modeling_custom.py

# modeling_custom.py
from transformers import XLMRobertaForSequenceClassification
import torch.nn as nn
import torch.nn.functional as F

class CustomModel(XLMRobertaForSequenceClassification):
    def __init__(self, config, num_emotion_labels):
        super(CustomModel, self).__init__(config)
        self.num_emotion_labels = num_emotion_labels

        # Freeze sentiment classifier parameters
        for param in self.classifier.parameters():
            param.requires_grad = False

        # Define emotion classifier
        self.dropout_emotion = nn.Dropout(config.hidden_dropout_prob)
        self.emotion_classifier = nn.Sequential(
            nn.Linear(config.hidden_size, 512),
            nn.Mish(),
            nn.Dropout(0.3),
            nn.Linear(512, num_emotion_labels)
        )

        # Initialize the weights of the new layers
        self._init_weights(self.emotion_classifier[0])
        self._init_weights(self.emotion_classifier[3])

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()

    def forward(self, input_ids=None, attention_mask=None, sentiment=None, labels=None):
        outputs = self.roberta(input_ids=input_ids, attention_mask=attention_mask)
        sequence_output = outputs[0]

        # Select the CLS token for emotion classification
        cls_hidden_states = sequence_output[:, 0, :]
        cls_hidden_states = self.dropout_emotion(cls_hidden_states)

        emotion_logits = self.emotion_classifier(cls_hidden_states)

        # Sentiment logits from the frozen classifier
        with torch.no_grad():
            cls_token_state = sequence_output[:, 0, :].unsqueeze(1)
            sentiment_logits = self.classifier(cls_token_state).squeeze(1)

        if labels is not None:
            class_weights = torch.tensor([1.0] * self.num_emotion_labels).to(labels.device)
            loss_fct = nn.BCEWithLogitsLoss(pos_weight=class_weights)
            loss = loss_fct(emotion_logits, labels)
            return {"loss": loss, "emotion_logits": emotion_logits, "sentiment_logits": sentiment_logits}

        return {"emotion_logits": emotion_logits, "sentiment_logits": sentiment_logits}