Caffe实现概述

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

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

29. def read():

30. lmdb_env = lmdb.open('lmdb_data')
31. lmdb_txt = lmdb_env.begin()
32.
33. datum = caffe_pb2.Datum()
34.
35. for key, value in lmdb_txt.cursor():
36.
37. datum.ParseFromString(value)
38.
39. label = datum.label
40.
41. data = caffe.io.datum_to_array(datum)
42.
43. print(label)
44. print(data)
45.
46.

47. if __name__ == '__main__':

48. write()
49. read()

demo_create_solver.py: 生成solver配置文件

from caffe.proto import caffe_pb2
s = caffe_pb2.SolverParameter()
s.train_net = "train.prototxt"
s.test_net.append("test.prototxt")
s.test_interval = 100
s.test_iter.append(10)
10.

11. s.max_iter = 1000

13. s.base_lr = 0.1

15. s.weight_decay = 5e-4

17. s.lr_policy = "step"

19. s.display = 10

21. s.snapshot = 10

23. s.snapshot_prefix = "model"

25. s.type = "SGD"

27. s.solver_mode = caffe_pb2.SolverParameter.GPU

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

30. f.write(str(s))
31.
32.
33.
34.

结果如下：

train_net: "/home/kuan/PycharmProjects/demo_cnn_net/net/train.prototxt"
test_net: "/home/kuan/PycharmProjects/demo_cnn_net/net/test.prototxt"
test_iter: 1000
test_interval: 100
base_lr: 0.10000000149
display: 100
max_iter: 100000
lr_policy: "step"
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

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

47. if __name__ == '__main__':

48. create_net()

生成结果如下：

layer {
name: "data"
type: "Data"
top: "data"
top: "label"
transform_param {
mirror: true
crop_size: 40
}
10. data_param {
11. source: "/home/kuan/PycharmProjects/demo_cnn_net/lmdb_data"
12. batch_size: 32
13. backend: LMDB
14. }

15. }

16. layer {

17. name: "conv1"
18. type: "Convolution"
19. bottom: "data"
20. top: "conv1"
21. convolution_param {
22. num_output: 20
23. kernel_size: 5
24. weight_filler {
25. type: "xavier"
26. }
27. bias_filler {
28. type: "xavier"
29. }
30. }

31. }

32. layer {

33. name: "relu1"
34. type: "ReLU"
35. bottom: "conv1"
36. top: "conv1"

37. }

38. layer {

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

48. }

49. layer {

50. name: "conv2"
51. type: "Convolution"
52. bottom: "pool1"
53. top: "conv2"
54. convolution_param {
55. num_output: 32
56. pad: 1
57. kernel_size: 3
58. weight_filler {
59. type: "xavier"
60. }
61. bias_filler {
62. type: "xavier"
63. }
64. }

65. }

66. layer {

67. name: "relu2"
68. type: "ReLU"
69. bottom: "conv2"
70. top: "conv2"

71. }

72. layer {

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

82. }

83. layer {

84. name: "fc3"
85. type: "InnerProduct"
86. bottom: "pool2"
87. top: "fc3"
88. inner_product_param {
89. num_output: 1024
90. weight_filler {
91. type: "xavier"
92. }
93. }

94. }

95. layer {

96. name: "relu3"
97. type: "ReLU"
98. bottom: "fc3"
99. top: "fc3"
}
layer {
name: "drop"
type: "Dropout"
bottom: "fc3"
top: "drop"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc4"
type: "InnerProduct"
bottom: "drop"
top: "fc4"
inner_product_param {
num_output: 10
weight_filler {
type: "xavier"
}
}
}
layer {
name: "loss"
type: "SoftmaxWithLoss"
bottom: "fc4"
bottom: "label"
top: "loss"
}

demo_train.py训练网络：

import sys
sys.path.append('/home/kuan/AM-softmax_caffe/python')
import caffe
solver = caffe.SGDSolver("/home/kuan/PycharmProjects/demo_cnn_net/cnn_net/alexnet/solver.prototxt")
solver.solve()

demo_test.py:测试网络

import sys
sys.path.append('/home/kuan/AM-softmax_caffe/python')
import caffe
import numpy as np
##caffemodel deploy.prototxt
deploy = "/home/kuan/PycharmProjects/demo_cnn_net/cnn_net/alexnet/deploy.prototxt"

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

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

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

17. net.forward()

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

21. print(prob)

人工智能芯片与自动驾驶