|
1
|
|
|
from keras.models import Sequential |
|
2
|
|
|
from keras.layers import Dense, Conv2D, MaxPooling2D, Flatten, Activation, Dropout |
|
3
|
|
|
from keras.datasets import cifar10 |
|
4
|
|
|
from keras.utils import to_categorical |
|
5
|
|
|
|
|
6
|
|
|
from hyperactive import Hyperactive |
|
7
|
|
|
|
|
8
|
|
|
(X_train, y_train), (X_test, y_test) = cifar10.load_data() |
|
9
|
|
|
|
|
10
|
|
|
y_train = to_categorical(y_train, 10) |
|
11
|
|
|
y_test = to_categorical(y_test, 10) |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
""" |
|
15
|
|
|
Efficient Neural Architecture Search via Parameter Sharing: |
|
16
|
|
|
https://arxiv.org/pdf/1802.03268.pdf |
|
17
|
|
|
""" |
|
18
|
|
|
|
|
19
|
|
|
|
|
20
|
|
|
def conv1(model): |
|
21
|
|
|
model.add(Conv2D(32, (3, 3))) |
|
22
|
|
|
model.add(Activation("relu")) |
|
23
|
|
|
model.add(MaxPooling2D(pool_size=(2, 2))) |
|
24
|
|
|
return model |
|
25
|
|
|
|
|
26
|
|
|
|
|
27
|
|
|
def conv2(model): |
|
28
|
|
|
model.add(Conv2D(32, (3, 3))) |
|
29
|
|
|
model.add(Activation("relu")) |
|
30
|
|
|
return model |
|
31
|
|
|
|
|
32
|
|
|
|
|
33
|
|
|
def conv3(model): |
|
34
|
|
|
return model |
|
35
|
|
|
|
|
36
|
|
|
|
|
37
|
|
|
model_pretrained = Sequential() |
|
38
|
|
|
model_pretrained.add(Conv2D(64, (3, 3), padding="same", input_shape=X_train.shape[1:])) |
|
39
|
|
|
model_pretrained.add(Activation("relu")) |
|
40
|
|
|
model_pretrained.add(Conv2D(32, (3, 3))) |
|
41
|
|
|
model_pretrained.add(Activation("relu")) |
|
42
|
|
|
model_pretrained.add(MaxPooling2D(pool_size=(2, 2))) |
|
43
|
|
|
model_pretrained.add(Dropout(0.25)) |
|
44
|
|
|
|
|
45
|
|
|
model_pretrained.add(Conv2D(32, (3, 3), padding="same")) |
|
46
|
|
|
model_pretrained.add(Activation("relu")) |
|
47
|
|
|
model_pretrained.add(Dropout(0.25)) |
|
48
|
|
|
|
|
49
|
|
|
model_pretrained.add(Flatten()) |
|
50
|
|
|
model_pretrained.add(Dense(200)) |
|
51
|
|
|
model_pretrained.add(Activation("relu")) |
|
52
|
|
|
model_pretrained.add(Dropout(0.5)) |
|
53
|
|
|
model_pretrained.add(Dense(10)) |
|
54
|
|
|
model_pretrained.add(Activation("softmax")) |
|
55
|
|
|
|
|
56
|
|
|
model_pretrained.compile( |
|
57
|
|
|
optimizer="adam", loss="categorical_crossentropy", metrics=["accuracy"] |
|
58
|
|
|
) |
|
59
|
|
|
model_pretrained.fit(X_train, y_train, epochs=50, batch_size=128) |
|
60
|
|
|
|
|
61
|
|
|
n_layers = len(model_pretrained.layers) |
|
62
|
|
|
|
|
63
|
|
|
for i in range(n_layers - 8): |
|
64
|
|
|
model_pretrained.pop() |
|
65
|
|
|
|
|
66
|
|
|
for layer in model_pretrained.layers: |
|
67
|
|
|
layer.trainable = False |
|
68
|
|
|
|
|
69
|
|
|
print(model_pretrained.summary()) |
|
70
|
|
|
|
|
71
|
|
|
|
|
72
|
|
View Code Duplication |
def cnn(para, X_train, y_train): |
|
|
|
|
|
|
73
|
|
|
""" |
|
74
|
|
|
model = Sequential() |
|
75
|
|
|
model.add( |
|
76
|
|
|
Conv2D(64, (3, 3), padding="same", input_shape=X_train.shape[1:]) |
|
77
|
|
|
) |
|
78
|
|
|
model.add(Activation("relu")) |
|
79
|
|
|
model.add(Conv2D(32, (3, 3))) |
|
80
|
|
|
model.add(Activation("relu")) |
|
81
|
|
|
|
|
82
|
|
|
|
|
83
|
|
|
model.add(MaxPooling2D(pool_size=(2, 2))) |
|
84
|
|
|
model.add(Dropout(0.25)) |
|
85
|
|
|
|
|
86
|
|
|
model.add(Conv2D(32, (3, 3), padding="same")) |
|
87
|
|
|
model.add(Activation("relu")) |
|
88
|
|
|
""" |
|
89
|
|
|
model = model_pretrained |
|
90
|
|
|
|
|
91
|
|
|
model = para["conv_layer.0"](model) |
|
92
|
|
|
model.add(Dropout(0.25)) |
|
93
|
|
|
|
|
94
|
|
|
model.add(Flatten()) |
|
95
|
|
|
model.add(Dense(para["neurons.0"])) |
|
96
|
|
|
model.add(Activation("relu")) |
|
97
|
|
|
model.add(Dropout(0.5)) |
|
98
|
|
|
model.add(Dense(10)) |
|
99
|
|
|
model.add(Activation("softmax")) |
|
100
|
|
|
|
|
101
|
|
|
model.compile( |
|
102
|
|
|
optimizer="adam", loss="categorical_crossentropy", metrics=["accuracy"] |
|
103
|
|
|
) |
|
104
|
|
|
model.fit(X_train, y_train, epochs=25, batch_size=128) |
|
105
|
|
|
|
|
106
|
|
|
_, score = model.evaluate(x=X_test, y=y_test) |
|
107
|
|
|
|
|
108
|
|
|
return score |
|
109
|
|
|
|
|
110
|
|
|
|
|
111
|
|
|
search_config = { |
|
112
|
|
|
cnn: {"conv_layer.0": [conv1, conv2, conv3], "neurons.0": range(100, 1000, 100)} |
|
113
|
|
|
} |
|
114
|
|
|
|
|
115
|
|
|
opt = Hyperactive(X_train, y_train) |
|
116
|
|
|
opt.search(search_config, n_iter=5) |
|
117
|
|
|
|
SimonBlanke /
Hyperactive
Loading history...