a question about loss for background

[liyunsheng13]

❓ Questions and Help

Hi

In your paper of Equation 6, you build the supervised loss for regions. I'm curious whether for the background regions, the numerator of Equation 6 should be 1, while the denominator reminds the same. Besides, I do not find the implementation of this part in your code. Could you point it out for me for further reference?

Thanks

[alirezazareian]

Hi Yunsheng,

Thanks for your comment. The loss is a regular CrossEntropyLoss, and does not need a custom implementation. The only difference is how the logits are computed. In open-vocabulary settings, we set the background embedding to an all-zero vector here. This means the exponential of dot product between background weights and any input will be exactly 1. Hence you are right, the numerator will be 1.