score broken if batch_iterator_test doesn't yield same order / size as y

[cancan101]

Right now score assumes that the values of X yielded by the batch_iterator_test are the same order / length as the y passed in.

Instead, I suggest the following modifications which uses the y's yielded by the iterator:

    def predict_proba(self, X, y=None):
        probas = []
        ys = []
        for Xb, yb in self.batch_iterator_test(X, y):
            probas.append(self.apply_batch_func(self.predict_iter_, Xb))
            ys.append(yb)
        return np.vstack(probas), np.hstack(ys)

    def predict(self, X, y=None):
        if self.regression:
            return self.predict_proba(X, y)
        else:
            predictions, y_actual = self.predict_proba(X, y)
            y_pred = np.argmax(predictions, axis=1)
            if self.use_label_encoder:
                y_pred = self.enc_.inverse_transform(y_pred)
            return y_pred, y_actual

    def score(self, X, y):
        score = mean_squared_error if self.regression else accuracy_score
        return float(score(*self.predict(X, y)))

In order to match the old signature, we might want to have the return look something like:

if y is not None:
        return np.vstack(probas), np.hstack(ys)
else:
       return np.vstack(probas)