DAE remember hand labeling training images at 3am, then finding out about half were wrong
Spent most of 2016 sorting 4,000 cat photos for a classifier at a lab in Ann Arbor, and a coworker ran a quick sanity check that showed my labels were a mess. What is the dumbest thing you have caught yourself doing on a dataset way too late?
Labeled 3,000 images of birds for a grad project once and realized around image 2,800 that I had been tagging every single one of them as "sparrow" out of pure autopilot. My accuracy was basically however many actual sparrows happened to be in there, which was maybe 12 percent. Spent a whole weekend redoing it while questioning every choice I had made up to that point.