This week I finally tracked down a bug in training DeOldify that I introduced back in July 2019. It didn't actually noticeably manifest until later in the fall due to ironically better training. Drove. Me. Nuts. Lesson: Ablations are great, but don't trust them 100%.
-
What sort of bug? I've been running into the same sort of thing. The trouble is, you often *have to* introduce these bugs. Because the only way to know whether something will work is to try it – and if it doesn't work, it's called a bug!
The general nature of it was that I was inadvertently building in criteria in training that was inherently conflicting, so the model "split the difference" and glitches resulted. Definitely agree on the "don't know until you try" philosophy- that's complex systems for you!
