Operators like batch norm contain non-visible outputs which, despite their invisibility, must be propagated to the final output in order to be included in the computation graph. Is there any plan for how this will be done in the new NNVM?
Currently, I’ve worked around the issue by having each node accumulate the “extra outputs” of its parents and just dumping them all into the final graph outputs. This isn’t a particular clean way to do things, though.