For single class segmentation , how to use num_classes , and mask ?

[Balakumaran-kandula]

I have a single class dataset , i am using num_classes here as 2 , and added the background also as a class (i,e) created mask with value 0 as the actual class and mask with value 1 as background class. So inside cross_entropy loss (batch_size,2,w,h) is the predicted output and (batch_size,w,h) as actual label with ignore_index being some random number other than 0 and 1. Whether this approach is correct or not ?

[MeowZheng]

I think it would be better to use binary cross-entropy and set num_classes as 1.

[tarunn2799]

@MeowZheng in that case how should I define my custom dataset, and what should the pixel values be for my images?

Will the classes be (background, foreground) and foreground will still have pixel values as 1?

[PeixinGuo3]

I think it would be better to use binary cross-entropy and set num_classes as 1.

I have a releated issue. When I replace "CrossEntropyLoss" to "binary_cross_entropy", output an error:
KeyError: 'EncoderDecoder: "DepthwiseSeparableASPPHead: 'binary_cross_entropy is not in the models registry'"'
Hope I didn't get the wrong name of loss function, I get "binary_cross_entropy" from "mmseg/models/losses/init.py"

[rouchoo]

I also have a releated issue. When I replace "CrossEntropyLoss" to "binary_cross_entropy", output an error:
KeyError: 'EncoderDecoder: "DepthwiseSeparableASPPHead: 'binary_cross_entropy is not in the models registry'"'
Hope I didn't get the wrong name of loss function, I get "binary_cross_entropy" from "mmseg/models/losses/init.py"

[fineday0505]

Even though I cannot point out what the problem is unless seeing your code,
you have to change the "use_sigmoid" from "False" to "True" in "CrossEntropyLoss" when you use binary_cross_entropy.
I wish this may help you.