-
Notifications
You must be signed in to change notification settings - Fork 3
/
stacked_ae.py
161 lines (127 loc) · 5.52 KB
/
stacked_ae.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
"""
Builds a stacked autoencoder.
Author(s): Wei Chen ([email protected])
"""
import numpy as np
from keras.models import Sequential
from keras.optimizers import Adagrad, SGD, Adadelta, Adam
from keras.regularizers import l2
from keras.layers import Input, Dense, noise
from keras.models import Model
from keras import backend as K
#from early_stopping import MyEarlyStopping
import ConfigParser
def save_decoder(model, ae_id, c):
# Get the file name
config = ConfigParser.ConfigParser()
config.read('config.ini')
source = config.get('Global', 'source')
noise_scale = config.getfloat('Global', 'noise_scale')
if source == 'sf':
alpha = config.getfloat('Superformula', 'nonlinearity')
beta = config.getint('Superformula', 'n_clusters')
sname = source + '-' + str(beta) + '-' + str(alpha)
elif source == 'glass' or source[:3] == 'sf-':
sname = source
fname = '%s_%.4f_%s_%d' % (sname, noise_scale, ae_id, c)
# Save model architecture and weights
json_string = model.to_json()
open('./trained_models/'+fname+'_architecture.json', 'w').write(json_string)
model.save_weights('./decoders/'+fname+'_weights.h5', overwrite=True)
def train_ae(data, feature_dim, hidden_sizes, l, p=0, batch_size=100, activation='tanh',
activity_regularizer=None, weights=None, nb_epoch=1000, loss='mse', verbose=False):
data_dim = data.shape[1]
inputs = Input(shape=(data_dim,))
sizes = [data_dim] + hidden_sizes + [feature_dim]
n_layers = len(sizes) - 1
# Encoder
x = noise.GaussianDropout(p)(inputs)
for i in range(n_layers):
x = Dense(sizes[i+1], activation=activation, W_regularizer=l2(l))(x)
# Decoder
for i in range(n_layers):
x = Dense(sizes[-i-2], activation=activation, W_regularizer=l2(l))(x)
decoded = x
model = Model(input=inputs, output=decoded)
if weights is not None:
model.set_weights(weights)
# optimizer = Adagrad(lr=lr, epsilon=epsilon)
optimizer = Adam()
model.compile(loss=loss, optimizer=optimizer)
# early_stopping = MyEarlyStopping(monitor='loss', patience=10, verbose=verbose, tol=1e-6)
model.fit(data, data, batch_size=batch_size, nb_epoch=nb_epoch, verbose=verbose)#, callbacks=[early_stopping])
if n_layers == 1:
W_en = model.layers[-2].get_weights()
W_de = model.layers[-1].get_weights()
else:
W_en = None
W_de = None
encode = K.function([model.layers[0].input, K.learning_phase()], [model.layers[-2].output])
a = encode([data, 0])[0] # hidden layer's activation
return a, W_en, W_de, model
def sae(data, c, feature_dim, train, test, hidden_size_l1=0, hidden_size_l2=0, hidden_size_l3=0, hidden_size_l4=0, p=0.3,
l=0, batch_size=100, evaluation=False, overwrite=True):
''' Select number of layers for autoencoder based on arguments
hidden_size_l1, hidden_size_l2, hidden_size_l3 and hidden_size_l4 '''
np.random.seed(0)
pre_training = False
verbose = 0
activation = 'tanh'
loss = 'mse'
nb_epoch = 5000 # maximum number of epochs
# p = 0.1 # dropout fraction for denoising autoencoders
if hidden_size_l1 == 0:
hidden_sizes = []
elif hidden_size_l2 == 0:
hidden_sizes = [hidden_size_l1]
elif hidden_size_l3 == 0:
hidden_sizes = [hidden_size_l1, hidden_size_l2]
elif hidden_size_l4 == 0:
hidden_sizes = [hidden_size_l1, hidden_size_l2, hidden_size_l3]
else:
hidden_sizes = [hidden_size_l1, hidden_size_l2, hidden_size_l3, hidden_size_l4]
data_dim = data.shape[1]
sizes = [data_dim] + hidden_sizes + [feature_dim]
n_layers = len(sizes) - 1
Ws = None
# Pre-training (greedy layer-wise training)
if pre_training:
Ws_en = []
Ws_de = []
a = data[train]
for i in range(n_layers):
if verbose:
print 'Pre-training for Layer %d ...' % (i+1)
a, W_en, W_de, _ = train_ae(a, sizes[i+1], [], l, p=p, batch_size=batch_size,
nb_epoch=nb_epoch, loss=loss, verbose=verbose)
Ws_en.append(W_en)
Ws_de.append(W_de)
Ws_de.reverse()
Ws = Ws_en + Ws_de
Ws = [item for sublist in Ws for item in sublist]
# Fine tuning
if verbose:
print 'Fine tuning ...'
_, _, _, model = train_ae(data[train], feature_dim, hidden_sizes, l, p=p, batch_size=batch_size,
nb_epoch=nb_epoch, loss=loss, verbose=verbose, weights=Ws)
if evaluation:
# Used for hyperparameter optimization
cost = model.evaluate(data[test], data[test], batch_size=len(test), verbose=verbose)
return cost
# Reconstruct using the decoder
decoder = Sequential()
for i in range(n_layers):
decoder.add(Dense(sizes[-i-2], input_dim=sizes[-i-1], activation=activation,
weights=model.layers[-n_layers+i].get_weights()))
decoder.compile(loss='mse', optimizer='sgd')
if p > 0:
name = 'SDAE-'+str(n_layers)
else:
name = 'SAE-'+str(n_layers)
if overwrite:
# Save the decoder
save_decoder(decoder, name, c)
encode = K.function([model.layers[0].input, K.learning_phase()], [model.layers[-n_layers-1].output])
features = np.zeros((data.shape[0],feature_dim))
features[train+test] = encode([data[train+test], 0])[0]
return features, name, decoder.predict