File size: 23,013 Bytes
2d06dcc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
from operator import mod
import torch
import torch.nn.functional as F
from torch import nn
from transformers import BertPreTrainedModel, BertModel,  AutoModelForMaskedLM, BertForMaskedLM
from torch.nn.parameter import Parameter
from .utils import PairEnum
from sentence_transformers import SentenceTransformer
from losses import SupConLoss

activation_map = {'relu': nn.ReLU(), 'tanh': nn.Tanh()}

class Bert_SCCL(BertPreTrainedModel):
    def __init__(self, config, args):
        super(Bert_SCCL, self).__init__(config)  
        self.bert = None
        self.contrast_head = None
        self.cluster_centers = None

    def init_model(self, cluster_centers=None, alpha=1.0):
        self.emb_size = self.bert.config.hidden_size
        self.alpha = alpha
        
        # Instance-CL head
        self.contrast_head = nn.Sequential(
            nn.Linear(self.emb_size, self.emb_size),
            nn.ReLU(inplace=True),
            nn.Linear(self.emb_size, 128))
        
        # Clustering head
        initial_cluster_centers = torch.tensor(
            cluster_centers, dtype=torch.float, requires_grad=True)
        self.cluster_centers = Parameter(initial_cluster_centers)
      
    def forward(self, input_ids, attention_mask, task_type):

        if task_type == "evaluate":
            return self.get_mean_embeddings(input_ids, attention_mask)
        
        elif task_type == "explicit":
            input_ids_1, input_ids_2, input_ids_3 = torch.unbind(input_ids, dim=1)
            attention_mask_1, attention_mask_2, attention_mask_3 = torch.unbind(attention_mask, dim=1) 
            
            mean_output_1 = self.get_mean_embeddings(input_ids_1, attention_mask_1)
            mean_output_2 = self.get_mean_embeddings(input_ids_2, attention_mask_2)
            mean_output_3 = self.get_mean_embeddings(input_ids_3, attention_mask_3)
           
            return mean_output_1, mean_output_2, mean_output_3
        
    def get_mean_embeddings(self, input_ids, attention_mask):
        bert_output = self.bert.forward(input_ids=input_ids, attention_mask=attention_mask)
        attention_mask = attention_mask.unsqueeze(-1)
        mean_output = torch.sum(bert_output[0]*attention_mask, dim=1) / torch.sum(attention_mask, dim=1)
        return mean_output

    def get_cluster_prob(self, embeddings):
        norm_squared = torch.sum((embeddings.unsqueeze(1) - self.cluster_centers) ** 2, 2)
        numerator = 1.0 / (1.0 + (norm_squared / self.alpha))
        power = float(self.alpha + 1) / 2
        numerator = numerator ** power
        return numerator / torch.sum(numerator, dim=1, keepdim=True)

    def local_consistency(self, embd0, embd1, embd2, criterion):
        p0 = self.get_cluster_prob(embd0)
        p1 = self.get_cluster_prob(embd1)
        p2 = self.get_cluster_prob(embd2)
        
        lds1 = criterion(p1, p0)
        lds2 = criterion(p2, p0)
        return lds1+lds2
    
    def contrast_logits(self, embd1, embd2=None):
        feat1 = F.normalize(self.contrast_head(embd1), dim=1)
        if embd2 != None:
            feat2 = F.normalize(self.contrast_head(embd2), dim=1)
            return feat1, feat2
        else: 
            return feat1

class BERT_MTP_Pretrain(nn.Module):
    
    def __init__(self,  args):

        super(BERT_MTP_Pretrain, self).__init__()
        self.num_labels = args.num_labels
        self.bert = AutoModelForMaskedLM.from_pretrained(args.pretrained_bert_model)
        self.dropout = nn.Dropout(0.1) #0.1
        self.classifier = nn.Linear(args.feat_dim, args.num_labels)      
    
    def forward(self, X, ):

        outputs = self.bert(**X,  output_hidden_states=True)
        
        CLSEmbedding = outputs.hidden_states[-1][:,0]
        CLSEmbedding = self.dropout(CLSEmbedding)
        logits = self.classifier(CLSEmbedding)
        output_dir = {"logits": logits}
        output_dir["hidden_states"] = outputs.hidden_states[-1][:, 0]
        
        return output_dir
    
    def mlmForward(self, X, Y = None):
        outputs = self.bert(**X,  labels = Y)
        return outputs.loss
        
    def loss_ce(self, logits, Y):
        loss = nn.CrossEntropyLoss()
        output = loss(logits, Y)
        return output

class BERT_MTP(nn.Module):
    def __init__(self,  args):
        super(BERT_MTP, self).__init__()
    
        self.bert = AutoModelForMaskedLM.from_pretrained(args.pretrained_bert_model)
        self.dropout = nn.Dropout(0.1)
        #self.classifier = nn.Linear(args.feat_dim, args.num_labels)      
        self.head = nn.Sequential(
            nn.Linear(args.feat_dim, args.feat_dim),
            nn.ReLU(inplace=True),
            nn.Dropout(0.1),
            nn.Linear(args.feat_dim, args.mlp_head_feat_dim)
        )

    def forward(self, X):
        """logits are not normalized by softmax in forward function"""
        outputs = self.bert(**X, output_hidden_states=True, output_attentions=True)
        cls_embed = outputs.hidden_states[-1][:,0]
        features = F.normalize(self.head(cls_embed), dim=1)
        output_dir = {"features": features}
        output_dir["hidden_states"] = cls_embed
        
        return output_dir

    def loss_cl(self, embds, label=None, mask=None, temperature=0.07, base_temperature=0.07, device=None):
        """compute contrastive loss"""
        loss = SupConLoss()
        output = loss(embds, labels=label, mask=mask, temperature = temperature, device=device)
        return output
    
    def save_backbone(self, save_path):
        self.bert.save_pretrained(save_path)

class BERT_GCD(BertPreTrainedModel):
    
    def __init__(self,config, args):

        super(BERT_GCD, self).__init__(config)
        self.num_labels = args.num_labels
        self.bert = BertModel(config)
        self.mlp_head = nn.Sequential(
            nn.Linear(args.feat_dim, args.feat_dim),
            nn.ReLU(inplace=True),
            nn.Linear(args.feat_dim, args.mlp_head_feat_dim)
        )
        self.init_weights()

    def forward(self, input_ids = None, token_type_ids = None, attention_mask=None , labels = None,
                feature_ext = False, mode = None, loss_fct = None):

        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)

        encoded_layer_12 = outputs.hidden_states
        last_output_tokens = encoded_layer_12[-1]     
        features = last_output_tokens.mean(dim = 1)
        
        return features 

class BERT_CC(BertPreTrainedModel):
    
    def __init__(self,config, args):

        super(BERT_CC, self).__init__(config)
        self.num_labels = args.num_labels
        self.bert = BertModel(config)
        self.cluster_num = args.num_labels
        
        self.instance_projector = nn.Sequential(
            nn.Linear(config.hidden_size, config.hidden_size),
            nn.ReLU(),
            nn.Linear(config.hidden_size, config.hidden_size),
        )

        self.cluster_projector = nn.Sequential(
            nn.Linear(config.hidden_size, config.hidden_size),
            nn.ReLU(),
            nn.Linear(config.hidden_size, self.cluster_num),
            nn.Softmax(dim=1)
        ) 
        
        self.init_weights()
        
    def get_features(self, h_i, h_j):
        
        z_i = F.normalize(self.instance_projector(h_i), dim=1)
        z_j = F.normalize(self.instance_projector(h_j), dim=1)

        c_i = self.cluster_projector(h_i)
        c_j = self.cluster_projector(h_j)

        return z_i, z_j, c_i, c_j

    def forward_cluster(self, x):
     
        c = self.cluster_projector(x)
        c = torch.argmax(c, dim=1)
        return c

    def forward(self, input_ids = None, token_type_ids = None, attention_mask=None , labels = None,
                feature_ext = False, mode = None, loss_fct = None):

        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)

        encoded_layer_12 = outputs.hidden_states
        last_output_tokens = encoded_layer_12[-1]     
        features = last_output_tokens.mean(dim = 1)

        return features
        
class BERTForDeepAligned(BertPreTrainedModel):
    
    def __init__(self,config, args):

        super(BERTForDeepAligned, self).__init__(config)
        self.num_labels = args.num_labels
        self.bert = BertModel(config)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = activation_map[args.activation]
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, args.num_labels)      
        self.init_weights()

    def forward(self, input_ids = None, token_type_ids = None, attention_mask=None , labels = None,
                feature_ext = False, mode = None, loss_fct = None):

        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)
        encoded_layer_12 = outputs.hidden_states
        pooled_output = outputs.pooler_output

        pooled_output = self.dense(encoded_layer_12[-1].mean(dim = 1))
        pooled_output = self.activation(pooled_output)
        pooled_output = self.dropout(pooled_output)

        logits = self.classifier(pooled_output)

        if feature_ext:
            return pooled_output
        else:
            if mode == 'train':
                
                loss = loss_fct(logits, labels)
                return loss
            else:
                return pooled_output, logits
       
class BERT_USNID(BertPreTrainedModel):
    
    def __init__(self, config, args):

        super(BERT_USNID, self).__init__(config)
        self.num_labels = args.num_labels
        self.bert = BertModel(config)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = activation_map[args.activation]
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.args = args
        
        if args.pretrain or (not args.wo_self):
            self.classifier = nn.Linear(config.hidden_size, args.num_labels)
                
        self.mlp_head = nn.Linear(config.hidden_size, args.num_labels)
            
        self.init_weights()

    def forward(self, input_ids = None, token_type_ids = None, attention_mask=None , feature_ext = False):

        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)

        encoded_layer_12 = outputs.hidden_states
        last_output_tokens = encoded_layer_12[-1]     
        features = last_output_tokens.mean(dim = 1)
            
        features = self.dense(features)
        pooled_output = self.activation(features)   
        pooled_output = self.dropout(features)
        
        if self.args.pretrain or (not self.args.wo_self):
            logits = self.classifier(pooled_output)
            
        mlp_outputs = self.mlp_head(pooled_output)
        
        if feature_ext:
            if self.args.pretrain or (not self.args.wo_self):
                return features, logits
            else:
                return features, mlp_outputs

        else:
            if self.args.pretrain or (not self.args.wo_self):
                return mlp_outputs, logits
            else:
                return mlp_outputs, mlp_outputs
            
class BERT_USNID_UNSUP(BertPreTrainedModel):
    
    def __init__(self, config, args):

        super(BERT_USNID_UNSUP, self).__init__(config)
        self.num_labels = args.num_labels
        self.bert = BertModel(config)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = activation_map[args.activation]
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.args = args
 
        self.classifier = nn.Linear(config.hidden_size, args.num_labels)
        self.mlp_head = nn.Linear(config.hidden_size, args.num_labels)
            
        self.init_weights()

    def forward(self, input_ids = None, token_type_ids = None, attention_mask=None , labels = None, weights = None,
                feature_ext = False, mode = None, loss_fct = None, aug_feats=None, use_aug = False):

        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)

        encoded_layer_12 = outputs.hidden_states
        last_output_tokens = encoded_layer_12[-1]     
        features = last_output_tokens.mean(dim = 1)
            
        features = self.dense(features)
        pooled_output = self.activation(features)   
        pooled_output = self.dropout(features)
        
        logits = self.classifier(pooled_output)
            
        mlp_outputs = self.mlp_head(pooled_output)
        
        if feature_ext:
            return features, mlp_outputs
        else:
            return mlp_outputs, logits
            
class BertForConstrainClustering(BertPreTrainedModel):
    def __init__(self, config, args):
        super(BertForConstrainClustering, self).__init__(config)
        self.num_labels = args.num_labels
        self.bert = BertModel(config)
        
        # train
        self.dense = nn.Linear(config.hidden_size, config.hidden_size) # Pooling-mean
        self.activation = activation_map[args.activation]
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, args.num_labels)
        self.init_weights()           
        # finetune
        self.alpha = 1.0
        self.cluster_layer = Parameter(torch.Tensor(args.num_labels, args.num_labels))
        torch.nn.init.xavier_normal_(self.cluster_layer.data)

    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None,
                feature_ext = False, u_threshold=None, l_threshold=None, mode=None,  semi=False):

        eps = 1e-10
        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)
        encoded_layer_12 = outputs.hidden_states
        pooled_output = outputs.pooler_output
        pooled_output = self.dense(encoded_layer_12[-1].mean(dim = 1))
        pooled_output = self.activation(pooled_output)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)

        if feature_ext:
            return logits
        else:
            if mode=='train':

                logits_norm = F.normalize(logits, p=2, dim=1)
                sim_mat = torch.matmul(logits_norm, logits_norm.transpose(0, -1))
                label_mat = labels.view(-1,1) - labels.view(1,-1)    
                label_mat[label_mat!=0] = -1 
                label_mat[label_mat==0] = 1 
                label_mat[label_mat==-1] = 0 

                if not semi:
                    pos_mask = (label_mat > u_threshold).type(torch.cuda.FloatTensor)
                    neg_mask = (label_mat < l_threshold).type(torch.cuda.FloatTensor)
                    pos_entropy = -torch.log(torch.clamp(sim_mat, eps, 1.0)) * pos_mask
                    neg_entropy = -torch.log(torch.clamp(1-sim_mat, eps, 1.0)) * neg_mask
                    loss = (pos_entropy.mean() + neg_entropy.mean()) * 5

                    return loss

                else:
                    label_mat[labels==-1, :] = -1
                    label_mat[:, labels==-1] = -1
                    label_mat[label_mat==0] = 0
                    label_mat[label_mat==1] = 1
                    pos_mask = (sim_mat > u_threshold).type(torch.cuda.FloatTensor)
                    neg_mask = (sim_mat < l_threshold).type(torch.cuda.FloatTensor)
                    pos_mask[label_mat==1] = 1
                    neg_mask[label_mat==0] = 1
                    pos_entropy = -torch.log(torch.clamp(sim_mat, eps, 1.0)) * pos_mask
                    neg_entropy = -torch.log(torch.clamp(1-sim_mat, eps, 1.0)) * neg_mask
                    loss = pos_entropy.mean() + neg_entropy.mean() + u_threshold - l_threshold

                    return loss

            else:
                q = 1.0 / (1.0 + torch.sum(torch.pow(logits.unsqueeze(1) - self.cluster_layer, 2), 2) / self.alpha)
                q = q.pow((self.alpha + 1.0) / 2.0)
                q = (q.t() / torch.sum(q, 1)).t() 
                return logits, q

class BertForDTC(BertPreTrainedModel):
    def __init__(self, config, args):

        super(BertForDTC, self).__init__(config)
        self.num_labels = args.num_labels
        self.bert = BertModel(config)

        #train
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = activation_map[args.activation]
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, args.num_labels)
        self.init_weights()

        #finetune
        self.alpha = 1.0
        self.cluster_layer = Parameter(torch.Tensor(args.num_labels, args.num_labels))
        torch.nn.init.xavier_normal_(self.cluster_layer.data)

    def forward(self, input_ids = None, token_type_ids = None, attention_mask=None , labels = None,
                feature_ext = False, mode = None, loss_fct=None):

        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)
        encoded_layer_12 = outputs.hidden_states
        pooled_output = outputs.pooler_output
        pooled_output = self.dense(encoded_layer_12[-1].mean(dim = 1))
        pooled_output = self.activation(pooled_output)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        if feature_ext:
            return pooled_output
        elif mode == 'train':
            loss = loss_fct(logits, labels)
            return loss
        else:
            q = 1.0 / (1.0 + torch.sum(torch.pow(logits.unsqueeze(1) - self.cluster_layer, 2), 2) / self.alpha)
            q = q.pow((self.alpha + 1.0) / 2.0)
            q = (q.t() / torch.sum(q, 1)).t() 
            return logits, q

class BertForKCL_Similarity(BertPreTrainedModel):
    def __init__(self, config, args):
        super(BertForKCL_Similarity,self).__init__(config)

        self.num_labels = args.num_labels
        self.bert = BertModel(config)

        self.dense = nn.Linear(config.hidden_size * 2, config.hidden_size * 4)
        self.normalization = nn.BatchNorm1d(config.hidden_size * 4)
        self.activation = activation_map[args.activation]
        
        self.classifier = nn.Linear(config.hidden_size * 4, args.num_labels)
        self.init_weights()
    
    def forward(self, input_ids, token_type_ids = None, attention_mask=None, labels=None, loss_fct=None, mode = None):
        
        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)
        encoded_layer_12 = outputs.hidden_states
        pooled_output = outputs.pooler_output
        feat1,feat2 = PairEnum(encoded_layer_12[-1].mean(dim = 1))
        feature_cat = torch.cat([feat1,feat2], 1)

        pooled_output = self.dense(feature_cat)
        pooled_output = self.normalization(pooled_output)
        pooled_output = self.activation(pooled_output)
        logits = self.classifier(pooled_output)
        
        if mode == 'train':    
            loss = loss_fct(logits.view(-1,self.num_labels), labels.view(-1))

            return loss
        else:
            return pooled_output, logits

class BertForKCL(BertPreTrainedModel):
    def __init__(self, config, args):
        super(BertForKCL, self).__init__(config)

        self.num_labels = args.num_labels
        self.bert = BertModel(config)

        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = activation_map[args.activation]
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

        self.classifier = nn.Linear(config.hidden_size, args.num_labels)
        self.init_weights()

    def forward(self, input_ids = None, token_type_ids = None, attention_mask=None , labels = None, mode = None, 
        simi = None, loss_fct = None):

        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)
        encoded_layer_12 = outputs.hidden_states
        pooled_output = outputs.pooler_output
        pooled_output = self.dense(encoded_layer_12[-1].mean(dim = 1))
        pooled_output = self.activation(pooled_output)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)

        if mode == 'train':    

            probs = F.softmax(logits,dim=1)
            prob1, prob2 = PairEnum(probs)

            loss_KCL = loss_fct(prob1, prob2, simi)
            flag = len(labels[labels != -1])

            if flag != 0:
                loss_ce = nn.CrossEntropyLoss()(logits[labels != -1], labels[labels != -1])
                loss = loss_ce + loss_KCL
            else:
                loss = loss_KCL

            return loss
        else:
            return pooled_output, logits

class BertForMCL(BertPreTrainedModel):
    def __init__(self, config, args):
        super(BertForMCL, self).__init__(config)

        self.num_labels = args.num_labels
        self.bert = BertModel(config)

        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = activation_map[args.activation]
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, args.num_labels)
        self.init_weights()

    def forward(self, input_ids = None, token_type_ids = None, attention_mask=None , labels = None, mode = None, loss_fct = None):

        outputs = self.bert(
            input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, output_hidden_states=True)
        encoded_layer_12 = outputs.hidden_states
        pooled_output = outputs.pooler_output
        pooled_output = self.dense(encoded_layer_12[-1].mean(dim = 1))
        pooled_output = self.activation(pooled_output)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        probs = F.softmax(logits, dim = 1)

        if mode == 'train':
            
            flag = len(labels[labels != -1])
            prob1, prob2 = PairEnum(probs)
            simi = torch.matmul(probs, probs.transpose(0, -1)).view(-1)

            simi[simi > 0.5] = 1
            simi[simi < 0.5] = -1
            loss_MCL = loss_fct(prob1, prob2, simi)

            if flag != 0:

                loss_ce = nn.CrossEntropyLoss()(logits[labels != -1], labels[labels != -1])
                loss = loss_ce + loss_MCL

            else:
                loss = loss_MCL

            return loss
            
        else:
            return pooled_output, logits