-
Notifications
You must be signed in to change notification settings - Fork 0
/
ar_predictor.py
223 lines (179 loc) · 8.6 KB
/
ar_predictor.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
217
218
219
220
221
# -*- coding: utf-8 -*-
from __future__ import division, print_function, absolute_import
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
import os
import time
import signal_provider
from signal_provider import SignalProvider
from hyper_parameter import HyperParameter
def get_weight_variable(input_dim, output_dim, regularizer):
shape = (input_dim, output_dim)
# weights = tf.get_variable("weights", shape, initializer=tf.truncated_normal_initializer(stddev=0.1))
weights = tf.get_variable("weights", shape, initializer=tf.contrib.layers.xavier_initializer())
if regularizer is not None:
tf.add_to_collection('reg_loss', regularizer(weights))
return weights
class ARSignalPredictor(object):
def __init__(self, params):
self._batch_size = params.ar_batch_size
self._input_steps = params.ar_input_steps
self._predict_steps = params.ar_predict_steps
self._input_depth = params.ar_input_depth
self._predict_depth = params.ar_predict_depth
self._train_epoch = params.ar_train_epoch
self._model_dir = params.ar_model_dir
if not os.path.exists(self._model_dir):
os.mkdir(self._model_dir)
self._data_provider = SignalProvider(self._batch_size,
input_steps=self._input_steps, predict_steps=self._predict_steps)
self._input_X = tf.placeholder(tf.float32, [self._batch_size, self._input_steps, self._input_depth])
self._truth_Y = tf.placeholder(tf.float32, [self._batch_size, self._predict_steps, self._predict_depth])
self._X = tf.placeholder(tf.float32, [None, self._input_steps, self._input_depth])
with tf.name_scope("train"):
with tf.variable_scope("ar", reuse=None):
self._trained_Y, self._loss, self._train_step = self._graph(self._input_X, self._truth_Y, True)
with tf.name_scope("eval"):
with tf.variable_scope("ar", reuse=True):
self._predict = self._graph(self._X, None, False)
self._sess = tf.InteractiveSession()
def _graph(self, inputX, truthY, is_train):
regularizer = tf.contrib.layers.l2_regularizer(0.0001)
input_dim = self._input_steps * self._input_depth
output_dim = self._predict_steps * self._predict_depth
weights = get_weight_variable(input_dim, output_dim, regularizer)
targets = tf.reshape(inputX, [-1, input_dim])
biases = tf.get_variable("biases", [self._predict_steps], initializer=tf.constant_initializer(0.0))
targets = tf.matmul(targets, weights) + biases
targets = tf.reshape(targets, [-1, self._predict_steps, self._predict_depth])
if is_train:
loss = tf.losses.mean_squared_error(targets, truthY)
# loss = tf.losses.absolute_difference(targets, truthY)
loss = loss + tf.add_n(tf.get_collection('reg_loss'))
train_step = tf.train.AdamOptimizer().minimize(loss)
# train_step = tf.train.RMSPropOptimizer(1e-3).minimize(loss)
return targets, loss, train_step
else:
return targets
def train(self):
saver = tf.train.Saver()
# start training
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=self._sess, coord=coord)
time_s = time.clock()
epoch = 0
i = 0
loss_avg = 0
while epoch < self._train_epoch:
x, y, nid = self._data_provider.get_next_batch()
p, loss, _ = self._sess.run([self._trained_Y, self._loss, self._train_step],
feed_dict={self._input_X: x, self._truth_Y: y})
loss_avg += loss
i += 1
if nid == 0:
epoch += 1
print("[INFO] after {} training steps, loss is {}, "
"time elapse {}".format(epoch, loss_avg / i, time.clock() - time_s))
i = 0
loss_avg = 0
time_s = time.clock()
saver.save(self._sess, os.path.join(self._model_dir, "ar_input{}.ckpt".format(
self._input_steps
)))
coord.request_stop()
coord.join(threads)
def evaluate_pn(self):
init_data = self._data_provider.evaluate_dat()
init_len = init_data.shape[1]
assert init_len >= self._input_steps
predicted = init_data
predicted_len = self._data_provider.evaluate_length()
while len(predicted[0]) < init_len + predicted_len:
x = predicted[:, len(predicted[0]) - self._input_steps:]
x = (x - self._data_provider.norm_mean) / self._data_provider.norm_std
y = self._sess.run(self._predict,
feed_dict={self._X: x})
y = y*self._data_provider.norm_std + self._data_provider.norm_mean
predicted = np.concatenate((predicted, y), axis=1)
predicted = predicted[:, init_len: init_len + predicted_len]
return self._data_provider.evaluate(predicted)
def evaluate_p1(self):
init_data = self._data_provider.evaluate_dat_v2()
init_len = init_data.shape[2]
assert init_len >= self._input_steps
predicted = []
for i in range(init_data.shape[1]):
x = init_data[:, i, init_len-self._input_steps:]
x = (x - self._data_provider.norm_mean) / self._data_provider.norm_std
y = self._sess.run(self._predict,
feed_dict={self._X: x})
y = y[:, 0]*self._data_provider.norm_std + self._data_provider.norm_mean
predicted.append(y)
predicted = np.stack(predicted, axis=1) # np.array(predicted, dtype=np.float32)
return self._data_provider.evaluate(predicted)
class ARSignalPredictorTF(object):
def __init__(self, params):
self._batch_size = params.ar_batch_size
self._input_steps = params.ar_input_steps
self._predict_steps = params.ar_predict_steps
self._s1, self._s2 = signal_provider.generate_signal()
self.prep_s = self._s2[:100]
self.eval_s = self._s2[100:]
x = np.array(range(len(self._s1)), dtype=np.float32)
data = {
tf.contrib.timeseries.TrainEvalFeatures.TIMES: x,
tf.contrib.timeseries.TrainEvalFeatures.VALUES: self._s1,
}
reader = tf.contrib.timeseries.NumpyReader(data)
self.train_input_fn = tf.contrib.timeseries.RandomWindowInputFn(
reader, batch_size=self._batch_size, window_size=self._input_steps+self._predict_steps)
self.ar = tf.contrib.timeseries.ARRegressor(
periodicities=628, input_window_size=self._input_steps, output_window_size=self._predict_steps,
num_features=1,
loss=tf.contrib.timeseries.ARModel.NORMAL_LIKELIHOOD_LOSS)
def train(self):
self.ar.train(input_fn=self.train_input_fn, steps=6000)
def evaluate(self):
# predicted = self.prep_s
# while len(predicted) < len(self._s2):
x = self._s1 # predicted[len(predicted) - self._input_steps:]
t = np.array(range(len(x)), dtype=np.float32)
data = {
tf.contrib.timeseries.TrainEvalFeatures.TIMES: t,
tf.contrib.timeseries.TrainEvalFeatures.VALUES: x,
}
reader = tf.contrib.timeseries.NumpyReader(data)
input_fn = tf.contrib.timeseries.WholeDatasetInputFn(reader)
y = self.ar.evaluate(input_fn=input_fn, steps=1)
(predictions,) = tuple(self.ar.predict(
input_fn=tf.contrib.timeseries.predict_continuation_input_fn(
y, steps=1000)))
plt.plot(data['times'].reshape(-1), data['values'].reshape(-1), label='origin')
plt.plot(y['times'].reshape(-1), y['mean'].reshape(-1), label='evaluation')
plt.plot(predictions['times'].reshape(-1), predictions['mean'].reshape(-1), label='prediction')
plt.xlabel('time_step')
plt.ylabel('values')
plt.legend(loc=4)
plt.show()
return np.sqrt(((np.squeeze(predictions['mean']) - self._s2) ** 2).mean(axis=0))
def train_once(param1):
tf.reset_default_graph()
with tf.device('/cpu:0'):
params = HyperParameter()
params.ar_input_steps = param1
sp = ARSignalPredictor(params)
sp.train()
rmse = sp.evaluate_pn()
return rmse
if __name__ == '__main__':
params = [10, 20]
res = np.zeros((len(params), 1), dtype=np.float32)
p = 0
for i in range(len(params)):
rmse = train_once(params[i])
# rmse = evaluate_once(params[i])
res[i] = rmse
print(res)
print(res)