Skip to content

Commit f2e2bef

Browse files
hunkimhunkim
committed
Added RNN basics
1 parent 692725b commit f2e2bef

1 file changed

Lines changed: 149 additions & 0 deletions

File tree

‎lab12-0-rnn_basics.py‎

Lines changed: 149 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,149 @@
1+
# http://www.wildml.com/2016/08/rnns-in-tensorflow-a-practical-guide-and-undocumented-features/
2+
# http://learningtensorflow.com/index.html
3+
# http://suriyadeepan.github.io/2016-12-31-practical-seq2seq/
4+
5+
import tensorflow as tf
6+
import numpy as np
7+
from tensorflow.contrib import rnn
8+
import pprint
9+
pp = pprint.PrettyPrinter(indent=4)
10+
tf.reset_default_graph()
11+
sess = tf.InteractiveSession()
12+
13+
14+
# One cell RNN input_dim (3) -> output_dim (5)
15+
hidden_size = 5
16+
cell = tf.contrib.rnn.BasicRNNCell(num_units=hidden_size)
17+
print(cell.output_size, cell.state_size)
18+
19+
x_data = np.array([[[1, 2, 3]]], dtype=np.float32)
20+
outputs, _states = tf.nn.dynamic_rnn(cell, x_data, dtype=tf.float32)
21+
22+
sess.run(tf.global_variables_initializer())
23+
pp.pprint(outputs.eval())
24+
sess.close()
25+
26+
states = 0
27+
28+
tf.reset_default_graph()
29+
sess = tf.InteractiveSession()
30+
31+
# One cell RNN input_dim (3) -> output_dim (5). sequence: 2
32+
hidden_size = 5
33+
cell = tf.contrib.rnn.BasicRNNCell(num_units=hidden_size)
34+
x_data = np.array([[[1, 2, 3],
35+
[4, 5, 6]]], dtype=np.float32)
36+
outputs, states = tf.nn.dynamic_rnn(cell, x_data, dtype=tf.float32)
37+
sess.run(tf.global_variables_initializer())
38+
pp.pprint(outputs.eval())
39+
sess.close()
40+
41+
42+
tf.reset_default_graph()
43+
sess = tf.InteractiveSession()
44+
# One cell RNN input_dim (3) -> output_dim (5). sequence: 2
45+
cell = rnn.BasicLSTMCell(num_units=5, state_is_tuple=True)
46+
x_data = np.array([[[1, 2, 3],
47+
[4, 5, 6]],
48+
49+
[[7, 8, 9],
50+
[10, 11, 12]],
51+
52+
[[13, 14, 15],
53+
[16, 17, 18]], ], dtype=np.float32)
54+
outputs, _states = tf.nn.dynamic_rnn(
55+
cell, x_data, sequence_length=[1,2,1], dtype=tf.float32)
56+
sess.run(tf.global_variables_initializer())
57+
pp.pprint(outputs.eval())
58+
sess.close()
59+
60+
61+
tf.reset_default_graph()
62+
sess = tf.InteractiveSession()
63+
# One cell RNN input_dim (3) -> output_dim (5). sequence: 2
64+
65+
batch_size = 3
66+
cell = rnn.BasicLSTMCell(num_units=5, state_is_tuple=True)
67+
initial_state = cell.zero_state(batch_size, tf.float32)
68+
69+
x_data = np.array([[[1, 2, 3],
70+
[4, 5, 6]],
71+
72+
[[7, 8, 9],
73+
[10, 11, 12]],
74+
75+
[[13, 14, 15],
76+
[16, 17, 18]], ], dtype=np.float32)
77+
outputs, _states = tf.nn.dynamic_rnn(cell, x_data,
78+
initial_state=initial_state, dtype=tf.float32)
79+
sess.run(tf.global_variables_initializer())
80+
pp.pprint(outputs.eval())
81+
sess.close()
82+
83+
84+
tf.reset_default_graph()
85+
sess = tf.InteractiveSession()
86+
# Create input data
87+
x_data = np.arange(24, dtype=np.float32).reshape(2, 4, 3)
88+
pp.pprint(x_data) # batch, sequence_length, input size
89+
90+
# Make rnn
91+
cell = rnn.BasicLSTMCell(num_units=5, state_is_tuple=True)
92+
cell = rnn.MultiRNNCell([cell] * 3, state_is_tuple=True)
93+
94+
# rnn in/out
95+
outputs, _states = tf.nn.dynamic_rnn(cell, x_data, dtype=tf.float32)
96+
print("dynamic rnn: ", outputs)
97+
sess.run(tf.global_variables_initializer())
98+
pp.pprint(outputs.eval()) # batch size, unrolling (time), hidden_size
99+
100+
cell = rnn.BasicLSTMCell(num_units=5, state_is_tuple=True)
101+
outputs, _states = tf.nn.dynamic_rnn(cell, x_data, dtype=tf.float32,
102+
sequence_length=[1, 2])
103+
print("dynamic rnn: ", outputs)
104+
sess.run(tf.global_variables_initializer())
105+
pp.pprint(outputs.eval()) # batch size, unrolling (time), hidden_size
106+
sess.close()
107+
108+
109+
tf.reset_default_graph()
110+
sess = tf.InteractiveSession()
111+
# bi-directional rnn
112+
cell_fw = rnn.BasicLSTMCell(num_units=5, state_is_tuple=True)
113+
cell_bw = rnn.BasicLSTMCell(num_units=5, state_is_tuple=True)
114+
115+
outputs, states = tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw, x_data,
116+
sequence_length=[2, 3],
117+
dtype=tf.float32)
118+
119+
sess.run(tf.global_variables_initializer())
120+
pp.pprint(sess.run(outputs))
121+
pp.pprint(sess.run(states))
122+
123+
# Broadcasting based softmax
124+
softmax_w = np.arange(12, dtype=np.float32).reshape(4, 3)
125+
outputs = x_data * softmax_w
126+
pp.pprint(softmax_w)
127+
pp.pprint(outputs)
128+
outputs = x_data * softmax_w + [1, 2, 3]
129+
pp.pprint(outputs)
130+
131+
# [batch_size, sequence_length, emb_dim ]
132+
prediction1 = tf.constant([[[0, 1], [0, 1], [0, 1]]], dtype=tf.float32)
133+
prediction2 = tf.constant([[[1, 0], [1, 0], [1, 0]]], dtype=tf.float32)
134+
prediction3 = tf.constant([[[0, 1], [1, 0], [0, 1]]], dtype=tf.float32)
135+
136+
# [batch_size, sequence_length]
137+
y_data = tf.constant([[1, 1, 1]])
138+
139+
# [batch_size * sequence_length]
140+
weights = tf.constant([[1, 1, 1]], dtype=tf.float32)
141+
142+
sequence_loss1 = tf.contrib.seq2seq.sequence_loss(prediction1, y_data, weights)
143+
sequence_loss2 = tf.contrib.seq2seq.sequence_loss(prediction2, y_data, weights)
144+
sequence_loss3 = tf.contrib.seq2seq.sequence_loss(prediction3, y_data, weights)
145+
146+
sess.run(tf.global_variables_initializer())
147+
print("Loss1: ", sequence_loss1.eval(),
148+
"Loss2: ", sequence_loss2.eval(),
149+
"Loss3: ", sequence_loss3.eval())

0 commit comments

Comments
 (0)