/
semantic_mnist_pairs.py
206 lines (164 loc) · 7.52 KB
/
semantic_mnist_pairs.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
from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf
import logictensornetworks as ltn
import numpy as np
from logictensornetworks import Forall,Exists, Implies, And, Or, Not
from datetime import datetime
import matplotlib.pyplot as plt
from termcolor import colored, cprint
# prepare data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
def overlay_pictures(pics,labels):
permutation = np.arange(len(labels))
np.random.shuffle(permutation)
pics_p = pics[permutation]
labels_p = labels[permutation]
ppics = np.zeros_like(pics)
llabels = np.zeros_like(np.concatenate([labels,labels_p], axis=1))
for i in range(len(pics)):
w = np.random.beta(2.0, 2.0)
if np.argmax(labels[i]) <= np.argmax(labels_p[i]):
ppics[i] = pics[i] - w*pics_p[i]
llabels[i] = np.concatenate([labels[i], labels_p[i]], axis=0)
else:
ppics[i] = pics_p[i] - w*pics[i]
llabels[i] = np.concatenate([labels_p[i], labels[i]], axis=0)
return ppics,llabels
ltn.LAYERS = 3
ltn.set_universal_aggreg("mean")
ltn.set_tnorm("prod")
logs_path = 'semantic_mnist_log/' + datetime.now().strftime("%Y%m%d-%H%M%S") + "/"
embedding_size = 10
train_batch_size = 50
test_batch_size = 30
# begin MNIST conv net
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
x = tf.placeholder(tf.float32, [None, 784],name="x")
y_ = tf.placeholder(tf.float32, shape=[None, embedding_size*2],name="y_")
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
x_image = tf.reshape(x, [-1, 28, 28, 1])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
keep_prob = tf.placeholder(tf.float32,name="keep_prob")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = weight_variable([1024, embedding_size*2])
b_fc2 = bias_variable([embedding_size*2])
y_conv_before_softmax = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
y_conv = tf.concat([
tf.nn.softmax(y_conv_before_softmax[:,:embedding_size]),
tf.nn.softmax(y_conv_before_softmax[:,embedding_size:])],axis=1)
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(
labels=tf.reshape(y_,(-1,embedding_size)),
logits=tf.reshape(y_conv_before_softmax[:train_batch_size],(-1,embedding_size))))
correct_prediction = tf.equal(
tf.argmax(tf.reshape(y_conv,(-1,embedding_size)), axis=1),
tf.argmax(tf.reshape(y_,(-1,embedding_size)), axis=1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# end of mnist
# begin of ltn
dd = ltn.variable("double_digit",y_conv)
def get_nth_element(n):
def result(p):
return p[:,n]
return result
IS1 = {n:ltn.predicate("is_equal_to_"+str(n),embedding_size*2,
pred_definition=get_nth_element(n)) for n in range(10)}
IS2 = {n:ltn.predicate("is_equal_to_"+str(n),embedding_size*2,
pred_definition=get_nth_element(10+n)) for n in range(10)}
examples_of_1 = {n:ltn.variable("examples_of_1_"+str(n),
tf.gather(y_conv[:train_batch_size],
tf.where(tf.equal(y_[:train_batch_size,tf.constant(n)],1))[:,0]))
for n in range(10)}
examples_of_2 = {n:ltn.variable("examples_of_2_"+str(n),
tf.gather(y_conv[:train_batch_size],
tf.where(tf.equal(y_[:train_batch_size,tf.constant(10+n)],1))[:,0]))
for n in range(10)}
examples_of_12 = {(m,n):ltn.variable("examples_of_12_"+str(m)+"_"+str(n),
tf.gather(y_conv[:train_batch_size],
tf.where(
tf.logical_and(
tf.equal(y_[:train_batch_size,tf.constant(10+n)],1),
tf.equal(y_[:train_batch_size,tf.constant(10+n)],1)))[:,0]))
for m in range(10) for n in range(10)}
constraints = tf.concat(
[Forall(examples_of_1[n],IS1[n](examples_of_1[n]))
for n in range(10)]+ \
[Forall(examples_of_2[n],IS2[n](examples_of_2[n]))
for n in range(10)] + \
[Exists(dd,IS1[n](dd)) for n in range(10)] + \
[Exists(dd,IS2[n](dd)) for n in range(10)] + \
[Forall(dd,Not(And(IS1[n](dd),IS2[m](dd))))
for n in range(10) for m in range(n)],0)
# [Forall(dd, And(Not(And(IS1[n](dd), IS1[m](dd))),
# Not(And(IS2[n](dd), IS2[m](dd)))))
# for m in range(10) for n in range(m) if m != n] + \
# [Forall(dd, And(Or(*[IS1[n](dd) for n in range(10)]),
# Or(*[IS2[n](dd) for n in range(10)])))] + \
# [Forall(examples_of_1[n],Not(IS1[m](examples_of_1[n])))
# for m in range(10)
# for n in range(10) if n != m] + \
# [Forall(examples_of_2[n], Not(IS2[m](examples_of_2[n])))
# for m in range(10)
# for n in range(10) if n != m] + \
constr_min = tf.reduce_min(constraints,axis=0)
constr_hmean = 1/tf.reduce_mean(1/constraints)
constr_mean = tf.reduce_mean(constraints,axis=0)
constr_sqmean = tf.reduce_mean(tf.square(constraints),axis=0)
constr_var = tf.reduce_mean(tf.square(constraints-constr_mean))
constraints_loss = (1.0-constr_mean)
optimize_only_conv_net = tf.train.AdamOptimizer(1e-3).minimize(cross_entropy)
optimize_conv_net_and_ltn = tf.train.AdamOptimizer(1e-3).minimize(cross_entropy+constraints_loss)
optimize_only_ltn = tf.train.AdamOptimizer(1e-3).minimize(constraints_loss)
init = tf.global_variables_initializer()
def get_feed_dict(train_size,test_size):
global x,y_keep_prob
train_x, train_y = mnist.train.next_batch(train_size)
train_x,train_y = overlay_pictures(train_x,train_y)
test_x = mnist.test.images[np.random.randint(len(mnist.test.images),size=test_size)]
test_x,_ = overlay_pictures(test_x,mnist.test.labels[:test_size])
feed_dict = {}
feed_dict[x] = np.concatenate([train_x,test_x],0)
feed_dict[y_] = train_y
feed_dict[keep_prob] = 1.0
return feed_dict
with tf.Session() as sess:
sess.run(init)
for i in range(20000):
if i%100 == 0:
train_feed_dict = get_feed_dict(train_batch_size,test_batch_size)
test_images,test_labels = overlay_pictures(mnist.test.images,mnist.test.labels)
train_feed_dict[keep_prob] = 1.0
sc = sess.run(constraints,feed_dict=train_feed_dict)
for s in sc:
cprint(s,"red")
print("accuracy at %d = %.5e"%
(i,sess.run(accuracy,
feed_dict={x:test_images,
y_:test_labels,
keep_prob:1.0})))
if i < 0:
train_feed_dict[keep_prob] = .5
sess.run(optimize_only_conv_net,feed_dict=train_feed_dict)
else:
train_feed_dict[keep_prob] = 1.0
sess.run(optimize_only_ltn,feed_dict=train_feed_dict)