models/roberta_model.py

# models/roberta_model.py

import torch
import torch.nn as nn
from transformers import RobertaModel
from config import DROPOUT_RATE, ROBERTA_MODEL_NAME # Import ROBERTA_MODEL_NAME

class RobertaMultiOutputModel(nn.Module):
    """
    RoBERTa-based model for multi-output classification.
    Uses a pre-trained RoBERTa model as its backbone. RoBERTa is an optimized
    version of BERT, often performing better.
    """
    # Statically set tokenizer name for easy access in main.py
    tokenizer_name = ROBERTA_MODEL_NAME

    def __init__(self, num_labels):
        """
        Initializes the RobertaMultiOutputModel.

        Args:
            num_labels (list): A list where each element is the number of classes
                                for a corresponding label column.
        """
        super(RobertaMultiOutputModel, self).__init__()
        # Load the pre-trained RoBERTa model.
        # RoBERTa's pooler_output typically corresponds to the hidden state of the
        # first token (<s>), which is often used for sequence classification.
        self.roberta = RobertaModel.from_pretrained(ROBERTA_MODEL_NAME)
        self.dropout = nn.Dropout(DROPOUT_RATE) # Dropout layer

        # Create classification heads for each label column.
        self.classifiers = nn.ModuleList([
            nn.Linear(self.roberta.config.hidden_size, n_classes) for n_classes in num_labels
        ])

    def forward(self, input_ids, attention_mask):
        """
        Performs the forward pass of the model.

        Args:
            input_ids (torch.Tensor): Tensor of token IDs.
            attention_mask (torch.Tensor): Tensor indicating attention.

        Returns:
            list: A list of logit tensors, one for each classification head.
        """
        # Pass input_ids and attention_mask through RoBERTa.
        # .pooler_output is used for classification.
        pooled_output = self.roberta(input_ids=input_ids, attention_mask=attention_mask).pooler_output

        # Apply dropout
        pooled_output = self.dropout(pooled_output)

        # Pass the pooled output through each classification head.
        return [classifier(pooled_output) for classifier in self.classifiers]