"cal" variable usage

[ser5kovskiy]

Thank you for your excellent work！ I have a question about condition encoding branch: Additionally to out (~uemb later) variable, cal variable is returned. As I understand, cal was used before and it is not used now. What was (maybe, is) purpose of cal? It seems like it was output of a decoder for condition encoder. This way, condition encoder was trained as part of whole separate segmentation network, yes?

[pedropesserl]

I'm confused about this as well. I'm using the model for segmentation of fire in images, and this variable "cal" appears to be (negatively) impacting the performance. For example, here's one prediction made with MedSegDiff (V1): https://imgur.com/a/EHQMYLw You can see that in every one of the 25 output images, the prediction (top) is more accurate to the original image (middle) than the one contained in the variable "cal" (bottom). Yet the ensemble seems to be "contaminated" with the "cal" because of that dark spot on the top left corner of the predicted mask.

[ZhangFfF1]

@pedropesserl hello, have you solved the problem of cal contaminating the final sampling results?Thank you！