Loading…
5
Showerthought: fixing a bad label in my training set took 11 days and I still feel dumb about it
Back in March I was building a simple image sorter for my cousin's bait shop down in Biloxi, just sorting photos of lures by type, and the model kept calling every red crankbait a hook. I figured it was a model problem, so I messed with layers and learning rates for over a week. Turned out 40 of the training images were mislabeled by me at 1 in the morning. The actual fix took 2 hours once I found it, but getting there ate 11 days of my evenings. What got me was that the old way of doing this by hand would have caught the bad names right away, because I would have seen each one as I filed it. Funny how faster tools let you make the same dumb mistake way quicker. I now keep a little checklist taped to my screen for label review. Has anyone else had a quiet data bug like that eat your whole schedule?
0 comments
Log in to join the discussion
Log In0 Comments
No comments yet
Be the first to share your thoughts on this discussion.