I cheated here (this _is_ a binary perceptron), but the conversion is super easy with one-vs-rest or one-vs-one (training all pairs and then doing pairwise comp

I cheated here (this _is_ a binary perceptron), but the conversion is super easy with one-vs-rest or one-vs-one (training all pairs and then doing pairwise comparison until we find the one that matches), that would still take ~20 seconds to train.