Loading…
7
Shoutout to the old timer at the Mesa meetup who told me to shrink my training set by 60 percent
Spent 3 weeks tuning a model on 40k labeled images from our warehouse cameras. 91 percent accuracy. Felt good.
He looks at my confusion matrix for 10 seconds and says your labels are junk, cut it down and fix what's left. So I pulled it to 16k and cleaned them by hand over a weekend.
Same accuracy in half the training time. Turns out 24,000 of those labels were near duplicates from the same 6 camera angles.
Anyone else learn the hard way that more data is not always the move?
1 comments
Log in to join the discussion
Log In1 Comment
henryreed12h ago
Cutting data is a risky habit to build though. You got lucky that your cameras were the problem, but in most cases those extra 24k near dupes still help the model see small changes in lighting, angle, and motion blur. Take them out and your 91 percent might hold on the test set while falling apart on a new camera you have not seen yet. A better move would have been keeping the data and just adding a dedupe step or weighting the samples, since storage is cheap and retraining is not.
2