modeling_bert.py

from typing import Optional, Union

import torch

from torch import nn
from transformers import (
    BertModel,
    BertPreTrainedModel,
)
from transformers.modeling_outputs import SequenceClassifierOutput
from transformers.models.bert.modeling_bert import BertOnlyMLMHead

from .configuration_bert import BertMultiTaskConfig

class BertForMultiTaskClassification(BertPreTrainedModel):
    config_class = BertMultiTaskConfig
    _tied_weights_keys = ["cls.predictions.decoder.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.tasks = config.tasks
        self.config = config

        self.bert = BertModel(config)
        classifier_dropout = (
            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
        )
        self.dropout = nn.Dropout(classifier_dropout)

        task_layers = {}
        for task_name, num_labels in self.tasks.items():
            if task_name.upper() == "MLM":
                self.cls = BertOnlyMLMHead(config)
            else:
                task_layers[task_name.upper()] = nn.Linear(config.hidden_size, num_labels)
        self.task_classifiers = nn.ModuleDict(task_layers)

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

    def get_output_embeddings(self):
        # This method tells the PreTrainedModel that self.cls.predictions.decoder is the output layer to be tied
        if hasattr(self, "cls"):
            return self.cls.predictions.decoder
        return None

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        task: str | None = None,  # For now the model will use single task per batch
    ) -> Union[tuple[torch.Tensor], SequenceClassifierOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        if task is None:
            raise ValueError(f"Task must be specified and one of {self.task_classifiers.keys()}")
        if task.upper() == "MLM":
            if not hasattr(self, "cls"):
                raise ValueError("Model was not initialized with an MLM head.")

        outputs = self.bert(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        loss = None
        logits = None
        num_labels = self.config.vocab_size if task.upper() == "MLM" else self.tasks[task]
        if task.upper() == "MLM":
            sequence_output = outputs[0]
            logits = self.cls(sequence_output)

        elif task.upper() in self.task_classifiers:
            pooled_output = outputs[1]
            pooled_output = self.dropout(pooled_output)
            logits = self.task_classifiers[task.upper()](pooled_output)

        else:
            raise ValueError(f"Invalid task: {task}")

        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, num_labels), labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

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


BertMultiTaskConfig.register_for_auto_class()
BertForMultiTaskClassification.register_for_auto_class("AutoModelForSequenceClassification")