returns: logits: tensor of shape (n,num classes,h,w) representing class scores at each pixel yhat: tensor of shape (n,h,w) representing predicted labels at each pixel main loss: loss computed on output of final classifier aux loss:loss computed on output of auxiliary classifier (from intermediate output). note: aux loss is set to a dummy value, since we are not using an auxiliary classifier here, but we keep the same api as pspnet in the next section