model.py

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

class BertLSTMConfig(PretrainedConfig):
    model_type = "bertLSTMForSequenceClassification"

    def __init__(self,
                 num_classes=2,
                 hidden_size=768,  # BERT hidden size
                 num_layers=12,
                 hidden_dim_lstm=256,  # New parameter for LSTM
                 hidden_dropout_prob=0.1,  # Changed from dropout_rate to hidden_dropout_prob
                 **kwargs):
        super().__init__(**kwargs)
        self.num_classes = num_classes
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.hidden_dim_lstm = hidden_dim_lstm  # Assign LSTM hidden dimension
        self.hidden_dropout_prob = hidden_dropout_prob  # Adjusted to BERT parameter name
        self.id2label = {
            0: "fake",
            1: "true",
        }
        self.label2id = {
            "fake": 0,
            "true": 1,
        }


class BertLSTMForSequenceClassification(PreTrainedModel, metaclass=ABCMeta):
    config_class = BertLSTMConfig

    def __init__(self, config):
        super(BertLSTMForSequenceClassification, self).__init__(config)
        self.num_classes = config.num_classes
        self.embed_dim = config.hidden_size  # BERT hidden size is used as the embedding dimension
        self.num_layers = config.num_layers
        self.hidden_dim_lstm = config.hidden_dim_lstm
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

        # Use BertModel instead of AlbertModel
        self.bert = BertModel.from_pretrained('bert-base-uncased', output_hidden_states=True, output_attentions=False)
        print("BERT Model Loaded")

        # Adjust the input dimension for LSTM
        self.lstm = nn.LSTM(self.embed_dim, self.hidden_dim_lstm, batch_first=True, num_layers=self.num_layers)

        # Adjust the output dimension for the linear layer
        self.fc = nn.Linear(self.hidden_dim_lstm, 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["hidden_states"]

        # Extract the embeddings from the last layer of BERT
        last_hidden_states = hidden_states[-1]

        # Apply dropout
        last_hidden_states = self.dropout(last_hidden_states)

        # Pass through LSTM
        lstm_output, _ = self.lstm(last_hidden_states, None)

        # Take the output from the last time step
        lstm_output = lstm_output[:, -1, :]

        # Apply dropout
        lstm_output = self.dropout(lstm_output)

        # Linear layer for classification
        logits = self.fc(lstm_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