-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathgenerator.py
164 lines (147 loc) · 8.14 KB
/
generator.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
# coding: utf-8
import os
import random
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from baseRNN import BaseRNN
from utils import ApproximateEmbeddingLayer
from utils import SYM_GO
class Generator(BaseRNN):
"""Generator """
# baseRNN中SYM_PAD = 0
def __init__(self, vocab_size, max_len, emb_dim, hidden_dim,
n_layers=1, rnn_cell='gru', dropout_p=0, use_attention=False):
super(Generator, self).__init__(vocab_size, hidden_dim, dropout_p, n_layers, rnn_cell)
self.max_len = max_len
self.hidden_dim = hidden_dim
# decoder一般都是单向的
self.rnn = self.rnn_cell(emb_dim, hidden_dim, n_layers,
batch_first=True, dropout=dropout_p, bidirectional=False)
# approximate embedding layer
self.ael = ApproximateEmbeddingLayer(hidden_dim, vocab_size)
def forward(self, init_state, word_embeddings):
"""
Args:
init_state: 如果encoder的rnn cell选用lstm，则(ht, ct)；如果用gru，则ht
shape都是(num_layers, batch, num_directions*hidden_size)
NOTED: 最好是encoder选用几层，decoder就选择几层；decoder设置其hidden_size=encoder hidden_size*2
传入word_embeddings矩阵后，首先得到go的输入vector，不断的rnn_cell生成出新词，decode到最大长度为止
mode: 有三种模式，
(1) 对抗训练: 这时decoder只需要y_input = [go]即可, 但reference = y
"""
if isinstance(init_state, tuple):
assert init_state[0].size(2) == self.hidden_dim, "If encoder is bidirectional, \
decoder hidden size should be 2 * encoder hidden state."
batch_size = init_state[0].size(1)
else:
assert init_state.size(2) == self.hidden_dim, "If encoder is bidirectional, \
decoder hidden size should be 2 * encoder hidden state."
batch_size = init_state.size(1)
go_inputs = Variable(torch.ones(batch_size, 1).long()*SYM_GO, requires_grad=False).cuda()
embedded_input = word_embeddings(go_inputs) # [B, 1, emb_dim]
state = init_state
outputs = [] # a list of approximate word embeddings(max_len)
for i in range(self.max_len):
output, state = self.rnn(embedded_input, state)
# output = [B, 1, hidden_dim], state有可能是tuple, 每个元素(num_layers, B, hidden_dim)
_, embedded_input = self.ael(output.squeeze(1), word_embeddings) # [B, emb_dim]
outputs.append(embedded_input)
embedded_input = embedded_input.unsqueeze(1) # [B, 1, emb_dim]
return torch.stack(outputs, dim=1) # [B, max_len, emb_dim]
# 输出的是每一时刻得到的approximate embedding
def inference(self, init_state, word_embeddings):
"""
Args:
init_state: 如果encoder的rnn cell选用lstm，则(ht, ct)；如果用gru，则ht
shape都是(num_layers, batch, num_directions*hidden_size)
NOTED: 最好是encoder选用几层，decoder就选择几层；decoder设置其hidden_size=encoder hidden_size*2
传入word_embeddings矩阵后，首先得到go的输入vector，不断的rnn_cell生成出新词，decode到最大长度为止
mode: 有三种模式，
(1) 对抗训练: 这时decoder只需要y_input = [go]即可, 但reference = y
(2) 有监督预训练: 这时y_input=[go] + y; reference = y + [eos]
(3) 预测: 这时只需要给[go], 不需要reference
"""
if isinstance(init_state, tuple):
assert init_state[0].size(2) == self.hidden_dim, "If encoder is bidirectional, \
decoder hidden size should be 2 * encoder hidden state."
batch_size = init_state[0].size(1)
else:
assert init_state.size(2) == self.hidden_dim, "If encoder is bidirectional, \
decoder hidden size should be 2 * encoder hidden state."
batch_size = init_state.size(1)
go_inputs = Variable(torch.ones(batch_size, 1).long()*SYM_GO, requires_grad=False).cuda()
embedded_input = word_embeddings(go_inputs) # [B, 1, emb_dim]
state = init_state
outputs = [] # a list of approximate word embeddings(max_len)
for i in range(self.max_len):
output, state = self.rnn(embedded_input, state)
# output = [B, 1, hidden_dim], state有可能是tuple, 每个元素(num_layers, B, hidden_dim)
word_distribution, embedded_input = self.ael(output.squeeze(1), word_embeddings) # [B, emb_dim]
# [B, vocab_size]
outputs.append(word_distribution)
embedded_input = embedded_input.unsqueeze(1)
return torch.stack(outputs, dim=1) # [B, max_len, vocab_size]
# 预测阶段，输出的是每一时刻最大概率出现的词
def update(self, s, embedded_i):
# s = (output, st)
# embedded_i = [B, T, emb_dim]
return self.rnn(embedded_i, s[1])
def predict(self, s, word_embeddings):
logp, _ = self.ael(s[0].squeeze(1), word_embeddings)
return logp # [B, vocab_size]
def supervise(self, dec_inputs, init_state, word_embeddings):
"""
Args:
init_state: 如果encoder的rnn cell选用lstm，则(ht, ct)；如果用gru，则ht
shape都是(num_layers, batch, num_directions*hidden_size)
NOTED: 最好是encoder选用几层，decoder就选择几层；decoder设置其hidden_size=encoder hidden_size*2
传入word_embeddings矩阵后，首先得到go的输入vector，不断的rnn_cell生成出新词，decode到最大长度为止
references = [B, T, emb_dim]
mode: 有三种模式，
(1) 对抗训练: 这时decoder只需要y_input = [go]即可, 但reference = y
(2) 有监督预训练: 这时y_input=[go] + y; reference = y + [eos]
(3) 预测: 这时只需要给[go], 不需要reference
"""
# assert references.size(1) == (self.max_len - 1), "When supervise learning, length of references should be (max_len - 1)."
if isinstance(init_state, tuple):
assert init_state[0].size(2) == self.hidden_dim, "If encoder is bidirectional, \
decoder hidden size should be 2 * encoder hidden state."
batch_size = init_state[0].size(1)
else:
assert init_state.size(2) == self.hidden_dim, "If encoder is bidirectional, \
decoder hidden size should be 2 * encoder hidden state."
batch_size = init_state.size(1)
ref_inputs = list(torch.split(dec_inputs, 1, dim=1)) # a list of [B, 1, emb_dim]
go_inputs = Variable(torch.ones(batch_size, 1).long()*SYM_GO, requires_grad=False).cuda()
embedded_inputs = [word_embeddings(go_inputs)] + ref_inputs
state = init_state
outputs = [] # a list of approximate word embeddings(max_len)
for i in range(self.max_len+1):
output, state = self.rnn(embedded_inputs[i], state)
# output = [B, 1, hidden_dim], state有可能是tuple, 每个元素(num_layers, B, hidden_dim)
word_distribution, _ = self.ael(output.squeeze(1), word_embeddings) # [B, emb_dim]
# [B, vocab_size]
outputs.append(word_distribution)
# embedded_input = embedded_input.unsqueeze(1)
return torch.stack(outputs, dim=1) # [B, max_len+1, vocab_size]
# 有监督阶段，输出是词向量分布概率，后续需要计算NLL
def init_params(self):
for param in self.parameters():
param.data.uniform_(-0.05, 0.05)
if __name__ == '__main__':
vocab_size = 10
max_len = 8
emb_dim = 6
hidden_dim = 5*2
n_layers = 3
# batch = 4
decoder = Generator(vocab_size, max_len, emb_dim, hidden_dim, n_layers, rnn_cell='gru')
init_state_ht = Variable(torch.rand(n_layers, 4, hidden_dim).uniform_(-0.05, 0.05))
# init_state_ct = Variable(torch.rand(3, 4, 5*2).uniform_(-0.05, 0.05))
embeddings = nn.Embedding(vocab_size, emb_dim)
# embedded_inputs = embeddings(inputs)
res = decoder(init_state_ht, embeddings)
print res.shape