Currently PROBE and CLUE have their "own" penalties/losses defined, but dispatched over gradient-based generators. Can move these into Objectives and thus make them accessible to all generators.
For PROBE, there probably should not be a separate convergence condition? Instead the invalidation rate loss should just be treated as a loss/penalty as always.
Currently PROBE and CLUE have their "own" penalties/losses defined, but dispatched over gradient-based generators. Can move these into
Objectivesand thus make them accessible to all generators.For PROBE, there probably should not be a separate convergence condition? Instead the invalidation rate loss should just be treated as a loss/penalty as always.