Question about loss about protoseg

[yannqi]

Hi @XiaoxxWang, our PPC loss takes the form of InfoNCE, which is in essence a cross-entropy loss. In our context, you can understand PPC as, for each pixel, we aim to identify its assigned (positive) prototype among a set of negative prototypes.

I also have a question, the PPC code does not seem to reflect the temperature coefficient setting? Or is the temperature coefficient set to 1 by default, which is different from the 0.1 given in the paper?

[tfzhou]

Hi @XiaoxxWang, our PPC loss takes the form of InfoNCE, which is in essence a cross-entropy loss. In our context, you can understand PPC as, for each pixel, we aim to identify its assigned (positive) prototype among a set of negative prototypes.

[XiaoxxWang]

Thanks for your reply, but I still can't understand the PPC loss well. The code shows it is defined as the cross entropy of the proto_logits with proto_targets. The proto_logits is the product of the feature and the prototype, but what does proto_targets mean? It seems it is not the groundtruth , which is not agree with infoNCE loss.

[wenguanwang]

@XiaoxxWang Just read the code more carefully... For, InforNCE, you also have the groundtruth -- you know which one is the positive sample, and which one is the negative sample, but the ground-truth is obtained by free.

[fangs99]

The temperature coefficient In paper is set as 0.1, howerver, in the source code it is neglected. In other words, it is set to 1 by default in source code. Does it have no influnce for the performance ?