temperature in logits vs targets

[EMCarrami]

Hi,

Below self.temperature is used in two different manners. While logits are divided by it, targets are multiplied by this value. Is this intentional or is it a mathematical mistake?

Thanks.

        # Calculating the Loss
        logits = (spot_embeddings @ image_embeddings.T) / self.temperature
        images_similarity = image_embeddings @ image_embeddings.T
        spots_similarity = spot_embeddings @ spot_embeddings.T
        targets = F.softmax(
            (images_similarity + spots_similarity) / 2 * self.temperature, dim=-1
        )
        spots_loss = cross_entropy(logits, targets, reduction='none')
        images_loss = cross_entropy(logits.T, targets.T, reduction='none')
        loss =  (images_loss + spots_loss) / 2.0 # shape: (batch_size)

[Xrioen]

Yes this should be a mistake. Good catch. I have now fixed the code. Luckily this shouldn't influence results since temperature was set to 1 for all experiments. Thanks for this.