Reading SET cards should be easy, but it isn't

I started taking the fast.ai course in 2018 and made good progress through some lessons and then I wanted to build my own app - one that would solve the card game SET. Each card has a shape, a number of shapes, a fill pattern, and a color. Each of those features has 3 values. I figured how hard can this possibly be for a modern deep learning algorithm! I spent days trying to tune it, but I could not get decent accuracy. I got some suggestions in the forums, but they didn’t seem to work. And then I just gave up. If I can’t solve this simple a problem, why bother to go on to the real things I hope to do… SO, fast forward to now and I’ve bought the new book and want to give this another go. But my concern is still there. This should be easy:

Those are my training images (and the repo has my previous attempts). Does anyone have thoughts on where I went wrong, whether anything has changed, or whether this problem really should be as easy as I think?

Thanks!

1 Like