What loss function and metric should I use for multi-label classification in keras?

Question

Model\'s final activation is softmax.(output means importance in my case.) I want to pick top 3, then I use categorical crossentropy for loss function / accuracy for metric.