I'm not experienced enough with ML to provide great feedback, but I did start getting segfault errors when I trained on too many images. I have multiple pictures of nearly 1000 different pinball machines, so right now it's a lot of trial/error as I determine how many pictures of each backglass I should train on and how diverse of angles/lighting produce the most accurate model. Right now I'm at about 15 different pinball machines, each with 20-30 pictures, and I get around 85% accuracy using my test set. I suspect/worry I won't be able to make a good model that can identify 1000 different pinball machines, but my patrons have asked for it so I'm giving it my all.
It might be an idea to train the same data with fast.ai and see where it leads you. If you have the data in the right structure, it should not take too long. Let me know if you have questions... :)