"""Partial dependence plots for regression and classification models."""

# Authors: Peter Prettenhofer
#          Trevor Stephens
#          Nicolas Hug
# License: BSD 3 clause

from collections.abc import Iterable

import numpy as np
from scipy import sparse
from scipy.stats.mstats import mquantiles

from ..base import is_classifier, is_regressor
from ..pipeline import Pipeline
from ..utils.extmath import cartesian
from ..utils import check_array
from ..utils import check_matplotlib_support  # noqa
from ..utils import _safe_indexing
from ..utils import _determine_key_type
from ..utils import _get_column_indices
from ..utils.validation import check_is_fitted
from ..utils.validation import _deprecate_positional_args
from ..tree import DecisionTreeRegressor
from ..ensemble import RandomForestRegressor
from ..exceptions import NotFittedError
from ..ensemble._gb import BaseGradientBoosting
from sklearn.ensemble._hist_gradient_boosting.gradient_boosting import (
    BaseHistGradientBoosting)


__all__ = [
    'partial_dependence',
]


def _grid_from_X(X, percentiles, grid_resolution):
    """Generate a grid of points based on the percentiles of X.

    The grid is a cartesian product between the columns of ``values``. The
    ith column of ``values`` consists in ``grid_resolution`` equally-spaced
    points between the percentiles of the jth column of X.
    If ``grid_resolution`` is bigger than the number of unique values in the
    jth column of X, then those unique values will be used instead.

    Parameters
    ----------
    X : ndarray, shape (n_samples, n_target_features)
        The data

    percentiles : tuple of floats
        The percentiles which are used to construct the extreme values of
        the grid. Must be in [0, 1].

    grid_resolution : int
        The number of equally spaced points to be placed on the grid for each
        feature.

    Returns
    -------
    grid : ndarray, shape (n_points, n_target_features)
        A value for each feature at each point in the grid. ``n_points`` is
        always ``<= grid_resolution ** X.shape[1]``.

    values : list of 1d ndarrays
        The values with which the grid has been created. The size of each
        array ``values[j]`` is either ``grid_resolution``, or the number of
        unique values in ``X[:, j]``, whichever is smaller.
    """
    if not isinstance(percentiles, Iterable) or len(percentiles) != 2:
        raise ValueError("'percentiles' must be a sequence of 2 elements.")
    if not all(0 <= x <= 1 for x in percentiles):
        raise ValueError("'percentiles' values must be in [0, 1].")
    if percentiles[0] >= percentiles[1]:
        raise ValueError('percentiles[0] must be strictly less '
                         'than percentiles[1].')

    if grid_resolution <= 1:
        raise ValueError("'grid_resolution' must be strictly greater than 1.")

    values = []
    for feature in range(X.shape[1]):
        uniques = np.unique(_safe_indexing(X, feature, axis=1))
        if uniques.shape[0] < grid_resolution:
            # feature has low resolution use unique vals
            axis = uniques
        else:
            # create axis based on percentiles and grid resolution
            emp_percentiles = mquantiles(
                _safe_indexing(X, feature, axis=1), prob=percentiles, axis=0
            )
            if np.allclose(emp_percentiles[0], emp_percentiles[1]):
                raise ValueError(
                    'percentiles are too close to each other, '
                    'unable to build the grid. Please choose percentiles '
                    'that are further apart.')
            axis = np.linspace(emp_percentiles[0],
                               emp_percentiles[1],
                               num=grid_resolution, endpoint=True)
        values.append(axis)

    return cartesian(values), values


def _partial_dependence_recursion(est, grid, features):
    averaged_predictions = est._compute_partial_dependence_recursion(grid,
                                                                     features)
    if averaged_predictions.ndim == 1:
        # reshape to (1, n_points) for consistency with
        # _partial_dependence_brute
        averaged_predictions = averaged_predictions.reshape(1, -1)

    return averaged_predictions


def _partial_dependence_brute(est, grid, features, X, response_method):
    averaged_predictions = []

    # define the prediction_method (predict, predict_proba, decision_function).
    if is_regressor(est):
        prediction_method = est.predict
    else:
        predict_proba = getattr(est, 'predict_proba', None)
        decision_function = getattr(est, 'decision_function', None)
        if response_method == 'auto':
            # try predict_proba, then decision_function if it doesn't exist
            prediction_method = predict_proba or decision_function
        else:
            prediction_method = (predict_proba if response_method ==
                                 'predict_proba' else decision_function)
        if prediction_method is None:
            if response_method == 'auto':
                raise ValueError(
                    'The estimator has no predict_proba and no '
                    'decision_function method.'
                )
            elif response_method == 'predict_proba':
                raise ValueError('The estimator has no predict_proba method.')
            else:
                raise ValueError(
                    'The estimator has no decision_function method.')

    for new_values in grid:
        X_eval = X.copy()
        for i, variable in enumerate(features):
            if hasattr(X_eval, 'iloc'):
                X_eval.iloc[:, variable] = new_values[i]
            else:
                X_eval[:, variable] = new_values[i]

        try:
            predictions = prediction_method(X_eval)
        except NotFittedError:
            raise ValueError(
                "'estimator' parameter must be a fitted estimator")

        # Note: predictions is of shape
        # (n_points,) for non-multioutput regressors
        # (n_points, n_tasks) for multioutput regressors
        # (n_points, 1) for the regressors in cross_decomposition (I think)
        # (n_points, 2) for binary classification
        # (n_points, n_classes) for multiclass classification

        # average over samples
        averaged_predictions.append(np.mean(predictions, axis=0))

    # reshape to (n_targets, n_points) where n_targets is:
    # - 1 for non-multioutput regression and binary classification (shape is
    #   already correct in those cases)
    # - n_tasks for multi-output regression
    # - n_classes for multiclass classification.
    averaged_predictions = np.array(averaged_predictions).T
    if is_regressor(est) and averaged_predictions.ndim == 1:
        # non-multioutput regression, shape is (n_points,)
        averaged_predictions = averaged_predictions.reshape(1, -1)
    elif is_classifier(est) and averaged_predictions.shape[0] == 2:
        # Binary classification, shape is (2, n_points).
        # we output the effect of **positive** class
        averaged_predictions = averaged_predictions[1]
        averaged_predictions = averaged_predictions.reshape(1, -1)

    return averaged_predictions


@_deprecate_positional_args
def partial_dependence(estimator, X, features, *, response_method='auto',
                       percentiles=(0.05, 0.95), grid_resolution=100,
                       method='auto'):
    """Partial dependence of ``features``.

    Partial dependence of a feature (or a set of features) corresponds to
    the average response of an estimator for each possible value of the
    feature.

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

    .. warning::

        For :class:`~sklearn.ensemble.GradientBoostingClassifier` and
        :class:`~sklearn.ensemble.GradientBoostingRegressor`, the
        'recursion' method (used by default) will not account for the `init`
        predictor of the boosting process. In practice, this will produce
        the same values as 'brute' up to a constant offset in the target
        response, provided that `init` is a constant estimator (which is the
        default). However, if `init` is not a constant estimator, the
        partial dependence values are incorrect for 'recursion' because the
        offset will be sample-dependent. It is preferable to use the 'brute'
        method. Note that this only applies to
        :class:`~sklearn.ensemble.GradientBoostingClassifier` and
        :class:`~sklearn.ensemble.GradientBoostingRegressor`, not to
        :class:`~sklearn.ensemble.HistGradientBoostingClassifier` and
        :class:`~sklearn.ensemble.HistGradientBoostingRegressor`.

    Parameters
    ----------
    estimator : BaseEstimator
        A fitted estimator object implementing :term:`predict`,
        :term:`predict_proba`, or :term:`decision_function`.
        Multioutput-multiclass classifiers are not supported.

    X : {array-like or dataframe} of shape (n_samples, n_features)
        ``X`` is used to generate a grid of values for the target
        ``features`` (where the partial dependence will be evaluated), and
        also to generate values for the complement features when the
        `method` is 'brute'.

    features : array-like of {int, str}
        The feature (e.g. `[0]`) or pair of interacting features
        (e.g. `[(0, 1)]`) for which the partial dependency should be computed.

    response_method : 'auto', 'predict_proba' or 'decision_function', \
            optional (default='auto')
        Specifies whether to use :term:`predict_proba` or
        :term:`decision_function` as the target response. For regressors
        this parameter is ignored and the response is always the output of
        :term:`predict`. By default, :term:`predict_proba` is tried first
        and we revert to :term:`decision_function` if it doesn't exist. If
        ``method`` is 'recursion', the response is always the output of
        :term:`decision_function`.

    percentiles : tuple of float, optional (default=(0.05, 0.95))
        The lower and upper percentile used to create the extreme values
        for the grid. Must be in [0, 1].

    grid_resolution : int, optional (default=100)
        The number of equally spaced points on the grid, for each target
        feature.

    method : str, optional (default='auto')
        The method used to calculate the averaged predictions:

        - 'recursion' is only supported for some tree-based estimators (namely
          :class:`~sklearn.ensemble.GradientBoostingClassifier`,
          :class:`~sklearn.ensemble.GradientBoostingRegressor`,
          :class:`~sklearn.ensemble.HistGradientBoostingClassifier`,
          :class:`~sklearn.ensemble.HistGradientBoostingRegressor`,
          :class:`~sklearn.tree.DecisionTreeRegressor`,
          :class:`~sklearn.ensemble.RandomForestRegressor`,
          )
          but is more efficient in terms of speed.
          With this method, the target response of a
          classifier is always the decision function, not the predicted
          probabilities.

        - 'brute' is supported for any estimator, but is more
          computationally intensive.

        - 'auto': the 'recursion' is used for estimators that support it,
          and 'brute' is used otherwise.

        Please see :ref:`this note <pdp_method_differences>` for
        differences between the 'brute' and 'recursion' method.

    Returns
    -------
    averaged_predictions : ndarray, \
            shape (n_outputs, len(values[0]), len(values[1]), ...)
        The predictions for all the points in the grid, averaged over all
        samples in X (or over the training data if ``method`` is
        'recursion'). ``n_outputs`` corresponds to the number of classes in
        a multi-class setting, or to the number of tasks for multi-output
        regression. For classical regression and binary classification
        ``n_outputs==1``. ``n_values_feature_j`` corresponds to the size
        ``values[j]``.

    values : seq of 1d ndarrays
        The values with which the grid has been created. The generated grid
        is a cartesian product of the arrays in ``values``. ``len(values) ==
        len(features)``. The size of each array ``values[j]`` is either
        ``grid_resolution``, or the number of unique values in ``X[:, j]``,
        whichever is smaller.

    Examples
    --------
    >>> X = [[0, 0, 2], [1, 0, 0]]
    >>> y = [0, 1]
    >>> from sklearn.ensemble import GradientBoostingClassifier
    >>> gb = GradientBoostingClassifier(random_state=0).fit(X, y)
    >>> partial_dependence(gb, features=[0], X=X, percentiles=(0, 1),
    ...                    grid_resolution=2) # doctest: +SKIP
    (array([[-4.52...,  4.52...]]), [array([ 0.,  1.])])

    See also
    --------
    sklearn.inspection.plot_partial_dependence: Plot partial dependence
    """
    if not (is_classifier(estimator) or is_regressor(estimator)):
        raise ValueError(
            "'estimator' must be a fitted regressor or classifier."
        )

    if isinstance(estimator, Pipeline):
        # TODO: to be removed if/when pipeline get a `steps_` attributes
        # assuming Pipeline is the only estimator that does not store a new
        # attribute
        for est in estimator:
            # FIXME: remove the None option when it will be deprecated
            if est not in (None, 'drop'):
                check_is_fitted(est)
    else:
        check_is_fitted(estimator)

    if (is_classifier(estimator) and
            isinstance(estimator.classes_[0], np.ndarray)):
        raise ValueError(
            'Multiclass-multioutput estimators are not supported'
        )

    # Use check_array only on lists and other non-array-likes / sparse. Do not
    # convert DataFrame into a NumPy array.
    if not(hasattr(X, '__array__') or sparse.issparse(X)):
        X = check_array(X, force_all_finite='allow-nan', dtype=np.object)

    accepted_responses = ('auto', 'predict_proba', 'decision_function')
    if response_method not in accepted_responses:
        raise ValueError(
            'response_method {} is invalid. Accepted response_method names '
            'are {}.'.format(response_method, ', '.join(accepted_responses)))

    if is_regressor(estimator) and response_method != 'auto':
        raise ValueError(
            "The response_method parameter is ignored for regressors and "
            "must be 'auto'."
        )