-
Notifications
You must be signed in to change notification settings - Fork 0
/
LDSGM.py
160 lines (143 loc) · 8.7 KB
/
LDSGM.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import RobertaModel
from Layers.gcn import GraphConvolution
import pickle
class Label2Context(nn.Module): # label attention in the encoder
def __init__(self, label_embedding_size, attn_hidden_size):
super(Label2Context,self).__init__()
self.proj_label_top = nn.Linear(label_embedding_size, attn_hidden_size, bias=False)
self.proj_label_sec = nn.Linear(label_embedding_size, attn_hidden_size, bias=False)
self.proj_label_conn = nn.Linear(label_embedding_size, attn_hidden_size, bias=False)
def forward(self, input_data, label_repr, use_label_num):
seq_len = input_data.shape[1]
lb_num = use_label_num[1] - use_label_num[0]
embedding_label = label_repr[use_label_num[0]:use_label_num[1]]
if lb_num == 4:
embedding_label = self.proj_label_top(embedding_label)
if lb_num == 11:
embedding_label = self.proj_label_sec(embedding_label)
if lb_num == 102:
embedding_label = self.proj_label_conn(embedding_label)
embedding_label = embedding_label.transpose(0, 1)
input_data = F.normalize(input_data, dim=-1, p=2)
embedding_label = F.normalize(embedding_label, dim=0, p=2)
G = torch.matmul(input_data.reshape(-1, input_data.shape[-1]), embedding_label)
G = G.reshape(-1, seq_len, G.shape[-1])
max_G = G.max(-1, keepdim=True)[0] # [batch_size, seq_length, 1]
soft_max_G = torch.softmax(max_G, dim=1)
output = torch.sum(input_data * soft_max_G, axis=1)
return output
class Decoder(nn.Module):
def __init__(self, label_embedding_size, hidden_size, enc_hidden_size, dec_hidden_size):
super().__init__()
self.hidden_size = hidden_size
self.rnn = nn.GRU(enc_hidden_size+100, dec_hidden_size)
self.proj_out = nn.Linear(dec_hidden_size*2+100, dec_hidden_size)
def forward(self, input, hidden, encoder_outputs):
attention_outputs = encoder_outputs
rnn_input = torch.cat([attention_outputs, input], dim=-1)
output, hidden = self.rnn(rnn_input.unsqueeze(0), hidden.unsqueeze(0))
assert (output == hidden).all()
output = output.squeeze(0)
hidden = hidden.squeeze(0)
pred = self.proj_out(torch.cat([attention_outputs, output, input], dim=1))
return pred, hidden
class Model(nn.Module):
def __init__(self, config):
super(Model, self).__init__()
with open('label_graph.g', 'rb') as f:
label_graph = pickle.load(f)
self.config = config
self.label_graph = torch.from_numpy(label_graph).float().to(config.device)
self.label_embedding = nn.Parameter(torch.randn(config.label_num, config.label_embedding_size, dtype=torch.float32))
nn.init.xavier_normal_(self.label_embedding.data)
self.bert = RobertaModel.from_pretrained(config.bert_path)
for param in self.bert.parameters():
param.requires_grad = config.finetune_bert
self.gcn_layers = nn.ModuleList()
for i in range(config.num_gcn_layer):
self.gcn_layers.append(GraphConvolution(config.label_embedding_size, config.label_embedding_size))
# top-down manner Decoder
self.decoder = Decoder(config.label_embedding_size, config.attn_hidden_size, config.enc_hidden_size, config.dec_hidden_size)
# bottom-up manner auxilary Decoder
self.decoder_reverse = Decoder(config.label_embedding_size, config.attn_hidden_size, config.enc_hidden_size,config.dec_hidden_size)
self.label2context = Label2Context(config.label_embedding_size, config.attn_hidden_size)
self.fc1_top = nn.Linear(config.dec_hidden_size, config.n_top)
self.fc1_sec = nn.Linear(config.dec_hidden_size, config.n_sec)
self.fc1_conn = nn.Linear(config.dec_hidden_size, config.n_conn)
self.fc2_top = nn.Linear(config.dec_hidden_size, config.n_top)
self.fc2_sec = nn.Linear(config.dec_hidden_size, config.n_sec)
self.fc2_conn = nn.Linear(config.dec_hidden_size, config.n_conn)
self.softmax = nn.Softmax(dim=1)
self.dropout = nn.Dropout(0.2)
def forward(self, x, argmask):
context = x[0] #
mask = x[2]
bert_out = self.bert(context, attention_mask=mask)
hidden_last, pooled, hidden_all = bert_out.last_hidden_state, bert_out.pooler_output, bert_out.hidden_states
hidden_cls, encode_out = hidden_last[:, 0, :], hidden_last[:, 1:, :]
encode_out = self.dropout(encode_out)
for gcn_layer in self.gcn_layers:
label_repr = F.relu(gcn_layer(self.label_embedding, self.label_graph))
bs = hidden_last.shape[0]
hidden_g = hidden_cls
hidden = hidden_cls
# input = torch.FloatTensor(torch.zeros(bs, self.label_embedding.shape[1])).to(self.config.device)
# input_reverse = torch.FloatTensor(torch.zeros(bs, self.label_embedding.shape[1])).to(self.config.device)
input = label_repr[0: 4].sum(dim=0, keepdim=True).repeat(bs, 1)
input_reverse = label_repr[15: 117].sum(dim=0, keepdim=True).repeat(bs, 1)
hidden_reverse = hidden_cls
out_top, out_sec, out_conn = 0, 0, 0
for i in range(3):
if i == 0: # top-level prediction
label2context_out = self.label2context(encode_out, label_repr, [0, 4])
pred, output = self.decoder(input, hidden.contiguous(), label2context_out)
hidden = output
out_top = self.fc1_top(pred)
t = self.softmax(out_top).unsqueeze(-1).expand(out_top.shape[0], out_top.shape[1], label_repr.shape[1])
input = (t * label_repr[0:4]).sum(dim=1, keepdim=False)
if i == 1: # second-level prediction
hidden = hidden + hidden_g
label2context_out = self.label2context(encode_out, label_repr, [4, 15])
pred, output = self.decoder(input, hidden.contiguous(), label2context_out)
hidden = output
out_sec = self.fc1_sec(pred)
t = self.softmax(out_sec).unsqueeze(-1).expand(out_sec.shape[0], out_sec.shape[1], label_repr.shape[1])
input = (t * label_repr[4:15]).sum(dim=1, keepdim=False)
if i == 2: # conn-level prediction
hidden = hidden + hidden_g
label2context_out = self.label2context(encode_out, label_repr, [15, 117])
pred, output = self.decoder(input, hidden.contiguous(), label2context_out)
out_conn = self.fc1_conn(pred)
t = self.softmax(out_conn).unsqueeze(-1).expand(out_conn.shape[0], out_conn.shape[1],
label_repr.shape[1])
input = (t * label_repr[15:117]).sum(dim=1, keepdim=False)
# the bottom-up manner auxilary decoder
for i in range(3):
if i == 0:
hidden_reverse = hidden_reverse.contiguous()
label2context_out = self.label2context(encode_out, label_repr, [15, 117])
pred_reverse, output_reverse = self.decoder_reverse(input_reverse, hidden_reverse,label2context_out)
hidden_reverse = output_reverse
out_conn_reverse = self.fc2_conn(pred_reverse)
t = self.softmax(out_conn_reverse).unsqueeze(-1).expand(out_conn_reverse.shape[0], out_conn_reverse.shape[1], label_repr.shape[1])
input_reverse = (t * label_repr[15:117]).sum(dim=1, keepdim=False)
if i == 1:
hidden_reverse = hidden_reverse + hidden_g
hidden_reverse = hidden_reverse.contiguous()
label2context_out = self.label2context(encode_out, label_repr, [4, 15])
pred_reverse, output_reverse = self.decoder_reverse(input_reverse, hidden_reverse, label2context_out)
hidden_reverse = output_reverse
out_sec_reverse = self.fc2_sec(pred_reverse)
t = self.softmax(out_sec_reverse).unsqueeze(-1).expand(out_sec_reverse.shape[0],
out_sec_reverse.shape[1], label_repr.shape[1])
input_reverse = (t * label_repr[4:15]).sum(dim=1, keepdim=False)
if i == 2:
hidden_reverse = hidden_reverse + hidden_g
hidden_reverse = hidden_reverse.contiguous()
label2context_out = self.label2context(encode_out, label_repr, [0, 4])
pred_reverse, output_reverse = self.decoder_reverse(input_reverse, hidden_reverse, label2context_out)
out_top_reverse = self.fc2_top(pred_reverse)
return out_top, out_sec, out_conn, out_top_reverse, out_sec_reverse, out_conn_reverse