CrossEntropyLoss doc say the target value must be:
0≤targets[i]≤C−1
And I have 4 type of objects which must be segmented,in this case C==4,and if I want to use single channel target image where all values must be <=4,if 0 for background,1 for the first type of objects…and How to encode the FOURTH type of objects?Or should outputs must be N+1(treating background as a class),in my case,the output channels will be 5(4+background)?
You would treat the background as a separate class, thus your target mask will have values in the range [0, 4]
.