Loss functions for classification - logits/probabilities (page 472)

[labdmitriy]

Hi Sebastian,

There is the same value on the picture on the page 472 for y_pred in the columns for probabilities (BCELoss) and logits (BCEWithLogitsLoss): 0.8
Probably the value for the first column (BCELoss) is 0.69, which is equal to sigmoid(0.8)?

Thank you.

[rasbt]

Thanks, again! Yeah, looking at it, it doesn't make sense to have the same values there. If we assume that the logits are 0.8, then the input for BCELoss should be 0.69.

Or, vice versa, if we assume that the predicted class membership proba from calling sigmoid(y_pred) = 0.8, then the logits should be 1.386.

Let me CC Hayden, who has the original figure. @haydenliu Can you update the figure and open a PR? Thanks!!

[labdmitriy]

Probably y_pred=0.8 for BCEWithLogitsLoss is more consistent with the code you provided in the book on the same page (and the second row for multiclass classification is already consistent with the code):

logits = torch.tensor([0.8])
probas = torch.sigmoid(logits)

[rasbt]

Good point

[haydenliu]

Thanks Dmitry and Seb.
Merge request: #33

[rasbt]

Thanks, just merged!