modeling_deberta_arg_classifier.py

from transformers import PreTrainedModel, AutoModel
import torch.nn as nn
import torch
from .configuration_deberta_arg_classifier import DebertaConfig


class DebertaArgClassifier(PreTrainedModel):

    config_class = DebertaConfig

    def __init__(self, config):
        super().__init__(config)
        self.bert = AutoModel.from_pretrained("microsoft/deberta-large")
        self.classifier = nn.Linear(self.bert.config.hidden_size, config.number_labels)
        self.criterion = nn.BCEWithLogitsLoss()


    def forward(self, input_ids, attention_mask, labels=None):
        output = self.bert(input_ids, attention_mask=attention_mask)
        output = self._cls_embeddings(output)
        output_cls = self.classifier(output)
        output = torch.sigmoid(output_cls)
        if labels is not None:
            loss = self.cirterion(output_cls, labels)
            return {"loss": loss, "logits": output}
        return {"logits": output}


    def _cls_embeddings(self, output):
        '''Returns the embeddings corresponding to the <CLS> token of each text. '''

        last_hidden_state = output[0]
        cls_embeddings = last_hidden_state[:, 0]
        return cls_embeddings