I just start reading about tensorflow.

Is it a crazy idea to train a model with about 40 classes?

How much images should I use at least for one class?

I trained it with pictures of semiauto pistols and revolvers, 1000 images for semi, ~200 for revolvers. Didn't make a big difference, but the semiauto was slightly more accurate - might just be the clearer shape though, not sure.

It ended up recognizing my monkey wrench as a revolver, no false negatives from the side and a few with "downrange" pictures.

Thanks. So I try to get 200 to 300 pictures for each category.

I plan to reduce it to 28x28 pixel and grayscale.

Good plan?

@JackToolsNet No idea there, but I had several resolutions starting from 300x300, in color. I'd say it depends on your use case, but that's beyond my knowledge.

@ilPinguino OK. Thanks for your help.