How to add a learnable weight for each sample when i compute loss?

1

Hi everyone. I am developing a new method to learn the weight of each sample(not each batch or each class). But i get some error message, so i try to get some help from you, thanks for your reading!

At first, i wanna to use the loss function pytorch officially provided

nn.BCEWithLogitsLoss(weight=weight_score)

so i write a critetion, when i want to compute the loss wight sample_weight, i will call the forward function

class criterion(nn.Module):
    def init(self):
        super(criterion, self).init()

    def forward(self, true, pred, score):
        loss_fuc = nn.BCEWithLogitsLoss(weight=score)
        return loss_fuc(pred, true)

But there is a error,

The size of tensor a (321) must match the size of tensor b (16) at non-singleton dimension 3

i don`t konw how to solve it, i get the shape of [true, pre, score] as follows:

torch.Size([16, 1, 321, 321])
torch.Size([16, 1, 321, 321])
torch.Size([16])
# batch_size = 16

So the key problem is “How to add a learnable weight of each sample when i compute loss?”

Thanks for your reading, i really want to solve this problem and continue my experiment

I will appreciate if there is another better solution for this problem, thanks again

2

Are you sure you are printing the shapes just before nn.bce… the true seems to be of shape 16?

3

yes, the shape of true is torch.Size([16, 1, 321, 321]). Beacause i`m try to build a semantic segmentation application, the label is a gray image and the shape is 1x321x321.