Loading…
14
Appreciation post: the 3 week fight with my image classifier that ended because of one typo in a folder name
I was building a small image classifier to sort recycling photos for a school project in Tacoma and it kept labeling everything as trash with 98% confidence no matter what I did. I tried new data, different models, and a bigger GPU before I finally noticed that one of my training folders was named "plactis" instead of "plastic," so half my labels pointed at a ghost class. What is the dumbest setup mistake that ate a big chunk of your time on an AI build?
1 comments
Log in to join the discussion
Log In1 Comment
ivan_harris1d ago
Mine was a CSV where the label column had a trailing space on half the rows, so the model learned two classes that looked the same to me. I spent two nights tuning learning rates before a friend glanced at the file and laughed at me. Check your labels first, always.
5