custom_model.py

import torch
from torch import nn
from transformers import BertPreTrainedModel, BertModel

class MyBERTClassifier(BertPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.bert = BertModel(config)
        hidden_size = config.hidden_size
        num_labels = config.num_labels

        self.classifier = nn.Sequential(
            nn.Linear(hidden_size, hidden_size // 2),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(hidden_size // 2, num_labels)
        )
        
        self.post_init()

    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, **kwargs):
        outputs = self.bert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            return_dict=True,
            **kwargs
        )
        pooled_output = outputs.pooler_output
        logits = self.classifier(pooled_output)
        return logits