Loading…
4
Hot take: the guy who said 'just fine tune it' about our image tagger in Austin was dead wrong
Three months ago a buddy from the Austin meetup swore retraining the last layer on 500 photos would fix our mislabeled drone shots, but accuracy went from 71% to 74% and we had to rebuild the whole dataset with 8,000 images instead, so has anyone else been burned by the "just fine tune it" advice?
1 comments
Log in to join the discussion
Log In1 Comment
jason111h ago
That 71 to 74 jump is the part that gets me, because it sounds like a win on paper until you run it on real footage. I hit the same wall last fall with a warehouse pallet counter, spent two weekends labeling 600 shots, only to watch it choke on anything with a weird shadow. The thing nobody tells you is that fine tuning only works if your training data actually looks like what the model sees in the wild. If your 500 drone photos were all clear skies and flat roofs, the last layer just learns those exact conditions, and the rest of the network (which is where the real features live) never gets a reason to change. So you end up with a model that's great at the 500 photos and clueless about everything else, which is exactly why 8,000 images and a full rebuild is what actually moved the needle. Your Austin buddy wasn't lying on purpose, he just left out the part where fine tuning is a data problem, not a tuning problem.
8