Per-class loss weights? Unlabeled/void class?

[svohara]

In caffe-segnet, you can specify different weights on the loss for each class, and are able to specify an "ignore class" to handle unlabeled/void areas of each image. Where in the neon_segnet is this done? Do you find that there's no need for per-class loss weighting to reproduce the accuracy reported in the paper?