text_analysis.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
#
# Copyright (C) 2013 Radim Rehurek <radimrehurek@seznam.cz>
# Licensed under the GNU LGPL v2.1 - http://www.gnu.org/licenses/lgpl.html

"""This module contains classes for analyzing the texts of a corpus to accumulate
statistical information about word occurrences."""

import itertools
import logging
import multiprocessing as mp
import sys
from collections import Counter

import numpy as np
import scipy.sparse as sps

from gensim import utils
from gensim.models.word2vec import Word2Vec

logger = logging.getLogger(__name__)


def _ids_to_words(ids, dictionary):
    """Convert an iterable of ids to their corresponding words using a dictionary.
    Abstract away the differences between the HashDictionary and the standard one.

    Parameters
    ----------
    ids: dict
        Dictionary of ids and their words.
    dictionary: :class:`~gensim.corpora.dictionary.Dictionary`
        Input gensim dictionary

    Returns
    -------
    set
        Corresponding words.

    Examples
    --------
    .. sourcecode:: pycon

        >>> from gensim.corpora.dictionary import Dictionary
        >>> from gensim.topic_coherence import text_analysis
        >>>
        >>> dictionary = Dictionary()
        >>> ids = {1: 'fake', 4: 'cats'}
        >>> dictionary.id2token = {1: 'fake', 2: 'tokens', 3: 'rabbids', 4: 'cats'}
        >>>
        >>> text_analysis._ids_to_words(ids, dictionary)
        set(['cats', 'fake'])

    """
    if not dictionary.id2token:  # may not be initialized in the standard gensim.corpora.Dictionary
        setattr(dictionary, 'id2token', {v: k for k, v in dictionary.token2id.items()})

    top_words = set()
    for word_id in ids:
        word = dictionary.id2token[word_id]
        if isinstance(word, set):
            top_words = top_words.union(word)
        else:
            top_words.add(word)

    return top_words


class BaseAnalyzer:
    """Base class for corpus and text analyzers.

    Attributes
    ----------
    relevant_ids : dict
        Mapping
    _vocab_size : int
        Size of vocabulary.
    id2contiguous : dict
        Mapping word_id -> number.
    log_every : int
        Interval for logging.
    _num_docs : int
        Number of documents.

    """
    def __init__(self, relevant_ids):
        """

        Parameters
        ----------
        relevant_ids : dict
            Mapping

        Examples
        --------
        .. sourcecode:: pycon

            >>> from gensim.topic_coherence import text_analysis
            >>> ids = {1: 'fake', 4: 'cats'}
            >>> base = text_analysis.BaseAnalyzer(ids)
            >>> # should return {1: 'fake', 4: 'cats'} 2 {1: 0, 4: 1} 1000 0
            >>> print(base.relevant_ids, base._vocab_size, base.id2contiguous, base.log_every, base._num_docs)
            {1: 'fake', 4: 'cats'} 2 {1: 0, 4: 1} 1000 0

        """
        self.relevant_ids = relevant_ids
        self._vocab_size = len(self.relevant_ids)
        self.id2contiguous = {word_id: n for n, word_id in enumerate(self.relevant_ids)}
        self.log_every = 1000
        self._num_docs = 0

    @property
    def num_docs(self):
        return self._num_docs

    @num_docs.setter
    def num_docs(self, num):
        self._num_docs = num
        if self._num_docs % self.log_every == 0:
            logger.info(
                "%s accumulated stats from %d documents",
                self.__class__.__name__, self._num_docs)

    def analyze_text(self, text, doc_num=None):
        raise NotImplementedError("Base classes should implement analyze_text.")

    def __getitem__(self, word_or_words):
        if isinstance(word_or_words, str) or not hasattr(word_or_words, '__iter__'):
            return self.get_occurrences(word_or_words)
        else:
            return self.get_co_occurrences(*word_or_words)

    def get_occurrences(self, word_id):
        """Return number of docs the word occurs in, once `accumulate` has been called."""
        return self._get_occurrences(self.id2contiguous[word_id])

    def _get_occurrences(self, word_id):
        raise NotImplementedError("Base classes should implement occurrences")

    def get_co_occurrences(self, word_id1, word_id2):
        """Return number of docs the words co-occur in, once `accumulate` has been called."""
        return self._get_co_occurrences(self.id2contiguous[word_id1], self.id2contiguous[word_id2])

    def _get_co_occurrences(self, word_id1, word_id2):
        raise NotImplementedError("Base classes should implement co_occurrences")


class UsesDictionary(BaseAnalyzer):
    """A BaseAnalyzer that uses a Dictionary, hence can translate tokens to counts.
    The standard BaseAnalyzer can only deal with token ids since it doesn't have the token2id
    mapping.

    Attributes
    ----------
    relevant_words : set
        Set of words that occurrences should be accumulated for.
    dictionary : :class:`~gensim.corpora.dictionary.Dictionary`
        Dictionary based on text
    token2id : dict
        Mapping from :class:`~gensim.corpora.dictionary.Dictionary`

    """
    def __init__(self, relevant_ids, dictionary):
        """

        Parameters
        ----------
        relevant_ids : dict
            Mapping
        dictionary : :class:`~gensim.corpora.dictionary.Dictionary`
            Dictionary based on text

        Examples
        --------
        .. sourcecode:: pycon

            >>> from gensim.topic_coherence import text_analysis
            >>> from gensim.corpora.dictionary import Dictionary
            >>>
            >>> ids = {1: 'foo', 2: 'bar'}
            >>> dictionary = Dictionary([['foo', 'bar', 'baz'], ['foo', 'bar', 'bar', 'baz']])
            >>> udict = text_analysis.UsesDictionary(ids, dictionary)
            >>>
            >>> print(udict.relevant_words)
            set([u'foo', u'baz'])

        """
        super(UsesDictionary, self).__init__(relevant_ids)
        self.relevant_words = _ids_to_words(self.relevant_ids, dictionary)
        self.dictionary = dictionary
        self.token2id = dictionary.token2id

    def get_occurrences(self, word):
        """Return number of docs the word occurs in, once `accumulate` has been called."""
        try:
            word_id = self.token2id[word]
        except KeyError:
            word_id = word
        return self._get_occurrences(self.id2contiguous[word_id])

    def _word2_contiguous_id(self, word):
        try:
            word_id = self.token2id[word]
        except KeyError:
            word_id = word
        return self.id2contiguous[word_id]

    def get_co_occurrences(self, word1, word2):
        """Return number of docs the words co-occur in, once `accumulate` has been called."""
        word_id1 = self._word2_contiguous_id(word1)
        word_id2 = self._word2_contiguous_id(word2)
        return self._get_co_occurrences(word_id1, word_id2)


class InvertedIndexBased(BaseAnalyzer):
    """Analyzer that builds up an inverted index to accumulate stats."""

    def __init__(self, *args):
        """

        Parameters
        ----------
        args : dict
            Look at :class:`~gensim.topic_coherence.text_analysis.BaseAnalyzer`

        Examples
        --------
        .. sourcecode:: pycon

            >>> from gensim.topic_coherence import text_analysis
            >>>
            >>> ids = {1: 'fake', 4: 'cats'}
            >>> ininb = text_analysis.InvertedIndexBased(ids)
            >>>
            >>> print(ininb._inverted_index)
            [set([]) set([])]

        """
        super(InvertedIndexBased, self).__init__(*args)
        self._inverted_index = np.array([set() for _ in range(self._vocab_size)])

    def _get_occurrences(self, word_id):
        return len(self._inverted_index[word_id])

    def _get_co_occurrences(self, word_id1, word_id2):
        s1 = self._inverted_index[word_id1]
        s2 = self._inverted_index[word_id2]
        return len(s1.intersection(s2))

    def index_to_dict(self):
        contiguous2id = {n: word_id for word_id, n in self.id2contiguous.items()}
        return {contiguous2id[n]: doc_id_set for n, doc_id_set in enumerate(self._inverted_index)}


class CorpusAccumulator(InvertedIndexBased):
    """Gather word occurrence stats from a corpus by iterating over its BoW representation."""

    def analyze_text(self, text, doc_num=None):
        """Build an inverted index from a sequence of corpus texts."""
        doc_words = frozenset(x[0] for x in text)
        top_ids_in_doc = self.relevant_ids.intersection(doc_words)
        for word_id in top_ids_in_doc:
            self._inverted_index[self.id2contiguous[word_id]].add(self._num_docs)

    def accumulate(self, corpus):
        for document in corpus:
            self.analyze_text(document)
            self.num_docs += 1
        return self


class WindowedTextsAnalyzer(UsesDictionary):
    """Gather some stats about relevant terms of a corpus by iterating over windows of texts."""

    def __init__(self, relevant_ids, dictionary):
        """

        Parameters
        ----------
        relevant_ids : set of int
            Relevant id
        dictionary : :class:`~gensim.corpora.dictionary.Dictionary`
            Dictionary instance with mappings for the relevant_ids.

        """
        super(WindowedTextsAnalyzer, self).__init__(relevant_ids, dictionary)
        self._none_token = self._vocab_size  # see _iter_texts for use of none token

    def accumulate(self, texts, window_size):
        relevant_texts = self._iter_texts(texts)
        windows = utils.iter_windows(
            relevant_texts, window_size, ignore_below_size=False, include_doc_num=True)

        for doc_num, virtual_document in windows:
            self.analyze_text(virtual_document, doc_num)
            self.num_docs += 1
        return self

    def _iter_texts(self, texts):
        dtype = np.uint16 if np.iinfo(np.uint16).max >= self._vocab_size else np.uint32
        for text in texts:
            if self.text_is_relevant(text):
                yield np.fromiter((
                    self.id2contiguous[self.token2id[w]] if w in self.relevant_words
                    else self._none_token
                    for w in text), dtype=dtype, count=len(text))

    def text_is_relevant(self, text):
        """Check if the text has any relevant words."""
        for word in text:
            if word in self.relevant_words:
                return True
        return False


class InvertedIndexAccumulator(WindowedTextsAnalyzer, InvertedIndexBased):
    """Build an inverted index from a sequence of corpus texts."""

    def analyze_text(self, window, doc_num=None):
        for word_id in window:
            if word_id is not self._none_token:
                self._inverted_index[word_id].add(self._num_docs)


class WordOccurrenceAccumulator(WindowedTextsAnalyzer):
    """Accumulate word occurrences and co-occurrences from a sequence of corpus texts."""

    def __init__(self, *args):
        super(WordOccurrenceAccumulator, self).__init__(*args)
        self._occurrences = np.zeros(self._vocab_size, dtype='uint32')
        self._co_occurrences = sps.lil_matrix((self._vocab_size, self._vocab_size), dtype='uint32')

        self._uniq_words = np.zeros((self._vocab_size + 1,), dtype=bool)  # add 1 for none token
        self._counter = Counter()

    def __str__(self):
        return self.__class__.__name__

    def accumulate(self, texts, window_size):
        self._co_occurrences = self._co_occurrences.tolil()
        self.partial_accumulate(texts, window_size)
        self._symmetrize()
        return self

    def partial_accumulate(self, texts, window_size):
        """Meant to be called several times to accumulate partial results.

        Notes
        -----
        The final accumulation should be performed with the `accumulate` method as opposed to this one.
        This method does not ensure the co-occurrence matrix is in lil format and does not
        symmetrize it after accumulation.

        """
        self._current_doc_num = -1
        self._token_at_edge = None
        self._counter.clear()

        super(WordOccurrenceAccumulator, self).accumulate(texts, window_size)
        for combo, count in self._counter.items():
            self._co_occurrences[combo] += count

        return self

    def analyze_text(self, window, doc_num=None):
        self._slide_window(window, doc_num)
        mask = self._uniq_words[:-1]  # to exclude none token
        if mask.any():
            self._occurrences[mask] += 1
            self._counter.update(itertools.combinations(np.nonzero(mask)[0], 2))

    def _slide_window(self, window, doc_num):
        if doc_num != self._current_doc_num:
            self._uniq_words[:] = False
            self._uniq_words[np.unique(window)] = True
            self._current_doc_num = doc_num
        else:
            self._uniq_words[self._token_at_edge] = False
            self._uniq_words[window[-1]] = True

        self._token_at_edge = window[0]

    def _symmetrize(self):
        """Word pairs may have been encountered in (i, j) and (j, i) order.

        Notes
        -----
        Rather than enforcing a particular ordering during the update process,
        we choose to symmetrize the co-occurrence matrix after accumulation has completed.

        """
        co_occ = self._co_occurrences
        co_occ.setdiag(self._occurrences)  # diagonal should be equal to occurrence counts
        self._co_occurrences = \
            co_occ + co_occ.T - sps.diags(co_occ.diagonal(), offsets=0, dtype='uint32')

    def _get_occurrences(self, word_id):
        return self._occurrences[word_id]

    def _get_co_occurrences(self, word_id1, word_id2):
        return self._co_occurrences[word_id1, word_id2]

    def merge(self, other):
        self._occurrences += other._occurrences
        self._co_occurrences += other._co_occurrences
        self._num_docs += other._num_docs


class PatchedWordOccurrenceAccumulator(WordOccurrenceAccumulator):
    """Monkey patched for multiprocessing worker usage, to move some of the logic to the master process."""
    def _iter_texts(self, texts):
        return texts  # master process will handle this


class ParallelWordOccurrenceAccumulator(WindowedTextsAnalyzer):
    """Accumulate word occurrences in parallel.

    Attributes
    ----------
    processes : int
        Number of processes to use; must be at least two.
    args :
        Should include `relevant_ids` and `dictionary` (see :class:`~UsesDictionary.__init__`).
    kwargs :
        Can include `batch_size`, which is the number of docs to send to a worker at a time.
        If not included, it defaults to 64.
    """

    def __init__(self, processes, *args, **kwargs):
        super(ParallelWordOccurrenceAccumulator, self).__init__(*args)
        if processes < 2:
            raise ValueError(
                "Must have at least 2 processes to run in parallel; got %d" % processes)
        self.processes = processes
        self.batch_size = kwargs.get('batch_size', 64)

    def __str__(self):
        return "%s(processes=%s, batch_size=%s)" % (
            self.__class__.__name__, self.processes, self.batch_size)

    def accumulate(self, texts, window_size):
        workers, input_q, output_q = self.start_workers(window_size)
        try:
            self.queue_all_texts(input_q, texts, window_size)
            interrupted = False
        except KeyboardInterrupt:
            logger.warn("stats accumulation interrupted; <= %d documents processed", self._num_docs)
            interrupted = True

        accumulators = self.terminate_workers(input_q, output_q, workers, interrupted)
        return self.merge_accumulators(accumulators)

    def start_workers(self, window_size):
        """Set up an input and output queue and start processes for each worker.

        Notes
        -----
        The input queue is used to transmit batches of documents to the workers.
        The output queue is used by workers to transmit the WordOccurrenceAccumulator instances.

        Parameters
        ----------
        window_size : int

        Returns
        -------
        (list of lists)
            Tuple of (list of workers, input queue, output queue).
        """
        input_q = mp.Queue(maxsize=self.processes)
        output_q = mp.Queue()
        workers = []
        for _ in range(self.processes):
            accumulator = PatchedWordOccurrenceAccumulator(self.relevant_ids, self.dictionary)
            worker = AccumulatingWorker(input_q, output_q, accumulator, window_size)
            worker.start()
            workers.append(worker)

        return workers, input_q, output_q

    def yield_batches(self, texts):
        """Return a generator over the given texts that yields batches of `batch_size` texts at a time."""
        batch = []
        for text in self._iter_texts(texts):
            batch.append(text)
            if len(batch) == self.batch_size:
                yield batch
                batch = []

        if batch:
            yield batch

    def queue_all_texts(self, q, texts, window_size):
        """Sequentially place batches of texts on the given queue until `texts` is consumed.
        The texts are filtered so that only those with at least one relevant token are queued.
        """
        for batch_num, batch in enumerate(self.yield_batches(texts)):
            q.put(batch, block=True)
            before = self._num_docs / self.log_every
            self._num_docs += sum(len(doc) - window_size + 1 for doc in batch)
            if before < (self._num_docs / self.log_every):
                logger.info(
                    "%d batches submitted to accumulate stats from %d documents (%d virtual)",
                    (batch_num + 1), (batch_num + 1) * self.batch_size, self._num_docs)

    def terminate_workers(self, input_q, output_q, workers, interrupted=False):
        """Wait until all workers have transmitted their WordOccurrenceAccumulator instances, then terminate each.

        Warnings
        --------
        We do not use join here because it has been shown to have some issues
        in Python 2.7 (and even in later versions). This method also closes both the input and output queue.
        If `interrupted` is False (normal execution), a None value is placed on the input queue for
        each worker. The workers are looking for this sentinel value and interpret it as a signal to
        terminate themselves. If `interrupted` is True, a KeyboardInterrupt occurred. The workers are
        programmed to recover from this and continue on to transmit their results before terminating.
        So in this instance, the sentinel values are not queued, but the rest of the execution
        continues as usual.

        """
        if not interrupted:
            for _ in workers:
                input_q.put(None, block=True)

        accumulators = []
        while len(accumulators) != len(workers):
            accumulators.append(output_q.get())
        logger.info("%d accumulators retrieved from output queue", len(accumulators))

        for worker in workers:
            if worker.is_alive():
                worker.terminate()

        input_q.close()
        output_q.close()
        return accumulators

    def merge_accumulators(self, accumulators):
        """Merge the list of accumulators into a single `WordOccurrenceAccumulator` with all
        occurrence and co-occurrence counts, and a `num_docs` that reflects the total observed
        by all the individual accumulators.

        """
        accumulator = WordOccurrenceAccumulator(self.relevant_ids, self.dictionary)
        for other_accumulator in accumulators:
            accumulator.merge(other_accumulator)
        # Workers do partial accumulation, so none of the co-occurrence matrices are symmetrized.
        # This is by design, to avoid unnecessary matrix additions/conversions during accumulation.
        accumulator._symmetrize()
        logger.info("accumulated word occurrence stats for %d virtual documents", accumulator.num_docs)
        return accumulator


class AccumulatingWorker(mp.Process):
    """Accumulate stats from texts fed in from queue."""

    def __init__(self, input_q, output_q, accumulator, window_size):
        super(AccumulatingWorker, self).__init__()
        self.input_q = input_q
        self.output_q = output_q
        self.accumulator = accumulator
        self.accumulator.log_every = sys.maxsize  # avoid logging in workers
        self.window_size = window_size

    def run(self):
        try:
            self._run()
        except KeyboardInterrupt:
            logger.info(
                "%s interrupted after processing %d documents",
                self.__class__.__name__, self.accumulator.num_docs)
        except Exception:
            logger.exception("worker encountered unexpected exception")
        finally:
            self.reply_to_master()

    def _run(self):
        batch_num = -1
        n_docs = 0
        while True:
            batch_num += 1
            docs = self.input_q.get(block=True)
            if docs is None:  # sentinel value
                logger.debug("observed sentinel value; terminating")
                break

            self.accumulator.partial_accumulate(docs, self.window_size)
            n_docs += len(docs)
            logger.debug(
                "completed batch %d; %d documents processed (%d virtual)",
                batch_num, n_docs, self.accumulator.num_docs)

        logger.debug(
            "finished all batches; %d documents processed (%d virtual)",
            n_docs, self.accumulator.num_docs)

    def reply_to_master(self):
        logger.info("serializing accumulator to return to master...")
        self.output_q.put(self.accumulator, block=False)
        logger.info("accumulator serialized")


class WordVectorsAccumulator(UsesDictionary):
    """Accumulate context vectors for words using word vector embeddings.

    Attributes
    ----------
    model: Word2Vec (:class:`~gensim.models.keyedvectors.KeyedVectors`)
        If None, a new Word2Vec model is trained on the given text corpus. Otherwise,
        it should be a pre-trained Word2Vec context vectors.
    model_kwargs:
        if model is None, these keyword arguments will be passed through to the Word2Vec constructor.
    """

    def __init__(self, relevant_ids, dictionary, model=None, **model_kwargs):
        super(WordVectorsAccumulator, self).__init__(relevant_ids, dictionary)
        self.model = model
        self.model_kwargs = model_kwargs

    def not_in_vocab(self, words):
        uniq_words = set(utils.flatten(words))
        return set(word for word in uniq_words if word not in self.model)

    def get_occurrences(self, word):
        """Return number of docs the word occurs in, once `accumulate` has been called."""
        try:
            self.token2id[word]  # is this a token or an id?
        except KeyError:
            word = self.dictionary.id2token[word]
        return self.model.get_vecattr(word, 'count')

    def get_co_occurrences(self, word1, word2):
        """Return number of docs the words co-occur in, once `accumulate` has been called."""
        raise NotImplementedError("Word2Vec model does not support co-occurrence counting")

    def accumulate(self, texts, window_size):
        if self.model is not None:
            logger.debug("model is already trained; no accumulation necessary")
            return self

        kwargs = self.model_kwargs.copy()
        if window_size is not None:
            kwargs['window'] = window_size
        kwargs['min_count'] = kwargs.get('min_count', 1)
        kwargs['sg'] = kwargs.get('sg', 1)
        kwargs['hs'] = kwargs.get('hw', 0)

        self.model = Word2Vec(**kwargs)
        self.model.build_vocab(texts)
        self.model.train(texts, total_examples=self.model.corpus_count, epochs=self.model.epochs)
        self.model = self.model.wv  # retain KeyedVectors
        return self

    def ids_similarity(self, ids1, ids2):
        words1 = self._words_with_embeddings(ids1)
        words2 = self._words_with_embeddings(ids2)
        return self.model.n_similarity(words1, words2)

    def _words_with_embeddings(self, ids):
        if not hasattr(ids, '__iter__'):
            ids = [ids]

        words = [self.dictionary.id2token[word_id] for word_id in ids]
        return [word for word in words if word in self.model]