B
26

Vent: That "AI expert" told me to train my model on 50 samples and it'd be fine

Some guy on a LinkedIn post swore 50 labeled images was enough for a defect detector. I burned 3 weeks of compute time and got a model that couldn't tell a scratch from a shadow lol. Has anyone else gotten bad advice from online "gurus" that just wasted your time?
3 comments

Log in to join the discussion

Log In
3 Comments
theagibson
Burned a week on a guru's "just use transfer learning" advice that failed completely.
7
ray_campbell46
Which specific dataset and base model were you trying to transfer from? I had a similar mess when I tried using a ResNet50 trained on ImageNet for a medical imaging problem, and those features were just useless for my tiny 200 image set. Did you try unfreezing just the last few layers first or did you go all in on full fine-tuning?
5
nancycooper
Those gurus have never shipped a model to production, guaranteed.
4