Loading…
8
A guy on a Discord said my detection model was just memorizing, so I ran a real test
I built a small image classifier last month on 4,000 photos and it hit 96% accuracy, so I felt pretty good about it. Someone in the server said "that's just your model memorizing the same backgrounds," so I re-split my data by location and accuracy fell to 61%, which was a rough pill to swallow. Now I split by place before I even look at results, but does anyone have a good rule for spotting this kind of leak before you waste two days?
0 comments
Log in to join the discussion
Log In0 Comments
No comments yet
Be the first to share your thoughts on this discussion.