Heya! New to the course. Started this week.
I managed to get everything working, so I decided to try the steps out on a small dataset. Currently I’m looking at about four categories of ~100-200 images each.
I went through once and had an error of approximately 50%. No big deal. I didn’t bother with any kind of image cleaning or anything.
After cleaning I’m down to around 30% error. So clearly it had an affect, but I’m missing some other kind of secret sauce.
My question is more of one of scale. I plan to go through and clean the images out again, but is it a case where I would just need more images? I figured that my dataset was pretty dang small, but I kind of want to have an idea of how many I should be aiming for with resnet34 and resnet50. Are we looking at 10^3? 10^4? Higher than that? Do we get diminishing returns after a certain point?
Thanks for any tips!