Adrian Tam does some iris categorizing:
Now you need to have a model that can take the input and predict the output, ideally in the form of one-hot vectors. There is no science behind the design of a perfect neural network model. But you know one thing, it has to take in a vector of 4 features and output a vector of 3 values. The 4 features corresponds to what you have in the dataset. The 3-value output is because we know the one-hot vector has 3 elements. Anything can be in between, and those are known as the “hidden layers” since they are neither input nor output.
Click through for the full tutorial.