onehot_transformer.py

import numpy as np

class OneHotTransformer:
    def __init__(self, func):
        self.f = func

    def fit(self, X, y=None):
        unseen = object()
        seen = set()
        for x in X:
            seen.add(self.f(x))
        self.seen = list(sorted(seen)) + [unseen]
        return self

    def transform(self, X):
        return np.array([self.transform_one(x) for x in X])

    def transform_one(self, x):
        result = [0] * len(self.seen)
        value = self.f(x)
        if value in self.seen:
            result[self.seen.index(value)] = 1
        else:
            result[-1] = 1
        return result

Using a list.index is slow, perhaps seen could be defined as: self.seen = {key: i for i, key in enumerate(seen)}

@jmansilla to use an extra column for unseen values is new to me: how has it worked out for you? have you found it useful?