forked from PaddlePaddle/models
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
202 lines (174 loc) · 7.64 KB
/
model.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
import six
import numpy as np
import paddle.fluid as fluid
import utils.layers as layers
class Net(object):
def __init__(self, max_turn_num, max_turn_len, vocab_size, emb_size,
stack_num, channel1_num, channel2_num):
self._max_turn_num = max_turn_num
self._max_turn_len = max_turn_len
self._vocab_size = vocab_size
self._emb_size = emb_size
self._stack_num = stack_num
self._channel1_num = channel1_num
self._channel2_num = channel2_num
self._feed_names = []
self.word_emb_name = "shared_word_emb"
self.use_stack_op = True
self.use_mask_cache = True
self.use_sparse_embedding = True
def create_py_reader(self, capacity, name):
# turns ids
shapes = [[-1, self._max_turn_len, 1]
for i in six.moves.xrange(self._max_turn_num)]
dtypes = ["int64" for i in six.moves.xrange(self._max_turn_num)]
# turns mask
shapes += [[-1, self._max_turn_len, 1]
for i in six.moves.xrange(self._max_turn_num)]
dtypes += ["float32" for i in six.moves.xrange(self._max_turn_num)]
# response ids, response mask, label
shapes += [[-1, self._max_turn_len, 1], [-1, self._max_turn_len, 1],
[-1, 1]]
dtypes += ["int64", "float32", "float32"]
py_reader = fluid.layers.py_reader(
capacity=capacity,
shapes=shapes,
lod_levels=[0] * (2 * self._max_turn_num + 3),
dtypes=dtypes,
name=name,
use_double_buffer=True)
data_vars = fluid.layers.read_file(py_reader)
self.turns_data = data_vars[0:self._max_turn_num]
self.turns_mask = data_vars[self._max_turn_num:2 * self._max_turn_num]
self.response = data_vars[-3]
self.response_mask = data_vars[-2]
self.label = data_vars[-1]
return py_reader
def create_data_layers(self):
self._feed_names = []
self.turns_data = []
for i in six.moves.xrange(self._max_turn_num):
name = "turn_%d" % i
turn = fluid.layers.data(
name=name, shape=[self._max_turn_len, 1], dtype="int64")
self.turns_data.append(turn)
self._feed_names.append(name)
self.turns_mask = []
for i in six.moves.xrange(self._max_turn_num):
name = "turn_mask_%d" % i
turn_mask = fluid.layers.data(
name=name, shape=[self._max_turn_len, 1], dtype="float32")
self.turns_mask.append(turn_mask)
self._feed_names.append(name)
self.response = fluid.layers.data(
name="response", shape=[self._max_turn_len, 1], dtype="int64")
self.response_mask = fluid.layers.data(
name="response_mask",
shape=[self._max_turn_len, 1],
dtype="float32")
self.label = fluid.layers.data(name="label", shape=[1], dtype="float32")
self._feed_names += ["response", "response_mask", "label"]
def get_feed_names(self):
return self._feed_names
def set_word_embedding(self, word_emb, place):
word_emb_param = fluid.global_scope().find_var(
self.word_emb_name).get_tensor()
word_emb_param.set(word_emb, place)
def create_network(self):
mask_cache = dict() if self.use_mask_cache else None
response_emb = fluid.layers.embedding(
input=self.response,
size=[self._vocab_size + 1, self._emb_size],
is_sparse=self.use_sparse_embedding,
param_attr=fluid.ParamAttr(
name=self.word_emb_name,
initializer=fluid.initializer.Normal(scale=0.1)))
# response part
Hr = response_emb
Hr_stack = [Hr]
for index in six.moves.xrange(self._stack_num):
Hr = layers.block(
name="response_self_stack" + str(index),
query=Hr,
key=Hr,
value=Hr,
d_key=self._emb_size,
q_mask=self.response_mask,
k_mask=self.response_mask,
mask_cache=mask_cache)
Hr_stack.append(Hr)
# context part
sim_turns = []
for t in six.moves.xrange(self._max_turn_num):
Hu = fluid.layers.embedding(
input=self.turns_data[t],
size=[self._vocab_size + 1, self._emb_size],
is_sparse=self.use_sparse_embedding,
param_attr=fluid.ParamAttr(
name=self.word_emb_name,
initializer=fluid.initializer.Normal(scale=0.1)))
Hu_stack = [Hu]
for index in six.moves.xrange(self._stack_num):
# share parameters
Hu = layers.block(
name="turn_self_stack" + str(index),
query=Hu,
key=Hu,
value=Hu,
d_key=self._emb_size,
q_mask=self.turns_mask[t],
k_mask=self.turns_mask[t],
mask_cache=mask_cache)
Hu_stack.append(Hu)
# cross attention
r_a_t_stack = []
t_a_r_stack = []
for index in six.moves.xrange(self._stack_num + 1):
t_a_r = layers.block(
name="t_attend_r_" + str(index),
query=Hu_stack[index],
key=Hr_stack[index],
value=Hr_stack[index],
d_key=self._emb_size,
q_mask=self.turns_mask[t],
k_mask=self.response_mask,
mask_cache=mask_cache)
r_a_t = layers.block(
name="r_attend_t_" + str(index),
query=Hr_stack[index],
key=Hu_stack[index],
value=Hu_stack[index],
d_key=self._emb_size,
q_mask=self.response_mask,
k_mask=self.turns_mask[t],
mask_cache=mask_cache)
t_a_r_stack.append(t_a_r)
r_a_t_stack.append(r_a_t)
t_a_r_stack.extend(Hu_stack)
r_a_t_stack.extend(Hr_stack)
if self.use_stack_op:
t_a_r = fluid.layers.stack(t_a_r_stack, axis=1)
r_a_t = fluid.layers.stack(r_a_t_stack, axis=1)
else:
for index in six.moves.xrange(len(t_a_r_stack)):
t_a_r_stack[index] = fluid.layers.unsqueeze(
input=t_a_r_stack[index], axes=[1])
r_a_t_stack[index] = fluid.layers.unsqueeze(
input=r_a_t_stack[index], axes=[1])
t_a_r = fluid.layers.concat(input=t_a_r_stack, axis=1)
r_a_t = fluid.layers.concat(input=r_a_t_stack, axis=1)
# sim shape: [batch_size, 2*(stack_num+1), max_turn_len, max_turn_len]
sim = fluid.layers.matmul(
x=t_a_r, y=r_a_t, transpose_y=True, alpha=1 / np.sqrt(200.0))
sim_turns.append(sim)
if self.use_stack_op:
sim = fluid.layers.stack(sim_turns, axis=2)
else:
for index in six.moves.xrange(len(sim_turns)):
sim_turns[index] = fluid.layers.unsqueeze(
input=sim_turns[index], axes=[2])
# sim shape: [batch_size, 2*(stack_num+1), max_turn_num, max_turn_len, max_turn_len]
sim = fluid.layers.concat(input=sim_turns, axis=2)
final_info = layers.cnn_3d(sim, self._channel1_num, self._channel2_num)
loss, logits = layers.loss(final_info, self.label)
return loss, logits