When using the BSELoss function, we pass the ground truth as y_train which contains either class 0 or 1 so the predictions should also be predicted classes and not class probabilities right?

[Blurr99]

So, in this video https://www.youtube.com/watch?v=LyJtbe__2i0, at 10:47:28, Daniel uses:

loss = loss_fn(torch.sigmoid(y_logits),  y_train)

but after the logits pass through the sigmoid function they turn into class probabilities right? So how can they be compared to y_train which contains either 0 or 1. Shouldn't it be:

loss = loss_fn(torch.round(torch.sigmoid(y_logits),  y_train))

[37Lime]

Greetings Blurr99,

The code loss = loss_fn(torch.sigmoid(y_logits), y_train) as Daniel used is correct.

You can see in the documentation (math included) that BCELoss measures the Binary Cross Entropy between the target and the input probabilities. Therefore, you shall feed the probabilities as they are (i.e., no rounding) into the loss function.

I hope this answers your question.

Best wishes.