-
Notifications
You must be signed in to change notification settings - Fork 2
/
tsae.py
163 lines (123 loc) · 6.88 KB
/
tsae.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
import numpy as np
import tensorflow as tf
from tensorflow.contrib.layers import fully_connected
import logging
class DataFeeder(object):
def get_random_block_from_data(self, data, batch_size):
start_index = np.random.randint(0, len(data) - batch_size)
return data[start_index:(start_index + batch_size)]
class TSAutoEncoder(object):
# This is to avoid tf var conflicts when we create many TSAutoEncoder instances
i = 0
def __init__(self, config, feeder, session=None):
self.config = config
self.num_features = len(self.config.feature_sizes)
self.feature_dim = sum(self.config.feature_sizes)
self.feeder = feeder
self._build_graph()
self._setup_tf(session)
TSAutoEncoder.i += 1
def _build_graph(self):
with tf.variable_scope('TSAE_%d' % TSAutoEncoder.i):
self._build_input()
self._build_layers()
self._build_train_op()
self._build_summary_op()
def _build_input(self):
with tf.variable_scope('Input'):
self.inputs = tf.placeholder(self._data_type(), [None, self.config.num_steps, self.num_features])
self.keep_prob = tf.placeholder(self._data_type())
self.inputs = tf.nn.dropout(self.inputs, self.keep_prob)
def _build_layers(self):
raise NotImplementedError('Must be defined by subclasses')
def _build_train_op(self):
self._lr = tf.Variable(0.0, trainable=False)
tf.summary.scalar("learning_rate", self._lr)
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(self._cost, tvars), self.config.max_grad_norm)
optimizer = tf.train.RMSPropOptimizer(self._lr)
self._global_step = tf.contrib.framework.get_or_create_global_step()
self._train_op = optimizer.apply_gradients(zip(grads, tvars),
global_step=self._global_step)
self._new_lr = tf.placeholder(tf.float32, shape=[], name="new_learning_rate")
self._lr_update = tf.assign(self._lr, self._new_lr)
def assign_lr(self, session, lr_value):
session.run(self._lr_update, feed_dict={self._new_lr: lr_value})
def _build_summary_op(self):
self._merged = tf.summary.merge_all()
def _setup_tf(self, session):
# This is for logging down the summaries for displaying in TensorBoard
self._writer = tf.summary.FileWriter(self.config.log_dir, graph=tf.get_default_graph())
# This is for checkpointing epoches
self._saver = tf.train.Saver(tf.global_variables())
gpu_options = tf.GPUOptions(allow_growth=True)
self.session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) if session is None else session
self.session.run(tf.global_variables_initializer())
def _data_type(self):
return tf.float32
def _run_epoch(self, X):
cost, summary, global_step, _ = self.session.run((self._cost, self._merged, self._global_step, self._train_op),
feed_dict={self.inputs: X,
self.keep_prob: self.config.keep_prob})
self._writer.add_summary(summary, global_step)
return cost
def fit(self, X, display_step=1):
n_samples = len(X)
for epoch in range(self.config.max_max_epoch):
lr_decay = self.config.lr_decay ** max(epoch - self.config.max_epoch, 0.0)
logging.debug("lr_decay = %s" % lr_decay)
self.assign_lr(self.session, self.config.learning_rate * lr_decay)
avg_cost = 0.
total_batch = int(n_samples / self.config.batch_size)
# Loop over batches
for i in range(total_batch):
batch_xs = self.feeder.get_random_block_from_data(X, self.config.batch_size)
cost = self._run_epoch(batch_xs)
avg_cost += cost / total_batch
if epoch % display_step == 0:
print("Epoch:", '%04d' % (epoch + 1), "cost=", "{:.9f}".format(avg_cost))
# Save a checkpoint
self._saver.save(self.session, self.config.log_dir + '/checkpoints-%d-%f' % (epoch, avg_cost))
def calc_total_cost(self, X):
return self.session.run(self._cost, feed_dict={self.inputs: X, self.keep_prob: 1.0})
def transform(self, X):
return self.session.run(self._final_state, feed_dict={self.inputs: X, self.keep_prob: 1.0})
def reconstruct(self, X):
return self.session.run(self.reconstruction, feed_dict={self.inputs: X, self.keep_prob: 1.0})
class SimpleTSAutoEncoder(TSAutoEncoder):
def _build_layers(self):
with tf.variable_scope('RNN'):
cell = tf.contrib.rnn.BasicLSTMCell(self.config.hidden_size, state_is_tuple=True)
cell = tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=self.keep_prob)
self.stacked_lstm = cell = tf.contrib.rnn.MultiRNNCell([cell] * self.config.num_layers,
state_is_tuple=True)
inputs = tf.unstack(self.inputs, num=self.config.num_steps, axis=1)
# outputs = [Tensor[batch_size, hidden_size]] of length num_steps
outputs, state = tf.contrib.rnn.static_rnn(cell, inputs, dtype=self._data_type())
with tf.variable_scope('Loss'):
loss = tf.zeros(shape=[])
output = tf.transpose(tf.stack(outputs, axis=0), [1, 0, 2]) # [batch_size, num_steps, hidden_size]
feature_sizes = self.config.feature_sizes
# loop over features and cont
feature_index = 0
features = []
for feature_size in feature_sizes:
feature_output = fully_connected(inputs=output,
num_outputs=feature_size,
activation_fn=tf.nn.relu,
biases_initializer=tf.constant_initializer(0.1),
weights_initializer=tf.contrib.layers.xavier_initializer()
)
feature_output = tf.reshape(feature_output, [-1, self.config.num_steps, feature_size])
features.append(feature_output)
feature_target = tf.slice(self.inputs, [0, 0, feature_index], [-1, -1, 1])
if feature_size == 1:
loss += tf.nn.l2_loss(feature_output - feature_target)
else:
# Categorical features
loss += tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(feature_output, feature_target))
feature_index += feature_size
self.reconstruction = tf.stack(features, axis=2)
self._cost = loss
tf.summary.scalar("loss", loss)
self._final_state = state