If Semantic segmentation is implemented in below three ways using BCEWithLogitsLoss and Dice score metric, will it give same outputs?
Assume Input Image Dimension [Batch,Channels,height,width] & number of classes is 21(VOC).
a) Converting the labels into one hot encoding where target label Dimension is [Batch,Num_classes,height,width].
b) Converting the labels into categorical values where target label Dimension is [Batch,1,height,width]. implemented repo https://github.com/mapillary/inplace_abn
c) Converting the labels into RGB values where target label Dimension is [Batch,3,height,width] implemented repo https://github.com/jfzhang95/pytorch-deeplab-xception
For all this methods whether prediction output shape should be [Batch,Channels,height,width] ?