/
UnitTest.py
167 lines (146 loc) · 6.28 KB
/
UnitTest.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
import random
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import tensorflow.contrib.layers as layers
import tensorflow.contrib.rnn as rnn
class AdditionGenerator:
def __init__(self, im, om, n_time_step):
self._im, self._om = im, om
self._base = self._om
self._n_time_step = n_time_step
def _op(self, seq):
return sum(seq)
def _gen_seq(self, n_time_step, tar):
seq = []
for _ in range(n_time_step):
seq.append(tar % self._base)
tar //= self._base
return seq
def _gen_targets(self, n_time_step):
return [int(random.randint(0, self._om ** n_time_step - 1) / self._im) for _ in range(self._im)]
def gen(self, batch_size):
x = np.empty([batch_size, self._n_time_step, self._im])
y = np.zeros([batch_size, self._n_time_step, self._om])
for i in range(batch_size):
targets = self._gen_targets(self._n_time_step)
sequences = [self._gen_seq(self._n_time_step, tar) for tar in targets]
for j in range(self._im):
x[i, ..., j] = sequences[j]
y[i, range(self._n_time_step), self._gen_seq(self._n_time_step, self._op(targets))] = 1
return x, y
class RNNWrapper:
def __init__(self):
self._log = {}
self._return_all_states = None
self._optimizer = self._generator = None
self._tfx = self._tfy = self._input = self._output = None
self._cell = self._im = self._om = None
self._sess = tf.Session()
def _verbose(self):
x_test, y_test = self._generator.gen(1)
ans = np.argmax(self._sess.run(self._output, {
self._tfx: x_test
}), axis=2).ravel()
x_test = x_test.astype(np.int)
print("I think {} = {}, answer: {}...".format(
" + ".join(
["".join(map(lambda n: str(n), x_test[0, ..., i][::-1])) for i in range(x_test.shape[2])]
),
"".join(map(lambda n: str(n), ans[::-1])),
"".join(map(lambda n: str(n), np.argmax(y_test, axis=2).ravel()[::-1]))))
def _get_output(self, rnn_outputs, rnn_states):
print("Outputs :", rnn_outputs)
print("States :", rnn_states)
if self._return_all_states:
outputs = tf.concat([rnn_outputs, rnn_states], axis=2)
else:
outputs = rnn_outputs
self._output = layers.fully_connected(
outputs, num_outputs=self._om, activation_fn=tf.nn.sigmoid)
def fit(self, im, om, generator, cell=rnn.BasicLSTMCell, return_all_states=False):
self._generator = generator
self._im, self._om = im, om
self._return_all_states = return_all_states
self._optimizer = tf.train.AdamOptimizer(0.01)
self._input = self._tfx = tf.placeholder(tf.float32, shape=[None, None, im])
self._tfy = tf.placeholder(tf.float32, shape=[None, None, om])
self._cell = cell(128)
rnn_outputs, rnn_states = tf.nn.dynamic_rnn(
self._cell, self._input, return_all_states=self._return_all_states,
initial_state=self._cell.zero_state(tf.shape(self._input)[0], tf.float32)
)
self._get_output(rnn_outputs, rnn_states)
loss = -tf.reduce_mean(
self._tfy * tf.log(self._output + 1e-8) + (1 - self._tfy) * tf.log(1 - self._output + 1e-8)
)
train_step = self._optimizer.minimize(loss)
self._log["iter_err"] = []
self._log["epoch_err"] = []
self._sess.run(tf.global_variables_initializer())
for _ in range(20):
epoch_err = 0
for __ in range(128):
x_batch, y_batch = self._generator.gen(64)
feed_dict = {self._tfx: x_batch, self._tfy: y_batch}
iter_err = self._sess.run([loss, train_step], feed_dict)[0]
self._log["iter_err"].append(iter_err)
epoch_err += iter_err
self._log["epoch_err"].append(epoch_err / 128)
self._verbose()
def predict(self, x):
x = np.atleast_3d(x)
output = self._sess.run(self._output, {self._tfx: x})
return np.argmax(output, axis=2).ravel()
def draw_err_logs(self):
ee, ie = self._log["epoch_err"], self._log["iter_err"]
ee_base = np.arange(len(ee))
ie_base = np.linspace(0, len(ee) - 1, len(ie))
plt.figure()
plt.plot(ie_base, ie, label="Iter error")
plt.plot(ee_base, ee, linewidth=3, label="Epoch error")
plt.legend()
plt.show()
if __name__ == '__main__':
random_scale = 2
digit_len = 4
digit_base = 10
n_digit = 2
_generator = AdditionGenerator(n_digit, digit_base, n_time_step=digit_len)
# Return final state only
# BasicRNNCell
print("=" * 60 + "\n" + "Return final state only (BasicRNNCell)\n" + "-" * 60)
lstm = RNNWrapper()
lstm.fit(n_digit, digit_base, _generator, rnn.BasicRNNCell)
lstm.draw_err_logs()
# BasicLSTMCell
print("=" * 60 + "\n" + "Return final state only (BasicLSTMCell)\n" + "-" * 60)
tf.reset_default_graph()
lstm = RNNWrapper()
lstm.fit(n_digit, digit_base, _generator)
lstm.draw_err_logs()
# LSTMCell
print("=" * 60 + "\n" + "Return final state only (LSTMCell)\n" + "-" * 60)
tf.reset_default_graph()
lstm = RNNWrapper()
lstm.fit(n_digit, digit_base, _generator, rnn.LSTMCell)
lstm.draw_err_logs()
# GRUCell
print("=" * 60 + "\n" + "Return final state only (GRUCell)\n" + "-" * 60)
tf.reset_default_graph()
lstm = RNNWrapper()
lstm.fit(n_digit, digit_base, _generator, rnn.GRUCell)
lstm.draw_err_logs()
# Return all states
# LSTMCell
print("=" * 60 + "\n" + "Return all states generated (LSTMCell)\n" + "-" * 60)
tf.reset_default_graph()
lstm = RNNWrapper()
lstm.fit(n_digit, digit_base, _generator, rnn.LSTMCell, return_all_states=True)
lstm.draw_err_logs()
# GRUCell
print("=" * 60 + "\n" + "Return all states generated (GRUCell)\n" + "-" * 60)
tf.reset_default_graph()
lstm = RNNWrapper()
lstm.fit(n_digit, digit_base, _generator, rnn.GRUCell, return_all_states=True)
lstm.draw_err_logs()