nn.BCEWithLogitsLoss() can't accept one-hot target

1

Hi,
My training loop looks something like this

loss_fn = nn.BCEWithLogitsLoss()
    for epoch in range(1, num_epochs+1):
        model.train()
        for X, y in train_loader:
            X, y = X.to(device), y.to(device)
            y_hot = F.one_hot(y, num_classes)
            output = model(X)
            optimizer.zero_grad()
            loss = loss_fn(output, y_hot)
            loss.backward()
            optimizer.step()

I am transforming the target to be a one-hot vector as expected by the BCEWithLogitsLoss().
But, I am still getting the following error, pointing at the loss function:

RuntimeError: result type Float can't be cast to the desired output type Long

I can’t really understand the error message. What am I doing wrong?

2 Likes
2

I think the problem is that y_hot is of type long and the BCE loss expects a Tensor of type float.
Can you try converting y_hot with y_hot = y_hot.type_as(output).

17 Likes
3

Thanks. It worked.

But, since it’s pretty common for BCELoss to have one-hot target, wouldn’t it be better if it could also handle long type?

4

Hi,

The general policy we have in pytorch is that we avoid doing any copy that are hidden from the user. (cpu/gpu or different types).
Even though you could argue that this can make the user’s code more verbose. It usually allows the user to have a better control on what happens in its code.

5 Likes