dungeon29 commited on
Commit
78920ba
·
verified ·
1 Parent(s): 8a1f01d

Upload model.py

Browse files
Files changed (1) hide show
  1. model.py +58 -0
model.py ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ import torch.nn.functional as F
4
+ from transformers import AutoModel
5
+
6
+ class DeBERTaLSTMClassifier(nn.Module):
7
+ def __init__(self, hidden_dim=128, num_labels=2):
8
+ super().__init__()
9
+ self.deberta = AutoModel.from_pretrained("microsoft/deberta-base")
10
+
11
+ # Đóng băng DeBERTa
12
+ for param in self.deberta.parameters():
13
+ param.requires_grad = False
14
+
15
+ self.lstm = nn.LSTM(
16
+ input_size=self.deberta.config.hidden_size,
17
+ hidden_size=hidden_dim,
18
+ batch_first=True,
19
+ bidirectional=True
20
+ )
21
+
22
+ # Lớp Attention: chuyển đổi hidden state thành điểm số quan trọng (score)
23
+ self.attention = nn.Linear(hidden_dim * 2, 1)
24
+
25
+ self.fc = nn.Linear(hidden_dim * 2, num_labels)
26
+
27
+ def forward(self, input_ids, attention_mask, return_attention=False):
28
+ # 1. DeBERTa
29
+ with torch.no_grad():
30
+ outputs = self.deberta(input_ids=input_ids, attention_mask=attention_mask, output_attentions=True)
31
+
32
+ # 2. LSTM
33
+ lstm_out, _ = self.lstm(outputs.last_hidden_state) # [batch, seq_len, hidden*2]
34
+
35
+ # 3. Tính Attention (Luôn luôn thực hiện)
36
+ # Tính score chưa qua softmax
37
+ attn_scores = self.attention(lstm_out).squeeze(-1) # [batch, seq_len]
38
+
39
+ # Masking chuẩn: Gán giá trị rất nhỏ (-inf) cho các vị trí padding trước khi Softmax
40
+ # Để đảm bảo padding có attention weight = 0 tuyệt đối
41
+ mask = attention_mask.float()
42
+ attn_scores = attn_scores.masked_fill(mask == 0, -1e9)
43
+
44
+ # Softmax để ra weights
45
+ attn_weights = F.softmax(attn_scores, dim=-1) # [batch, seq_len]
46
+
47
+ # Tính Context Vector (Weighted Sum)
48
+ # [batch, seq_len, 1] * [batch, seq_len, hidden*2] -> sum -> [batch, hidden*2]
49
+ context_vector = torch.sum(attn_weights.unsqueeze(-1) * lstm_out, dim=1)
50
+
51
+ # 4. Classification
52
+ logits = self.fc(context_vector)
53
+
54
+ # 5. Return tùy theo yêu cầu
55
+ if return_attention:
56
+ return logits, attn_weights, outputs.attentions
57
+ else:
58
+ return logits