-
Notifications
You must be signed in to change notification settings - Fork 13
/
model.py
165 lines (127 loc) · 6.7 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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.modules import Transformer
class SynPG(nn.Module):
def __init__(self, vocab_size, em_size, word_dropout=0.4, dropout=0.1):
super(SynPG, self).__init__()
self.vocab_size = vocab_size
self.em_size = em_size
self.word_dropout = word_dropout
self.dropout = dropout
# vcocabulary embedding
self.embedding_encoder = nn.Embedding(vocab_size, em_size)
self.embedding_decoder = nn.Embedding(vocab_size, em_size)
# positional encoding
self.pos_encoder = PositionalEncoding(em_size, dropout=0.0)
self.transformer = Transformer(d_model=em_size, nhead=6, dropout=dropout)
# linear Transformation
self.linear = nn.Linear(em_size, vocab_size)
self.init_weights()
def init_weights(self):
initrange = 0.1
# initialize cocabulary matrix weight
self.embedding_encoder.weight.data.uniform_(-initrange, initrange)
self.embedding_decoder.weight.data.uniform_(-initrange, initrange)
# initialize linear weight
self.linear.weight.data.uniform_(-initrange, initrange)
self.linear.bias.data.fill_(0)
def load_embedding(self, embedding):
self.embedding_encoder.weight.data.copy_(torch.from_numpy(embedding))
self.embedding_decoder.weight.data.copy_(torch.from_numpy(embedding))
def store_grad_norm(self, grad):
norm = torch.norm(grad, 2, 1)
self.grad_norm = norm.detach().data.mean()
return grad
def generate_square_mask(self, max_sent_len, max_synt_len):
size = max_sent_len + max_synt_len + 2
mask = torch.zeros((size, size))
mask[:max_sent_len, max_sent_len:] = float("-inf")
mask[max_sent_len:, :max_sent_len] = float("-inf")
return mask
def forward(self, sents, synts, targs):
batch_size = sents.size(0)
max_sent_len = sents.size(1)
max_synt_len = synts.size(1) - 2 # count without <sos> and <eos>
max_targ_len = targs.size(1) - 2 # count without <sos> and <eos>
# apply word dropout
drop_mask = torch.bernoulli(self.word_dropout*torch.ones(max_sent_len)).bool().cuda()
sents = sents.masked_fill(drop_mask, 0)
# sentence, syntax => embedding
sent_embeddings = self.embedding_encoder(sents).transpose(0, 1) * np.sqrt(self.em_size)
synt_embeddings = self.embedding_encoder(synts).transpose(0, 1) * np.sqrt(self.em_size)
synt_embeddings = self.pos_encoder(synt_embeddings)
en_embeddings = torch.cat((sent_embeddings, synt_embeddings), dim=0)
# record gradient
if en_embeddings.requires_grad:
en_embeddings.register_hook(self.store_grad_norm)
# do not allow cross attetion
src_mask = self.generate_square_mask(max_sent_len, max_synt_len).cuda()
# target => embedding
de_embeddings = self.embedding_decoder(targs[:, :-1]).transpose(0, 1) * np.sqrt(self.em_size)
de_embeddings = self.pos_encoder(de_embeddings)
# sequential mask
tgt_mask = self.transformer.generate_square_subsequent_mask(max_targ_len+1).cuda()
# forward
outputs = self.transformer(en_embeddings, de_embeddings, src_mask=src_mask, tgt_mask=tgt_mask)
# apply linear layer to vcocabulary size
outputs = outputs.transpose(0, 1)
outputs = self.linear(outputs.contiguous().view(-1, self.em_size))
outputs = outputs.view(batch_size, max_targ_len+1, self.vocab_size)
return outputs
def generate(self, sents, synts, max_len, sample=True, temp=0.5):
batch_size = sents.size(0)
max_sent_len = sents.size(1)
max_synt_len = synts.size(1) - 2 # count without <sos> and <eos>
max_targ_len = max_len
# output index starts with <sos>
idxs = torch.zeros((batch_size, max_targ_len+2), dtype=torch.long).cuda()
idxs[:, 0] = 1
# sentence, syntax => embedding
sent_embeddings = self.embedding_encoder(sents).transpose(0, 1) * np.sqrt(self.em_size)
synt_embeddings = self.embedding_encoder(synts).transpose(0, 1) * np.sqrt(self.em_size)
synt_embeddings = self.pos_encoder(synt_embeddings)
en_embeddings = torch.cat((sent_embeddings, synt_embeddings), dim=0)
# do not allow cross attetion
src_mask = self.generate_square_mask(max_sent_len, max_synt_len).cuda()
# starting index => embedding
de_embeddings = self.embedding_decoder(idxs[:, :1]).transpose(0, 1) * np.sqrt(self.em_size)
de_embeddings = self.pos_encoder(de_embeddings)
# sequential mask
tgt_mask = self.transformer.generate_square_subsequent_mask(de_embeddings.size(0)).cuda()
# encode
memory = self.transformer.encoder(en_embeddings, mask=src_mask)
# auto-regressively generate output
for i in range(1, max_targ_len+2):
# decode
outputs = self.transformer.decoder(de_embeddings, memory, tgt_mask=tgt_mask)
outputs = self.linear(outputs[-1].contiguous().view(-1, self.em_size))
# get argmax index or sample index
if not sample:
values, idx = torch.max(outputs, 1)
else:
probs = F.softmax(outputs/temp, dim=1)
idx = torch.multinomial(probs, 1).squeeze(1)
# save to output index
idxs[:, i] = idx
# concatenate index to decoding
de_embeddings = self.embedding_decoder(idxs[:, :i+1]).transpose(0, 1) * np.sqrt(self.em_size)
de_embeddings = self.pos_encoder(de_embeddings)
# new sequential mask
tgt_mask = self.transformer.generate_square_subsequent_mask(de_embeddings.size(0)).cuda()
return idxs[:, 1:]
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout=0.1, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-np.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0).transpose(0, 1)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:x.size(0), :]
return self.dropout(x)