Loading…
16
Shoutout to the guy at Micro Center who saved me from a fatal mistake with my training data
I used to label every image myself, one by one, and thought that was just the job. Then a data guy at Micro Center in Chicago told me my labels were dirty and that's why my model stalled at 78 percent for three weeks. He showed me how to run a quick check that flags mismatched labels before training even starts. I reran the same batch and accuracy jumped to 91 percent in two days. It's not a fix-all, but if your model keeps failing on the same class, ask yourself if the labels are actually correct. Anyone else find bad labels were the real bottleneck?
0 comments
Log in to join the discussion
Log In0 Comments
No comments yet
Be the first to share your thoughts on this discussion.