learners.py

import numpy as np
import scipy
import time
from data.embeddings import WordEmbeddings
from transformers.clesa import CLESA
from transformers.riboc import RandomIndexingBoC
from transformers.dci import DistributionalCorrespondenceIndexing
from scipy.sparse import issparse, csr_matrix
from sklearn.multiclass import OneVsRestClassifier
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import KFold
from sklearn.externals.joblib import Parallel, delayed
from sklearn.feature_extraction.text import TfidfVectorizer, CountVectorizer


class FunnellingPolylingualClassifier:
    """
    This classifier projects each document d into a language-independent feature space where each dimension fi is the
    decision score phi_l(d,ci) of an auxiliary classifier phi_l trained on category ci for documents in language l;
    then trains one single classifier for all documents in this space, irrespective of their originary language
    """
    def __init__(self, first_tier_learner, meta_learner, first_tier_parameters=None, meta_parameters=None, folded_projections=1,
                 calmode='cal', n_jobs=-1):
        """
        :param first_tier_learner: the learner used in the first-tier level
        :param meta_learner: the learner used in the second-tier level
        :param first_tier_parameters: parameters for the learner in the doc_projector
        :param meta_parameters: parameters for the learner in the z-space
        :param folded_projections: if 1 then the model trains the auxiliar classifiers with all training data and
        :param calmode: 'cal' to calibrate the base classifiers, 'nocal' to use the decision_function instead, or
        :param n_jobs: number of parallel threads
        'sigmoid' to use the sigmoid of the decision_function
        projects the data before training the final classifier; if greater than one, the training set is split in as
        many folds as indicated, and the projected space is composed by concatenating each fold prediction based on
        models trained on the remaining folds. This should increase the generality of the space to unseen data.
        """
        assert folded_projections>0, "positive number of folds expected"
        assert calmode in ['cal','nocal','sigmoid'], 'unknown calmode'
        assert calmode!='cal' or first_tier_learner.probability, 'calmode=cal requires the learner to have probability=True'

        self.fist_tier_learner = first_tier_learner
        self.meta_learner = meta_learner
        self.fist_tier_parameters=first_tier_parameters
        self.meta_parameters = meta_parameters
        self.doc_projector = NaivePolylingualClassifier(self.fist_tier_learner, self.fist_tier_parameters, n_jobs=n_jobs)
        self.doc_projector_bu = NaivePolylingualClassifier(self.fist_tier_learner, self.fist_tier_parameters, n_jobs=n_jobs)
        self.folded_projections = folded_projections
        self.n_jobs = n_jobs
        self.calmode = calmode

    def _projection(self, doc_projector, lX):
        """
        Decides the projection function to be applied; predict_proba if the base classifiers are calibrated or
        decision_function if otherwise
        :param doc_projector: the document projector (a NaivePolylingualClassifier)
        :param lX: {lang:matrix} to train
        :return: the projection, applied with predict_proba or decision_function
        """
        if self.calmode=='cal':
            return doc_projector.predict_proba(lX)
        else:
            l_decision_scores = doc_projector.decision_function(lX)
            if self.calmode=='sigmoid':
                def sigmoid(x): return 1 / (1 + np.exp(-x))
                for lang in l_decision_scores.keys():
                    l_decision_scores[lang] = sigmoid(l_decision_scores[lang])
            return l_decision_scores

    def _get_zspace(self, lXtr, lYtr, lXproj=None, lYproj=None):
        """
        Produces the vector space of posterior probabilities (in case the first-tier is calibrated) or of
        decision scores (if otherwise). This space is here named zspace.
        :param lXtr: {lang:matrix} to train
        :param lYtr: {lang:labels} to train
        :param lXproj: {lang:matrix} to project (if None, then projects the lXtr)
        :param lYproj: {lang:labels} to stack in the same order (if None, then lYtr will be stacked)
        :return: the projection of lXproj documents into the Z-space defined by the confidence scores of language-specific
        models trained on lXtr, and the lYproj labels stacked consistently
        """
        repair_empty_folds = True
        if lXproj is None and lYproj is None:
            lXproj, lYproj = lXtr, lYtr
            repair_empty_folds = False

        print('fitting the projectors... {}'.format(lXtr.keys()))
        self.doc_projector.fit(lXtr, lYtr)

        print('projecting the documents')
        langs = list(lXtr.keys())
        lZ = self._projection(self.doc_projector, lXproj)

        if repair_empty_folds: #empty folds are replaced by the posterior probabilities generated by the non-folded version
            empty_categories = self.doc_projector.empty_categories
            lZ_bu = self._projection(self.doc_projector_bu, lXproj)

            for lang in langs:
                repair = empty_categories[lang]
                lZ[lang][:,repair] = lZ_bu[lang][:,repair]

        Z = np.vstack([lZ[lang] for lang in langs])  # Z is the language independent space
        zy = np.vstack([lYproj[lang] for lang in langs])
        return Z, zy

    def _get_zspace_folds(self, lX, ly):
        self.doc_projector_bu.fit(lX, ly)

        print('split of {} folds'.format(self.folded_projections))
        skf = KFold(n_splits=self.folded_projections, shuffle=True)

        Z, zy = [], []
        lfold = {lang: list(skf.split(lX[lang], ly[lang])) for lang in lX.keys()}
        for fold in range(self.folded_projections):
            print('fitting the projectors ({}/{})...'.format(fold + 1, self.folded_projections))
            lfoldXtr, lfoldYtr = {}, {}
            lfoldXte, lfoldYte = {}, {}
            for lang in lX.keys():
                train, test = lfold[lang][fold]
                lfoldXtr[lang] = lX[lang][train]
                lfoldYtr[lang] = ly[lang][train]
                lfoldXte[lang] = lX[lang][test]
                lfoldYte[lang] = ly[lang][test]
            Zfold, zYfold = self._get_zspace(lfoldXtr, lfoldYtr, lfoldXte, lfoldYte)
            Z.append(Zfold)
            zy.append(zYfold)
        # compose the Z-space as the union of all folded predictions
        Z = np.vstack(Z)
        zy = np.vstack(zy)
        # refit the document projector with all examples to have a more reliable projector for test data
        self.doc_projector = self.doc_projector_bu
        return Z, zy

    def fit(self, lX, ly, lZ=None, lzy=None):
        tinit = time.time()
        Z, zy = self._get_zspace(lX, ly) if self.folded_projections == 1 else self._get_zspace_folds(lX, ly)

        #experimental: adds the posterior probabilities (computed outside) to the meta-classifier
        if lZ is not None and lzy is not None:
            zlangs = list(lZ.keys())
            Z = np.vstack((Z, *[lZ[l] for l in zlangs]))
            zy = np.vstack((zy, *[lzy[l] for l in zlangs]))

        print('fitting the Z-space of shape={}'.format(Z.shape))
        self.model = MonolingualClassifier(base_learner=self.meta_learner, parameters=self.meta_parameters, n_jobs=self.n_jobs)
        self.model.fit(Z, zy)
        self.time = time.time() - tinit

        return self

    def predict(self, lX, lZ=None):
        """
        :param lX: a dictionary {language_label: X csr-matrix}
        :param lZ: a dictionary {language_label: Z matrix}; if specified, concats this representation
        :return: a dictionary of predictions
        """
        lZ_ = self._projection(self.doc_projector, lX)
        if lZ is not None:
            lZ_ = {**lZ_, **lZ}
        return _joblib_transform_multiling(self.model.predict, lZ_, n_jobs=self.n_jobs)

    def best_params(self):
        params = self.doc_projector.best_params()
        params['meta'] = self.model.best_params()
        return params


class NaivePolylingualClassifier:
    """
    Is a mere set of independet MonolingualClassifiers
    """
    def __init__(self, base_learner, parameters=None, n_jobs=-1):
        self.base_learner = base_learner
        self.parameters = parameters
        self.model = None
        self.n_jobs = n_jobs

    def fit(self, lX, ly):
        """
        trains the independent monolingual classifiers
        :param lX: a dictionary {language_label: X csr-matrix}
        :param ly: a dictionary {language_label: y np.array}
        :return: self
        """
        tinit = time.time()
        assert set(lX.keys()) == set(ly.keys()), 'inconsistent language mappings in fit'
        langs = list(lX.keys())
        for lang in langs: _sort_if_sparse(lX[lang])
        models = Parallel(n_jobs=self.n_jobs)\
            (delayed(MonolingualClassifier(self.base_learner, parameters=self.parameters).fit)
             (lX[lang],ly[lang]) for lang in langs)
        self.model = {lang: models[i] for i, lang in enumerate(langs)}
        self.empty_categories = {lang:self.model[lang].empty_categories for lang in langs}
        self.time = time.time() - tinit
        return self

    def decision_function(self, lX):
        """
        :param lX: a dictionary {language_label: X csr-matrix}
        :return: a dictionary of classification scores for each class
        """
        assert self.model is not None, 'predict called before fit'
        assert set(lX.keys()).issubset(set(self.model.keys())), 'unknown languages requested in decision function'
        langs=list(lX.keys())
        scores = Parallel(n_jobs=self.n_jobs)(delayed(self.model[lang].decision_function)(lX[lang]) for lang in langs)
        return {lang:scores[i] for i,lang in enumerate(langs)}

    def predict_proba(self, lX):
        """
        :param lX: a dictionary {language_label: X csr-matrix}
        :return: a dictionary of probabilities that each document belongs to each class
        """
        assert self.model is not None, 'predict called before fit'
        assert set(lX.keys()).issubset(set(self.model.keys())), 'unknown languages requested in decision function'
        langs=list(lX.keys())
        scores = Parallel(n_jobs=self.n_jobs, max_nbytes=None)(delayed(self.model[lang].predict_proba)(lX[lang]) for lang in langs)
        return {lang:scores[i] for i,lang in enumerate(langs)}

    def predict(self, lX):
        """
        :param lX: a dictionary {language_label: X csr-matrix}
        :return: a dictionary of predictions
        """
        assert self.model is not None, 'predict called before fit'
        assert set(lX.keys()).issubset(set(self.model.keys())), 'unknown languages requested in predict'
        if self.n_jobs == 1:
            return {lang:self.model[lang].predict(lX[lang]) for lang in lX.keys()}
        else:
            langs = list(lX.keys())
            scores = Parallel(n_jobs=self.n_jobs)(delayed(self.model[lang].predict)(lX[lang]) for lang in langs)
            return {lang: scores[i] for i, lang in enumerate(langs)}

    def best_params(self):
        return {l:model.best_params() for l,model in self.model.items()}


class CLESAPolylingualClassifier:
    """
    A polylingual classifier based on the cross-lingual ESA method
    """
    def __init__(self, base_learner, lW, z_parameters=None, similarity='dot', post=False, n_jobs=-1):
        """
        :param base_learner: the learner operating in the clesa space
        :param lW: a dictionary {lang : wikipedia doc-by-term csr_matrix}
        :param z_parameters: the parameters of the learner to optimize for via 5-fold cv in the z-space
        :param similarity: the similarity measure between documents to be used ('dot' or 'cosine')
        :param post: any valid sklearn normalization method to be applied to the resulting doc embeddings, or None (default)
        :param n_jobs: number of parallel threads
        """
        self.base_learner = base_learner
        self.z_parameters=z_parameters
        self.doc_projector = CLESA(similarity=similarity, post=post).fit(lW)
        self.n_jobs = n_jobs
        self.time = 0

    def fit(self, lX, ly):
        """
        :param lX: a dictionary {language_label: X csr-matrix}
        :param ly: a dictionary {language_label: y np.array}
        """
        assert set(lX.keys()) == set(ly.keys()), 'inconsistent dictionaries in fit'

        print('projecting the documents')
        lZ = self.transform(lX)
        return self.fit_from_transformed(lZ, ly)

    def fit_from_transformed(self, lZ, ly):
        tinit = time.time()
        langs = list(lZ.keys())
        Z = np.vstack([lZ[lang] for lang in langs])  # Z is the language independent space
        zy = np.vstack([ly[lang] for lang in langs])

        print('fitting the Z-space of shape={}'.format(Z.shape))
        self.model = MonolingualClassifier(base_learner=self.base_learner, parameters=self.z_parameters, n_jobs=self.n_jobs)
        self.model.fit(Z, zy)
        self.time = self.time + (time.time() - tinit)
        return self

    def transform(self, lX, accum_time=True):
        tinit = time.time()
        lZ = self.doc_projector.transform(lX)
        if accum_time:
            self.time = self.time + (time.time() - tinit)
        return lZ

    def predict_from_transformed(self, lZ):
        assert self.model is not None, 'predict called before fit'
        return _joblib_transform_multiling(self.model.predict, lZ, n_jobs=self.n_jobs)

    def predict(self, lX):
        """
        :param lX: a dictionary {language_label: X csr-matrix}
        :return: a dictionary of predictions
        """
        assert self.doc_projector is not None, 'clesa transformed disabled'
        lZ = self.transform(lX)
        return self.predict_from_transformed(lZ)

    def clear_transformer(self):
        print('clearing CLESA projector (freeing memory for all language-matrices) '
              '[warning: no further transformation will be allowed]')
        self.doc_projector = None

    def best_params(self):
        return self.model.best_params_


class KCCAPolylingualClassifier:
    """
    A polylingual classifier based on the KCCA method. Using:
    @article{bilenko2016pyrcca,
      title={Pyrcca: regularized kernel canonical correlation analysis in python and its applications to neuroimaging},
      author={Bilenko, Natalia Y and Gallant, Jack L},
      journal={Frontiers in neuroinformatics},
      volume={10},
      pages={49},
      year={2016},
      publisher={Frontiers}
    }
    Notes: other implementations (e.g., https://github.com/lorenzoriano/PyKCCA/blob/master/kcca.py) work only for
    2 languages
    """
    def __init__(self, base_learner, lW, z_parameters=None, kernel='linear', numCC=100, reg=0.0001,  max_wiki=-1, n_jobs=-1, dopickle=True):
        """
        :param base_learner: the learner operating in the kcca space
        :param lW: a dictionary {lang : wikipedia doc-by-term csr_matrix}
        :param z_parameters: the parameters of the learner to optimize for via 5-fold cv in the z-space
        :param similarity: the similarity measure between documents to be used ('dot' or 'cosine')
        :param post: any valid sklearn normalization method to be applied to the resulting doc embeddings, or None (default)
        :param n_jobs: number of parallel threads


        :param lW: a dictionary {lang : wikipedia doc-by-term matrix}
        :param z_parameters: the parameters of the learner to optimize for via 5-fold cv in the z-space
        """
        self.base_learner = base_learner
        self.z_parameters=z_parameters
        self.lW = lW
        self.kernel = kernel
        self.n_jobs = n_jobs
        self.time = 0
        self.langs = sorted(list(lW.keys()))
        self.kcca = None
        self.numCC = numCC
        self.reg = reg
        self.max_wiki = max_wiki
        self.dopickle = dopickle

    def fit(self, lX, ly):
        assert set(lX.keys()).issubset(set(self.lW.keys())), 'not all languages in scope'
        assert set(lX.keys()) == set(ly.keys()), 'inconsistent dictionaries in fit'

        from pyrcca.rcca import CCA

        tinit = time.time()
        nlangs = len(self.langs)
        nWdocs = self.lW[self.langs[0]].shape[0]
        if self.max_wiki > -1 and nWdocs > self.max_wiki:
            nWdocs = self.max_wiki
            self.lW = {l: self.lW[l][:nWdocs] for l in self.langs}
        print('Considering {} Wikipedia articles'.format(nWdocs))

        self.kcca = CCA(kernelcca=True, reg=self.reg, numCC=self.numCC)
        self.kcca = CCA(kernelcca=True, reg=self.reg, numCC=self.numCC)
        self.kcca.train([self.lW[l].toarray() for l in self.langs])
        print('kcca train took {:.2f} s'.format(self.kcca.train_time))

        print('projecting the documents')
        projections = self.kcca.project([lX[l].toarray() for l in self.langs])
        lZ = {l:projections[i] for i,l in enumerate(self.langs)}

        self.time=time.time()-tinit
        return self.fit_from_transformed(lZ, ly)

    def fit_from_transformed(self, lZ, ly):
        tinit = time.time()
        langs = list(lZ.keys())
        Z = np.vstack([lZ[lang] for lang in langs])  # Z is the language independent space
        zy = np.vstack([ly[lang] for lang in langs])

        print('fitting the Z-space of shape={}'.format(Z.shape))
        self.model = MonolingualClassifier(base_learner=self.base_learner, parameters=self.z_parameters, n_jobs=self.n_jobs)
        self.model.fit(Z, zy)
        self.time = self.time + (time.time() - tinit)
        print('SVM fit done in {:.2f} s'.format(self.time))
        return self

    def transform(self, lX, accum_time=True):
        tinit = time.time()
        projections = self.kcca.project([lX[l].toarray() for l in self.langs])
        lZ = {l: projections[i] for i, l in enumerate(self.langs)}
        if accum_time:
            self.time = self.time + (time.time() - tinit)
        return lZ

    def predict_from_transformed(self, lZ):
        assert self.model is not None, 'predict called before fit'
        return _joblib_transform_multiling(self.model.predict, lZ, n_jobs=self.n_jobs)

    def predict(self, lX):
        assert self.kcca is not None, 'KCCA predict called before fit'
        lZ = self.transform(lX, accum_time=False)
        return self.predict_from_transformed(lZ)

    def best_params(self):
        return self.model.best_params_



class DCIPolylingualClassifier:
    """
    An instantiation of Distributional Correspondence Indexing in polylingual documents using categories as pivots.
    @article{fernandez2016distributional,
      title={Distributional Correspondence Indexing for Cross-Lingual and Cross-Domain Sentiment Classification.},
      author={Fern{\'a}ndez, Alejandro Moreo and Esuli, Andrea and Sebastiani, Fabrizio},
      journal={Journal of artificial intelligence research},
      volume={55},
      pages={131--163},
      year={2016}
    }
    """
    def __init__(self, base_learner, dcf='linear', z_parameters=None, n_jobs=-1):
        self.base_learner = base_learner
        self.z_parameters=z_parameters
        self.n_jobs = n_jobs
        self.doc_projector = DistributionalCorrespondenceIndexing(dcf=dcf, post='normal', n_jobs=n_jobs)
        self.model = None

    def fit(self, lX, ly):
        tinit = time.time()
        print('fitting the projectors...')
        self.doc_projector.fit(lX,ly)

        print('projecting the documents')
        langs = list(lX.keys())
        lZ = self.doc_projector.transform(lX)

        Z = np.vstack([lZ[lang] for lang in langs]) # Z is the language independent space
        zy = np.vstack([ly[lang] for lang in langs])

        print('fitting the Z-space of shape={}'.format(Z.shape))
        self.model = MonolingualClassifier(base_learner=self.base_learner, parameters=self.z_parameters, n_jobs=self.n_jobs)
        self.model.fit(Z, zy)
        self.time = time.time() - tinit
        return self

    def predict(self, lX):
        """
        :param lX: a dictionary {language_label: X csr-matrix}
        :return: a dictionary of predictions
        """
        assert self.model is not None, 'predict called before fit'
        lZ = self.doc_projector.transform(lX)
        return _joblib_transform_multiling(self.model.predict, lZ, n_jobs=self.n_jobs)

    def best_params(self):
        return self.model.best_params()


class LRIPolylingualClassifier:
    """
    Performs Random Indexing (Bag-of-Concepts) in the juxtaposed representation, see:
    @article{fernandez2016lightweight,
      title={Lightweight random indexing for polylingual text classification},
      author={Moreo, Alejandro and Esuli, Andrea and Sebastiani, Fabrizio},
      journal={Journal of Artificial Intelligence Research},
      volume={57},
      pages={151--185},
      year={2016}
    }
    """
    def __init__(self, base_learner, parameters=None, reduction=0., n_jobs=-1):
        """
        :param base_learner: the learner in the random indexed space
        :param parameters: the parameters of the learner to optimize for via 5-fold cv
        :param reduction: the ratio of reduction of the dimensionality
        :param n_jobs: number of parallel threads
        """
        assert (isinstance(reduction, float) and 0 <= reduction < 1)\
            or (isinstance(reduction, int) and reduction > 0), \
            'reduction ratio should be in range [0,1) or positive integer'
        self.base_learner = base_learner
        self.parameters = parameters
        self.model = None
        self.reduction = reduction
        self.n_jobs = n_jobs

    def fit(self, lX, ly):
        """
        trains one classifiers in the juxtaposed random indexed feature space
        :param lX: a dictionary {language_label: X csr-matrix}; the feature space is assumed to be juxtaposed
        :param ly: a dictionary {language_label: y np.array}
        :return: self
        """
        tinit = time.time()
        assert set(lX.keys()) == set(ly.keys()), 'inconsistent language mappings in fit'
        assert len(np.unique([X.shape[1] for X in lX.values()]))==1, \
            'feature-spaces in the juxtaposed representation should be equal'

        langs = list(lX.keys())
        Xtr = scipy.sparse.vstack([lX[lang] for lang in langs])
        Ytr = np.vstack([ly[lang] for lang in langs])
        nF = Xtr.shape[1]

        print('random projection with LRI')
        if isinstance(self.reduction, float):
            dimensions = int(nF * (1.-self.reduction))
        else:
            dimensions = self.reduction
        self.BoC = RandomIndexingBoC(latent_dimensions=dimensions, non_zeros=2) # extremely sparse
        Xtr = self.BoC.fit_transform(Xtr)

        print('model fit')
        self.model = MonolingualClassifier(self.base_learner, parameters=self.parameters, n_jobs=self.n_jobs)
        self.model.fit(Xtr, Ytr)
        self.time = time.time() - tinit
        return self

    def transform(self, lX):
        assert self.model is not None, 'predict called before fit'
        return _joblib_transform_multiling(self.BoC.transform, lX, n_jobs=self.n_jobs)

    def predict(self, lX):
        """
        :param lX: a dictionary {language_label: X csr-matrix}; the feature space is assumed to be juxtaposed
        :return: a dictionary of predictions
        """
        assert self.model is not None, 'predict called before fit'
        lZ = self.transform(lX)
        return _joblib_transform_multiling(self.model.predict, lZ, n_jobs=self.n_jobs)

    def best_params(self):
        return self.model.best_params()


class MonolingualClassifier:

    def __init__(self, base_learner, parameters=None, n_jobs=-1):
        self.learner = base_learner
        self.parameters = parameters
        self.model = None
        self.n_jobs = n_jobs
        self.best_params_ = None

    def fit(self, X, y):
        if X.shape[0] == 0:
            print('Warning: X has 0 elements, a trivial rejector will be created')
            self.model = TrivialRejector().fit(X,y)
            self.empty_categories = np.arange(y.shape[1])
            return self

        tinit = time.time()
        _sort_if_sparse(X)
        self.empty_categories = np.argwhere(np.sum(y, axis=0)==0).flatten()

        # multi-class format
        if len(y.shape) == 2:
            if self.parameters is not None:
                self.parameters = [{'estimator__' + key: params[key] for key in params.keys()}
                                   for params in self.parameters]
            self.model = OneVsRestClassifier(self.learner, n_jobs=self.n_jobs)
        else:
            self.model = self.learner
            raise NotImplementedError('not working as a base-classifier for funneling if there are gaps in the labels across languages')

        # parameter optimization?
        if self.parameters:
            print('debug: optimizing parameters:', self.parameters)
            self.model = GridSearchCV(self.model, param_grid=self.parameters, refit=True, cv=5, n_jobs=self.n_jobs,
                                      error_score=0, verbose=10)

        print('fitting:',self.model)
        self.model.fit(X,y)
        if isinstance(self.model, GridSearchCV):
            self.best_params_ = self.model.best_params_
            print('best parameters: ', self.best_params_)
        self.time=time.time()-tinit
        return self

    def decision_function(self, X):
        assert self.model is not None, 'predict called before fit'
        _sort_if_sparse(X)
        return self.model.decision_function(X)

    def predict_proba(self, X):
        assert self.model is not None, 'predict called before fit'
        assert hasattr(self.model, 'predict_proba'), 'the probability predictions are not enabled in this model'
        _sort_if_sparse(X)
        return self.model.predict_proba(X)

    def predict(self, X):
        assert self.model is not None, 'predict called before fit'
        _sort_if_sparse(X)
        return self.model.predict(X)

    def best_params(self):
        return self.best_params_


class TrivialRejector:
    def fit(self, X, y):
        self.cats = y.shape[1]
        return self
    def decision_function(self, X): return np.zeros((X.shape[0],self.cats))
    def predict(self, X): return np.zeros((X.shape[0],self.cats))
    def predict_proba(self, X): return np.zeros((X.shape[0],self.cats))
    def best_params(self): return {}


class PolylingualEmbeddingsClassifier:
    """
    This classifier creates document embeddings by a tfidf weighted average of polylingual embeddings from the article
    @article{conneau2017word,
      title={Word translation without parallel data},
      author={Conneau, Alexis and Lample, Guillaume and Ranzato, Marc'Aurelio and Denoyer, Ludovic and J{\'e}gou, Herv{\'e}},
      journal={arXiv preprint arXiv:1710.04087},
      year={2017}
    }
    url: https://github.com/facebookresearch/MUSE
    """
    def __init__(self, wordembeddings_path, learner, c_parameters=None, n_jobs=-1):
        """
        :param wordembeddings_path: the path to the directory containing the polylingual embeddings
        :param learner: the learner
        :param c_parameters: parameters for learner
        :param n_jobs: the number of concurrent threads
        """
        self.wordembeddings_path = wordembeddings_path
        self.learner = learner
        self.c_parameters=c_parameters
        self.n_jobs = n_jobs
        self.lang_tfidf = {}
        self.model = None

    def fit_vectorizers(self, lX):
        for lang in lX.keys():
            if lang not in self.lang_tfidf:
                tfidf = TfidfVectorizer(sublinear_tf=True, use_idf=True)  # text is already processed
                docs = lX[lang]
                tfidf.fit(docs)
                self.lang_tfidf[lang] = tfidf

    def embed(self, docs, lang):
        assert lang in self.lang_tfidf, 'unknown language'
        tfidf_vectorizer = self.lang_tfidf[lang]
        V = tfidf_vectorizer.vocabulary_
        Xweights = tfidf_vectorizer.transform(docs)

        print('loading word embeddings for ' + lang)
        we = WordEmbeddings.load(self.wordembeddings_path, lang)

        nD = len(docs)
        doc_vecs = np.zeros((nD, we.dim()))

        for i, doc in enumerate(docs):
            print('\r\tcomplete {:.3f}%'.format(100 * (i + 1) / nD), end='')
            # averaging with tfidf (summing each word only once, since the frequency is already controlled)
            for w in set(doc.split()):
                if w in we and w in V:
                    doc_vecs[i] += (we[w] * Xweights[i, V[w]])
            # works much worse with idf; works much worse with document l2-normalization
        print()

        return doc_vecs

    def fit(self, lX, ly):
        """
        :param lX: a dictionary {language_label: [list of preprocessed documents]}
        :param ly: a dictionary {language_label: ndarray of shape (ndocs, ncats) binary labels}
        :return: self
        """
        tinit = time.time()
        langs = list(lX.keys())
        WEtr, Ytr = [], []
        self.fit_vectorizers(lX) # if already fit, does nothing
        for lang in langs:
            WEtr.append(self.embed(lX[lang], lang))
            Ytr.append(ly[lang])

        WEtr = np.vstack(WEtr)
        Ytr = np.vstack(Ytr)
        self.embed_time = time.time() - tinit

        print('fitting the WE-space of shape={}'.format(WEtr.shape))
        self.model = MonolingualClassifier(base_learner=self.learner, parameters=self.c_parameters, n_jobs=self.n_jobs)
        self.model.fit(WEtr, Ytr)
        self.time = time.time() - tinit
        return self

    def predict(self, lX):
        """
        :param lX: a dictionary {language_label: [list of preprocessed documents]}
        """
        assert self.model is not None, 'predict called before fit'
        langs = list(lX.keys())
        lWEte = {lang:self.embed(lX[lang], lang) for lang in langs} # parallelizing this may consume too much memory
        return _joblib_transform_multiling(self.model.predict, lWEte, n_jobs=self.n_jobs)

    def predict_proba(self, lX):
        """
        :param lX: a dictionary {language_label: [list of preprocessed documents]}
        """
        assert self.model is not None, 'predict called before fit'
        langs = list(lX.keys())
        lWEte = {lang:self.embed(lX[lang], lang) for lang in langs} # parallelizing this may consume too much memory
        return _joblib_transform_multiling(self.model.predict_proba, lWEte, n_jobs=self.n_jobs)

    def best_params(self):
        return self.model.best_params()


class FunnellingEmbeddingPolylingualClassifier:
    """ Simulated: this setting is merely for testing purposes, and is not realistic. We here assume to have a tfidf
    vectorizer for the out-of-scope languages (which is not fair)."""
    def __init__(self, first_tier_learner, embed_learner, meta_learner, wordembeddings_path, training_languages,
                 first_tier_parameters = None, embed_parameters = None, meta_parameters = None, n_jobs=-1):

        assert first_tier_learner.probability==True and embed_learner.probability==True, \
            'both the first-tier classifier and the polyembedding classifier shoud allow calibration'

        self.training_languages = training_languages

        self.PLE = PolylingualEmbeddingsClassifier(wordembeddings_path, embed_learner,
                                                   c_parameters=embed_parameters, n_jobs=n_jobs)

        self.Funnelling = FunnellingPolylingualClassifier(first_tier_learner, meta_learner,
                                                          first_tier_parameters=first_tier_parameters,
                                                          meta_parameters=meta_parameters, n_jobs=n_jobs)
        self.n_jobs = n_jobs

    def vectorize(self, lX):
        return {l:self.PLE.lang_tfidf[l].transform(lX[l]) for l in lX.keys()}

    def fit(self, lX, ly):
        """
        :param lX: a dictionary {language_label: [list of preprocessed documents]}
        :param ly: a dictionary {language_label: ndarray of shape (ndocs, ncats) binary labels}
        :return:
        """
        self.PLE.fit_vectorizers(lX)
        tinit = time.time()
        lX = {l: lX[l] for l in lX.keys() if l in self.training_languages}
        ly = {l: ly[l] for l in lX.keys() if l in self.training_languages}
        self.PLE.fit(lX, ly)
        lZ = self.PLE.predict_proba(lX)
        self.Funnelling.fit(self.vectorize(lX),ly,lZ,ly)
        self.time = time.time() - tinit
        return self

    def predict(self, lX):
        """
        :param lX: a dictionary {language_label: [list of preprocessed documents]}
        """
        lXin = {l: lX[l] for l in lX.keys() if l in self.training_languages}
        lXout = {l: lX[l] for l in lX.keys() if l not in self.training_languages}

        lZ = self.PLE.predict_proba(lXout)

        return self.Funnelling.predict(self.vectorize(lXin), lZ)


    def best_params(self):
        return {'PLE':self.PLE.best_params(), 'Funnelling':self.Funnelling.best_params()}


def _sort_if_sparse(X):
    if issparse(X) and not X.has_sorted_indices:
        X.sort_indices()


def _joblib_transform_multiling(transformer, lX, n_jobs=-1):
    if n_jobs == 1:
        return {lang:transformer(lX[lang]) for lang in lX.keys()}
    else:
        langs = list(lX.keys())
        transformations = Parallel(n_jobs=n_jobs)(delayed(transformer)(lX[lang]) for lang in langs)
        return {lang: transformations[i] for i, lang in enumerate(langs)}