Logits or softmax probabilities

[YanWang2014]

In the paper,

Through this loss function, we aim to directly penalize the distance between the predicted output logits.

So for the PairwiseConfusion, we are using logits? which is the direct output of pytorch models.

But for EntropicConfusion, obviously we should use softmax probabilities, which is obtained by feeding logits through a softmax function.

Am I right? Thank you

[abhimanyudubey]

Hi Yan,

For both models of confusion, the preferred usage is softmax probabilities (obtained after feeding the logits through a softmax). You can, alternatively, try the logits for pairwise confusion, but the loss weight will have to be scaled to a very small value to prevent oscillations, and hence we recommend operating on the softmax probabilities themselves.

Abhi

[YanWang2014]

Thank you!

[twmht]

@abhimanyudubey

In the Entropy confusion (https://github.com/abhimanyudubey/confusion/blob/master/confusion_pytorch/__init__.py#L15), it seems that you missed a negative sign according to the paper. isn't this a loss function?