-
Notifications
You must be signed in to change notification settings - Fork 0
/
memN2N.py
159 lines (128 loc) · 7.4 KB
/
memN2N.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
import tensorflow as tf
import numpy as np
import pdb
class MemN2N(object):
"""End-To-End Memory Networks"""
def __init__(self, config, sess):
self.sess = sess
self.batch_size = config.batch_size
self.n_emb = config.emb_size
self.n_memory = config.memory_size
self.n_sentence = config.description_length
self.n_vocab = config.vocab_size
self.learning_rate = config.learning_rate
self.num_epoch = config.num_epoch
self.max_grad_norm = 40
self.set_inputs_and_variables()
self.logs_path = "/tmp/memnet"
def set_inputs_and_variables(self):
# define inputs
self.in_history = tf.placeholder(tf.int32, [None, self.n_memory, self.n_sentence], name = "history")
self.in_query = tf.placeholder(tf.int32, [None, self.n_sentence], name = "query")
self.in_answer = tf.placeholder(tf.int32, [None, self.n_vocab], name = "answer")
# define variables
self.weights = {}
def concat_nil_random():
nil_embedding = tf.zeros([1, self.n_emb])
concat_emb = tf.concat(0, [nil_embedding, tf.random_normal([self.n_vocab - 1, self.n_emb], stddev=0.1)])
# if you meet problem within this sentence,
# please change the position of two parameters of tf.concat,
# this is because the interface changes among versions
return concat_emb
self.weights["embA"] = tf.Variable(concat_nil_random(), name = "embA")
self.weights["embB"] = self.weights["embA"]
self.weights["embC"] = tf.Variable(concat_nil_random(), name = "embC")
self.weights["W"] = tf.Variable(concat_nil_random(), name = "W")
def forward(self, history, query):
m_embA = tf.nn.embedding_lookup(self.weights["embA"], history) # A * x_{ij}
u_emb = tf.nn.embedding_lookup(self.weights["embB"], query) # B * q_{ij}
"""
shape of m_embA and u_emb:
m_embA: (batch_size, self.n_memory, self. self.n_sentence, self.n_emb)
u_emb: (batch_size, self.n_sentence, self.n_emb)
should be:
mA: (batch_size, self.n_memory, self.n_emb)
u: (batch_size, self.n_emb)
"""
mA = tf.reduce_sum(m_embA, 2) # history embedding: m_i=\sum_j{A x_{ij}} shape:(batch_size, self.n_memory, self.n_emb)
u = tf.reduce_sum(u_emb, 1) # query embedding: u=\sum_j{B q_{ij}} shape:(batch_size, self.n_emb)
u_tmp = tf.transpose(tf.expand_dims(u, -1), [0, 2, 1]) # shape:(batch_size, 1, self.n_emb)
# get probability: match between utterance and memory
probs = tf.nn.softmax(tf.reduce_sum(mA * u_tmp, 2)) # shape of m * u_tmp: (batch_size, self.n_memory, self.n_emb), shape of probs: (batch_size, self.n_memory)
# get memory representation
m_embC = tf.nn.embedding_lookup(self.weights["embC"], history) # C * x_{ij}
mC = tf.reduce_sum(m_embC, 2) # shape:(batch_size, self.n_memory, self.n_emb)
o = tf.reduce_sum(tf.expand_dims(probs, -1) * mC, 1) # shape:(batch_size, self.n_emb)
# get prediction before softmax
a = tf.matmul(o + u, tf.transpose(self.weights["W"], [1, 0]))
return a
def defgradient(self):
# define cost
self.prediction = self.forward(self.in_history, self.in_query)
self.cost = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(
logits = self.prediction,
labels = tf.cast(self.in_answer, dtype=tf.float32)), name = "cost")
correct_prediction = tf.equal(tf.argmax(self.in_answer, 1), tf.argmax(self.prediction, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, dtype = tf.float32))
# cannot minimize cost directly, because nil items should be fixed
self._nil_vars = set([self.weights["embA"].name, self.weights["embC"].name, self.weights['W'].name])
# define gradient
def fillnil(grad):
grad = tf.convert_to_tensor(grad)
nil_embedding = tf.zeros([1, int(grad.get_shape()[1])])
grad_rest = tf.slice(grad, [1,0], [-1, -1])
return tf.concat(0, [nil_embedding, grad_rest])
self.finalgrads = []
self.gradvalue = {}
gradFromCost = tf.train.GradientDescentOptimizer(self.learning_rate).compute_gradients(self.cost)
#gradFromCost = [(tf.clip_by_norm(g, self.max_grad_norm), v) for g, v in gradFromCost]
for g, v in gradFromCost:
if v.name in self._nil_vars:
print 'v.name: ', v.name
g_ = fillnil(g)
self.gradvalue[v.name] = tf.reduce_mean(g_)
self.finalgrads.append((g_, v))
else:
self.finalgrads.append((g, v))
self.train_opt = tf.train.GradientDescentOptimizer(self.learning_rate).apply_gradients(self.finalgrads)
def train(self, data):
# run
init = tf.initialize_all_variables()
self.sess.run(init)
self.defgradient()
writer = tf.train.SummaryWriter(self.logs_path, graph=tf.get_default_graph())
tf.scalar_summary("accuracy", self.accuracy)
tf.scalar_summary("cost", self.cost)
tf.scalar_summary("embAgrad", self.gradvalue[self.weights["embA"].name])
tf.scalar_summary("embCgrad", self.gradvalue[self.weights["embC"].name])
summary_op = tf.merge_all_summaries()
for epoch in range(self.num_epoch):
epochCost = 0
batch_idx = 0
while True:
batch_idx += 1
batchh, batchq, batcha, finishOneEpoch = data.train.next_batch(self.batch_size)
if finishOneEpoch:
break
dataTofeed = {self.in_history: batchh,
self.in_query: batchq,
self.in_answer: batcha}
#batchcost, _ = self.sess.run([self.cost, self.train_opt], feed_dict = dataTofeed)
_, summary = self.sess.run([self.train_opt, summary_op], feed_dict = dataTofeed)
writer.add_summary(summary, epoch * self.batch_size + batch_idx)
epochCost += self.sess.run(self.cost, feed_dict = dataTofeed)
epochTrainAccuracy = self.sess.run(self.accuracy, feed_dict = {self.in_history: data.train.h,
self.in_query: data.train.q,
self.in_answer: data.train.a})
epochTestAccuracy = self.sess.run(self.accuracy, feed_dict = {self.in_history: data.test.h,
self.in_query: data.test.q,
self.in_answer: data.test.a})
print 'Epoch {0}: \nTraining Accuracy: {1} Test Accuracy: {2} Cost: {3}'.format(epoch,
round(epochTrainAccuracy, 4),
round(epochTestAccuracy, 4),
round(epochCost, 4))
#with self.sess.as_default():
# wA = self.sess.run(self.weights["embA"])
# wC = self.sess.run(self.weights["embC"])
# w = self.sess.run(self.weights["W"])
# print 'Average embA: {0}, embC: {1}, W: {2}'.format(tf.reduce_mean(wA).eval(), tf.reduce_mean(wC).eval(), tf.reduce_mean(w).eval())