view iraps_classifier.py @ 8:1c4a241bef5c draft

planemo upload for repository https://github.com/bgruening/galaxytools/tree/master/tools/sklearn commit c0a3a186966888e5787335a7628bf0a4382637e7

"""
class IRAPSCore
class IRAPSClassifier
class BinarizeTargetClassifier
class BinarizeTargetRegressor
class _BinarizeTargetScorer
class _BinarizeTargetProbaScorer

binarize_auc_scorer
binarize_average_precision_scorer

binarize_accuracy_scorer
binarize_balanced_accuracy_scorer
binarize_precision_scorer
binarize_recall_scorer
"""


import numpy as np
import random
import warnings

from abc import ABCMeta
from scipy.stats import ttest_ind
from sklearn import metrics
from sklearn.base import BaseEstimator, clone, RegressorMixin
from sklearn.externals import six
from sklearn.feature_selection.univariate_selection import _BaseFilter
from sklearn.metrics.scorer import _BaseScorer
from sklearn.pipeline import Pipeline
from sklearn.utils import as_float_array, check_X_y
from sklearn.utils._joblib import Parallel, delayed
from sklearn.utils.validation import (check_array, check_is_fitted,
                                      check_memory, column_or_1d)


VERSION = '0.1.1'


class IRAPSCore(six.with_metaclass(ABCMeta, BaseEstimator)):
    """
    Base class of IRAPSClassifier
    From sklearn BaseEstimator:
        get_params()
        set_params()

    Parameters
    ----------
    n_iter : int
        sample count

    positive_thres : float
        z_score shreshold to discretize positive target values

    negative_thres : float
        z_score threshold to discretize negative target values

    verbose : int
        0 or geater, if not 0, print progress

    n_jobs : int, default=1
        The number of CPUs to use to do the computation.

    pre_dispatch : int, or string.
        Controls the number of jobs that get dispatched during parallel
        execution. Reducing this number can be useful to avoid an
        explosion of memory consumption when more jobs get dispatched
        than CPUs can process. This parameter can be:
            - None, in which case all the jobs are immediately
              created and spawned. Use this for lightweight and
              fast-running jobs, to avoid delays due to on-demand
              spawning of the jobs
            - An int, giving the exact number of total jobs that are
              spawned
            - A string, giving an expression as a function of n_jobs,
              as in '2*n_jobs'

    random_state : int or None
    """

    def __init__(self, n_iter=1000, positive_thres=-1, negative_thres=0,
                 verbose=0, n_jobs=1, pre_dispatch='2*n_jobs',
                 random_state=None):
        """
        IRAPS turns towwards general Anomaly Detection
        It comapares positive_thres with negative_thres,
        and decide which portion is the positive target.
        e.g.:
        (positive_thres=-1, negative_thres=0)
                 => positive = Z_score of target < -1
        (positive_thres=1, negative_thres=0)
                 => positive = Z_score of target > 1

        Note: The positive targets here is always the
            abnormal minority group.
        """
        self.n_iter = n_iter
        self.positive_thres = positive_thres
        self.negative_thres = negative_thres
        self.verbose = verbose
        self.n_jobs = n_jobs
        self.pre_dispatch = pre_dispatch
        self.random_state = random_state

    def fit(self, X, y):
        """
        X: array-like (n_samples x n_features)
        y: 1-d array-like (n_samples)
        """
        X, y = check_X_y(X, y, ['csr', 'csc'], multi_output=False)

        def _stochastic_sampling(X, y, random_state=None, positive_thres=-1,
                                 negative_thres=0):
            # each iteration select a random number of random subset of
            # training samples. this is somewhat different from the original
            # IRAPS method, but effect is almost the same.
            SAMPLE_SIZE = [0.25, 0.75]
            n_samples = X.shape[0]

            if random_state is None:
                n_select = random.randint(int(n_samples * SAMPLE_SIZE[0]),
                                          int(n_samples * SAMPLE_SIZE[1]))
                index = random.sample(list(range(n_samples)), n_select)
            else:
                n_select = random.Random(random_state).randint(
                                    int(n_samples * SAMPLE_SIZE[0]),
                                    int(n_samples * SAMPLE_SIZE[1]))
                index = random.Random(random_state).sample(
                                    list(range(n_samples)), n_select)

            X_selected, y_selected = X[index], y[index]

            # Spliting by z_scores.
            y_selected = (y_selected - y_selected.mean()) / y_selected.std()
            if positive_thres < negative_thres:
                X_selected_positive = X_selected[y_selected < positive_thres]
                X_selected_negative = X_selected[y_selected > negative_thres]
            else:
                X_selected_positive = X_selected[y_selected > positive_thres]
                X_selected_negative = X_selected[y_selected < negative_thres]

            # For every iteration, at least 5 responders are selected
            if X_selected_positive.shape[0] < 5:
                warnings.warn("Warning: fewer than 5 positives were selected!")
                return

            # p_values
            _, p = ttest_ind(X_selected_positive, X_selected_negative,
                             axis=0, equal_var=False)

            # fold_change == mean change?
            # TODO implement other normalization method
            positive_mean = X_selected_positive.mean(axis=0)
            negative_mean = X_selected_negative.mean(axis=0)
            mean_change = positive_mean - negative_mean
            # mean_change = np.select(
            #       [positive_mean > negative_mean,
            #           positive_mean < negative_mean],
            #       [positive_mean / negative_mean,
            #           -negative_mean / positive_mean])
            # mean_change could be adjusted by power of 2
            # mean_change = 2**mean_change \
            #       if mean_change>0 else -2**abs(mean_change)

            return p, mean_change, negative_mean

        parallel = Parallel(n_jobs=self.n_jobs, verbose=self.verbose,
                            pre_dispatch=self.pre_dispatch)
        if self.random_state is None:
            res = parallel(delayed(_stochastic_sampling)(
                    X, y, random_state=None,
                    positive_thres=self.positive_thres,
                    negative_thres=self.negative_thres)
                        for i in range(self.n_iter))
        else:
            res = parallel(delayed(_stochastic_sampling)(
                    X, y, random_state=seed,
                    positive_thres=self.positive_thres,
                    negative_thres=self.negative_thres)
                        for seed in range(self.random_state,
                                          self.random_state+self.n_iter))
        res = [_ for _ in res if _]
        if len(res) < 50:
            raise ValueError("too few (%d) valid feature lists "
                             "were generated!" % len(res))
        pvalues = np.vstack([x[0] for x in res])
        fold_changes = np.vstack([x[1] for x in res])
        base_values = np.vstack([x[2] for x in res])

        self.pvalues_ = np.asarray(pvalues)
        self.fold_changes_ = np.asarray(fold_changes)
        self.base_values_ = np.asarray(base_values)

        return self


def _iraps_core_fit(iraps_core, X, y):
    return iraps_core.fit(X, y)


class IRAPSClassifier(six.with_metaclass(ABCMeta, _BaseFilter,
                                         BaseEstimator, RegressorMixin)):
    """
    Extend the bases of both sklearn feature_selector and classifier.
    From sklearn BaseEstimator:
        get_params()
        set_params()
    From sklearn _BaseFilter:
        get_support()
        fit_transform(X)
        transform(X)
    From sklearn RegressorMixin:
        score(X, y): R2
    New:
        predict(X)
        predict_label(X)
        get_signature()
    Properties:
        discretize_value

    Parameters
    ----------
    iraps_core: object
    p_thres: float, threshold for p_values
    fc_thres: float, threshold for fold change or mean difference
    occurrence: float, occurrence rate selected by set of p_thres and fc_thres
    discretize: float, threshold of z_score to discretize target value
    memory: None, str or joblib.Memory object
    min_signature_features: int, the mininum number of features in a signature
    """

    def __init__(self, iraps_core, p_thres=1e-4, fc_thres=0.1,
                 occurrence=0.8, discretize=-1, memory=None,
                 min_signature_features=1):
        self.iraps_core = iraps_core
        self.p_thres = p_thres
        self.fc_thres = fc_thres
        self.occurrence = occurrence
        self.discretize = discretize
        self.memory = memory
        self.min_signature_features = min_signature_features

    def fit(self, X, y):
        memory = check_memory(self.memory)
        cached_fit = memory.cache(_iraps_core_fit)
        iraps_core = clone(self.iraps_core)
        # allow pre-fitted iraps_core here
        if not hasattr(iraps_core, 'pvalues_'):
            iraps_core = cached_fit(iraps_core, X, y)
        self.iraps_core_ = iraps_core

        pvalues = as_float_array(iraps_core.pvalues_, copy=True)
        # why np.nan is here?
        pvalues[np.isnan(pvalues)] = np.finfo(pvalues.dtype).max

        fold_changes = as_float_array(iraps_core.fold_changes_, copy=True)
        fold_changes[np.isnan(fold_changes)] = 0.0

        base_values = as_float_array(iraps_core.base_values_, copy=True)

        p_thres = self.p_thres
        fc_thres = self.fc_thres
        occurrence = self.occurrence

        mask_0 = np.zeros(pvalues.shape, dtype=np.int32)
        # mark p_values less than the threashold
        mask_0[pvalues <= p_thres] = 1
        # mark fold_changes only when greater than the threashold
        mask_0[abs(fold_changes) < fc_thres] = 0

        # count the occurrence and mask greater than the threshold
        counts = mask_0.sum(axis=0)
        occurrence_thres = int(occurrence * iraps_core.n_iter)
        mask = np.zeros(counts.shape, dtype=bool)
        mask[counts >= occurrence_thres] = 1

        # generate signature
        fold_changes[mask_0 == 0] = 0.0
        signature = fold_changes[:, mask].sum(axis=0) / counts[mask]
        signature = np.vstack((signature, base_values[:, mask].mean(axis=0)))
        # It's not clearn whether min_size could impact prediction
        # performance
        if signature is None\
                or signature.shape[1] < self.min_signature_features:
            raise ValueError("The classifier got None signature or the number "
                             "of sinature feature is less than minimum!")

        self.signature_ = np.asarray(signature)
        self.mask_ = mask
        # TODO: support other discretize method: fixed value, upper
        # third quater, etc.
        self.discretize_value = y.mean() + y.std() * self.discretize
        if iraps_core.negative_thres > iraps_core.positive_thres:
            self.less_is_positive = True
        else:
            self.less_is_positive = False

        return self

    def _get_support_mask(self):
        """
        return mask of feature selection indices
        """
        check_is_fitted(self, 'mask_')

        return self.mask_

    def get_signature(self):
        """
        return signature
        """
        check_is_fitted(self, 'signature_')

        return self.signature_

    def predict(self, X):
        """
        compute the correlation coefficient with irpas signature
        """
        signature = self.get_signature()

        X = as_float_array(X)
        X_transformed = self.transform(X) - signature[1]
        corrcoef = np.array(
            [np.corrcoef(signature[0], e)[0][1] for e in X_transformed])
        corrcoef[np.isnan(corrcoef)] = np.finfo(np.float32).min

        return corrcoef

    def predict_label(self, X, clf_cutoff=0.4):
        return self.predict(X) >= clf_cutoff


class BinarizeTargetClassifier(BaseEstimator, RegressorMixin):
    """
    Convert continuous target to binary labels (True and False)
    and apply a classification estimator.

    Parameters
    ----------
    classifier: object
        Estimator object such as derived from sklearn `ClassifierMixin`.

    z_score: float, default=-1.0
        Threshold value based on z_score. Will be ignored when
        fixed_value is set

    value: float, default=None
        Threshold value

    less_is_positive: boolean, default=True
        When target is less the threshold value, it will be converted
        to True, False otherwise.

    Attributes
    ----------
    classifier_: object
        Fitted classifier

    discretize_value: float
        The threshold value used to discretize True and False targets
    """

    def __init__(self, classifier, z_score=-1, value=None,
                 less_is_positive=True):
        self.classifier = classifier
        self.z_score = z_score
        self.value = value
        self.less_is_positive = less_is_positive

    def fit(self, X, y, sample_weight=None):
        """
        Convert y to True and False labels and then fit the classifier
        with X and new y

        Returns
        ------
        self: object
        """
        y = check_array(y, accept_sparse=False, force_all_finite=True,
                        ensure_2d=False, dtype='numeric')
        y = column_or_1d(y)

        if self.value is None:
            discretize_value = y.mean() + y.std() * self.z_score
        else:
            discretize_value = self.Value
        self.discretize_value = discretize_value

        if self.less_is_positive:
            y_trans = y < discretize_value
        else:
            y_trans = y > discretize_value

        self.classifier_ = clone(self.classifier)

        if sample_weight is not None:
            self.classifier_.fit(X, y_trans, sample_weight=sample_weight)
        else:
            self.classifier_.fit(X, y_trans)

        if hasattr(self.classifier_, 'feature_importances_'):
            self.feature_importances_ = self.classifier_.feature_importances_
        if hasattr(self.classifier_, 'coef_'):
            self.coef_ = self.classifier_.coef_
        if hasattr(self.classifier_, 'n_outputs_'):
            self.n_outputs_ = self.classifier_.n_outputs_
        if hasattr(self.classifier_, 'n_features_'):
            self.n_features_ = self.classifier_.n_features_

        return self

    def predict(self, X):
        """
        Predict class probabilities of X.
        """
        check_is_fitted(self, 'classifier_')
        proba = self.classifier_.predict_proba(X)
        return proba[:, 1]

    def predict_label(self, X):
        """Predict class label of X
        """
        check_is_fitted(self, 'classifier_')
        return self.classifier_.predict(X)


class _BinarizeTargetProbaScorer(_BaseScorer):
    """
    base class to make binarized target specific scorer
    """

    def __call__(self, clf, X, y, sample_weight=None):
        clf_name = clf.__class__.__name__
        # support pipeline object
        if isinstance(clf, Pipeline):
            main_estimator = clf.steps[-1][-1]
        # support stacking ensemble estimators
        # TODO support nested pipeline/stacking estimators
        elif clf_name in ['StackingCVClassifier', 'StackingClassifier']:
            main_estimator = clf.meta_clf_
        elif clf_name in ['StackingCVRegressor', 'StackingRegressor']:
            main_estimator = clf.meta_regr_
        else:
            main_estimator = clf

        discretize_value = main_estimator.discretize_value
        less_is_positive = main_estimator.less_is_positive

        if less_is_positive:
            y_trans = y < discretize_value
        else:
            y_trans = y > discretize_value

        y_pred = clf.predict(X)
        if sample_weight is not None:
            return self._sign * self._score_func(y_trans, y_pred,
                                                 sample_weight=sample_weight,
                                                 **self._kwargs)
        else:
            return self._sign * self._score_func(y_trans, y_pred,
                                                 **self._kwargs)


# roc_auc
binarize_auc_scorer =\
        _BinarizeTargetProbaScorer(metrics.roc_auc_score, 1, {})

# average_precision_scorer
binarize_average_precision_scorer =\
        _BinarizeTargetProbaScorer(metrics.average_precision_score, 1, {})

# roc_auc_scorer
iraps_auc_scorer = binarize_auc_scorer

# average_precision_scorer
iraps_average_precision_scorer = binarize_average_precision_scorer


class BinarizeTargetRegressor(BaseEstimator, RegressorMixin):
    """
    Extend regression estimator to have discretize_value

    Parameters
    ----------
    regressor: object
        Estimator object such as derived from sklearn `RegressionMixin`.

    z_score: float, default=-1.0
        Threshold value based on z_score. Will be ignored when
        fixed_value is set

    value: float, default=None
        Threshold value

    less_is_positive: boolean, default=True
        When target is less the threshold value, it will be converted
        to True, False otherwise.

    Attributes
    ----------
    regressor_: object
        Fitted regressor

    discretize_value: float
        The threshold value used to discretize True and False targets
    """

    def __init__(self, regressor, z_score=-1, value=None,
                 less_is_positive=True):
        self.regressor = regressor
        self.z_score = z_score
        self.value = value
        self.less_is_positive = less_is_positive

    def fit(self, X, y, sample_weight=None):
        """
        Calculate the discretize_value fit the regressor with traning data

        Returns
        ------
        self: object
        """
        y = check_array(y, accept_sparse=False, force_all_finite=True,
                        ensure_2d=False, dtype='numeric')
        y = column_or_1d(y)

        if self.value is None:
            discretize_value = y.mean() + y.std() * self.z_score
        else:
            discretize_value = self.Value
        self.discretize_value = discretize_value

        self.regressor_ = clone(self.regressor)

        if sample_weight is not None:
            self.regressor_.fit(X, y, sample_weight=sample_weight)
        else:
            self.regressor_.fit(X, y)

        # attach classifier attributes
        if hasattr(self.regressor_, 'feature_importances_'):
            self.feature_importances_ = self.regressor_.feature_importances_
        if hasattr(self.regressor_, 'coef_'):
            self.coef_ = self.regressor_.coef_
        if hasattr(self.regressor_, 'n_outputs_'):
            self.n_outputs_ = self.regressor_.n_outputs_
        if hasattr(self.regressor_, 'n_features_'):
            self.n_features_ = self.regressor_.n_features_

        return self

    def predict(self, X):
        """Predict target value of X
        """
        check_is_fitted(self, 'regressor_')
        y_pred = self.regressor_.predict(X)
        if not np.all((y_pred >= 0) & (y_pred <= 1)):
            y_pred = (y_pred - y_pred.min()) / (y_pred.max() - y_pred.min())
        if self.less_is_positive:
            y_pred = 1 - y_pred
        return y_pred


# roc_auc_scorer
regression_auc_scorer = binarize_auc_scorer

# average_precision_scorer
regression_average_precision_scorer = binarize_average_precision_scorer