keras实现常用深度学习模型LeNet，AlexNet，ZFNet，VGGNet，GoogleNet，Resnet

转载自：点击打开链接

LeNet

[python]  view plain  copy

1. #coding=utf-8  
2. from keras.models import Sequential  
3. from keras.layers import Dense,Flatten  
4. from keras.layers.convolutional import Conv2D,MaxPooling2D  
5. from keras.utils.np_utils import to_categorical  
6. import cPickle  
7. import gzip  
8. import numpy as np  
9. seed = 7  
10. np.random.seed(seed)  
11.   
12. data = gzip.open(r'/media/wmy/document/BigData/kaggle/Digit Recognizer/mnist.pkl.gz')  
13. train_set,valid_set,test_set = cPickle.load(data)  
14. #train_x is [0,1]  
15. train_x = train_set[0].reshape((-1,28,28,1))  
16. train_y = to_categorical(train_set[1])  
17.   
18. valid_x = valid_set[0].reshape((-1,28,28,1))  
19. valid_y = to_categorical(valid_set[1])  
20.   
21. test_x = test_set[0].reshape((-1,28,28,1))  
22. test_y = to_categorical(test_set[1])  
23.   
24. model = Sequential()  
25. model.add(Conv2D(32,(5,5),strides=(1,1),input_shape=(28,28,1),padding='valid',activation='relu',kernel_initializer='uniform'))  
26. model.add(MaxPooling2D(pool_size=(2,2)))  
27. model.add(Conv2D(64,(5,5),strides=(1,1),padding='valid',activation='relu',kernel_initializer='uniform'))  
28. model.add(MaxPooling2D(pool_size=(2,2)))  
29. model.add(Flatten())  
30. model.add(Dense(100,activation='relu'))  
31. model.add(Dense(10,activation='softmax'))  
32. model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])  
33. model.summary()  
34.   
35. model.fit(train_x,train_y,validation_data=(valid_x,valid_y),batch_size=20,epochs=20,verbose=2)  
36. #[0.031825309940411217, 0.98979999780654904]  
37. print model.evaluate(test_x,test_y,batch_size=20,verbose=2)  

AlexNet

[python]  view plain  copy

1. #coding=utf-8  
2. from keras.models import Sequential  
3. from keras.layers import Dense,Flatten,Dropout  
4. from keras.layers.convolutional import Conv2D,MaxPooling2D  
5. from keras.utils.np_utils import to_categorical  
6. import numpy as np  
7. seed = 7  
8. np.random.seed(seed)  
9.   
10. model = Sequential()  
11. model.add(Conv2D(96,(11,11),strides=(4,4),input_shape=(227,227,3),padding='valid',activation='relu',kernel_initializer='uniform'))  
12. model.add(MaxPooling2D(pool_size=(3,3),strides=(2,2)))  
13. model.add(Conv2D(256,(5,5),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
14. model.add(MaxPooling2D(pool_size=(3,3),strides=(2,2)))  
15. model.add(Conv2D(384,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
16. model.add(Conv2D(384,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
17. model.add(Conv2D(256,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
18. model.add(MaxPooling2D(pool_size=(3,3),strides=(2,2)))  
19. model.add(Flatten())  
20. model.add(Dense(4096,activation='relu'))  
21. model.add(Dropout(0.5))  
22. model.add(Dense(4096,activation='relu'))  
23. model.add(Dropout(0.5))  
24. model.add(Dense(1000,activation='softmax'))  
25. model.compile(loss='categorical_crossentropy',optimizer='sgd',metrics=['accuracy'])  
26. model.summary()  

ZFNet

[python]  view plain  copy

1. #coding=utf-8  
2. from keras.models import Sequential  
3. from keras.layers import Dense,Flatten,Dropout  
4. from keras.layers.convolutional import Conv2D,MaxPooling2D  
5. from keras.utils.np_utils import to_categorical  
6. import numpy as np  
7. seed = 7  
8. np.random.seed(seed)  
9.   
10. model = Sequential()  
11. model.add(Conv2D(96,(7,7),strides=(2,2),input_shape=(224,224,3),padding='valid',activation='relu',kernel_initializer='uniform'))  
12. model.add(MaxPooling2D(pool_size=(3,3),strides=(2,2)))  
13. model.add(Conv2D(256,(5,5),strides=(2,2),padding='same',activation='relu',kernel_initializer='uniform'))  
14. model.add(MaxPooling2D(pool_size=(3,3),strides=(2,2)))  
15. model.add(Conv2D(384,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
16. model.add(Conv2D(384,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
17. model.add(Conv2D(256,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
18. model.add(MaxPooling2D(pool_size=(3,3),strides=(2,2)))  
19. model.add(Flatten())  
20. model.add(Dense(4096,activation='relu'))  
21. model.add(Dropout(0.5))  
22. model.add(Dense(4096,activation='relu'))  
23. model.add(Dropout(0.5))  
24. model.add(Dense(1000,activation='softmax'))  
25. model.compile(loss='categorical_crossentropy',optimizer='sgd',metrics=['accuracy'])  
26. model.summary()  

VGG-13

[python]  view plain  copy

1. #coding=utf-8  
2. from keras.models import Sequential  
3. from keras.layers import Dense,Flatten,Dropout  
4. from keras.layers.convolutional import Conv2D,MaxPooling2D  
5. import numpy as np  
6. seed = 7  
7. np.random.seed(seed)  
8.   
9. model = Sequential()  
10. model.add(Conv2D(64,(3,3),strides=(1,1),input_shape=(224,224,3),padding='same',activation='relu',kernel_initializer='uniform'))  
11. model.add(Conv2D(64,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
12. model.add(MaxPooling2D(pool_size=(2,2)))  
13. model.add(Conv2D(128,(3,2),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
14. model.add(Conv2D(128,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
15. model.add(MaxPooling2D(pool_size=(2,2)))  
16. model.add(Conv2D(256,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
17. model.add(Conv2D(256,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
18. model.add(MaxPooling2D(pool_size=(2,2)))  
19. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
20. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
21. model.add(MaxPooling2D(pool_size=(2,2)))  
22. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
23. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
24. model.add(MaxPooling2D(pool_size=(2,2)))  
25. model.add(Flatten())  
26. model.add(Dense(4096,activation='relu'))  
27. model.add(Dropout(0.5))  
28. model.add(Dense(4096,activation='relu'))  
29. model.add(Dropout(0.5))  
30. model.add(Dense(1000,activation='softmax'))  
31. model.compile(loss='categorical_crossentropy',optimizer='sgd',metrics=['accuracy'])  
32. model.summary()  

VGG-16

[python]  view plain  copy

1. #coding=utf-8  
2. from keras.models import Sequential  
3. from keras.layers import Dense,Flatten,Dropout  
4. from keras.layers.convolutional import Conv2D,MaxPooling2D  
5. import numpy as np  
6. seed = 7  
7. np.random.seed(seed)  
8.   
9. model = Sequential()  
10. model.add(Conv2D(64,(3,3),strides=(1,1),input_shape=(224,224,3),padding='same',activation='relu',kernel_initializer='uniform'))  
11. model.add(Conv2D(64,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
12. model.add(MaxPooling2D(pool_size=(2,2)))  
13. model.add(Conv2D(128,(3,2),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
14. model.add(Conv2D(128,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
15. model.add(MaxPooling2D(pool_size=(2,2)))  
16. model.add(Conv2D(256,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
17. model.add(Conv2D(256,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
18. model.add(Conv2D(256,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
19. model.add(MaxPooling2D(pool_size=(2,2)))  
20. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
21. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
22. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
23. model.add(MaxPooling2D(pool_size=(2,2)))  
24. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
25. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
26. model.add(Conv2D(512,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer='uniform'))  
27. model.add(MaxPooling2D(pool_size=(2,2)))  
28. model.add(Flatten())  
29. model.add(Dense(4096,activation='relu'))  
30. model.add(Dropout(0.5))  
31. model.add(Dense(4096,activation='relu'))  
32. model.add(Dropout(0.5))  
33. model.add(Dense(1000,activation='softmax'))  
34. model.compile(loss='categorical_crossentropy',optimizer='sgd',metrics=['accuracy'])  
35. model.summary()  

GoogleNet

[python]  view plain  copy

1. #coding=utf-8  
2. from keras.models import Model  
3. from keras.layers import Input,Dense,Dropout,BatchNormalization,Conv2D,MaxPooling2D,AveragePooling2D,concatenate  
4. from keras.layers.convolutional import Conv2D,MaxPooling2D,AveragePooling2D  
5. import numpy as np  
6. seed = 7  
7. np.random.seed(seed)  
8.   
9. def Conv2d_BN(x, nb_filter,kernel_size, padding='same',strides=(1,1),name=None):  
10.     if name is not None:  
11.         bn_name = name + '_bn'  
12.         conv_name = name + '_conv'  
13.     else:  
14.         bn_name = None  
15.         conv_name = None  
16.   
17.     x = Conv2D(nb_filter,kernel_size,padding=padding,strides=strides,activation='relu',name=conv_name)(x)  
18.     x = BatchNormalization(axis=3,name=bn_name)(x)  
19.     return x  
20.   
21. def Inception(x,nb_filter):  
22.     branch1x1 = Conv2d_BN(x,nb_filter,(1,1), padding='same',strides=(1,1),name=None)  
23.   
24.     branch3x3 = Conv2d_BN(x,nb_filter,(1,1), padding='same',strides=(1,1),name=None)  
25.     branch3x3 = Conv2d_BN(branch3x3,nb_filter,(3,3), padding='same',strides=(1,1),name=None)  
26.   
27.     branch5x5 = Conv2d_BN(x,nb_filter,(1,1), padding='same',strides=(1,1),name=None)  
28.     branch5x5 = Conv2d_BN(branch5x5,nb_filter,(1,1), padding='same',strides=(1,1),name=None)  
29.   
30.     branchpool = MaxPooling2D(pool_size=(3,3),strides=(1,1),padding='same')(x)  
31.     branchpool = Conv2d_BN(branchpool,nb_filter,(1,1),padding='same',strides=(1,1),name=None)  
32.   
33.     x = concatenate([branch1x1,branch3x3,branch5x5,branchpool],axis=3)  
34.   
35.     return x  
36.   
37. inpt = Input(shape=(224,224,3))  
38. #padding = 'same'，填充为(步长-1）/2,还可以用ZeroPadding2D((3,3))  
39. x = Conv2d_BN(inpt,64,(7,7),strides=(2,2),padding='same')  
40. x = MaxPooling2D(pool_size=(3,3),strides=(2,2),padding='same')(x)  
41. x = Conv2d_BN(x,192,(3,3),strides=(1,1),padding='same')  
42. x = MaxPooling2D(pool_size=(3,3),strides=(2,2),padding='same')(x)  
43. x = Inception(x,64)#256  
44. x = Inception(x,120)#480  
45. x = MaxPooling2D(pool_size=(3,3),strides=(2,2),padding='same')(x)  
46. x = Inception(x,128)#512  
47. x = Inception(x,128)  
48. x = Inception(x,128)  
49. x = Inception(x,132)#528  
50. x = Inception(x,208)#832  
51. x = MaxPooling2D(pool_size=(3,3),strides=(2,2),padding='same')(x)  
52. x = Inception(x,208)  
53. x = Inception(x,256)#1024  
54. x = AveragePooling2D(pool_size=(7,7),strides=(7,7),padding='same')(x)  
55. x = Dropout(0.4)(x)  
56. x = Dense(1000,activation='relu')(x)  
57. x = Dense(1000,activation='softmax')(x)  
58. model = Model(inpt,x,name='inception')  
59. model.compile(loss='categorical_crossentropy',optimizer='sgd',metrics=['accuracy'])  
60. model.summary()  

Resnet-34

[python]  view plain  copy

1. #coding=utf-8  
2. from keras.models import Model  
3. from keras.layers import Input,Dense,Dropout,BatchNormalization,Conv2D,MaxPooling2D,AveragePooling2D,concatenate,Activation,ZeroPadding2D  
4. from keras.layers import add,Flatten  
5. #from keras.layers.convolutional import Conv2D,MaxPooling2D,AveragePooling2D  
6. import numpy as np  
7. seed = 7  
8. np.random.seed(seed)  
9.   
10. def Conv2d_BN(x, nb_filter,kernel_size, strides=(1,1), padding='same',name=None):  
11.     if name is not None:  
12.         bn_name = name + '_bn'  
13.         conv_name = name + '_conv'  
14.     else:  
15.         bn_name = None  
16.         conv_name = None  
17.   
18.     x = Conv2D(nb_filter,kernel_size,padding=padding,strides=strides,activation='relu',name=conv_name)(x)  
19.     x = BatchNormalization(axis=3,name=bn_name)(x)  
20.     return x  
21.   
22. def Conv_Block(inpt,nb_filter,kernel_size,strides=(1,1), with_conv_shortcut=False):  
23.     x = Conv2d_BN(inpt,nb_filter=nb_filter,kernel_size=kernel_size,strides=strides,padding='same')  
24.     x = Conv2d_BN(x, nb_filter=nb_filter, kernel_size=kernel_size,padding='same')  
25.     if with_conv_shortcut:  
26.         shortcut = Conv2d_BN(inpt,nb_filter=nb_filter,strides=strides,kernel_size=kernel_size)  
27.         x = add([x,shortcut])  
28.         return x  
29.     else:  
30.         x = add([x,inpt])  
31.         return x  
32.   
33. inpt = Input(shape=(224,224,3))  
34. x = ZeroPadding2D((3,3))(inpt)  
35. x = Conv2d_BN(x,nb_filter=64,kernel_size=(7,7),strides=(2,2),padding='valid')  
36. x = MaxPooling2D(pool_size=(3,3),strides=(2,2),padding='same')(x)  
37. #(56,56,64)  
38. x = Conv_Block(x,nb_filter=64,kernel_size=(3,3))  
39. x = Conv_Block(x,nb_filter=64,kernel_size=(3,3))  
40. x = Conv_Block(x,nb_filter=64,kernel_size=(3,3))  
41. #(28,28,128)  
42. x = Conv_Block(x,nb_filter=128,kernel_size=(3,3),strides=(2,2),with_conv_shortcut=True)  
43. x = Conv_Block(x,nb_filter=128,kernel_size=(3,3))  
44. x = Conv_Block(x,nb_filter=128,kernel_size=(3,3))  
45. x = Conv_Block(x,nb_filter=128,kernel_size=(3,3))  
46. #(14,14,256)  
47. x = Conv_Block(x,nb_filter=256,kernel_size=(3,3),strides=(2,2),with_conv_shortcut=True)  
48. x = Conv_Block(x,nb_filter=256,kernel_size=(3,3))  
49. x = Conv_Block(x,nb_filter=256,kernel_size=(3,3))  
50. x = Conv_Block(x,nb_filter=256,kernel_size=(3,3))  
51. x = Conv_Block(x,nb_filter=256,kernel_size=(3,3))  
52. x = Conv_Block(x,nb_filter=256,kernel_size=(3,3))  
53. #(7,7,512)  
54. x = Conv_Block(x,nb_filter=512,kernel_size=(3,3),strides=(2,2),with_conv_shortcut=True)  
55. x = Conv_Block(x,nb_filter=512,kernel_size=(3,3))  
56. x = Conv_Block(x,nb_filter=512,kernel_size=(3,3))  
57. x = AveragePooling2D(pool_size=(7,7))(x)  
58. x = Flatten()(x)  
59. x = Dense(1000,activation='softmax')(x)  
60.   
61. model = Model(inputs=inpt,outputs=x)  
62. model.compile(loss='categorical_crossentropy',optimizer='sgd',metrics=['accuracy'])  
63. model.summary()  

Resnet-50

[python]  view plain  copy

1. #coding=utf-8  
2. from keras.models import Model  
3. from keras.layers import Input,Dense,BatchNormalization,Conv2D,MaxPooling2D,AveragePooling2D,ZeroPadding2D  
4. from keras.layers import add,Flatten  
5. #from keras.layers.convolutional import Conv2D,MaxPooling2D,AveragePooling2D  
6. from keras.optimizers import SGD  
7. import numpy as np  
8. seed = 7  
9. np.random.seed(seed)  
10.   
11. def Conv2d_BN(x, nb_filter,kernel_size, strides=(1,1), padding='same',name=None):  
12.     if name is not None:  
13.         bn_name = name + '_bn'  
14.         conv_name = name + '_conv'  
15.     else:  
16.         bn_name = None  
17.         conv_name = None  
18.   
19.     x = Conv2D(nb_filter,kernel_size,padding=padding,strides=strides,activation='relu',name=conv_name)(x)  
20.     x = BatchNormalization(axis=3,name=bn_name)(x)  
21.     return x  
22.   
23. def Conv_Block(inpt,nb_filter,kernel_size,strides=(1,1), with_conv_shortcut=False):  
24.     x = Conv2d_BN(inpt,nb_filter=nb_filter[0],kernel_size=(1,1),strides=strides,padding='same')  
25.     x = Conv2d_BN(x, nb_filter=nb_filter[1], kernel_size=(3,3), padding='same')  
26.     x = Conv2d_BN(x, nb_filter=nb_filter[2], kernel_size=(1,1), padding='same')  
27.     if with_conv_shortcut:  
28.         shortcut = Conv2d_BN(inpt,nb_filter=nb_filter[2],strides=strides,kernel_size=kernel_size)  
29.         x = add([x,shortcut])  
30.         return x  
31.     else:  
32.         x = add([x,inpt])  
33.         return x  
34.   
35. inpt = Input(shape=(224,224,3))  
36. x = ZeroPadding2D((3,3))(inpt)  
37. x = Conv2d_BN(x,nb_filter=64,kernel_size=(7,7),strides=(2,2),padding='valid')  
38. x = MaxPooling2D(pool_size=(3,3),strides=(2,2),padding='same')(x)  
39.   
40. x = Conv_Block(x,nb_filter=[64,64,256],kernel_size=(3,3),strides=(1,1),with_conv_shortcut=True)  
41. x = Conv_Block(x,nb_filter=[64,64,256],kernel_size=(3,3))  
42. x = Conv_Block(x,nb_filter=[64,64,256],kernel_size=(3,3))  
43.   
44. x = Conv_Block(x,nb_filter=[128,128,512],kernel_size=(3,3),strides=(2,2),with_conv_shortcut=True)  
45. x = Conv_Block(x,nb_filter=[128,128,512],kernel_size=(3,3))  
46. x = Conv_Block(x,nb_filter=[128,128,512],kernel_size=(3,3))  
47. x = Conv_Block(x,nb_filter=[128,128,512],kernel_size=(3,3))  
48.   
49. x = Conv_Block(x,nb_filter=[256,256,1024],kernel_size=(3,3),strides=(2,2),with_conv_shortcut=True)  
50. x = Conv_Block(x,nb_filter=[256,256,1024],kernel_size=(3,3))  
51. x = Conv_Block(x,nb_filter=[256,256,1024],kernel_size=(3,3))  
52. x = Conv_Block(x,nb_filter=[256,256,1024],kernel_size=(3,3))  
53. x = Conv_Block(x,nb_filter=[256,256,1024],kernel_size=(3,3))  
54. x = Conv_Block(x,nb_filter=[256,256,1024],kernel_size=(3,3))  
55.   
56. x = Conv_Block(x,nb_filter=[512,512,2048],kernel_size=(3,3),strides=(2,2),with_conv_shortcut=True)  
57. x = Conv_Block(x,nb_filter=[512,512,2048],kernel_size=(3,3))  
58. x = Conv_Block(x,nb_filter=[512,512,2048],kernel_size=(3,3))  
59. x = AveragePooling2D(pool_size=(7,7))(x)  
60. x = Flatten()(x)  
61. x = Dense(1000,activation='softmax')(x)  
62.   
63. model = Model(inputs=inpt,outputs=x)  
64. sgd = SGD(decay=0.0001,momentum=0.9)  
65. model.compile(loss='categorical_crossentropy',optimizer=sgd,metrics=['accuracy'])  
66. model.summary()  
67.