roberta.py

import math
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss
from dataclasses import dataclass
from typing import Optional, Tuple, Union
from transformers.models.roberta.modeling_roberta import RobertaPreTrainedModel, RobertaModel
from models.output import QuestionAnsweringV2ModelOutput


class RobertaForV2QuestionAnswering(RobertaPreTrainedModel):
    _keys_to_ignore_on_load_unexpected = [r"pooler"]
    _keys_to_ignore_on_load_missing = [r"position_ids"]

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

        self.roberta = RobertaModel(config, add_pooling_layer=False)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
        self.impossible_outputs = nn.Sequential(nn.Dropout(p=0.1), nn.Linear(config.hidden_size, 1))

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        is_impossibles: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        start_positions: Optional[torch.LongTensor] = None,
        end_positions: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], QuestionAnsweringV2ModelOutput]:

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.roberta(
            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,
        )

        sequence_output = outputs[0]  # (batch_size, seq_size, hidden_size)

        sentence_output = sequence_output[:, 0, :]  # (batch_size, hidden_size)
        sentence_logits = self.impossible_outputs(sentence_output)

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)

            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            loss_fct_start = CrossEntropyLoss(ignore_index=ignored_index)
            loss_fct_end = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct_start(start_logits, start_positions)
            end_loss = loss_fct_end(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        if is_impossibles is not None:
            loss_fct = BCEWithLogitsLoss()
            is_impossible_loss = loss_fct(
                sentence_logits.view(-1,), is_impossibles.float().view(-1,)
            )
            total_loss += is_impossible_loss * 0.1

        if not return_dict:
            output = (start_logits, end_logits) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return QuestionAnsweringV2ModelOutput(
            loss=total_loss,
            is_impossible_logits=sentence_logits,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )