1
|
|
|
""" |
2
|
|
|
This script describes how to save time during the optimization by |
3
|
|
|
using a pretrained model. It is similar to the transer learning example, |
4
|
|
|
but here you do the training and model creation of the pretrained model |
5
|
|
|
yourself. |
6
|
|
|
|
7
|
|
|
The problem is that most of the optimization time is "waisted" by |
8
|
|
|
training the model. The time to find a new position to explore by |
9
|
|
|
Hyperactive is very small compared to the training time of |
10
|
|
|
neural networks. This means, that we can do more optimization |
11
|
|
|
if we keep the training time as little as possible. |
12
|
|
|
|
13
|
|
|
The idea of pretrained neural architecture search is to pretrain a complete model one time. |
14
|
|
|
In the next step we remove the layers that should be optimized |
15
|
|
|
and make the remaining layers not-trainable. |
16
|
|
|
|
17
|
|
|
This results in a partial, pretrained, not-trainable model that will be |
18
|
|
|
used during the Hyperactive optimization. |
19
|
|
|
|
20
|
|
|
You can now add layers to the partial model in the objective function |
21
|
|
|
and add the parameters or layers that will be optimized by Hyperactive. |
22
|
|
|
|
23
|
|
|
With each iteration of the optimization run we are only training |
24
|
|
|
the added layers of the model. This saves a lot of training time. |
25
|
|
|
|
26
|
|
|
""" |
27
|
|
|
|
28
|
|
|
|
29
|
|
|
import numpy as np |
30
|
|
|
from keras.models import Sequential |
31
|
|
|
from keras.layers import ( |
32
|
|
|
Dense, |
33
|
|
|
Conv2D, |
34
|
|
|
MaxPooling2D, |
35
|
|
|
Flatten, |
36
|
|
|
Activation, |
37
|
|
|
Dropout, |
38
|
|
|
) |
39
|
|
|
from keras.datasets import cifar10 |
40
|
|
|
from keras.utils import to_categorical |
41
|
|
|
|
42
|
|
|
from hyperactive import Hyperactive |
43
|
|
|
|
44
|
|
|
(X_train, y_train), (X_test, y_test) = cifar10.load_data() |
45
|
|
|
|
46
|
|
|
y_train = to_categorical(y_train, 10) |
47
|
|
|
y_test = to_categorical(y_test, 10) |
48
|
|
|
|
49
|
|
|
# to make the example quick |
50
|
|
|
X_train = X_train[0:1000] |
51
|
|
|
y_train = y_train[0:1000] |
52
|
|
|
|
53
|
|
|
X_test = X_test[0:1000] |
54
|
|
|
y_test = y_test[0:1000] |
55
|
|
|
|
56
|
|
|
|
57
|
|
|
# create model and train it |
58
|
|
|
model_pretrained = Sequential() |
59
|
|
|
model_pretrained.add(Conv2D(64, (3, 3), padding="same", input_shape=X_train.shape[1:])) |
60
|
|
|
model_pretrained.add(Activation("relu")) |
61
|
|
|
model_pretrained.add(Conv2D(32, (3, 3))) |
62
|
|
|
model_pretrained.add(Activation("relu")) |
63
|
|
|
model_pretrained.add(MaxPooling2D(pool_size=(2, 2))) |
64
|
|
|
model_pretrained.add(Dropout(0.25)) |
65
|
|
|
|
66
|
|
|
model_pretrained.add(Conv2D(32, (3, 3), padding="same")) |
67
|
|
|
model_pretrained.add(Activation("relu")) |
68
|
|
|
model_pretrained.add(Dropout(0.25)) |
69
|
|
|
model_pretrained.add(Flatten()) |
70
|
|
|
model_pretrained.add(Dense(200)) |
71
|
|
|
model_pretrained.add(Activation("relu")) |
72
|
|
|
model_pretrained.add(Dropout(0.5)) |
73
|
|
|
model_pretrained.add(Dense(10)) |
74
|
|
|
model_pretrained.add(Activation("softmax")) |
75
|
|
|
|
76
|
|
|
model_pretrained.compile( |
77
|
|
|
optimizer="adam", loss="categorical_crossentropy", metrics=["accuracy"] |
78
|
|
|
) |
79
|
|
|
model_pretrained.fit(X_train, y_train, epochs=5, batch_size=256) |
80
|
|
|
|
81
|
|
|
|
82
|
|
|
def cnn(opt): |
83
|
|
|
model = model_pretrained |
84
|
|
|
n_layers = len(model.layers) |
85
|
|
|
|
86
|
|
|
# delete the last 9 layers |
87
|
|
|
for i in range(n_layers - 9): |
88
|
|
|
model.pop() |
89
|
|
|
|
90
|
|
|
# set remaining layers to not-trainable |
91
|
|
|
for layer in model.layers: |
92
|
|
|
layer.trainable = False |
93
|
|
|
|
94
|
|
|
model = opt["conv_layer.0"](model) |
95
|
|
|
model.add(Dropout(0.25)) |
96
|
|
|
|
97
|
|
|
model.add(Flatten()) |
98
|
|
|
model.add(Dense(opt["neurons.0"])) |
99
|
|
|
model.add(Activation("relu")) |
100
|
|
|
model.add(Dropout(0.5)) |
101
|
|
|
model.add(Dense(10)) |
102
|
|
|
model.add(Activation("softmax")) |
103
|
|
|
|
104
|
|
|
model.compile( |
105
|
|
|
optimizer="adam", loss="categorical_crossentropy", metrics=["accuracy"] |
106
|
|
|
) |
107
|
|
|
model.fit(X_train, y_train, epochs=5, batch_size=256) |
108
|
|
|
|
109
|
|
|
_, score = model.evaluate(x=X_test, y=y_test) |
110
|
|
|
|
111
|
|
|
return score |
112
|
|
|
|
113
|
|
|
|
114
|
|
|
# conv 1, 2, 3 are functions that adds layers. We want to know which function is the best |
115
|
|
|
def conv1(model): |
116
|
|
|
model.add(Conv2D(64, (3, 3))) |
117
|
|
|
model.add(Activation("relu")) |
118
|
|
|
model.add(MaxPooling2D(pool_size=(2, 2))) |
119
|
|
|
return model |
120
|
|
|
|
121
|
|
|
|
122
|
|
|
def conv2(model): |
123
|
|
|
model.add(Conv2D(64, (3, 3))) |
124
|
|
|
model.add(Activation("relu")) |
125
|
|
|
return model |
126
|
|
|
|
127
|
|
|
|
128
|
|
|
def conv3(model): |
129
|
|
|
return model |
130
|
|
|
|
131
|
|
|
|
132
|
|
|
search_space = { |
133
|
|
|
"conv_layer.0": [conv1, conv2, conv3], |
134
|
|
|
"neurons.0": list(range(100, 1000, 100)), |
135
|
|
|
} |
136
|
|
|
|
137
|
|
|
|
138
|
|
|
hyper = Hyperactive() |
139
|
|
|
hyper.add_search(cnn, search_space, n_iter=5) |
140
|
|
|
hyper.run() |
141
|
|
|
|