model.py

import torch.nn as nn
from transformers import XLMRobertaModel
from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaPreTrainedModel
from transformers.modeling_outputs import SequenceClassifierOutput

class Smish(nn.Module):
    def __init__(self):
        super().__init__()
    def forward(self, x):
        return x * (x.sigmoid() + 1).log().tanh()

class NoRefER(XLMRobertaPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        hidden_size = 32
        self.config = config
        self.roberta = XLMRobertaModel(config)
        self.dense = nn.Sequential(
            nn.Dropout(config.hidden_dropout_prob), 
            nn.Linear(config.hidden_size, hidden_size, bias = False),
            nn.Dropout(config.hidden_dropout_prob), Smish(), 
            nn.Linear(hidden_size, 1, bias = False)
        )

        self.post_init()

    def forward(self, positive_input_ids, positive_attention_mask, negative_input_ids, negative_attention_mask, labels, weight=None):
        # positive processing
        positive_inputs = {
            "input_ids": positive_input_ids #, "attention_mask": positive_attention_mask
        }
        positive = self.dense(self.roberta(**positive_inputs).pooler_output).squeeze(-1)

        # negative processing
        negative_inputs = {
            "input_ids": negative_input_ids #, "attention_mask": negative_attention_mask
        }
        negative = self.dense(self.roberta(**negative_inputs).pooler_output).squeeze(-1)

        if weight is None:
            bce = nn.BCEWithLogitsLoss()
        else:
            bs = len(positive)
            weights = (weight.float() * bs) / weight.sum()
            bce = nn.BCEWithLogitsLoss(weight = weights)
        loss = bce(positive - negative, labels.float())
        return SequenceClassifierOutput(
            loss=loss,
            logits=positive.sigmoid()-negative.sigmoid(),
        )
    
    def score(
        self,
        input_ids,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        output_attentions=None,
        output_hidden_states=None,
    ):
        h = 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,).pooler_output

        return self.dense(h).sigmoid().squeeze(-1)