caffe用到的命令和零碎知识

这些东西总是忘记来回查，特此记录一下：

---

1. caffe标注txt文件的读取与保存（使用pandas.DataFrame）

读取:

train_data = pd.read_csv('Train.txt', sep=' ',encoding='gbk',index_col=0, header=None)

保存

test1.to_csv('test1.txt', sep=' ',encoding='gbk', header=None)

实例：

2. LMDB数据读取Data层：

caffe生成的lmdb数据，有生产2文件夹和4文件夹版本， 对应data层如下：

windows版本：

layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TRAIN
  }
   transform_param {
    mirror: true
    crop_size: 224
    mean_file: "mean.binaryproto"
    contrast_brightness_adjustment: true
    smooth_filtering: true
    max_color_shift: 10
    max_smooth: 6
    apply_probability: 0.5
  }
  data_param {
    source: "CaffeLMDB/TrainDataDB"
    batch_size: 24
    backend: LMDB
  }
}
layer {
  name: "label"
  type: "Data"
  top: "label"
  include {
    phase: TRAIN
  }
  data_param {
    source: "CaffeLMDB/TrainlableDB"
    batch_size: 24
    backend: LMDB
  }
}
layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TEST
  }
   transform_param {
    mirror: false
    crop_size: 224
    mean_file: "mean.binaryproto"
  }
  data_param {
    source: "CaffeLMDB/ValDataDB"
    batch_size: 1
    backend: LMDB
  }
}
layer {
  name: "label"
  type: "Data"
  top: "label"
  include {
    phase: TEST
  }
  data_param {
    source: "CaffeLMDB/VallableDB"
    batch_size: 1
    backend: LMDB
  }
}

我的linux版本：

layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    mirror: true
    crop_size: 331
    mean_file: "./lmdb_data/img_test_lmdb/mean.binaryproto"
    contrast_brightness_adjustment: true
    smooth_filtering: true
    max_color_shift: 10
    max_smooth: 6
    apply_probability: 0.5
  }
  data_param {
    source: "./lmdb_data/img_test_lmdb"
    batch_size: 8
    backend: LMDB
  }
}

layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    mirror: false
    crop_size: 331
    mean_file: "./lmdb_data/img_train_lmdb/mean.binaryproto"
  }
  data_param {
    source: "./lmdb_data/img_train_lmdb"
    batch_size: 1
    backend: LMDB
  }
}

3. caffe训练命令

#!/bin/bash
set -e 
# 设置环境变量, 制定caffe路径                                                                                                                                                                                
export PATH=/home/zhuoshi/ZSZT/Geoffrey/caffe/caffe-master/build/tools:$PATH

# 设置开启日志
GLOG_logtostderr=0 
GLOG_log_dir='./log' 
caffe train 
     --solver=solver.prototxt 
     --weights=snapshot/solver_iter_50000.caffemodel # 这里指定caffe模型文件
     --snapshot=snapshot/solver_iter_20000.solverstate # 这里指定caffe中间态文件
     --gpu=0 # 可以指定序号或者`all`

snapshot和weights只能指定一个用于继续训练和finetune

4. caffe的solver文件解释

# 网络的模型文件
net: "./train.prototxt"
# test_iter * 测试batch_size = 测试集大小
test_iter: 2000
# 每迭代多少次测试一遍
test_interval: 25000
base_lr: 0.001
lr_policy: "step"
gamma: 0.1
# 每隔多少次衰减学习率
stepsize: 50000
# 迭代多少次显示一次日志
display: 100
# 最大迭代次数
max_iter: 500000
momentum: 0.9
weight_decay: 0.0005
# 每隔多少次保存
snapshot: 20000 
snapshot_prefix: "./snapshot"
solver_mode: GPU

5. txt2lmdb脚本

地址

#!/usr/bin/env sh
# Create the imagenet lmdb inputs
 
####################################################
# 配置
####################################################
EXAMPLE=/home/zhuoshi/ZSZT/Geoffrey/Person-resnet18/lmdb_data       # lmdb存储位置
DATA=/home/zhuoshi/ZSZT/Geoffrey/Person-resnet18/data                   # txt文件所在文件夹 - 同时也是.txt相对路径的起始点(图片绝对路径=$DATA+txt中相对路径)
CAFFE_HOME=/home/zhuoshi/ZSZT/Geoffrey/caffe/caffe-master               # caffe的工具库
HEIGHT=256
WIDTH=256
####################################################
# 处理train
####################################################
echo "Creating train lmdb..."
TRAIN_PATH=$EXAMPLE/img_train_lmdb
# 如果存在,删除原数据
if [ ! -d "$TRAIN_PATH/" ];then
    echo "文件不存在"
    mkdir -p $TRAIN_PATH/
else
    echo "$TRAIN_PATH文件夹已存在"
    rm -rf $TRAIN_PATH/
    # mkdir -p $TRAIN_PATH/
fi
# 生成lmdb
$CAFFE_HOME/build/tools/convert_imageset --shuffle --resize_height=$HEIGHT --resize_width=$WIDTH  $DATA/  $DATA/train.txt  $TRAIN_PATH  # 
echo "Creating train lmdb Done!, Create mean.binaryproto..."
# # 计算图片均值
$CAFFE_HOME/build/tools/compute_image_mean $TRAIN_PATH/ $TRAIN_PATH/mean.binaryproto
echo "train Done!"

####################################################
# 处理test
####################################################
echo "Creating test lmdb ..."
TEST_PATH=$EXAMPLE/img_test_lmdb
# 如果存在,删除原数据
if [ ! -d "$TEST_PATH/" ];then
    echo "文件不存在"
    mkdir $TEST_PATH
else
    echo "$TEST_PATH文件夹已存在"
    rm -rf $TEST_PATH/
    # mkdir $TEST_PATH/
fi
# 生成lmdb
$CAFFE_HOME/build/tools/convert_imageset --shuffle --resize_height=256 --resize_width=256  $DATA/  $DATA/test.txt  $TEST_PATH  # 
echo "Creating test lmdb Done!, Create mean.binaryproto..."
# # 计算图片均值
$CAFFE_HOME/build/tools/compute_image_mean $TEST_PATH/ $TEST_PATH/mean.binaryproto
echo "test Done!"

####################################################
echo "Done."

windows版本的bat命令：

echo "Creating train Mulit-Out CaffeLMDB..."

del "E:/CaffeLMDB/TrainDataDB*.*" /f /s 
del "E:/CaffeLMDB/TrainlableDB*.*" /f /s 

del "E:/CaffeLMDB/ValDataDB*.*" /f /s 
del "E:/CaffeLMDB/VallableDB*.*" /f /s 

rd /s /q "E:/CaffeLMDB/TrainDataDB"
rd /s /q "E:/CaffeLMDB/TrainlableDB"

rd /s /q "E:/CaffeLMDB/ValDataDB"
rd /s /q "E:/CaffeLMDB/VallableDB"

"E:/CaffeTool/Caffe_Release/convert_imageset" --gray=false --shuffle --resize_height=224 --resize_width=224 "" "E:/CaffeTrain/PersonProperty/data/Train.txt"  "E:/CaffeLMDB/TrainDataDB" "E:/CaffeLMDB/TrainlableDB" 20

"E:/CaffeTool/Caffe_Release/convert_imageset" --gray=false --shuffle --resize_height=224 --resize_width=224 "" "E:/CaffeTrain/PersonProperty/data/Validate.txt"  "E:/CaffeLMDB/ValDataDB" "E:/CaffeLMDB/VallableDB" 20

pause

6. 编译错误

caffe编译错误记录

7. anaconda环境编译caffe

使用Anaconda虚拟环境编译caffe-gpu pycaffe