# DNCoreLSTM 回归测试

``````

In [1]:

import numpy as np
import pandas as pd
import TAF
import datetime
import talib
import matplotlib.pylab as plt
import seaborn as sns
% matplotlib inline

``````
``````

In [2]:

index = factors['index']
High = factors.high.values
Low = factors.low.values
Close = factors.close.values
Open = factors.open.values
Volume = factors.volume.values

factors = TAF.get_factors(index, Open, Close, High, Low, Volume, drop=True)

factors = factors.iloc[-700 * 16 - 11 * 16:]

``````
``````

In [3]:

start_date = factors.index[11*16][:10]
end_date = factors.index[-1][:10]

print ('开始时间', start_date)
print ('结束时间', end_date)

``````
``````

``````
``````

In [4]:

rolling = 88

targets['returns'] = targets.close.shift(-5)/ targets.close - 1.

targets= targets.loc[start_date:end_date, 'returns']

``````

#### 输入数据

``````

In [5]:

inputs = np.array(factors).reshape(-1, 1, 58)

targets = np.expand_dims(targets, axis=1)
targets = np.expand_dims(targets, axis=1)

train_inputs = inputs[:-100*16]
test_inputs = inputs[-100*16 - 11 * 16:]

train_targets = targets[:-100]
test_targets = targets[-100:]

train_gather_list = np.arange(train_inputs.shape[0])
train_gather_list = train_gather_list.reshape([-1,16])[11:]
train_gather_list = train_gather_list[:,-1]

test_gather_list = np.arange(test_inputs.shape[0])
test_gather_list = test_gather_list.reshape([-1,16])[11:]
test_gather_list = test_gather_list[:,-1]

``````

#### DNCoreLSTM 回归测试

``````

In [6]:

import tensorflow as tf
from DNCore import DNCoreLSTM

class Regression_DNCoreLSTM(object):

def __init__(self,
inputs,
targets,
gather_list=None,
batch_size=1,
hidden_size=50,
memory_size=50,
num_writes=1,
learning_rate = 1e-4,
optimizer_epsilon = 1e-10,
max_gard_norm = 50,
reset_graph = True):

if reset_graph:
tf.reset_default_graph()
# 控制参数
self._tmp_inputs = inputs
self._tmp_targets = targets
self._in_length = None
self._in_width = inputs.shape[2]
self._out_length = None
self._out_width = targets.shape[2]
self._batch_size = batch_size

# 声明会话
self._sess = tf.InteractiveSession()

self._inputs = tf.placeholder(
dtype=tf.float32,
shape=[self._in_length, self._batch_size, self._in_width],
name='inputs')
self._targets = tf.placeholder(
dtype=tf.float32,
shape=[self._out_length, self._batch_size, self._out_width],
name='targets')

self._RNNCoreCell = DNCoreLSTM(
dnc_output_size=self._out_width,
hidden_size=hidden_size,
memory_size=memory_size,
word_size=self._in_width,

self._initial_state = \
self._RNNCoreCell.initial_state(batch_size)

output_sequences, _ = \
tf.nn.dynamic_rnn(cell= self._RNNCoreCell,
inputs=self._inputs,
initial_state=self._initial_state,
time_major=True)

self._original_output_sequences = output_sequences
if gather_list is not None:
output_sequences = tf.gather(output_sequences, gather_list)
output_sequences = tf.tanh(output_sequences)

cost = tf.square(output_sequences-self._targets)
self._cost = tf.reduce_mean(cost)

# Set up optimizer with global norm clipping.
trainable_variables = tf.trainable_variables()
global_step = tf.get_variable(
name="global_step",
shape=[],
dtype=tf.int64,
initializer=tf.zeros_initializer(),
trainable=False,
collections=[tf.GraphKeys.GLOBAL_VARIABLES, tf.GraphKeys.GLOBAL_STEP])

optimizer = tf.train.RMSPropOptimizer(
learning_rate=learning_rate, epsilon=optimizer_epsilon)

self._sess.run(tf.global_variables_initializer())
self._variables_saver = tf.train.Saver()

def fit(self,
training_iters =1e2,
display_step = 5,
save_path = None,
restore_path = None):

if restore_path is not None:
self._variables_saver.restore(self._sess, restore_path)

for scope in range(np.int(training_iters)):
self._sess.run([self._train_step],
feed_dict = {self._inputs:self._tmp_inputs, self._targets:self._tmp_targets})

if scope % display_step == 0:
loss = self._sess.run(
self._cost,
feed_dict = {self._inputs:self._tmp_inputs, self._targets:self._tmp_targets})
print (scope, '  loss--', loss)

print ("Optimization Finished!")
loss = self._sess.run(
self._cost, feed_dict = {self._inputs:self._tmp_inputs, self._targets:self._tmp_targets})
print ('Model assessment  loss--', '  loss--', loss)

# 保存模型可训练变量
if save_path is not None:
self._variables_saver.save(self._sess, save_path)

def close(self):
self._sess.close()
print ('结束进程，清理tensorflow内存/显存占用')

def pred(self, inputs, gather_list=None, restore_path=None):
output_sequences = self._original_output_sequences
if gather_list is not None:
output_sequences = tf.gather(output_sequences, gather_list)
outputs = tf.tanh(output_sequences)
return self._sess.run(outputs, feed_dict = {self._inputs:inputs})

def restore_trainable_variables(self, restore_path):
self._variables_saver.restore(self._sess, restore_path)

``````
``````

In [7]:

a = Regression_DNCoreLSTM(train_inputs, train_targets, train_gather_list)
a.fit(training_iters = 500, save_path='models/R1.ckpt')
a.close()

``````
``````

C:\ProgramData\Anaconda3\lib\site-packages\tensorflow\python\ops\gradients_impl.py:93: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
"Converting sparse IndexedSlices to a dense Tensor of unknown shape. "

0   loss-- 0.0149116
5   loss-- 0.0148272
10   loss-- 0.0147316
15   loss-- 0.0146004
20   loss-- 0.0144359
25   loss-- 0.0142147
30   loss-- 0.0139486
35   loss-- 0.0136023
40   loss-- 0.0131801
45   loss-- 0.0126733
50   loss-- 0.0120637
55   loss-- 0.0113554
60   loss-- 0.010562
65   loss-- 0.00970413
70   loss-- 0.00881338
75   loss-- 0.00794179
80   loss-- 0.00712633
85   loss-- 0.00642303
90   loss-- 0.0058539
95   loss-- 0.00543427
100   loss-- 0.00514201
105   loss-- 0.00493222
110   loss-- 0.00476724
115   loss-- 0.00462523
120   loss-- 0.00447889
125   loss-- 0.0043294
130   loss-- 0.00417228
135   loss-- 0.00400143
140   loss-- 0.00381986
145   loss-- 0.00363136
150   loss-- 0.00343413
155   loss-- 0.00324269
160   loss-- 0.00309782
165   loss-- 0.00289509
170   loss-- 0.00274539
175   loss-- 0.00262085
180   loss-- 0.00246892
185   loss-- 0.00235731
190   loss-- 0.00224838
195   loss-- 0.00213627
200   loss-- 0.00205065
205   loss-- 0.00195998
210   loss-- 0.00188252
215   loss-- 0.00180501
220   loss-- 0.00175475
225   loss-- 0.00167092
230   loss-- 0.001633
235   loss-- 0.00156636
240   loss-- 0.00152514
245   loss-- 0.00147259
250   loss-- 0.00143725
255   loss-- 0.00139229
260   loss-- 0.00135657
265   loss-- 0.0013233
270   loss-- 0.0012851
275   loss-- 0.00125996
280   loss-- 0.00122647
285   loss-- 0.00120017
290   loss-- 0.00116794
295   loss-- 0.00115305
300   loss-- 0.00111746
305   loss-- 0.00110359
310   loss-- 0.00106962
315   loss-- 0.00106154
320   loss-- 0.00103117
325   loss-- 0.00101729
330   loss-- 0.00099108
335   loss-- 0.000983903
340   loss-- 0.000952042
345   loss-- 0.000946161
350   loss-- 0.000917308
355   loss-- 0.000916552
360   loss-- 0.000888917
365   loss-- 0.000881377
370   loss-- 0.000857288
375   loss-- 0.000858365
380   loss-- 0.000829304
385   loss-- 0.000827135
390   loss-- 0.000803459
395   loss-- 0.000803866
400   loss-- 0.000781087
405   loss-- 0.000774746
410   loss-- 0.000756085
415   loss-- 0.000757288
420   loss-- 0.000732103
425   loss-- 0.000732525
430   loss-- 0.000709998
435   loss-- 0.000715097
440   loss-- 0.00069051
445   loss-- 0.000691042
450   loss-- 0.000670274
455   loss-- 0.000673939
460   loss-- 0.000652421
465   loss-- 0.000654323
470   loss-- 0.000632649
475   loss-- 0.000640306
480   loss-- 0.000616631
485   loss-- 0.000619596
490   loss-- 0.000601106
495   loss-- 0.000605031
Optimization Finished!
Model assessment  loss--   loss-- 0.00058358

``````
``````

In [8]:

a = Regression_DNCoreLSTM(train_inputs, train_targets, train_gather_list)
a.fit(training_iters = 500,
save_path='models/R2.ckpt',
restore_path='models/R1.ckpt')
a.close()

``````
``````

C:\ProgramData\Anaconda3\lib\site-packages\tensorflow\python\ops\gradients_impl.py:93: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
"Converting sparse IndexedSlices to a dense Tensor of unknown shape. "

INFO:tensorflow:Restoring parameters from models/R1.ckpt
0   loss-- 0.000584173
5   loss-- 0.000590047
10   loss-- 0.000567725
15   loss-- 0.00057557
20   loss-- 0.000555442
25   loss-- 0.000557696
30   loss-- 0.000540423
35   loss-- 0.000546677
40   loss-- 0.000524918
45   loss-- 0.000533925
50   loss-- 0.000510787
55   loss-- 0.000521456
60   loss-- 0.000498488
65   loss-- 0.000507708
70   loss-- 0.000486787
75   loss-- 0.000493403
80   loss-- 0.000476066
85   loss-- 0.000481163
90   loss-- 0.00046324
95   loss-- 0.000473
100   loss-- 0.000451606
105   loss-- 0.000459443
110   loss-- 0.000440397
115   loss-- 0.000451587
120   loss-- 0.000429613
125   loss-- 0.000439785
130   loss-- 0.000418625
135   loss-- 0.000430468
140   loss-- 0.000409455
145   loss-- 0.000418764
150   loss-- 0.00039949
155   loss-- 0.00041189
160   loss-- 0.000389385
165   loss-- 0.000401372
170   loss-- 0.000379985
175   loss-- 0.000394522
180   loss-- 0.000370459
185   loss-- 0.000385208
190   loss-- 0.000359228
195   loss-- 0.000379497
200   loss-- 0.000352105
205   loss-- 0.000370216
210   loss-- 0.000343124
215   loss-- 0.000364151
220   loss-- 0.000335226
225   loss-- 0.000353414
230   loss-- 0.00032796
235   loss-- 0.000346219
240   loss-- 0.000320001
245   loss-- 0.000338879
250   loss-- 0.000312291
255   loss-- 0.000332826
260   loss-- 0.000303887
265   loss-- 0.000326207
270   loss-- 0.000297274
275   loss-- 0.000317167
280   loss-- 0.000292035
285   loss-- 0.000308789
290   loss-- 0.00028473
295   loss-- 0.000301311
300   loss-- 0.000279718
305   loss-- 0.000293514
310   loss-- 0.000274261
315   loss-- 0.000285039
320   loss-- 0.000269482
325   loss-- 0.000277628
330   loss-- 0.0002635
335   loss-- 0.000268799
340   loss-- 0.000260747
345   loss-- 0.000257114
350   loss-- 0.000262042
355   loss-- 0.000248687
360   loss-- 0.000256312
365   loss-- 0.000244192
370   loss-- 0.000251556
375   loss-- 0.000234246
380   loss-- 0.000256098
385   loss-- 0.000226217
390   loss-- 0.000253778
395   loss-- 0.000219956
400   loss-- 0.000250336
405   loss-- 0.00021661
410   loss-- 0.00024236
415   loss-- 0.000210551
420   loss-- 0.000240523
425   loss-- 0.000205184
430   loss-- 0.000231786
435   loss-- 0.000203871
440   loss-- 0.000218571
445   loss-- 0.00020412
450   loss-- 0.000210869
455   loss-- 0.000202422
460   loss-- 0.00019981
465   loss-- 0.000205192
470   loss-- 0.000190601
475   loss-- 0.000207638
480   loss-- 0.000182452
485   loss-- 0.000211279
490   loss-- 0.000175791
495   loss-- 0.000213278
Optimization Finished!
Model assessment  loss--   loss-- 0.000174389

``````
``````

In [9]:

a = Regression_DNCoreLSTM(train_inputs, train_targets, train_gather_list)
a.restore_trainable_variables(restore_path='models/R2.ckpt')
b = a.pred(train_inputs, train_gather_list)

``````
``````

C:\ProgramData\Anaconda3\lib\site-packages\tensorflow\python\ops\gradients_impl.py:93: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
"Converting sparse IndexedSlices to a dense Tensor of unknown shape. "

INFO:tensorflow:Restoring parameters from models/R2.ckpt

``````
``````

In [10]:

tmp = pd.DataFrame([train_targets.flatten(), b.flatten()]).T
tmp.columns = ['targets','dnc_tanh']

``````
``````

In [11]:

tmp.corr()

``````
``````

Out[11]:

targets
dnc_tanh

targets
1.000000
0.965116

dnc_tanh
0.965116
1.000000

``````
``````

In [12]:

tmp.plot(figsize=(60,6))

``````
``````

Out[12]:

<matplotlib.axes._subplots.AxesSubplot at 0x110921d0>

``````