# -*- coding: utf8
"""Random Projection transformers

Random Projections are a simple and computationally efficient way to
reduce the dimensionality of the data by trading a controlled amount
of accuracy (as additional variance) for faster processing times and
smaller model sizes.

The dimensions and distribution of Random Projections matrices are
controlled so as to preserve the pairwise distances between any two
samples of the dataset.

The main theoretical result behind the efficiency of random projection is the
`Johnson-Lindenstrauss lemma (quoting Wikipedia)
<http://en.wikipedia.org/wiki/Johnson%E2%80%93Lindenstrauss_lemma>`_:

  In mathematics, the Johnson-Lindenstrauss lemma is a result
  concerning low-distortion embeddings of points from high-dimensional
  into low-dimensional Euclidean space. The lemma states that a small set
  of points in a high-dimensional space can be embedded into a space of
  much lower dimension in such a way that distances between the points are
  nearly preserved. The map used for the embedding is at least Lipschitz,
  and can even be taken to be an orthogonal projection.

"""
# Authors: Olivier Grisel <olivier.grisel@ensta.org>,
#          Arnaud Joly <a.joly@ulg.ac.be>
# License: BSD 3 clause

from __future__ import division
import warnings
from abc import ABCMeta, abstractmethod

import numpy as np
from numpy.testing import assert_equal
import scipy.sparse as sp

from .base import BaseEstimator, TransformerMixin
from .externals import six
from .externals.six.moves import xrange
from .utils import check_random_state
from .utils.extmath import safe_sparse_dot
from .utils.random import sample_without_replacement
from .utils.validation import check_array, NotFittedError
from .utils import DataDimensionalityWarning


__all__ = ["SparseRandomProjection",
           "GaussianRandomProjection",
           "johnson_lindenstrauss_min_dim"]


def johnson_lindenstrauss_min_dim(n_samples, eps=0.1):
    """Find a 'safe' number of components to randomly project to

    The distortion introduced by a random projection `p` only changes the
    distance between two points by a factor (1 +- eps) in an euclidean space
    with good probability. The projection `p` is an eps-embedding as defined
    by:

      (1 - eps) ||u - v||^2 < ||p(u) - p(v)||^2 < (1 + eps) ||u - v||^2

    Where u and v are any rows taken from a dataset of shape [n_samples,
    n_features], eps is in ]0, 1[ and p is a projection by a random Gaussian
    N(0, 1) matrix with shape [n_components, n_features] (or a sparse
    Achlioptas matrix).

    The minimum number of components to guarantee the eps-embedding is
    given by:

      n_components >= 4 log(n_samples) / (eps^2 / 2 - eps^3 / 3)

    Note that the number of dimensions is independent of the original
    number of features but instead depends on the size of the dataset:
    the larger the dataset, the higher is the minimal dimensionality of
    an eps-embedding.

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

    Parameters
    ----------
    n_samples : int or numpy array of int greater than 0,
        Number of samples. If an array is given, it will compute
        a safe number of components array-wise.

    eps : float or numpy array of float in ]0,1[, optional (default=0.1)
        Maximum distortion rate as defined by the Johnson-Lindenstrauss lemma.
        If an array is given, it will compute a safe number of components
        array-wise.

    Returns
    -------
    n_components : int or numpy array of int,
        The minimal number of components to guarantee with good probability
        an eps-embedding with n_samples.

    Examples
    --------

    >>> johnson_lindenstrauss_min_dim(1e6, eps=0.5)
    663

    >>> johnson_lindenstrauss_min_dim(1e6, eps=[0.5, 0.1, 0.01])
    array([    663,   11841, 1112658])

    >>> johnson_lindenstrauss_min_dim([1e4, 1e5, 1e6], eps=0.1)
    array([ 7894,  9868, 11841])

    References
    ----------

    .. [1] http://en.wikipedia.org/wiki/Johnson%E2%80%93Lindenstrauss_lemma

    .. [2] Sanjoy Dasgupta and Anupam Gupta, 1999,
           "An elementary proof of the Johnson-Lindenstrauss Lemma."
           http://citeseer.ist.psu.edu/viewdoc/summary?doi=10.1.1.45.3654

    """
    eps = np.asarray(eps)
    n_samples = np.asarray(n_samples)

    if np.any(eps <= 0.0) or np.any(eps >= 1):
        raise ValueError(
            "The JL bound is defined for eps in ]0, 1[, got %r" % eps)

    if np.any(n_samples) <= 0:
        raise ValueError(
            "The JL bound is defined for n_samples greater than zero, got %r"
            % n_samples)

    denominator = (eps ** 2 / 2) - (eps ** 3 / 3)
    return (4 * np.log(n_samples) / denominator).astype(np.int)


def _check_density(density, n_features):
    """Factorize density check according to Li et al."""
    if density == 'auto':
        density = 1 / np.sqrt(n_features)

    elif density <= 0 or density > 1:
        raise ValueError("Expected density in range ]0, 1], got: %r"
                         % density)
    return density


def _check_input_size(n_components, n_features):
    """Factorize argument checking for random matrix generation"""
    if n_components <= 0:
        raise ValueError("n_components must be strictly positive, got %d" %
                         n_components)
    if n_features <= 0:
        raise ValueError("n_features must be strictly positive, got %d" %
                         n_components)


def gaussian_random_matrix(n_components, n_features, random_state=None):
    """ Generate a dense Gaussian random matrix.

    The components of the random matrix are drawn from

        N(0, 1.0 / n_components).

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

    Parameters
    ----------
    n_components : int,
        Dimensionality of the target projection space.

    n_features : int,
        Dimensionality of the original source space.

    random_state : int, RandomState instance or None (default=None)
        Control the pseudo random number generator used to generate the
        matrix at fit time.

    Returns
    -------
    components : numpy array of shape [n_components, n_features]
        The generated Gaussian random matrix.

    See Also
    --------
    GaussianRandomProjection
    sparse_random_matrix
    """
    _check_input_size(n_components, n_features)
    rng = check_random_state(random_state)
    components = rng.normal(loc=0.0,
                            scale=1.0 / np.sqrt(n_components),
                            size=(n_components, n_features))
    return components


def sparse_random_matrix(n_components, n_features, density='auto',
                         random_state=None):
    """Generalized Achlioptas random sparse matrix for random projection

    Setting density to 1 / 3 will yield the original matrix by Dimitris
    Achlioptas while setting a lower value will yield the generalization
    by Ping Li et al.

    If we note :math:`s = 1 / density`, the components of the random matrix are
    drawn from:

      - -sqrt(s) / sqrt(n_components)   with probability 1 / 2s
      -  0                              with probability 1 - 1 / s
      - +sqrt(s) / sqrt(n_components)   with probability 1 / 2s

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

    Parameters
    ----------
    n_components : int,
        Dimensionality of the target projection space.

    n_features : int,
        Dimensionality of the original source space.

    density : float in range ]0, 1] or 'auto', optional (default='auto')
        Ratio of non-zero component in the random projection matrix.

        If density = 'auto', the value is set to the minimum density
        as recommended by Ping Li et al.: 1 / sqrt(n_features).

        Use density = 1 / 3.0 if you want to reproduce the results from
        Achlioptas, 2001.

    random_state : integer, RandomState instance or None (default=None)
        Control the pseudo random number generator used to generate the
        matrix at fit time.

    Returns
    -------
    components: numpy array or CSR matrix with shape [n_components, n_features]
        The generated Gaussian random matrix.

    See Also
    --------
    SparseRandomProjection
    gaussian_random_matrix

    References
    ----------

    .. [1] Ping Li, T. Hastie and K. W. Church, 2006,
           "Very Sparse Random Projections".
           http://www.stanford.edu/~hastie/Papers/Ping/KDD06_rp.pdf

    .. [2] D. Achlioptas, 2001, "Database-friendly random projections",
           http://www.cs.ucsc.edu/~optas/papers/jl.pdf

    """
    _check_input_size(n_components, n_features)
    density = _check_density(density, n_features)
    rng = check_random_state(random_state)

    if density == 1:
        # skip index generation if totally dense
        components = rng.binomial(1, 0.5, (n_components, n_features)) * 2 - 1
        return 1 / np.sqrt(n_components) * components

    else:
        # Generate location of non zero elements
        indices = []
        offset = 0
        indptr = [offset]
        for i in xrange(n_components):
            # find the indices of the non-zero components for row i
            n_nonzero_i = rng.binomial(n_features, density)
            indices_i = sample_without_replacement(n_features, n_nonzero_i,
                                                   random_state=rng)
            indices.append(indices_i)
            offset += n_nonzero_i
            indptr.append(offset)

        indices = np.concatenate(indices)

        # Among non zero components the probability of the sign is 50%/50%
        data = rng.binomial(1, 0.5, size=np.size(indices)) * 2 - 1

        # build the CSR structure by concatenating the rows
        components = sp.csr_matrix((data, indices, indptr),
                                   shape=(n_components, n_features))

        return np.sqrt(1 / density) / np.sqrt(n_components) * components


class BaseRandomProjection(six.with_metaclass(ABCMeta, BaseEstimator,
                                              TransformerMixin)):
    """Base class for random projections.

    Warning: This class should not be used directly.
    Use derived classes instead.
    """

    @abstractmethod
    def __init__(self, n_components='auto', eps=0.1, dense_output=False,
                 random_state=None):
        self.n_components = n_components
        self.eps = eps
        self.dense_output = dense_output
        self.random_state = random_state

        self.components_ = None
        self.n_components_ = None

    @abstractmethod
    def _make_random_matrix(n_components, n_features):
        """ Generate the random projection matrix

        Parameters
        ----------
        n_components : int,
            Dimensionality of the target projection space.

        n_features : int,
            Dimensionality of the original source space.

        Returns
        -------
        components : numpy array or CSR matrix [n_components, n_features]
            The generated random matrix.

        """

    def fit(self, X, y=None):
        """Generate a sparse random projection matrix

        Parameters
        ----------
        X : numpy array or scipy.sparse of shape [n_samples, n_features]
            Training set: only the shape is used to find optimal random
            matrix dimensions based on the theory referenced in the
            afore mentioned papers.

        y : is not used: placeholder to allow for usage in a Pipeline.

        Returns
        -------
        self

        """
        X = check_array(X, accept_sparse=['csr', 'csc'])