What does pre_classifier in distilbert do?

Question

I was reading huggingface's DistilBertForSequenceClassification implementation code and noticed that they create a classifier and a pre_classifier when initiating the object. Later in the forward method they send the pooled output to the pre_classifier before preparing and sending it to the classifier. Unfortunately, I'm having a hard time understanding what the pre_classifier is meant to achieve here. Nor have I been able to find much info on it. Does anyone know what it is supposed to do?

class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
    def __init__(self, config: PretrainedConfig):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.config = config

        self.distilbert = DistilBertModel(config)
        self.pre_classifier = nn.Linear(config.dim, config.dim)
        self.classifier = nn.Linear(config.dim, config.num_labels)
        self.dropout = nn.Dropout(config.seq_classif_dropout)

        # Initialize weights and apply final processing
        self.post_init()


    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[SequenceClassifierOutput, Tuple[torch.Tensor, ...]]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        distilbert_output = self.distilbert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_state = distilbert_output[0]  # (bs, seq_len, dim)
        pooled_output = hidden_state[:, 0]  # (bs, dim)
        pooled_output = self.pre_classifier(pooled_output)  # (bs, dim)
        pooled_output = nn.ReLU()(pooled_output)  # (bs, dim)
        pooled_output = self.dropout(pooled_output)  # (bs, dim)
        logits = self.classifier(pooled_output)  # (bs, num_labels)

        loss = None
        if labels is not None:
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

        if not return_dict:
            output = (logits,) + distilbert_output[1:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=distilbert_output.hidden_states,
            attentions=distilbert_output.attentions,
        )

Answer 1

I found the answer myself. It was rather difficult to find but that is actually the dense layer. Since DistilBert does not have a pooler, it does not need a dense layer. But for sequence classification a pooler is added so the dense layer is also needed. In other words self.pre_classifer in DistilBert is the same as self.dense in bert. Now I have to figure out why distilbert does not need a dense.