AnnyNguyen commited on
Commit
d9c317f
·
verified ·
1 Parent(s): 008b518

Upload 2 files

Browse files
Files changed (2) hide show
  1. __init__.py +0 -0
  2. models.py +70 -0
__init__.py ADDED
File without changes
models.py ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import os
3
+ import torch.nn as nn
4
+ from transformers import RobertaPreTrainedModel, RobertaModel, AutoConfig
5
+ from transformers.modeling_outputs import SequenceClassifierOutput
6
+
7
+ class TransformerForABSA(RobertaPreTrainedModel):
8
+ base_model_prefix = "roberta"
9
+
10
+ def __init__(self, config):
11
+ super().__init__(config)
12
+ self.roberta = RobertaModel(config)
13
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
14
+ # Thêm lớp "none" vào num_sentiments
15
+ self.sentiment_classifiers = nn.ModuleList([
16
+ nn.Linear(config.hidden_size, config.num_sentiments + 1) # +1 cho "none"
17
+ for _ in range(config.num_aspects)
18
+ ])
19
+ self.init_weights()
20
+
21
+ def forward(
22
+ self,
23
+ input_ids=None,
24
+ attention_mask=None,
25
+ labels=None,
26
+ return_dict=None
27
+ ):
28
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
29
+ outputs = self.roberta(input_ids, attention_mask=attention_mask, return_dict=return_dict)
30
+ pooled = self.dropout(outputs.pooler_output) # [B, H]
31
+ all_logits = torch.stack([cls(pooled) for cls in self.sentiment_classifiers], dim=1) # [B, A, S+1]
32
+
33
+ loss = None
34
+ if labels is not None:
35
+ # labels: [B, A], với lớp "none" thay vì -100
36
+ B, A, _ = all_logits.size()
37
+ logits_flat = all_logits.view(-1, all_logits.size(-1))
38
+ targets_flat = labels.view(-1)
39
+ loss_fct = nn.CrossEntropyLoss() # Không dùng ignore_index
40
+ loss = loss_fct(logits_flat, targets_flat)
41
+
42
+ if not return_dict:
43
+ return ((loss, all_logits) + outputs[2:]) if loss is not None else (all_logits,) + outputs[2:]
44
+ return SequenceClassifierOutput(
45
+ loss=loss,
46
+ logits=all_logits,
47
+ hidden_states=outputs.hidden_states,
48
+ attentions=outputs.attentions,
49
+ )
50
+
51
+ def save_pretrained(self, save_directory: str, **kwargs):
52
+ """
53
+ HuggingFace Trainer đôi khi truyền vào state_dict=..., nên ta
54
+ chấp nhận thêm **kwargs để không vướng lỗi.
55
+ """
56
+ # 1) Lưu phần backbone (encoder)
57
+ self.roberta.save_pretrained(save_directory, **kwargs)
58
+
59
+ # 2) Cập nhật config rồi lưu
60
+ config = self.roberta.config
61
+ config.num_aspects = len(self.sentiment_classifiers)
62
+ config.num_sentiments = self.sentiment_classifiers[0].out_features
63
+ config.auto_map = {"AutoModel": "models.TransformerForABSA"}
64
+ config.save_pretrained(save_directory, **kwargs)
65
+
66
+ # 3) Lưu toàn bộ state_dict (bao gồm cả 2 head) —
67
+ # nếu Trainer đã truyền state_dict trong kwargs, có thể dùng luôn,
68
+ # nếu không, lấy từ self.state_dict()
69
+ sd = kwargs.get("state_dict", None) or self.state_dict()
70
+ torch.save(sd, os.path.join(save_directory, "pytorch_model.bin"))