-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
130 lines (110 loc) · 5.72 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
from transformers import BertPreTrainedModel
import torch.nn.functional as F
from csrl_bert import CSRLBert
from model_utils import *
class CSAGN(BertPreTrainedModel):
def __init__(self, config, wp, wf, n_speakers=2, mode="max", g_dim=100, num_layers=4, intra_loss=False,
inter_loss=False, inter_relation_num=4):
super().__init__(config)
self.config = config
self.bert = CSRLBert(config)
self.wp = wp
self.wf = wf
self.mode = mode
self.g_dim = g_dim
self.intra_loss = intra_loss
self.inter_loss = inter_loss
self.inter_relation_num = inter_relation_num
self.edge_att = EdgeAtt(input_dim=config.hidden_size, wp=wp, wf=wf)
self.gcn = GraphNet(config.hidden_size, g_dim, n_speakers)
self.pred_aware_att = TencentSelfAtt(hidden_size=config.hidden_size, num_hidden_layers=num_layers, in_config=config)
self.token_embedding_proj = nn.Linear(2 * config.hidden_size, config.hidden_size)
self.token_fusion_layer = MHA(config)
self.speaker_embedding_proj = nn.Linear(config.hidden_size + g_dim, config.hidden_size, bias=True)
self.speaker_fusion_layer = MHA(config)
edge_type_to_idx = {}
for j in range(1, n_speakers + 1):
for k in range(1, n_speakers + 1):
edge_type_to_idx[str(j) + str(k) + '0'] = len(edge_type_to_idx)
edge_type_to_idx[str(j) + str(k) + '1'] = len(edge_type_to_idx)
self.edge_type_to_idx = edge_type_to_idx
self.classifier = nn.Sequential(
nn.Linear(2 * config.hidden_size, config.hidden_size),
nn.ReLU(),
nn.Dropout(config.hidden_dropout_prob),
nn.Linear(config.hidden_size, config.num_labels)
)
self.inter_classifier = nn.Linear(config.hidden_size + g_dim, inter_relation_num)
self.intra_proj = nn.Linear(4 * config.hidden_size, 2 * config.hidden_size)
def forward(
self,
input_ids=None,
token_type_ids=None,
attention_mask=None,
text_lens=None,
position_ids=None,
pred_ids=None,
head_mask=None,
inputs_embeds=None,
speaker_ids=None,
labels=None,
cls_vec=None,
utt_labels=None,
utt_mask=None,
last_label=None,
turn_ids=None
):
context_features = self.bert(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
pred_ids=pred_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
turn_ids=turn_ids
) # (bsz, token_num, dim)
bsz, seq_len, _ = context_features.shape
attention_mask = attention_mask.unsqueeze(1).unsqueeze(-1).expand((bsz, 1, seq_len, seq_len))
attention_mask = (1 - attention_mask) * -10000.0
# token level
last_utt_embed = self.pred_aware_att(context_features, attention_mask=utt_mask)[0]
y = self.token_embedding_proj(torch.cat([context_features, last_utt_embed], dim=-1))
token_level_feat = self.token_fusion_layer(y, y, attention_mask)[0]
# utterance level
utt_features, cls_list = extract_utt_feat(token_level_feat, cls_vec, mode=self.mode) # (bsz, turn_num, dim)
utt_node_feat, utt_edge_idx, utt_edge_norm, utt_edge_type, edge_index_lengths = batch_graphify(
utt_features, speaker_ids, text_lens, self.wp, self.wf, self.edge_type_to_idx, self.edge_att,
self.device
) # (turn_num, g_dim)
utt_graph_out = self.gcn(utt_node_feat, utt_edge_idx, utt_edge_norm, utt_edge_type) # (turn_num, g_dim)
# (bsz, seq_len, hidden_size)
speaker_gcn_embed = flatten_graph_out(utt_graph_out, cls_list, context_feat=context_features,
text_len=text_lens) # (bsz, seq_len, g_dim)
y = self.speaker_embedding_proj(torch.cat([token_level_feat, speaker_gcn_embed], dim=-1))
utt_level_feat = self.speaker_fusion_layer(y, y, attention_mask)[0]
logits = self.classifier(torch.cat([token_level_feat, utt_level_feat], dim=-1)) # (bsz, token_num, num_label)
prob = F.softmax(logits, dim=-1)
_, prediction = torch.max(prob, dim=-1)
if labels is not None:
log_probs = F.log_softmax(logits, dim=-1)
loss = F.nll_loss(log_probs.view(-1, self.config.num_labels), labels.view(-1), reduction='mean',
ignore_index=-100)
if self.inter_loss:
utt_labels = utt_labels[utt_labels > 0]
inter_log_probs = F.log_softmax(
self.inter_classifier(torch.cat([utt_node_feat, utt_graph_out], dim=-1)), dim=-1)
inter_loss = F.nll_loss(inter_log_probs.view(-1, self.inter_relation_num), utt_labels.view(-1),
reduction='mean', ignore_index=0)
# print("inter_loss: {}".format(inter_loss))
loss += inter_loss
if self.intra_loss:
intra_log_probs = F.log_softmax(self.classifier(self.intra_proj(torch.cat(
[last_utt_embed, last_utt_embed - token_level_feat, last_utt_embed * token_level_feat,
token_level_feat], dim=-1))), dim=-1)
intra_loss = F.nll_loss(intra_log_probs.view(-1, self.config.num_labels), last_label.view(-1),
reduction='mean', ignore_index=-100)
# print("intra_loss: {}".format(intra_loss))
loss += intra_loss
return loss, prediction
return logits, prediction