forked from jramapuram/LSTM_Anomaly_Detector
-
Notifications
You must be signed in to change notification settings - Fork 0
/
autoencoder.py
216 lines (180 loc) · 9.12 KB
/
autoencoder.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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
__author__ = 'jramapuram'
import os.path
import numpy as np
from data_manipulator import elementwise_square, roll_rows
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, AutoEncoder, Activation
from keras.layers.recurrent import LSTM, GRU
from keras.regularizers import l2
from convolutional import Convolution1D, MaxPooling1D
class TimeDistributedAutoEncoder:
def __init__(self, conf):
self.conf = conf
self.model_dir = ''
self.model_name = ''
self.encoder_sizes = []
self.decoder_sizes = []
self.models = []
self.compiled = False
@property
def get_model_name(self):
if not self.compiled:
raise Exception("Cannot determine model name without it first being compiled");
model_structure = 'weights_[%s]Enc_[%s]Dec_%dbatch_%s_autoencoder.dat'
model_name = model_structure % ('_'.join(str(e) for e in self.encoder_sizes)
, '_'.join(str(d) for d in self.decoder_sizes)
, int(self.conf['--batch_size'])
, self.conf['--model_type'])
model_dir = model_name.replace("weights_", "").replace(".dat", "")
from data_manipulator import create_dir
create_dir(model_dir)
return model_dir, model_name
def compile(self, optimizer=None):
for model in self.models:
print model.get_config(verbose=1)
if optimizer is not None:
model.compile(loss=self.conf['--loss'], optimizer=optimizer)
else:
model.compile(loss=self.conf['--loss'], optimizer=self.conf['--optimizer'])
self.compiled = True
def add_autoencoder(self, encoder_sizes=[], decoder_sizes=[]):
assert(len(encoder_sizes) != 0 and len(decoder_sizes) != 0)
assert(len(encoder_sizes) == len(decoder_sizes))
self.encoder_sizes = encoder_sizes
self.decoder_sizes = decoder_sizes
# self.models = [Sequential() for i in range(len(encoder_sizes))]
self.models = [Sequential()]
encoders = Sequential()
decoders = Sequential()
for i in range(0, len(encoder_sizes) - 1):
encoders.add(Dense(encoder_sizes[i], encoder_sizes[i + 1]
, init=self.conf['--initialization']
, activation=self.conf['--activation']
, W_regularizer=l2()))
decoders.add(Dense(decoder_sizes[i], decoder_sizes[i + 1]
, init=self.conf['--initialization']
, activation=self.conf['--activation']
, W_regularizer=l2()))
self.models[0].add(AutoEncoder(encoder=encoders
, decoder=decoders
, output_reconstruction=(i == 0)))
return self.models
# TODO: This doesnt work yet
# (batch size, stack size, nb row, nb col)
def add_conv_autoencoder(self, encoder_sizes=[], decoder_sizes=[]):
assert(len(encoder_sizes) != 0 and len(decoder_sizes) != 0)
assert(len(encoder_sizes) == len(decoder_sizes))
self.encoder_sizes = encoder_sizes
self.decoder_sizes = decoder_sizes
# self.models = [Sequential() for i in range(len(encoder_sizes))]
self.models = [Sequential()]
encoders = Sequential()
decoders = Sequential()
for i in range(0, len(encoder_sizes) - 1):
encoders.add(Convolution1D(32, 3, 3
, activation=self.conf['--activation']
, init=self.conf['--initialization']
, border_mode='valid'))
encoders.add(Activation('relu'))
encoders.add(MaxPooling1D())
encoders.add(Convolution1D(32, 1, 1
, activation=self.conf['--activation']
, init=self.conf['--initialization']
, border_mode='valid'))
decoders.add(Convolution1D(32, 1, 1
, activation=self.conf['--activation']
, init=self.conf['--initialization']
, border_mode='valid'))
decoders.add(Activation('relu'))
decoders.add(MaxPooling1D())
self.models[0].add(AutoEncoder(encoder=encoders
, decoder=decoders
, output_reconstruction=(i == 0)))
return self.models
def add_lstm_autoencoder(self, encoder_sizes=[], decoder_sizes=[]):
assert(len(encoder_sizes) != 0 and len(decoder_sizes) != 0)
assert(len(encoder_sizes) == len(decoder_sizes))
self.encoder_sizes = encoder_sizes
self.decoder_sizes = decoder_sizes
# self.models = [Sequential() for i in range(len(encoder_sizes))]
self.models = [Sequential()]
encoders = Sequential()
decoders = Sequential()
for i in range(0, len(encoder_sizes) - 1):
encoders.add(LSTM(encoder_sizes[i], encoder_sizes[i + 1]
, activation=self.conf['--activation']
, inner_activation=self.conf['--inner_activation']
, init=self.conf['--initialization']
, inner_init=self.conf['--inner_init']
, truncate_gradient=int(self.conf['--truncated_gradient'])
, return_sequences=True))
decoders.add(LSTM(decoder_sizes[i], decoder_sizes[i + 1]
, activation=self.conf['--activation']
, inner_activation=self.conf['--inner_activation']
, init=self.conf['--initialization']
, inner_init=self.conf['--inner_init']
, truncate_gradient=int(int(self.conf['--truncated_gradient']))
, return_sequences=not (i == len(encoder_sizes) - 1)))
self.models[0].add(AutoEncoder(encoder=encoders
, decoder=decoders
, output_reconstruction=(i == 0)))
return self.models
def format_lstm_data(self, x):
# Need to create a 3d vector [samples, timesteps, input_dim]
if self.conf['--model_type'].strip().lower() == 'lstm':
x = x[:, np.newaxis, :]
print 'modified training data to fit LSTM: ', x.shape
return x
def unformat_lstm_data(self, x):
# Need to create a 2d vector [samples, input_dim]
if self.conf['--model_type'].strip().lower() == 'lstm':
x = x[:, np.newaxis, :]
return x
@staticmethod
def softmax(x):
e_x = np.exp(x - np.max(x))
out = e_x / e_x.sum()
return out
@staticmethod
def sigmoid(x):
return 1 / (1 + np.exp(-x))
def train_and_predict(self, x):
self.compile()
self.model_dir, self.model_name = self.get_model_name
_ = self.load_model(os.path.join(self.model_dir, self.model_name), self.models[0])
x = self.format_lstm_data(x)
predictions = []
# Keras LSTMs are NOT stateful and thus need to be trained in batch
if self.conf['--model_type'].strip().lower() == 'lstm':
self.models[0].train_on_batch(x, roll_rows(x, -1) , accuracy=True)
for i in xrange(x.shape[0] - 1):
if self.conf['--model_type'].strip().lower() == 'lstm':
predictions.append(self.models[0].predict(x[i:i+1, :, :], verbose=False))
#self.models[0].train_on_batch(x[i:i+1, :, :], x[i:i+1, :, :], accuracy=True)
else:
predictions.append(self.models[0].predict(x[i:i+1, :], verbose=False))
self.models[0].train_on_batch(x[i:i+1, :], x[i:i+1, :], accuracy=True)
predictions.append(predictions[-1]) # XXX
print 'saving model to %s...' % os.path.join(self.model_dir, self.model_name)
self.models[0].save_weights(os.path.join(self.model_dir, self.model_name), overwrite=True)
predictions = np.array(predictions)
if len(predictions.shape) > 3:
predictions = np.squeeze(np.squeeze(np.array(predictions), (1,)), (1,))
elif len(predictions.shape) == 3:
predictions = np.squeeze(np.array(predictions), (1,))
print 'predictions.shape: ', predictions.shape
np.savetxt(os.path.join(self.model_dir, 'outputs.csv'), predictions, delimiter=',')
return predictions
def get_model(self):
return self.models
def get_model_type(self):
return self.conf['--model_type']
@staticmethod
def load_model(path_str, model):
if os.path.isfile(path_str):
print 'model found, loading existing model...'
model.load_weights(path_str)
return True
else:
print 'model does not exist...'
return False