This repository holds the demo code for this article I written on this.
As you can see, I take images for the three classes. One neutral, one with my phone in my hand and one with a coffee in my hand. Then when I click “Train & Predict” it trains in a fraction of a second. Finally, for each class, in real time you can see how confident it is that the class matches. Naturally, you can provide more diverse images so it gets better and better and classifying.
You can try the demo out yourself here.