-
Notifications
You must be signed in to change notification settings - Fork 3
/
gnn_transformer.py
205 lines (166 loc) · 8.51 KB
/
gnn_transformer.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 xmlrpc.client import FastParser
import torch.nn as nn
import torch.nn.functional as F
import math
import torch
from math import *
from combination_layer import CombinationLayer
def position_encoding(length, dmodel):
pos = []
for i in range(length):
pos_cur = []
for j in range(dmodel // 2):
pos_cur.append(math.sin(i/(10000 ** (2 * j / dmodel))))
pos_cur.append(math.cos(i/(10000 ** (2 * j / dmodel))))
pos.append(pos_cur)
pos_tensor = torch.tensor(pos)
return pos_tensor
class Encoder(nn.Module):
def __init__(self, args, pad_token_id):
super(Encoder, self).__init__()
self.dropout_rate = args.dropout_rate
self.sou_len = args.sou_len
self.att_len = args.att_len
self.ast_change_len = args.ast_change_len
self.sub_token_len = args.sub_token_len
self.embedding_dim = args.embedding_dim
self.pad_token_id = pad_token_id
self.embedding = nn.Embedding(
num_embeddings=args.vocab_size, embedding_dim=args.embedding_dim, padding_idx=pad_token_id)
self.ast_change_embedding = nn.Embedding(
num_embeddings=args.ast_change_vocab_size, embedding_dim=args.embedding_dim, padding_idx=pad_token_id)
self.pos_encode = position_encoding(args.sou_len, self.embedding_dim)
self.mark_embedding = nn.Embedding(num_embeddings=4, embedding_dim=args.embedding_dim, padding_idx=0)
self.lstm = nn.LSTM(input_size=args.embedding_dim, hidden_size=args.embedding_dim, num_layers=3, batch_first=True)
self.combination_list1 = nn.ModuleList([Combination(h=args.num_head, d_model=args.embedding_dim) for i in range(6)])
self.combination_list2 = nn.ModuleList([Combination(h=args.num_head, d_model=args.embedding_dim) for i in range(6)])
self.gcn_list = nn.ModuleList([GCN(args.embedding_dim, dropout_rate=0.2) for i in range(6)])
def forward(self, input_token, sou_mask, attr, mark, ast_change, edge, sub_token):
input_em = self.embedding(input_token) + self.pos_encode.to(input_token.device)
# batch * len * emdbedding
mark_em = self.mark_embedding(mark)
ast_change_em = self.ast_change_embedding(ast_change)
sub_token_em = self.embedding(sub_token)
# batch
for i in range(len(self.gcn_list)):
input_em = self.combination_list2[i](input_em, input_em, mark_em)
graph_em = torch.cat((input_em, sub_token_em, ast_change_em), dim = 1)
# batch * (code_len + sub_token_len + change_len) * embedding
input_em, sub_token_em, ast_change_em= self.gcn_list[i](graph_em, edge, self.sou_len, self.sub_token_len, self.ast_change_len)
return input_em, sub_token_em
class GCN(nn.Module):
def __init__(self, dmodel, dropout_rate=0.1):
super(GCN, self).__init__()
self.dmodel = dmodel
self.fc1 = nn.Linear(dmodel, dmodel)
self.fc2 = nn.Linear(dmodel, dmodel)
self.dropout = nn.Dropout(dropout_rate)
self.layernorm = nn.LayerNorm(dmodel)
def forward(self, graph_em, edge, code_len, sub_token_len, ast_change_len):
# graph_em: batch * len * embeding
# edge: batch * len * len
total_len = graph_em.size(1)
x = self.fc1(graph_em)
x = torch.bmm(edge.float(), x)
x = self.fc2(x)
res = self.layernorm(self.dropout(x) + graph_em)
assert res.size(1) == code_len + sub_token_len + ast_change_len
return res[:,:code_len], res[:,code_len:code_len + sub_token_len], res[:, code_len + sub_token_len:]
class Decoder(nn.Module):
def __init__(self, args, pad_token_id):
super(Decoder, self).__init__()
self.embedding_dim = args.embedding_dim
self.pad_token_id = pad_token_id
self.embedding = nn.Embedding(
num_embeddings=args.vocab_size, embedding_dim=args.embedding_dim)
self.pos_encode = position_encoding(args.tar_len, self.embedding_dim)
self.tar_mask_pos = torch.ones(args.tar_len, args.tar_len)
for i in range(args.tar_len):
self.tar_mask_pos[i][i+1:] = 0
self.attention_list = nn.ModuleList(
[Attention(dmodel=args.embedding_dim, num_head=args.num_head) for i in range(6)])
self.cross_attention_list = nn.ModuleList(
[Attention(dmodel=args.embedding_dim, num_head=args.num_head) for i in range(6)])
self.feed_forward_list = nn.ModuleList(
[FeedForward(args.embedding_dim) for i in range(6)])
def forward(self, output_token, input_em, sou_mask, tar_mask_pad):
# input: batch * len
if torch.cuda.is_available():
output_em = self.embedding(output_token) + self.pos_encode.cuda(output_token.device)
else:
output_em = self.embedding(output_token) + self.pos_encode
# batch * len * emdbedding
self.tar_mask_pos = self.tar_mask_pos.to(output_token.device)
tar_mask = torch.logical_and(tar_mask_pad.unsqueeze(1).unsqueeze(1), self.tar_mask_pos.unsqueeze(0).unsqueeze(0))
for i in range(len(self.attention_list)):
output_em = self.attention_list[i](output_em, output_em, output_em, tar_mask)
output_em = self.cross_attention_list[i](output_em, input_em, input_em, sou_mask)
output_em = self.feed_forward_list[i](output_em)
return output_em
class Attention(nn.Module):
def __init__(self, dmodel, num_head, dropout_rate=0.1):
super(Attention, self).__init__()
self.fc_q = nn.Linear(dmodel, dmodel)
self.fc_k = nn.Linear(dmodel, dmodel)
self.fc_v = nn.Linear(dmodel, dmodel)
self.fc_o = nn.Linear(dmodel, dmodel)
self.layernorm = nn.LayerNorm(dmodel)
self.dropout = nn.Dropout(dropout_rate)
self.num_head = num_head
assert dmodel % self.num_head == 0
self.dhead = dmodel // self.num_head
def forward(self, query, key, value, mask):
old_query = query
batch_size = query.size(0)
q_len = query.size(1)
kv_len = key.size(1)
query = self.fc_q(query)
key = self.fc_k(key)
value = self.fc_v(value)
# batch * len * dmodel
query = query.view(batch_size, q_len, self.num_head, self.dhead).transpose(1,2)
key = key.view(batch_size, kv_len, self.num_head, self.dhead).transpose(1,2)
value = value.view(batch_size, kv_len, self.num_head, self.dhead).transpose(1,2)
weight = torch.matmul(query, key.transpose(-2,-1)) / sqrt(self.dhead)
# batch * num_head * len * len
if len(mask.shape) < 4:
mask = mask.unsqueeze(1).unsqueeze(1)
weight = weight.masked_fill(mask == 0, -1e9)
weight_softmax = F.softmax(weight, dim = -1)
weighted_sum = torch.matmul(weight_softmax, value)
# batch * num_head * len *embedding
output = weighted_sum.transpose(1,2).contiguous().view(batch_size,q_len,self.num_head * self.dhead)
output = self.fc_o(output)
return self.layernorm(self.dropout(output) + old_query)
class FeedForward(nn.Module):
def __init__(self, dmodel, dropout_rate=0.1):
super(FeedForward, self).__init__()
self.fc1 = nn.Linear(dmodel, 4 * dmodel)
self.fc2 = nn.Linear(4 * dmodel, dmodel)
self.dropout = nn.Dropout(dropout_rate)
self.layernorm = nn.LayerNorm(dmodel)
def forward(self, input_em):
x = self.fc1(input_em)
x = F.relu(x)
x = self.fc2(x)
return self.layernorm(self.dropout(x) + input_em)
class Combination(nn.Module):
def __init__(self, h, d_model, dropout_rate=0.1):
super().__init__()
assert d_model % h == 0
self.d_k = d_model // h
self.h = h
self.linear_layers = nn.ModuleList([nn.Linear(d_model, d_model) for _ in range(3)])
self.output_linear = nn.Linear(d_model, d_model)
self.combination = CombinationLayer()
self.dropout = nn.Dropout(p=dropout_rate)
self.layernorm = nn.LayerNorm(d_model)
def forward(self, query, key, value, mask=None):
old_query = query
batch_size = query.size(0)
query, key, value = [l(x).view(batch_size, -1, self.h, self.d_k).transpose(1, 2)
for l, x in zip(self.linear_layers, (query, key, value))]
x = self.combination(query, key, value, dropout=self.dropout)
x = x.transpose(1, 2).contiguous().view(batch_size, -1, self.h * self.d_k)
output = self.output_linear(x)
return self.layernorm(self.dropout(output) + old_query)