的路径
EXAMPLE=/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa/lmdb_webface
#均值文件保存的路径
DATA=/media/gk/44CA719BCA718A46  
#转换图片的工具所在的目录 
TOOLS=/home/gk/caffe-master/build/tools  
  
$TOOLS/compute_image_mean $EXAMPLE/train_lmdb \  
  $DATA/mean.binaryproto  
  
echo "Done."  

五：训练模型

./build/tools/caffe train --solver=models/bvlc_reference_caffenet/solver.prototxt

slover.prototxt

net: "/home/gk/caffe-master/examples/vgg/train.prototxt"
test_iter: 100
test_interval: 1000


base_lr: 0.001
lr_policy: "step"
gamma: 0.95
stepsize:  100000
momentum: 0.9
weight_decay: 0.0005
display: 100
max_iter:  5000000
snapshot:  50000
snapshot_prefix: "/home/gk/caffe-master/examples/DeepID/snapshot"
solver_mode: GPU
device_id:0
#debug_info: true

tarin.prototxt

name: "VGG_FACE_16_layers"
layer {
  top: "data_1"
  top: "label_1"
  name: "data_1"
  type: "Data"
  data_param {
    source: "/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa/webface_lmdb/train"
    backend:LMDB
    batch_size: 128
  }
  transform_param {
     mean_file: "/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa/webface_lmdb/mean.binaryproto"
     mirror: true
  }
  include: { phase: TRAIN }
}


layer {
  top: "data_1"
  top: "label_1"
  name: "data_1"
  type: "Data"
  data_param {
    source: "/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa/webface_lmdb/val"
    backend:LMDB
    batch_size: 128
  }
  transform_param {
    mean_file: "/media/gk/9ec75485-26b1-471f-9b7b-d18554ca3fdd/aa/webface_lmdb/mean.binaryproto"
    mirror: true
  }
  include: { 
    phase: TEST 
  }
}


layers {
  bottom: "data"
  top: "conv1_1"
  name: "conv1_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv1_1"
  top: "conv1_1"
  name: "relu1_1"
  type: RELU
}
layers {
  bottom: "conv1_1"
  top: "conv1_2"
  name: "conv1_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv1_2"
  top: "conv1_2"
  name: "relu1_2"
  type: RELU
}
layers {
  bottom: "conv1_2"
  top: "pool1"
  name: "pool1"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool1"
  top: "conv2_1"
  name: "conv2_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv2_1"
  top: "conv2_1"
  name: "relu2_1"
  type: RELU
}
layers {
  bottom: "conv2_1"
  top: "conv2_2"
  name: "conv2_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv2_2"
  top: "conv2_2"
  name: "relu2_2"
  type: RELU
}
layers {
  bottom: "conv2_2"
  top: "pool2"
  name: "pool2"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool2"
  top: "conv3_1"
  name: "conv3_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv3_1"
  top: "conv3_1"
  name: "relu3_1"
  type: RELU
}
layers {
  bottom: "conv3_1"
  top: "conv3_2"
  name: "conv3_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv3_2"
  top: "conv3_2"
  name: "relu3_2"
  type: RELU
}
layers {
  bottom: "conv3_2"
  top: "conv3_3"
  name: "conv3_3"
  type: CONVOLUTION
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv3_3"
  top: "conv3_3"
  name: "relu3_3"
  type: RELU
}
layers {
  bottom: "conv3_3"
  top: "pool3"
  name: "pool3"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool3"
  top: "conv4_1"
  name: "conv4_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv4_1"
  top: "conv4_1"
  name: "relu4_1"
  type: RELU
}
layers {
  bottom: "conv4_1"
  top: "conv4_2"
  name: "conv4_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv4_2"
  top: "conv4_2"
  name: "relu4_2"
  type: RELU
}
layers {
  bottom: "conv4_2"
  top: "conv4_3"
  name: "conv4_3"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv4_3"
  top: "conv4_3"
  name: "relu4_3"
  type: RELU
}
layers {
  bottom: "conv4_3"
  top: "pool4"
  name: "pool4"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool4"
  top: "conv5_1"
  name: "conv5_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv5_1"
  top: "conv5_1"
  name: "relu5_1"
  type: RELU
}
layers {
  bottom: "conv5_1"
  top: "conv5_2"
  name: "conv5_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv5_2"
  top: "conv5_2"
  name: "relu5_2"
  type: RELU
}
layers {
  bottom: "conv5_2"
  top: "conv5_3"
  name: "conv5_3"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}