Open Keinitial opened 3 months ago
Excellent job! Hello! Do you know why the model suddenly gets stuck and stops training during the training process? Thank you for your attention.
hello!Glad you're looking into the author's code too. I now have doubts about the author's niid_extend part of the code, because the original text for the niid partition of the cityscapes dataset is as follows: “Cityscapes [9] and CamVID [3] are two semantic segmentation datasets of street view with 19 and 11 semantic classes, respectively. Unlike classification, an image from semantic segmentation datasets contains objects of many classes that are hard to split. To generate the class-heterogeneous data partition among clients, we split Cityscapes and CamVID into K subsets. Each subset maintains one or two semantic classes and sets other classes as background. K is set to 19 and 11 for Cityscapes and CamVID, respectively. In this setting, there exists an inconsistent foreground-background problem for different clients. "But I did not understand how the author reflected the distinction between foreground and background in the code
And I don't know why, after using the BackCE loss function, the mIoU of FedAvg is always about 20%, can't rise
Excellent job! Hello! Do you know why the model suddenly gets stuck and stops training during the training process? Thank you for your attention.