⼀、前⾔
⼆、训练模型之前的准备⼯作
由于主要是使⽤lenet模型训练⾃⼰的图⽚数据,我的图像数据共有10个类别,分别是0~9,相应地保存在名为0~9的⽂件夹,在/home/您的⽤户名/下新建⼀⽂件夹char_images,⽤于保存图像数据,在/home/您的⽤户名/char_images/下新建两个⽂件夹,名字分别为train和val,各⾃都包含了名为0~9的⽂件夹,例如⽂件夹0内存放的是字符”0”的图像,我的⽂件夹 如下: (2)对图像数据作统⼀缩放⾄28*28,并⽣成txt标签
为了计算均值⽂件,需要将所有图⽚缩放⾄统⼀的尺⼨,在train和val⽂件夹所在路径下创建python⽂件,命名getPath.py,并写⼊以下内容:
1. #coding:utf-8
2.
3. import cv2
4. import os
5.
6. def IsSubString( SubStrList , Str): #判断SubStrList的元素
7. flag = True #是否在Str内
8. for substr in SubStrList:
9. if not ( substr in Str):
10. flag = False
11.
12. return flag
13.
14. def GetFileList(FindPath,FlagStr=[]): #搜索⽬录下的⼦⽂件路径
15. FileList=[]
16. FileNames=os.listdir(FindPath)
17. if len(FileNames)>0:
18. for fn in FileNames:
19. if len(FlagStr)>0:
20. if IsSubString(FlagStr,fn): #不明⽩这⾥判断是为了啥
21. fullfilename=os.path.join(FindPath,fn)
22. FileList.append(fullfilename)
23. else:
24. fullfilename=os.path.join(FindPath,fn)
25. FileList.append(fullfilename)
26.狮龙音响
28. FileList.sort()
29.
30. return FileList
31.
32.
33. train_txt = open('' , 'w') #制作标签数据
34. classList =['0','1','2','3','4','5','6','7','8','9']
35. for idx in range(len(classList)) :
36. imgfile=GetFileList('train/'+ classList[idx])#将数据集放在与.py⽂件相同⽬录下
37. for img in imgfile:
38. srcImg = cv2.imread( img);
39. resizedImg = size(srcImg , (28,28))
40. cv2.imwrite( img ,resizedImg)
41. strTemp=img+' '+classList[idx]+'\n'#⽤空格代替转义字符 \t
42. train_txt.writelines(strTemp)
43. train_txt.close()
44.
45.
46. test_txt = open('' , 'w') #制作标签数据
47. for idx in range(len(classList)) :
48. imgfile=GetFileList('val/'+ classList[idx])
49. for img in imgfile:
50. srcImg = cv2.imread( img);
51. resizedImg = size(srcImg , (28,28))
52. cv2.imwrite( img ,resizedImg)
53. strTemp=img+' '+classList[idx]+'\n'#⽤空格代替转义字符 \t
54. test_txt.writelines(strTemp)
55. test_txt.close()
56.
57. print("成功⽣成⽂件列表")
运⾏该py⽂件,可将所有图⽚缩放⾄28*28⼤⼩,并且在rain和val⽂件夹所在路径下⽣成训练和测试图像数据的标签txt⽂件,⽂件内容为:
(3)⽣成lmdb格式的数据集
⾸先于caffe路径下新建⼀⽂件夹My_File,并在My_File下新建两个⽂件夹Build_lmdb和Data_label,将(2)中⽣成⽂本⽂件和搬⾄Data_label 下
将caffe路径下 examples/imagenet/create_imagenet.sh 复制⼀份到Build_lmdb⽂件夹下
打开create_imagenet.sh ,并做如下修改:
1. #!/usr/bin/env sh
2. # Create the imagenet lmdb inputs
3. # N.B. set the path to the imagenet train + val data dirs
4. set -e
5.
6. EXAMPLE=/home/你的⽤户名/caffe/My_File/Build_lmdb #⽣成的lmdb格式数据保存地址
7. DATA=/home/你的⽤户名/caffe/My_File/Data_label #两个txt标签⽂件所在路径
8. TOOLS=/home/你的⽤户名/caffe/build/tools #caffe⾃带⼯具,不⽤管,最好加绝对路径
9.
10. TRAIN_DATA_ROOT=/home/你的⽤户名/char_images/ #预先准备的训练图⽚路径,该路径和上写的路径合起来是图⽚完整路径
11. VAL_DATA_ROOT=/home/你的⽤户名/char_images/ #预先准备的测试图⽚路径,...
12.
13. # Set RESIZE=true to resize the images to 256x256. Leave as false if images have
14. # already been resized using another tool.
15. RESIZE=false #因为在这之前已经将图像归⼀化为28*28,所以设置为false
16. if $RESIZE; then
17. RESIZE_HEIGHT=28
18. RESIZE_WIDTH=28
19. else
20. RESIZE_HEIGHT=0
21. RESIZE_WIDTH=0
22. fi
23.
24. if [ ! -d "$TRAIN_DATA_ROOT" ]; then
25. echo "Error: TRAIN_DATA_ROOT is not a path to a directory: $TRAIN_DATA_ROOT"
26. echo "Set the TRAIN_DATA_ROOT variable in create_imagenet.sh to the path" \
27. "where the ImageNet training data is stored."
28. exit 1
29. fi
30.
31. if [ ! -d "$VAL_DATA_ROOT" ]; then
32. echo "Error: VAL_DATA_ROOT is not a path to a directory: $VAL_DATA_ROOT"
33. echo "Set the VAL_DATA_ROOT variable in create_imagenet.sh to the path" \
34. "where the ImageNet validation data is stored."
35. exit 1
36. fi
37.
38. echo "Creating "
39.
40. GLOG_logtostderr=1 $TOOLS/convert_imageset \
41. --resize_height=$RESIZE_HEIGHT \
42. --resize_width=$RESIZE_WIDTH \
44. --gray \ #灰度图像加上这个
45. $TRAIN_DATA_ROOT \
46. $ \
47. $EXAMPLE/train_lmdb #⽣成的lmdb格式训练数据集所在的⽂件夹
48.
49. echo "Creating "
50.
51. GLOG_logtostderr=1 $TOOLS/convert_imageset \
52. --resize_height=$RESIZE_HEIGHT \
53. --resize_width=$RESIZE_WIDTH \
54. --shuffle \
55. --gray \ #灰度图像加上这个
56. $VAL_DATA_ROOT \
57. $ \
58. $EXAMPLE/val_lmdb #⽣成的lmdb格式训练数据集所在的⽂件夹
液基细胞学59.
60. echo "Done."
以上只是为了说明修改的地⽅才添加汉字注释,实际时sh⽂件不要出现汉字,运⾏该sh⽂件,可在Build_lmdb⽂件夹内⽣成2个⽂件夹train_lmdb和val_lmdb,⾥⾯各有2个lmdb格式的⽂件
(4)更改lenet_solver.prototxt和lenet_train_test.prototxt
将caffe/examples/mnist下的 train_lenet.sh 、lenet_solver.prototxt 、lenet_train_test.prototxt 这三个⽂件复制⾄ My_File,⾸先修改train_lenet.sh 如下,只改了solver.prototxt的路径
1. #!/usr/bin/env sh
2. set -e
病毒唑注射液
3.
4. ./build/tools/caffe train --solver=My_File/lenet_solver.prototxt $@ #改路径
然后再更改lenet_solver.prototxt,如下:
1. # The train/test net protocol buffer definition
2. net: "My_File/lenet_train_test.prototxt"#改这⾥
3. # test_iter specifies how many forward passes the test should carry out.
4. # In the case of MNIST, we have test batch size 100 and 100 test iterations,
5. # covering the full 10,000 testing images.
6. test_iter: 100
7. # Carry out testing every 500 training iterations.
8. test_interval: 500
9. # The base learning rate, momentum and the weight decay of the network.
10. base_lr: 0.01
11. momentum: 0.9
12. weight_decay: 0.0005
13. # The learning rate policy
14. lr_policy: "inv"
15. gamma: 0.0001
16. power: 0.75
17. # Display every 100 iterations
18. display: 100
19. # The maximum number of iterations
21. # snapshot intermediate results
22. snapshot: 5000
23. snapshot_prefix: "My_File/"#改这⾥
24. # solver mode: CPU or GPU
25. solver_mode: GPU (/CPU)
最后修改lenet_train_test.prototxt ,如下:
1. name: "LeNet"
2. layer {
空间
3. name: "mnist"
4. type: "Data"
5. top: "data"
6. top: "label"
7. include {
8. phase: TRAIN
9. }
10. transform_param {
11. scale: 0.00390625
12. }
13. data_param {
朵康14. source: "My_File/Build_lmdb/train_lmdb"#改成⾃⼰的
15. batch_size: 64
16. backend: LMDB
17. }
18. }
19. layer {
20. name: "mnist"
21. type: "Data"
22. top: "data"
23. top: "label"
24. include {
25. phase: TEST
神龙赋
26. }
27. transform_param {
28. scale: 0.00390625
29. }
30. data_param {
31. source: "My_File/Build_lmdb/val_lmdb"#改成⾃⼰的
32. batch_size: 100
33. backend: LMDB
34. }
35. }
36. layer {
37. name: "conv1"
38. type: "Convolution"
39. bottom: "data"
40. top: "conv1"
41. param {
42. lr_mult: 1
43. }
44. param {
45. lr_mult: 2
46. }
47. convolution_param {