model.py

from abc import ABCMeta
import torch
from transformers.pytorch_utils import nn
import torch.nn.functional as F
from transformers import BertModel, BertForSequenceClassification, PreTrainedModel
from transformers.modeling_outputs import SequenceClassifierOutput
from transformers import BertConfig
from transformers import PretrainedConfig

class BertAttentionConfig(PretrainedConfig):
    model_type = "bertAttentionForSequenceClassification"  # Update the model type

    def __init__(self,
                 num_classes=2,
                 hidden_size=768,   # Update embed_dim to hidden_size
                 fc_hidden=128,     # New parameter for FC layer
                 num_layers=12,
                 dropout_rate=0.1,
                 **kwargs):
        super().__init__(**kwargs)
        self.num_classes = num_classes
        self.hidden_size = hidden_size  # Update embed_dim to hidden_size
        self.fc_hidden = fc_hidden      # Assign FC layer hidden units
        self.num_layers = num_layers
        self.dropout_rate = dropout_rate
        self.id2label = {
            0: "fake",
            1: "true",
        }
        self.label2id = {
            "fake": 0,
            "true": 1,
        }


class BertAttentionForSequenceClassification(PreTrainedModel, metaclass=ABCMeta):
    config_class = BertAttentionConfig  # Use the appropriate BERT configuration class

    def __init__(self, config):
        super(BertAttentionForSequenceClassification, self).__init__(config)
        self.num_classes = config.num_classes
        self.embed_dim = config.hidden_size  # Hidden size is the BERT embedding dimension
        self.bert = BertModel.from_pretrained('bert-base-uncased', output_hidden_states=True, output_attentions=True)
        print("BERT Model Loaded")
        self.fc = nn.Linear(config.hidden_size, self.num_classes)

    def forward(self, input_ids, attention_mask, token_type_ids, labels=None):
        bert_output = self.bert(input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)
        hidden_states = bert_output.last_hidden_state  # Use the last hidden state

        # Apply self-attention (scaled dot-product attention)
        attention_scores = torch.matmul(hidden_states, hidden_states.transpose(1, 2))
        attention_scores = attention_scores / (self.embed_dim ** 0.5)
        attention_probs = F.softmax(attention_scores, dim=-1)
        attention_output = torch.matmul(attention_probs, hidden_states)

        # Pool over the sequence length to get the final representation
        pooled_output = torch.mean(attention_output, dim=1)

        logits = self.fc(pooled_output)
        loss = None
        if labels is not None:
            loss = F.cross_entropy(logits, labels)

        out = SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=bert_output.hidden_states,
            attentions=bert_output.attentions,
        )
        return out