Caffe实现概述

Caffe实现概述

目录

一、caffe配置文件介绍

二、标准层的定义

 三、网络微调技巧

四、Linux脚本使用及LMDB文件生成

五、带你设计一个Caffe网络，用于分类任务

一、caffe配置文件介绍

 

 

 

   

 

 

 二、标准层的定义

 

 

 三、网络微调技巧

 

 

 其中，multistep最为常用

 

 

 四、Linux脚本使用及LMDB文件生成

 

 

 五、带你设计一个Caffe网络，用于分类任务

 

 

 下面：

使用pycaffe生成solver配置

使用pycaffe生成caffe测试网络和训练网络

 

数据集下载：

# demoCaffe

数据集下载，cifar mnist:
百度云盘:

链接: https://pan.baidu.com/s/1bHFQUz7Q6BMBZv25AhsXKQ 密码: dva9
链接: https://pan.baidu.com/s/1rPRjf2hanlYYjBQQDmIjNQ 密码: 5nhv

1. lmdb数据制作：

手动实现： https://blog.csdn.net/yx2017/article/details/72953537   

               https://www.jianshu.com/p/9d7ed35960cb

代码实现：https://www.cnblogs.com/leemo-o/p/4990021.html

                 https://www.jianshu.com/p/ef84715e0fdc

以下仅供对比阅读：

demo_lmdb.py：  生成lmdb格式数据

 

1. import lmdb
2. import numpy as np
3. import cv2
4. import caffe
5. from caffe.proto import caffe_pb2
6.  
7. def write():
8.     # basic setting
9.  
10. 10.     lmdb_file = 'lmdb_data'
11. 11.     batch_size = 256
12. 12.  
13. 13.  
14. 14.     lmdb_env = lmdb.open(lmdb_file, map_size = int(1e12))
15. 15.  
16. 16.     lmdb_txn = lmdb_env.begin(write = True)
17. 17.  
18. 18.     for x in range(batch_size):
19. 19.         data = np.ones((3, 64, 64), np.uint8)
20. 20.         label = x
21. 21.  
22. 22.         datum = caffe.io.array_to_datum(data,label)
23. 23.         keystr = "{:0>8d}".format(x)
24. 24.  
25. 25.         lmdb_txn.put(keystr, datum.SerializeToString())
26. 26.  
27. 27.     lmdb_txn.commit()
28. 28.  

29. def read():

1. 30.     lmdb_env = lmdb.open('lmdb_data')
2. 31.     lmdb_txt = lmdb_env.begin()
3. 32.  
4. 33.     datum = caffe_pb2.Datum()
5. 34.  
6. 35.     for key, value in lmdb_txt.cursor():
7. 36.  
8. 37.         datum.ParseFromString(value)
9. 38.  
10. 39.         label = datum.label
11. 40.  
12. 41.         data = caffe.io.datum_to_array(datum)
13. 42.  
14. 43.         print(label)
15. 44.         print(data)
16. 45.  
17. 46.  

47. if __name__ == '__main__':

1. 48.     write()
2. 49.     read()

demo_create_solver.py:  生成solver配置文件

1. from caffe.proto import caffe_pb2
2.  
3. s = caffe_pb2.SolverParameter()
4.  
5. s.train_net = "train.prototxt"
6. s.test_net.append("test.prototxt")
7.  
8. s.test_interval = 100
9. s.test_iter.append(10)
10. 10.  

11. s.max_iter = 1000

1. 12.  

13. s.base_lr = 0.1

1. 14.  

15. s.weight_decay = 5e-4

1. 16.  

17. s.lr_policy = "step"

1. 18.  

19. s.display = 10

1. 20.  

21. s.snapshot = 10

1. 22.  

23. s.snapshot_prefix = "model"

1. 24.  

25. s.type = "SGD"

1. 26.  

27. s.solver_mode = caffe_pb2.SolverParameter.GPU

1. 28.  

29. with open("net/s.prototxt", "w") as f:

1. 30.     f.write(str(s))
2. 31.  
3. 32.  
4. 33.  
5. 34.  

结果如下：

1. train_net: "/home/kuan/PycharmProjects/demo_cnn_net/net/train.prototxt"
2. test_net: "/home/kuan/PycharmProjects/demo_cnn_net/net/test.prototxt"
3. test_iter: 1000
4. test_interval: 100
5. base_lr: 0.10000000149
6. display: 100
7. max_iter: 100000
8. lr_policy: "step"
9. weight_decay: 0.000500000023749

10. snapshot: 100

11. snapshot_prefix: "/home/kuan/PycharmProjects/demo_cnn_net/cnn_model/mnist/lenet/"

12. solver_mode: GPU

13. type: "SGD"

demo_creat_net.py:    创建网络

1. import caffe
2.  
3. def create_net():
4.     net = caffe.NetSpec()
5.  
6.     net.data, net.label = caffe.layers.Data(source="data.lmdb",
7.                                             backend=caffe.params.Data.LMDB,
8.                                             batch_size=32,
9.                                             ntop=2,  #数据层数据个数，分别为data,label
10. 10.                                             transform_param=dict(crop_size=40, mirror=True)
11. 11.                                             )
12. 12.  
13. 13.     net.conv1 = caffe.layers.Convolution(net.data, num_output=20, kernel_size=5,
14. 14.                                          weight_filler={"type": "xavier"},
15. 15.                                          bias_filler={"type":"xavier"})  #卷积核参数
16. 16.  
17. 17.     net.relu1 = caffe.layers.ReLU(net.conv1, in_place=True)
18. 18.  
19. 19.     net.pool1 = caffe.layers.Pooling(net.relu1, pool=caffe.params.Pooling.MAX,
20. 20.                                      kernel_size=3, stride=2)
21. 21.  
22. 22.     net.conv2 = caffe.layers.Convolution(net.pool1, num_output=32, kernel_size=3,
23. 23.                                          pad=1,
24. 24.                                          weight_filler={"type": "xavier"},
25. 25.                                          bias_filler={"type": "xavier"})
26. 26.  
27. 27.     net.relu2 = caffe.layers.ReLU(net.conv2, in_place=True)
28. 28.  
29. 29.     net.pool2 = caffe.layers.Pooling(net.relu2, pool=caffe.params.Pooling.MAX,
30. 30.                                      kernel_size=3, stride=2)
31. 31.     #下面为全连接层
32. 32.     net.fc3 = caffe.layers.InnerProduct(net.pool2, num_output=1024, weight_filler=dict(type='xavier'))
33. 33.  
34. 34.     net.relu3 = caffe.layers.ReLU(net.fc3, in_place=True)
35. 35.  
36. 36.     ##drop
37. 37.     net.drop = caffe.layers.Dropout(net.relu3, dropout_param=dict(dropout_ratio=0.5))
38. 38.  
39. 39.     net.fc4 = caffe.layers.InnerProduct(net.drop, num_output=10, weight_filler=dict(type='xavier'))
40. 40.  
41. 41.     net.loss = caffe.layers.SoftmaxWithLoss(net.fc4, net.label)
42. 42.  
43. 43.     with open("net/tt.prototxt", 'w') as f:
44. 44.         f.write(str(net.to_proto()))
45. 45.  
46. 46.  

47. if __name__ == '__main__':

1. 48.     create_net()

生成结果如下：

1. layer {
2.   name: "data"
3.   type: "Data"
4.   top: "data"
5.   top: "label"
6.   transform_param {
7.     mirror: true
8.     crop_size: 40
9.   }
10. 10.   data_param {
11. 11.     source: "/home/kuan/PycharmProjects/demo_cnn_net/lmdb_data"
12. 12.     batch_size: 32
13. 13.     backend: LMDB
14. 14.   }

15. }

16. layer {

1. 17.   name: "conv1"
2. 18.   type: "Convolution"
3. 19.   bottom: "data"
4. 20.   top: "conv1"
5. 21.   convolution_param {
6. 22.     num_output: 20
7. 23.     kernel_size: 5
8. 24.     weight_filler {
9. 25.       type: "xavier"
10. 26.     }
11. 27.     bias_filler {
12. 28.       type: "xavier"
13. 29.     }
14. 30.   }

31. }

32. layer {

1. 33.   name: "relu1"
2. 34.   type: "ReLU"
3. 35.   bottom: "conv1"
4. 36.   top: "conv1"

37. }

38. layer {

1. 39.   name: "pool1"
2. 40.   type: "Pooling"
3. 41.   bottom: "conv1"
4. 42.   top: "pool1"
5. 43.   pooling_param {
6. 44.     pool: MAX
7. 45.     kernel_size: 3
8. 46.     stride: 2
9. 47.   }

48. }

49. layer {

1. 50.   name: "conv2"
2. 51.   type: "Convolution"
3. 52.   bottom: "pool1"
4. 53.   top: "conv2"
5. 54.   convolution_param {
6. 55.     num_output: 32
7. 56.     pad: 1
8. 57.     kernel_size: 3
9. 58.     weight_filler {
10. 59.       type: "xavier"
11. 60.     }
12. 61.     bias_filler {
13. 62.       type: "xavier"
14. 63.     }
15. 64.   }

65. }

66. layer {

1. 67.   name: "relu2"
2. 68.   type: "ReLU"
3. 69.   bottom: "conv2"
4. 70.   top: "conv2"

71. }

72. layer {

1. 73.   name: "pool2"
2. 74.   type: "Pooling"
3. 75.   bottom: "conv2"
4. 76.   top: "pool2"
5. 77.   pooling_param {
6. 78.     pool: MAX
7. 79.     kernel_size: 3
8. 80.     stride: 2
9. 81.   }

82. }

83. layer {

1. 84.   name: "fc3"
2. 85.   type: "InnerProduct"
3. 86.   bottom: "pool2"
4. 87.   top: "fc3"
5. 88.   inner_product_param {
6. 89.     num_output: 1024
7. 90.     weight_filler {
8. 91.       type: "xavier"
9. 92.     }
10. 93.   }

94. }

95. layer {

1. 96.   name: "relu3"
2. 97.   type: "ReLU"
3. 98.   bottom: "fc3"
4. 99.   top: "fc3"
5. }
6. layer {
7.   name: "drop"
8.   type: "Dropout"
9.   bottom: "fc3"
10.   top: "drop"
11.   dropout_param {
12.     dropout_ratio: 0.5
13.   }
14. }
15. layer {
16.   name: "fc4"
17.   type: "InnerProduct"
18.   bottom: "drop"
19.   top: "fc4"
20.   inner_product_param {
21.     num_output: 10
22.     weight_filler {
23.       type: "xavier"
24.     }
25.   }
26. }
27. layer {
28.   name: "loss"
29.   type: "SoftmaxWithLoss"
30.   bottom: "fc4"
31.   bottom: "label"
32.   top: "loss"
33. }

demo_train.py训练网络：

1. import sys
2. sys.path.append('/home/kuan/AM-softmax_caffe/python')
3. import caffe
4.  
5. solver = caffe.SGDSolver("/home/kuan/PycharmProjects/demo_cnn_net/cnn_net/alexnet/solver.prototxt")
6.  
7. solver.solve()

demo_test.py:测试网络

1. import sys
2. sys.path.append('/home/kuan/AM-softmax_caffe/python')
3. import caffe
4. import numpy as np
5.  
6. ##caffemodel deploy.prototxt
7.  
8. deploy = "/home/kuan/PycharmProjects/demo_cnn_net/cnn_net/alexnet/deploy.prototxt"
9.  

10. model = "/home/kuan/PycharmProjects/demo_cnn_net/cnn_model/cifar/alexnet/alexnet_iter_110.caffemodel"

1. 11.  

12. net = caffe.Net(deploy, model, caffe.TEST)

1. 13.  
2. 14.  

15. net.blobs["data"].data[...] = np.ones((3,32,32),np.uint8)

1. 16.  

17. net.forward()

1. 18.  

19. prob = net.blobs["prob"].data[0]

1. 20.  

21. print(prob)

1. 22.