-
Notifications
You must be signed in to change notification settings - Fork 0
/
mnist_rnn.py
261 lines (205 loc) · 11.6 KB
/
mnist_rnn.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
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
''''
Task 1. Write simple RNN to recognize MNIST digits.
The image is 28x28. Flatten it to a 784 vector.
Pick some divisior d of 784, e.g. d = 28.
At each timestep the input will be d bits of the image.
Thus the sequence length will be 784 / d
You should be able to get over 93% accuracy
Write your own implementation of RNN, you can look at the one from the slide,
but do not copy it blindly.
Task 2.
Same, but use LSTM instead of simple RNN.
What accuracy do you get.
Experiment with choosing d, compare RNN and LSTM.
Again do not use builtin Tensorflow implementation. Write your own :)
Task 3*.
Make LSTM a deep bidirectional, multilayer LSTM.
'''
class SingleLSTMLayer(object):
def __init__(self, x, d, steps_n, hidden_n=100):
self.d = d
self.steps_n = steps_n
self.hidden_n = hidden_n
stddev=0.1
with tf.variable_scope('lstm'):
self.h0 = tf.Variable(name='h0', initial_value=tf.truncated_normal(shape=[hidden_n], stddev=stddev), dtype=tf.float32)
self.c0 = tf.Variable(name='c0', initial_value=tf.truncated_normal(shape=[hidden_n], stddev=stddev), dtype=tf.float32)
self.W = tf.Variable(name='W', initial_value=tf.truncated_normal(shape=[self.d + hidden_n, 4 * hidden_n], stddev=stddev), dtype=tf.float32)
self.bias = tf.Variable(name='bias', initial_value=tf.truncated_normal(shape=[4 * hidden_n], stddev=stddev), dtype=tf.float32)
self.h = {}
self.h[-1] = tf.tile(tf.reshape(self.h0, (1, -1)), [tf.shape(x)[0], 1])
self.c = tf.tile(tf.reshape(self.c0, (1, -1)), [tf.shape(x)[0], 1])
for step in range(self.steps_n):
step_x = x[:, step, :]
x_h = tf.concat([step_x, self.h[step-1]], axis=1)
# h shape: bs x steps x d
ih = tf.matmul(x_h, self.W) + self.bias
i = tf.sigmoid(ih[:, 0*hidden_n:1*hidden_n])
f = tf.sigmoid(ih[:, 1*hidden_n:2*hidden_n])
o = tf.sigmoid(ih[:, 2*hidden_n:3*hidden_n])
g = tf.tanh(ih[:, 3*hidden_n:4*hidden_n])
self.c = f * self.c + i * g
self.h[step] = o * tf.tanh(self.c)
def combined_h(self):
n = max(self.h.keys())+1
to_concat = []
for step in range(n):
to_concat.append(tf.reshape(self.h[step], (-1, 1, self.hidden_n)))
tf.assert_equal(tf.shape(to_concat[-1])[0], 128)
concat = tf.concat(to_concat, axis=1)
tf.assert_equal(tf.shape(to_concat[-1])[1], n)
return concat
class BiLSTMLayer(object):
def __init__(self, x, d, steps_n, hidden_n=100):
self.forward = SingleLSTMLayer(x, d, steps_n, hidden_n)
rev_x = tf.reverse(x, axis=[1])
self.reverse = SingleLSTMLayer(rev_x, d, steps_n, hidden_n)
def combined_h(self):
fwh = self.forward.combined_h()
rwh = self.reverse.combined_h()
return tf.concat([fwh, tf.reverse(rwh, axis=[1])], axis=2)
def last_h(self):
return self.combined_h()[:,-1,:]
class MnistTrainer(object):
def train_on_batch(self, batch_xs, batch_ys):
transf_x = np.reshape(batch_xs, (-1, self.steps_n, self.d))
return self.sess.run([self.loss, self.accuracy, self.train_step],
feed_dict={self.x: transf_x,
self.y_target: batch_ys})[:2]
def create_simple_rnn_model(self):
self.d = 28
self.steps_n = 784 / self.d
HSZ = 100
self.x = tf.placeholder(dtype=tf.float32, shape=[None, self.steps_n, self.d])
self.y_target = tf.placeholder(dtype=tf.float32, shape=[None, 10])
stddev=0.1
with tf.variable_scope('rnn'):
self.h0 = tf.Variable(name='h0',
initial_value=tf.truncated_normal(shape=[HSZ], stddev=stddev), dtype=tf.float32)
self.bias = tf.Variable(name='bias',
initial_value=tf.truncated_normal(shape=[HSZ], stddev=stddev), dtype=tf.float32)
self.W = tf.Variable(name='W',
initial_value=tf.truncated_normal(shape=[self.d + HSZ, HSZ], stddev=stddev), dtype=tf.float32)
self.out_W = tf.Variable(name='out_W',
initial_value=tf.truncated_normal(shape=(HSZ, 10), stddev=stddev), dtype=tf.float32)
self.out_bias = tf.Variable(name='out_bias',
initial_value=tf.truncated_normal(shape=(10,), stddev=stddev), dtype=tf.float32)
self.h = tf.tile(tf.reshape(self.h0, (1,-1)), [tf.shape(self.x)[0], 1]) # to jest sliskie
for step in range(self.steps_n):
step_x = self.x[:, step, :]
x_h = tf.concat([step_x, self.h], axis=1)
# h shape: bs x steps x d
self.h = tf.tanh(tf.matmul(x_h, self.W) + self.bias)
signal = tf.matmul(self.h, self.out_W) + self.out_bias
self.loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=self.y_target, logits=signal))
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.y_target, axis=1), tf.argmax(signal, axis=1)), tf.float32))
self.train_step = tf.train.AdamOptimizer(0.001).minimize(self.loss)
def create_lstm_model(self):
self.d = 28
self.steps_n = 784 / self.d
hidden_n = 100
self.x = tf.placeholder(dtype=tf.float32, shape=[None, self.steps_n, self.d])
self.y_target = tf.placeholder(dtype=tf.float32, shape=[None, 10])
stddev=0.1
with tf.variable_scope('lstm'):
self.out_W = tf.Variable(name='out_W', initial_value=tf.truncated_normal(shape=[hidden_n, 10], stddev=stddev), dtype=tf.float32)
self.out_bias = tf.Variable(name='out_bias', initial_value=tf.truncated_normal(shape=[10], stddev=stddev), dtype=tf.float32)
layer = SingleLSTMLayer(self.x, self.d, self.steps_n, hidden_n=hidden_n)
last_h = layer.h[max(layer.h.keys())]
signal = tf.matmul(last_h, self.out_W) + self.out_bias
self.loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=self.y_target, logits=signal))
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.y_target, axis=1), tf.argmax(signal, axis=1)), tf.float32))
self.train_step = tf.train.AdamOptimizer(0.001).minimize(self.loss)
def create_bi_lstm_model(self):
self.d = 28
self.steps_n = 784 / self.d
hidden_n = 100
self.x = tf.placeholder(dtype=tf.float32, shape=[None, self.steps_n, self.d])
self.y_target = tf.placeholder(dtype=tf.float32, shape=[None, 10])
stddev=0.1
with tf.variable_scope('lstm'):
self.out_W = tf.Variable(name='out_W', initial_value=tf.truncated_normal(shape=[hidden_n*2, 10], stddev=stddev), dtype=tf.float32)
self.out_bias = tf.Variable(name='out_bias', initial_value=tf.truncated_normal(shape=[10], stddev=stddev), dtype=tf.float32)
layer = BiLSTMLayer(self.x, self.d, self.steps_n, hidden_n=hidden_n)
last_h = layer.last_h()
signal = tf.matmul(last_h, self.out_W) + self.out_bias
self.loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=self.y_target, logits=signal))
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.y_target, axis=1), tf.argmax(signal, axis=1)), tf.float32))
self.train_step = tf.train.AdamOptimizer(0.001).minimize(self.loss)
def create_deep_lstm_model(self):
self.d = 28
self.steps_n = 784 / self.d
hidden_n = 100
self.x = tf.placeholder(dtype=tf.float32, shape=[None, self.steps_n, self.d])
self.y_target = tf.placeholder(dtype=tf.float32, shape=[None, 10])
stddev=0.1
with tf.variable_scope('lstm'):
self.out_W = tf.Variable(name='out_W', initial_value=tf.truncated_normal(shape=[hidden_n, 10], stddev=stddev), dtype=tf.float32)
self.out_bias = tf.Variable(name='out_bias', initial_value=tf.truncated_normal(shape=[10], stddev=stddev), dtype=tf.float32)
layer = SingleLSTMLayer(self.x, self.d, self.steps_n, hidden_n=hidden_n)
layer1 = SingleLSTMLayer(layer.combined_h(), hidden_n, self.steps_n, hidden_n=hidden_n)
layer2 = SingleLSTMLayer(layer1.combined_h(), hidden_n, self.steps_n, hidden_n=hidden_n)
last_h = layer2.h[max(layer2.h.keys())]
signal = tf.matmul(last_h, self.out_W) + self.out_bias
self.loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=self.y_target, logits=signal))
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.y_target, axis=1), tf.argmax(signal, axis=1)), tf.float32))
self.train_step = tf.train.AdamOptimizer(0.001).minimize(self.loss)
def create_deep_bi_lstm_model(self):
self.d = 28
self.steps_n = 784 / self.d
hidden_n = 100
self.x = tf.placeholder(dtype=tf.float32, shape=[None, self.steps_n, self.d])
self.y_target = tf.placeholder(dtype=tf.float32, shape=[None, 10])
stddev=0.1
with tf.variable_scope('lstm'):
self.out_W = tf.Variable(name='out_W', initial_value=tf.truncated_normal(shape=[hidden_n*2, 10], stddev=stddev), dtype=tf.float32)
self.out_bias = tf.Variable(name='out_bias', initial_value=tf.truncated_normal(shape=[10], stddev=stddev), dtype=tf.float32)
layer = BiLSTMLayer(self.x, self.d, self.steps_n, hidden_n=hidden_n)
layer2 = BiLSTMLayer(layer.combined_h(), 2*hidden_n, self.steps_n, hidden_n=hidden_n)
last_h = layer2.last_h()
signal = tf.matmul(last_h, self.out_W) + self.out_bias
self.loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=self.y_target, logits=signal))
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.y_target, axis=1), tf.argmax(signal, axis=1)), tf.float32))
self.train_step = tf.train.AdamOptimizer(0.001).minimize(self.loss)
def train(self):
# self.create_deep_lstm_model()
self.create_deep_bi_lstm_model()
# self.create_bi_lstm_model()
# self.create_lstm_model()
# self.create_simple_rnn_model()
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
with tf.Session() as self.sess:
tf.global_variables_initializer().run() # initialize variables
batches_n = 100000
mb_size = 128
losses = []
try:
for batch_idx in range(batches_n):
batch_xs, batch_ys = mnist.train.next_batch(mb_size)
loss, accuracy = self.train_on_batch(batch_xs, batch_ys)
losses.append(loss)
if batch_idx % 100 == 0:
print('Batch {batch_idx}: loss {loss}, mean_loss {mean_loss}'.format(
batch_idx=batch_idx, loss=loss, mean_loss=np.mean(losses[-200:]))
)
except KeyboardInterrupt:
print('Stopping training!')
pass
test_x, test_y = mnist.test.next_batch(2000)
transf_x = np.reshape(test_x, (-1, self.steps_n, self.d))
loss, acc = self.sess.run([self.loss, self.accuracy],
feed_dict={self.x: transf_x,
self.y_target: test_y})
print "Test set loss {loss} accuracy {acc}".format(
loss=loss, acc=acc)
if __name__ == '__main__':
trainer = MnistTrainer()
trainer.train()