code/utils/classifier_loss.py

import torch
import torch.nn as nn
import torch.nn.functional as F

# CrossEntropyLoss = softmax + log + NLLLoss

class FocalLoss(nn.Module):
    def __init__(self, weight=None, gamma=0.5, reduction=None):
        super(FocalLoss, self).__init__()

        self.weight = weight
        self.gamma = gamma
        self.reduction = reduction
    
    def forward(self, input_tensor, target_tensor):
        assert input_tensor.shape[0] == target_tensor.shape[0]
        
        prob = F.softmax(input_tensor, dim = -1)
        log_prob = torch.log(prob + 1e-8)
        
        loss = F.nll_loss(
            ((1 - prob) ** self.gamma) * log_prob, 
            target_tensor, 
            weight=self.weight,
            reduction=self.reduction
        )

        return loss

class ClassifierLoss():
    def __init__(self, alpha, gamma):
        weight = torch.FloatTensor([1-alpha, alpha])
        if torch.cuda.is_available():
            weight = weight.cuda()
        
        self.focal_loss = FocalLoss(
            weight = weight,
            gamma = gamma,
            reduction = 'mean'
        )

        # self.ce_loss = nn.CrossEntropyLoss(weight = weight, reduction = "mean")

    def compute_batch_loss(self, batch_logits, batch_labels, batch_size):
        loss = 0
        for logits, labels in zip(batch_logits, batch_labels):
            loss += self.focal_loss(logits, labels)
        
        return loss/batch_size
    
    def compute_loss(
        self,
        batch_table_name_cls_logits,
        batch_table_labels,
        batch_column_info_cls_logits,
        batch_column_labels
    ):
        batch_size = len(batch_table_labels)

        table_loss = self.compute_batch_loss(batch_table_name_cls_logits, batch_table_labels, batch_size)
        column_loss = self.compute_batch_loss(batch_column_info_cls_logits, batch_column_labels, batch_size)

        return table_loss + column_loss