mlen superlearner for MIMO multi-input multi-output about mlens

[flennerhag]

Hi! Thanks for raising this issue - this would indeed be a good feature to have. For the time being, it isn't possible with the ready-made ensemble classes we have. We can easily fix that with some tweaks. Basically, you need to set the multiplier argument in the Learner class when assigning columns. Currently, this is defaulting to 1, and this argument is not exposed at the ensemble level (which we probably should).

For the time being, you need to set up a custom class

from mlens.parallel import Learner, Transformer, Pipeline, Group, Layer, make_group
from mlens.ensemble import BaseEnsemble
from mlens.index import FoldIndex, FullIndex
from mlens.utils import check_instances
from mlens.ensemble.base import check_kwargs

# First, your new Learner. ``num_targets`` will be your multi-output dimensionality.

class MultiLearner(Learner):

    def __init__(self, estimator, num_targets, **kwargs):
        super(MultiLearner, self).__init__(estimator, **kwargs)
        self.num_targets = num_targets

    def _get_multiplier(self, X, y):
        return self.num_targets


def make_multi_group(indexer, estimators, preprocessing,
                     learner_kwargs=None, transformer_kwargs=None, name=None):
    preprocessing, estimators = check_instances(estimators, preprocessing)

    if learner_kwargs is None:
        learner_kwargs = {}
    if transformer_kwargs is None:
        transformer_kwargs = {}

    transformers = [Transformer(estimator=Pipeline(tr, return_y=True),
                                name=case_name, **transformer_kwargs)
                    for case_name, tr in preprocessing]

    # We use your new MultiLearner class here
    learners = [MultiLearner(estimator=est, preprocess=case_name,
                        name=learner_name, **learner_kwargs)
                for case_name, learner_name, est in estimators]

    group = Group(indexer=indexer, learners=learners,
                  transformers=transformers, name=name)
    return group


# Change the make_group function in the base ensemble class


class MultiBaseEnsemble(BaseEnsemble):

    def _build_layer(self, estimators, indexer, preprocessing, **kwargs):
        check_kwargs(kwargs, ['backend', 'n_jobs'])
        verbose = kwargs.pop('verbose', max(self._backend.verbose - 1, 0))
        dtype = kwargs.pop('dtype', self._backend.dtype)
        propagate = kwargs.pop('propagate_features', None)
        shuffle = kwargs.pop('shuffle', self.shuffle)
        random_state = kwargs.pop('random_state', self.random_state)
        rs = kwargs.pop('raise_on_exception', self.raise_on_exception)
        if random_state:
            random_state = check_random_state(random_state).randint(0, 10000)

        kwargs['verbose'] = max(verbose - 1, 0)
        kwargs['scorer'] = kwargs.pop('scorer', self.scorer)

        # We use your make_multi_group function from above
        group = make_multi_group(indexer, estimators, preprocessing, kwargs)

        name = "layer-%i" % (len(self._backend.stack) + 1)  # Start count at 1
        lyr = Layer(
            name=name, dtype=dtype, shuffle=shuffle,
            random_state=random_state, verbose=verbose,
            raise_on_exception=rs, propagate_features=propagate)
        lyr.push(group)
        return lyr
    

# Finally, build the SuperLearner (or similar)

class MultiSuperLearner(MultiBaseEnsemble):

    def __init__(
            self, folds=2, shuffle=False, random_state=None, scorer=None,
            raise_on_exception=True, array_check=None, verbose=False, n_jobs=-1,
            backend='threading', model_selection=False, sample_size=20, layers=None):
        super(MultiSuperLearner, self).__init__(
            shuffle=shuffle, random_state=random_state, scorer=scorer,
            raise_on_exception=raise_on_exception, verbose=verbose,
            n_jobs=n_jobs, layers=layers, backend=backend,
            array_check=array_check, model_selection=model_selection,
            sample_size=sample_size)

        self.__initialized__ = 0  # Unlock parameter setting
        self.folds = folds
        self.__initialized__ = 1  # Protect against param resets

    def add_meta(self, estimator, **kwargs):
        return self.add(estimators=estimator, meta=True, **kwargs)

    def add(self, estimators, num_targets=1, preprocessing=None,
            proba=False, meta=False, propagate_features=None, **kwargs):
        c = kwargs.pop('folds', self.folds)

        if meta:
            idx = FullIndex()
        else:
            idx = FoldIndex(c, raise_on_exception=self.raise_on_exception)

        return super(MultiSuperLearner, self).add(
            estimators=estimators, num_targets=num_targets, indexer=idx, preprocessing=preprocessing,
            proba=proba, propagate_features=propagate_features, **kwargs)

You can now use this class as you normally would:

X, y = make_regression(n_samples=1000, n_features=10, n_informative=5, n_targets=2, random_state=1, noise=0.5)

ensemble = MultiSuperLearner()
ensemble.add([LinearRegression()], num_targets=2)
ensemble.add_meta(LinearRegression(), num_targets=2)

ensemble.fit(X, y)

[Bahar1978]

Thank you very much for the clarification in details. I have reused this code and used my data(after normalization and applying PCA). In my case the num_targets is number of PCA features forexample num_targets=7.
It gives me answers but I still have some questions:
1- In order to see the score I added the score = rmse, is this the correct way?
return super(MultiSuperLearner, self).add( scorer=rmse, estimators=estimators, num_targets=num_targets, indexer=idx, preprocessing=preprocessing, proba=proba, propagate_features=propagate_features, **kwargs)
2- Is num_targets = 1 is correct or it should be equal to num_targets?
def add(self, estimators, num_targets=1, preprocessing=None,

3- Using superlearner or multisuperlearner we should have the best of all the models in terms of score, yes? but it does not give the Min RMSE:
there are my 2 output using 2 different inputs:

4- could you please also let me know where/how can we set the ft-m/ft-s ? in the aforementioned example it these values are not correct.

again thanks for your time and kindness.