"""Stacking classifier and regressor."""

# Authors: Guillaume Lemaitre <g.lemaitre58@gmail.com>
# License: BSD 3 clause

from abc import ABCMeta, abstractmethod
from copy import deepcopy

import numpy as np
from joblib import Parallel, delayed
import scipy.sparse as sparse

from ..base import clone
from ..base import ClassifierMixin, RegressorMixin, TransformerMixin
from ..base import is_classifier, is_regressor
from ..utils._estimator_html_repr import _VisualBlock

from ._base import _fit_single_estimator
from ._base import _BaseHeterogeneousEnsemble

from ..linear_model import LogisticRegression
from ..linear_model import RidgeCV

from ..model_selection import cross_val_predict
from ..model_selection import check_cv

from ..preprocessing import LabelEncoder

from ..utils import Bunch
from ..utils.metaestimators import if_delegate_has_method
from ..utils.multiclass import check_classification_targets
from ..utils.validation import check_is_fitted
from ..utils.validation import column_or_1d
from ..utils.validation import _deprecate_positional_args


class _BaseStacking(TransformerMixin, _BaseHeterogeneousEnsemble,
                    metaclass=ABCMeta):
    """Base class for stacking method."""

    @abstractmethod
    def __init__(self, estimators, final_estimator=None, *, cv=None,
                 stack_method='auto', n_jobs=None, verbose=0,
                 passthrough=False):
        super().__init__(estimators=estimators)
        self.final_estimator = final_estimator
        self.cv = cv
        self.stack_method = stack_method
        self.n_jobs = n_jobs
        self.verbose = verbose
        self.passthrough = passthrough

    def _clone_final_estimator(self, default):
        if self.final_estimator is not None:
            self.final_estimator_ = clone(self.final_estimator)
        else:
            self.final_estimator_ = clone(default)

    def _concatenate_predictions(self, X, predictions):
        """Concatenate the predictions of each first layer learner and
        possibly the input dataset `X`.

        If `X` is sparse and `self.passthrough` is False, the output of
        `transform` will be dense (the predictions). If `X` is sparse
        and `self.passthrough` is True, the output of `transform` will
        be sparse.

        This helper is in charge of ensuring the predictions are 2D arrays and
        it will drop one of the probability column when using probabilities
        in the binary case. Indeed, the p(y|c=0) = 1 - p(y|c=1)
        """
        X_meta = []
        for est_idx, preds in enumerate(predictions):
            # case where the the estimator returned a 1D array
            if preds.ndim == 1:
                X_meta.append(preds.reshape(-1, 1))
            else:
                if (self.stack_method_[est_idx] == 'predict_proba' and
                        len(self.classes_) == 2):
                    # Remove the first column when using probabilities in
                    # binary classification because both features are perfectly
                    # collinear.
                    X_meta.append(preds[:, 1:])
                else:
                    X_meta.append(preds)
        if self.passthrough:
            X_meta.append(X)
            if sparse.issparse(X):
                return sparse.hstack(X_meta, format=X.format)

        return np.hstack(X_meta)

    @staticmethod
    def _method_name(name, estimator, method):
        if estimator == 'drop':
            return None
        if method == 'auto':
            if getattr(estimator, 'predict_proba', None):
                return 'predict_proba'
            elif getattr(estimator, 'decision_function', None):
                return 'decision_function'
            else:
                return 'predict'
        else:
            if not hasattr(estimator, method):
                raise ValueError('Underlying estimator {} does not implement '
                                 'the method {}.'.format(name, method))
            return method

    def fit(self, X, y, sample_weight=None):
        """Fit the estimators.

        Parameters
        ----------
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training vectors, where `n_samples` is the number of samples and
            `n_features` is the number of features.

        y : array-like of shape (n_samples,)
            Target values.

        sample_weight : array-like of shape (n_samples,) or default=None
            Sample weights. If None, then samples are equally weighted.
            Note that this is supported only if all underlying estimators
            support sample weights.

            .. versionchanged:: 0.23
               when not None, `sample_weight` is passed to all underlying
               estimators

        Returns
        -------
        self : object
        """
        # all_estimators contains all estimators, the one to be fitted and the
        # 'drop' string.
        names, all_estimators = self._validate_estimators()
        self._validate_final_estimator()

        stack_method = [self.stack_method] * len(all_estimators)

        # Fit the base estimators on the whole training data. Those
        # base estimators will be used in transform, predict, and
        # predict_proba. They are exposed publicly.
        self.estimators_ = Parallel(n_jobs=self.n_jobs)(
            delayed(_fit_single_estimator)(clone(est), X, y, sample_weight)
            for est in all_estimators if est != 'drop'
        )
        self.n_features_in_ = self.estimators_[0].n_features_in_

        self.named_estimators_ = Bunch()
        est_fitted_idx = 0
        for name_est, org_est in zip(names, all_estimators):
            if org_est != 'drop':
                self.named_estimators_[name_est] = self.estimators_[
                    est_fitted_idx]
                est_fitted_idx += 1
            else:
                self.named_estimators_[name_est] = 'drop'

        # To train the meta-classifier using the most data as possible, we use
        # a cross-validation to obtain the output of the stacked estimators.

        # To ensure that the data provided to each estimator are the same, we
        # need to set the random state of the cv if there is one and we need to
        # take a copy.
        cv = check_cv(self.cv, y=y, classifier=is_classifier(self))
        if hasattr(cv, 'random_state') and cv.random_state is None:
            cv.random_state = np.random.RandomState()

        self.stack_method_ = [
            self._method_name(name, est, meth)
            for name, est, meth in zip(names, all_estimators, stack_method)
        ]
        fit_params = ({"sample_weight": sample_weight}
                      if sample_weight is not None
                      else None)
        predictions = Parallel(n_jobs=self.n_jobs)(
            delayed(cross_val_predict)(clone(est), X, y, cv=deepcopy(cv),
                                       method=meth, n_jobs=self.n_jobs,
                                       fit_params=fit_params,
                                       verbose=self.verbose)
            for est, meth in zip(all_estimators, self.stack_method_)
            if est != 'drop'
        )

        # Only not None or not 'drop' estimators will be used in transform.
        # Remove the None from the method as well.
        self.stack_method_ = [
            meth for (meth, est) in zip(self.stack_method_, all_estimators)
            if est != 'drop'
        ]

        X_meta = self._concatenate_predictions(X, predictions)
        _fit_single_estimator(self.final_estimator_, X_meta, y,
                              sample_weight=sample_weight)

        return self

    def _transform(self, X):
        """Concatenate and return the predictions of the estimators."""
        check_is_fitted(self)
        predictions = [
            getattr(est, meth)(X)
            for est, meth in zip(self.estimators_, self.stack_method_)
            if est != 'drop'
        ]
        return self._concatenate_predictions(X, predictions)

    @if_delegate_has_method(delegate='final_estimator_')
    def predict(self, X, **predict_params):
        """Predict target for X.

        Parameters
        ----------
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training vectors, where n_samples is the number of samples and
            n_features is the number of features.

        **predict_params : dict of str -> obj
            Parameters to the `predict` called by the `final_estimator`. Note
            that this may be used to return uncertainties from some estimators
            with `return_std` or `return_cov`. Be aware that it will only
            accounts for uncertainty in the final estimator.

        Returns
        -------
        y_pred : ndarray of shape (n_samples,) or (n_samples, n_output)
            Predicted targets.
        """

        check_is_fitted(self)
        return self.final_estimator_.predict(
            self.transform(X), **predict_params
        )

    def _sk_visual_block_(self, final_estimator):
        names, estimators = zip(*self.estimators)
        parallel = _VisualBlock('parallel', estimators, names=names,
                                dash_wrapped=False)
        serial = _VisualBlock('serial', (parallel, final_estimator),
                              dash_wrapped=False)
        return _VisualBlock('serial', [serial])


class StackingClassifier(ClassifierMixin, _BaseStacking):
    """Stack of estimators with a final classifier.

    Stacked generalization consists in stacking the output of individual
    estimator and use a classifier to compute the final prediction. Stacking
    allows to use the strength of each individual estimator by using their
    output as input of a final estimator.

    Note that `estimators_` are fitted on the full `X` while `final_estimator_`
    is trained using cross-validated predictions of the base estimators using
    `cross_val_predict`.

    .. versionadded:: 0.22

    Read more in the :ref:`User Guide <stacking>`.

    Parameters
    ----------
    estimators : list of (str, estimator)
        Base estimators which will be stacked together. Each element of the
        list is defined as a tuple of string (i.e. name) and an estimator
        instance. An estimator can be set to 'drop' using `set_params`.

    final_estimator : estimator, default=None
        A classifier which will be used to combine the base estimators.
        The default classifier is a `LogisticRegression`.

    cv : int, cross-validation generator or an iterable, default=None
        Determines the cross-validation splitting strategy used in
        `cross_val_predict` to train `final_estimator`. Possible inputs for
        cv are:

        * None, to use the default 5-fold cross validation,
        * integer, to specify the number of folds in a (Stratified) KFold,
        * An object to be used as a cross-validation generator,
        * An iterable yielding train, test splits.

        For integer/None inputs, if the estimator is a classifier and y is
        either binary or multiclass, `StratifiedKFold` is used. In all other
        cases, `KFold` is used.

        Refer :ref:`User Guide <cross_validation>` for the various
        cross-validation strategies that can be used here.

        .. note::
           A larger number of split will provide no benefits if the number
           of training samples is large enough. Indeed, the training time
           will increase. ``cv`` is not used for model evaluation but for
           prediction.

    stack_method : {'auto', 'predict_proba', 'decision_function', 'predict'}, \
            default='auto'
        Methods called for each base estimator. It can be:

        * if 'auto', it will try to invoke, for each estimator,
          `'predict_proba'`, `'decision_function'` or `'predict'` in that
          order.
        * otherwise, one of `'predict_proba'`, `'decision_function'` or
          `'predict'`. If the method is not implemented by the estimator, it
          will raise an error.

    n_jobs : int, default=None
        The number of jobs to run in parallel all `estimators` `fit`.
        `None` means 1 unless in a `joblib.parallel_backend` context. -1 means
        using all processors. See Glossary for more details.

    passthrough : bool, default=False
        When False, only the predictions of estimators will be used as
        training data for `final_estimator`. When True, the
        `final_estimator` is trained on the predictions as well as the
        original training data.

    verbose : int, default=0
        Verbosity level.

    Attributes
    ----------
    classes_ : ndarray of shape (n_classes,)
        Class labels.

    estimators_ : list of estimators
        The elements of the estimators parameter, having been fitted on the
        training data. If an estimator has been set to `'drop'`, it
        will not appear in `estimators_`.

    named_estimators_ : :class:`~sklearn.utils.Bunch`
        Attribute to access any fitted sub-estimators by name.

    final_estimator_ : estimator
        The classifier which predicts given the output of `estimators_`.

    stack_method_ : list of str
        The method used by each base estimator.

    Notes
    -----
    When `predict_proba` is used by each estimator (i.e. most of the time for
    `stack_method='auto'` or specifically for `stack_method='predict_proba'`),
    The first column predicted by each estimator will be dropped in the case
    of a binary classification problem. Indeed, both feature will be perfectly