commit the model file.
Browse files
model.py
ADDED
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import AutoModelForTokenClassification, AutoModelForSequenceClassification, AdamW, get_linear_schedule_with_warmup, AutoModel
|
3 |
+
from transformers import BertForTokenClassification, BertForSequenceClassification,BertPreTrainedModel, BertModel
|
4 |
+
import torch.nn as nn
|
5 |
+
from .utils import *
|
6 |
+
import torch.nn.functional as F
|
7 |
+
|
8 |
+
from ekphrasis.classes.preprocessor import TextPreProcessor
|
9 |
+
from ekphrasis.classes.tokenizer import SocialTokenizer
|
10 |
+
from ekphrasis.dicts.emoticons import emoticons
|
11 |
+
import re
|
12 |
+
from transformers import AutoTokenizer, AutoModelForTokenClassification, AdamW, get_linear_schedule_with_warmup
|
13 |
+
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
|
14 |
+
|
15 |
+
|
16 |
+
class Model_Rational_Label(BertPreTrainedModel):
|
17 |
+
def __init__(self,config,params):
|
18 |
+
super().__init__(config)
|
19 |
+
self.num_labels=params['num_classes']
|
20 |
+
self.num_targets=params['targets_num']
|
21 |
+
self.impact_factor=params['rationale_impact']
|
22 |
+
self.target_factor=params['target_impact']
|
23 |
+
self.bert = BertModel(config,add_pooling_layer=False)
|
24 |
+
self.pooler=BertPooler(config)
|
25 |
+
self.token_dropout = nn.Dropout(0.2)
|
26 |
+
self.token_classifier = nn.Linear(config.hidden_size, 2)
|
27 |
+
self.dropout = nn.Dropout(0.2)
|
28 |
+
self.classifier = nn.Linear(config.hidden_size, self.num_labels)
|
29 |
+
self.target_dropout = nn.Dropout(0.2)
|
30 |
+
self.target_classifier = nn.Linear(config.hidden_size, self.num_targets)
|
31 |
+
self.init_weights()
|
32 |
+
# self.embeddings = AutoModelForTokenClassification.from_pretrained(params['model_path'], cache_dir=params['cache_path'])
|
33 |
+
|
34 |
+
def forward(self, input_ids=None, mask=None, attn=None, labels=None, targets=None):
|
35 |
+
outputs = self.bert(input_ids, mask)
|
36 |
+
# out = outputs.last_hidden_state
|
37 |
+
out=outputs[0]
|
38 |
+
logits = self.token_classifier(self.token_dropout(out))
|
39 |
+
|
40 |
+
|
41 |
+
# mean_pooling = torch.mean(out, 1)
|
42 |
+
# max_pooling, _ = torch.max(out, 1)
|
43 |
+
# embed = torch.cat((mean_pooling, max_pooling), 1)
|
44 |
+
embed=self.pooler(outputs[0])
|
45 |
+
y_pred = self.classifier(self.dropout(embed))
|
46 |
+
y_pred_target = torch.sigmoid(self.target_classifier(self.target_dropout(embed)))
|
47 |
+
|
48 |
+
loss_token = None
|
49 |
+
loss_target= None
|
50 |
+
loss_label = None
|
51 |
+
loss_total = None
|
52 |
+
|
53 |
+
if attn is not None:
|
54 |
+
loss_fct = nn.CrossEntropyLoss()
|
55 |
+
### Adding weighted
|
56 |
+
|
57 |
+
# Only keep active parts of the loss
|
58 |
+
if mask is not None:
|
59 |
+
class_weights=torch.tensor([1.0,1.0],dtype=torch.float).to(input_ids.device)
|
60 |
+
loss_funct = nn.CrossEntropyLoss(class_weights)
|
61 |
+
active_loss = mask.view(-1) == 1
|
62 |
+
active_logits = logits.view(-1, 2)
|
63 |
+
active_labels = torch.where(
|
64 |
+
active_loss, attn.view(-1), torch.tensor(loss_fct.ignore_index).type_as(attn)
|
65 |
+
)
|
66 |
+
loss_token = loss_funct(active_logits, active_labels)
|
67 |
+
else:
|
68 |
+
loss_token = loss_funct(logits.view(-1, 2), attn.view(-1))
|
69 |
+
|
70 |
+
loss_total=self.impact_factor*loss_token
|
71 |
+
|
72 |
+
if targets is not None:
|
73 |
+
loss_funct = nn.BCELoss()
|
74 |
+
loss_logits = loss_funct(y_pred_target.view(-1, self.num_targets), targets.view(-1, self.num_targets))
|
75 |
+
loss_targets= loss_logits
|
76 |
+
loss_total+=self.target_factor*loss_targets
|
77 |
+
|
78 |
+
|
79 |
+
if labels is not None:
|
80 |
+
loss_funct = nn.CrossEntropyLoss()
|
81 |
+
loss_logits = loss_funct(y_pred.view(-1, self.num_labels), labels.view(-1))
|
82 |
+
loss_label= loss_logits
|
83 |
+
if(loss_total is not None):
|
84 |
+
loss_total+=loss_label
|
85 |
+
else:
|
86 |
+
loss_total=loss_label
|
87 |
+
if(loss_total is not None):
|
88 |
+
return y_pred,y_pred_target,logits, loss_total
|
89 |
+
else:
|
90 |
+
return y_pred,y_pred_target,logits
|