Loading…
30
Rant: my little plant leaf spotter broke on day 3 and the fix taught me more than the tool did
So I built this tiny image classifier to spot leaf spots on my tomato plants, nothing fancy, just a small model I trained on maybe 200 photos I took on my phone in the backyard here in Boise. On day 3 it started labeling every single leaf as "healthy" even the ones with clear brown patches, and I sat there staring at it feeling like a total idiot. Turns out I had shuffled my training data wrong and all the sick leaf photos ended up in the test folder, so the model never actually learned what a spot looked like. I spent about 4 hours re-sorting 200 images into a clean 80/20 split and retrained it on my old laptop, and the thing jumped from 50% accuracy (basically a coin flip) to 87%. The weird part is I learned more in those 4 hours of breaking it than in the 2 weeks of reading tutorials before that. I also added a simple check that prints out how many images are in each folder before training, which would have saved me the whole mess. If you have ever trained a small model on a tiny dataset, what is your go-to sanity check before you hit run?
1 comments
Log in to join the discussion
Log In1 Comment
stone.evan3d ago
Wow, a buddy did the same thing with his fern pics.
7