Source code for gensim.utils

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Copyright (C) 2010 Radim Rehurek <>
# Licensed under the GNU LGPL v2.1 -

This module contains various general utility functions.

from __future__ import with_statement

import logging

logger = logging.getLogger(__name__)

    from html.entities import name2codepoint as n2cp
except ImportError:
    from htmlentitydefs import name2codepoint as n2cp
    import cPickle as _pickle
except ImportError:
    import pickle as _pickle

import re
import unicodedata
import os
import random
import itertools
import tempfile
from functools import wraps  # for `synchronous` function lock
import multiprocessing
import shutil
import sys
from contextlib import contextmanager
import subprocess

import numpy
import scipy.sparse

if sys.version_info[0] >= 3:
    unicode = str

from six import iteritems, u, string_types, unichr
from six.moves import xrange

    from smart_open import smart_open
except ImportError:"smart_open library not found; falling back to local-filesystem-only")

    def make_closing(base, **attrs):
        Add support for `with Base(attrs) as fout:` to the base class if it's missing.
        The base class' `close()` method will be called on context exit, to always close the file properly.

        This is needed for gzip.GzipFile, bz2.BZ2File etc in older Pythons (<=2.6), which otherwise
        raise "AttributeError: GzipFile instance has no attribute '__exit__'".

        if not hasattr(base, '__enter__'):
            attrs['__enter__'] = lambda self: self
        if not hasattr(base, '__exit__'):
            attrs['__exit__'] = lambda self, type, value, traceback: self.close()
        return type('Closing' + base.__name__, (base, object), attrs)

    def smart_open(fname, mode='rb'):
        _, ext = os.path.splitext(fname)
        if ext == '.bz2':
            from bz2 import BZ2File
            return make_closing(BZ2File)(fname, mode)
        if ext == '.gz':
            from gzip import GzipFile
            return make_closing(GzipFile)(fname, mode)
        return open(fname, mode)

PAT_ALPHABETIC = re.compile('(((?![\d])\w)+)', re.UNICODE)
RE_HTML_ENTITY = re.compile(r'&(#?)([xX]?)(\w{1,8});', re.UNICODE)

[docs]def synchronous(tlockname): """ A decorator to place an instance-based lock around a method. Adapted from """ def _synched(func): @wraps(func) def _synchronizer(self, *args, **kwargs): tlock = getattr(self, tlockname) logger.debug("acquiring lock %r for %s" % (tlockname, func.__name__)) with tlock: # use lock as a context manager to perform safe acquire/release pairs logger.debug("acquired lock %r for %s" % (tlockname, func.__name__)) result = func(self, *args, **kwargs) logger.debug("releasing lock %r for %s" % (tlockname, func.__name__)) return result return _synchronizer return _synched
[docs]class NoCM(object):
[docs] def acquire(self): pass
[docs] def release(self): pass
def __enter__(self): pass def __exit__(self, type, value, traceback): pass
nocm = NoCM() @contextmanager
[docs]def file_or_filename(input): """ Return a file-like object ready to be read from the beginning. `input` is either a filename (gz/bz2 also supported) or a file-like object supporting seek. """ if isinstance(input, string_types): # input was a filename: open as file yield smart_open(input) else: # input already a file-like object; just reset to the beginning yield input
[docs]def deaccent(text): """ Remove accentuation from the given string. Input text is either a unicode string or utf8 encoded bytestring. Return input string with accents removed, as unicode. >>> deaccent("Šéf chomutovských komunistů dostal poštou bílý prášek") u'Sef chomutovskych komunistu dostal postou bily prasek' """ if not isinstance(text, unicode): # assume utf8 for byte strings, use default (strict) error handling text = text.decode('utf8') norm = unicodedata.normalize("NFD", text) result = u('').join(ch for ch in norm if unicodedata.category(ch) != 'Mn') return unicodedata.normalize("NFC", result)
[docs]def tokenize(text, lowercase=False, deacc=False, errors="strict", to_lower=False, lower=False): """ Iteratively yield tokens as unicode strings, removing accent marks and optionally lowercasing the unidoce string by assigning True to one of the parameters, lowercase, to_lower, or lower. Input text may be either unicode or utf8-encoded byte string. The tokens on output are maximal contiguous sequences of alphabetic characters (no digits!). >>> list(tokenize('Nic nemůže letět rychlostí vyšší, než 300 tisíc kilometrů za sekundu!', deacc = True)) [u'Nic', u'nemuze', u'letet', u'rychlosti', u'vyssi', u'nez', u'tisic', u'kilometru', u'za', u'sekundu'] """ lowercase = lowercase or to_lower or lower text = to_unicode(text, errors=errors) if lowercase: text = text.lower() if deacc: text = deaccent(text) for match in PAT_ALPHABETIC.finditer(text): yield
[docs]def simple_preprocess(doc, deacc=False, min_len=2, max_len=15): """ Convert a document into a list of tokens. This lowercases, tokenizes, de-accents (optional). -- the output are final tokens = unicode strings, that won't be processed any further. """ tokens = [ token for token in tokenize(doc, lower=True, deacc=deacc, errors='ignore') if min_len <= len(token) <= max_len and not token.startswith('_') ] return tokens
[docs]def any2utf8(text, errors='strict', encoding='utf8'): """Convert a string (unicode or bytestring in `encoding`), to bytestring in utf8.""" if isinstance(text, unicode): return text.encode('utf8') # do bytestring -> unicode -> utf8 full circle, to ensure valid utf8 return unicode(text, encoding, errors=errors).encode('utf8')
to_utf8 = any2utf8
[docs]def any2unicode(text, encoding='utf8', errors='strict'): """Convert a string (bytestring in `encoding` or unicode), to unicode.""" if isinstance(text, unicode): return text return unicode(text, encoding, errors=errors)
to_unicode = any2unicode
[docs]class SaveLoad(object): """ Objects which inherit from this class have save/load functions, which un/pickle them to disk. This uses pickle for de/serializing, so objects must not contain unpicklable attributes, such as lambda functions etc. """ @classmethod
[docs] def load(cls, fname, mmap=None): """ Load a previously saved object from file (also see `save`). If the object was saved with large arrays stored separately, you can load these arrays via mmap (shared memory) using `mmap='r'`. Default: don't use mmap, load large arrays as normal objects. If the file being loaded is compressed (either '.gz' or '.bz2'), then `mmap=None` must be set. Load will raise an `IOError` if this condition is encountered. """"loading %s object from %s" % (cls.__name__, fname)) compress, subname = SaveLoad._adapt_by_suffix(fname) obj = unpickle(fname) obj._load_specials(fname, mmap, compress, subname) return obj
def _load_specials(self, fname, mmap, compress, subname): """ Loads any attributes that were stored specially, and gives the same opportunity to recursively included SaveLoad instances. """ mmap_error = lambda x, y: IOError( 'Cannot mmap compressed object %s in file %s. ' % (x, y) + 'Use `load(fname, mmap=None)` or uncompress files manually.') for attrib in getattr(self, '__recursive_saveloads', []): cfname = '.'.join((fname, attrib))"loading %s recursively from %s.* with mmap=%s" % ( attrib, cfname, mmap)) getattr(self, attrib)._load_specials(cfname, mmap, compress, subname) for attrib in getattr(self, '__numpys', []):"loading %s from %s with mmap=%s" % ( attrib, subname(fname, attrib), mmap)) if compress: if mmap: raise mmap_error(attrib, subname(fname, attrib)) val = numpy.load(subname(fname, attrib))['val'] else: val = numpy.load(subname(fname, attrib), mmap_mode=mmap) setattr(self, attrib, val) for attrib in getattr(self, '__scipys', []):"loading %s from %s with mmap=%s" % ( attrib, subname(fname, attrib), mmap)) sparse = unpickle(subname(fname, attrib)) if compress: if mmap: raise mmap_error(attrib, subname(fname, attrib)) with numpy.load(subname(fname, attrib, 'sparse')) as f: = f['data'] sparse.indptr = f['indptr'] sparse.indices = f['indices'] else: = numpy.load(subname(fname, attrib, 'data'), mmap_mode=mmap) sparse.indptr = numpy.load(subname(fname, attrib, 'indptr'), mmap_mode=mmap) sparse.indices = numpy.load(subname(fname, attrib, 'indices'), mmap_mode=mmap) setattr(self, attrib, sparse) for attrib in getattr(self, '__ignoreds', []):"setting ignored attribute %s to None" % (attrib)) setattr(self, attrib, None) @staticmethod def _adapt_by_suffix(fname): """Give appropriate compress setting and filename formula""" if fname.endswith('.gz') or fname.endswith('.bz2'): compress = True subname = lambda *args: '.'.join(list(args) + ['npz']) else: compress = False subname = lambda *args: '.'.join(list(args) + ['npy']) return (compress, subname) def _smart_save(self, fname, separately=None, sep_limit=10 * 1024**2, ignore=frozenset(), pickle_protocol=2): """ Save the object to file (also see `load`). If `separately` is None, automatically detect large numpy/scipy.sparse arrays in the object being stored, and store them into separate files. This avoids pickle memory errors and allows mmap'ing large arrays back on load efficiently. You can also set `separately` manually, in which case it must be a list of attribute names to be stored in separate files. The automatic check is not performed in this case. `ignore` is a set of attribute names to *not* serialize (file handles, caches etc). On subsequent load() these attributes will be set to None. `pickle_protocol` defaults to 2 so the pickled object can be imported in both Python 2 and 3. """ "saving %s object under %s, separately %s" % ( self.__class__.__name__, fname, separately)) compress, subname = SaveLoad._adapt_by_suffix(fname) restores = self._save_specials(fname, separately, sep_limit, ignore, pickle_protocol, compress, subname) try: pickle(self, fname, protocol=pickle_protocol) finally: # restore attribs handled specially for obj, asides in restores: for attrib, val in iteritems(asides): setattr(obj, attrib, val) def _save_specials(self, fname, separately, sep_limit, ignore, pickle_protocol, compress, subname): """ Save aside any attributes that need to be handled separately, including by recursion any attributes that are themselves SaveLoad instances. Returns a list of (obj, {attrib: value, ...}) settings that the caller should use to restore each object's attributes that were set aside during the default pickle(). """ asides = {} sparse_matrices = (scipy.sparse.csr_matrix, scipy.sparse.csc_matrix) if separately is None: separately = [] for attrib, val in iteritems(self.__dict__): if isinstance(val, numpy.ndarray) and val.size >= sep_limit: separately.append(attrib) elif isinstance(val, sparse_matrices) and val.nnz >= sep_limit: separately.append(attrib) # whatever's in `separately` or `ignore` at this point won't get pickled for attrib in separately + list(ignore): if hasattr(self, attrib): asides[attrib] = getattr(self, attrib) delattr(self, attrib) recursive_saveloads = [] restores = [] for attrib, val in iteritems(self.__dict__): if hasattr(val, '_save_specials'): # better than 'isinstance(val, SaveLoad)' if IPython reloading recursive_saveloads.append(attrib) cfname = '.'.join((fname,attrib)) restores.extend(val._save_specials(cfname, None, sep_limit, ignore, pickle_protocol, compress, subname)) try: numpys, scipys, ignoreds = [], [], [] for attrib, val in iteritems(asides): if isinstance(val, numpy.ndarray) and attrib not in ignore: numpys.append(attrib)"storing numpy array '%s' to %s" % ( attrib, subname(fname, attrib))) if compress: numpy.savez_compressed(subname(fname, attrib), val=numpy.ascontiguousarray(val)) else:, attrib), numpy.ascontiguousarray(val)) elif isinstance(val, (scipy.sparse.csr_matrix, scipy.sparse.csc_matrix)) and attrib not in ignore: scipys.append(attrib)"storing scipy.sparse array '%s' under %s" % ( attrib, subname(fname, attrib))) if compress: numpy.savez_compressed(subname(fname, attrib, 'sparse'),, indptr=val.indptr, indices=val.indices) else:, attrib, 'data'),, attrib, 'indptr'), val.indptr), attrib, 'indices'), val.indices) data, indptr, indices =, val.indptr, val.indices, val.indptr, val.indices = None, None, None try: # store array-less object pickle(val, subname(fname, attrib), protocol=pickle_protocol) finally:, val.indptr, val.indices = data, indptr, indices else:"not storing attribute %s" % (attrib)) ignoreds.append(attrib) self.__dict__['__numpys'] = numpys self.__dict__['__scipys'] = scipys self.__dict__['__ignoreds'] = ignoreds self.__dict__['__recursive_saveloads'] = recursive_saveloads except: # restore the attributes if exception-interrupted for attrib, val in iteritems(asides): setattr(self, attrib, val) raise return restores + [(self, asides)]
[docs] def save(self, fname_or_handle, separately=None, sep_limit=10 * 1024**2, ignore=frozenset(), pickle_protocol=2): """ Save the object to file (also see `load`). `fname_or_handle` is either a string specifying the file name to save to, or an open file-like object which can be written to. If the object is a file handle, no special array handling will be performed; all attributes will be saved to the same file. If `separately` is None, automatically detect large numpy/scipy.sparse arrays in the object being stored, and store them into separate files. This avoids pickle memory errors and allows mmap'ing large arrays back on load efficiently. You can also set `separately` manually, in which case it must be a list of attribute names to be stored in separate files. The automatic check is not performed in this case. `ignore` is a set of attribute names to *not* serialize (file handles, caches etc). On subsequent load() these attributes will be set to None. `pickle_protocol` defaults to 2 so the pickled object can be imported in both Python 2 and 3. """ try: _pickle.dump(self, fname_or_handle, protocol=pickle_protocol)"saved %s object" % self.__class__.__name__) except TypeError: # `fname_or_handle` does not have write attribute self._smart_save(fname_or_handle, separately, sep_limit, ignore, pickle_protocol=pickle_protocol)
#endclass SaveLoad
[docs]def identity(p): """Identity fnc, for flows that don't accept lambda (pickling etc).""" return p
[docs]def get_max_id(corpus): """ Return the highest feature id that appears in the corpus. For empty corpora (no features at all), return -1. """ maxid = -1 for document in corpus: maxid = max(maxid, max([-1] + [fieldid for fieldid, _ in document])) # [-1] to avoid exceptions from max(empty) return maxid
[docs]class FakeDict(object): """ Objects of this class act as dictionaries that map integer->str(integer), for a specified range of integers <0, num_terms). This is meant to avoid allocating real dictionaries when `num_terms` is huge, which is a waste of memory. """
[docs] def __init__(self, num_terms): self.num_terms = num_terms
def __str__(self): return "FakeDict(num_terms=%s)" % self.num_terms def __getitem__(self, val): if 0 <= val < self.num_terms: return str(val) raise ValueError("internal id out of bounds (%s, expected <0..%s))" % (val, self.num_terms))
[docs] def iteritems(self): for i in xrange(self.num_terms): yield i, str(i)
[docs] def keys(self): """ Override the dict.keys() function, which is used to determine the maximum internal id of a corpus = the vocabulary dimensionality. HACK: To avoid materializing the whole `range(0, self.num_terms)`, this returns the highest id = `[self.num_terms - 1]` only. """ return [self.num_terms - 1]
def __len__(self): return self.num_terms
[docs] def get(self, val, default=None): if 0 <= val < self.num_terms: return str(val) return default
[docs]def dict_from_corpus(corpus): """ Scan corpus for all word ids that appear in it, then construct and return a mapping which maps each `wordId -> str(wordId)`. This function is used whenever *words* need to be displayed (as opposed to just their ids) but no wordId->word mapping was provided. The resulting mapping only covers words actually used in the corpus, up to the highest wordId found. """ num_terms = 1 + get_max_id(corpus) id2word = FakeDict(num_terms) return id2word
[docs]def is_corpus(obj): """ Check whether `obj` is a corpus. Return (is_corpus, new) 2-tuple, where `new is obj` if `obj` was an iterable, or `new` yields the same sequence as `obj` if it was an iterator. `obj` is a corpus if it supports iteration over documents, where a document is in turn anything that acts as a sequence of 2-tuples (int, float). Note: An "empty" corpus (empty input sequence) is ambiguous, so in this case the result is forcefully defined as `is_corpus=False`. """ try: if 'Corpus' in obj.__class__.__name__: # the most common case, quick hack return True, obj except: pass try: if hasattr(obj, 'next') or hasattr(obj, '__next__'): # the input is an iterator object, meaning once we call next() # that element could be gone forever. we must be careful to put # whatever we retrieve back again doc1 = next(obj) obj = itertools.chain([doc1], obj) else: doc1 = next(iter(obj)) # empty corpus is resolved to False here if len(doc1) == 0: # sparse documents must have a __len__ function (list, tuple...) return True, obj # the first document is empty=>assume this is a corpus id1, val1 = next(iter(doc1)) # if obj is a numpy array, it resolves to False here id1, val1 = int(id1), float(val1) # must be a 2-tuple (integer, float) except Exception: return False, obj return True, obj
[docs]def get_my_ip(): """ Try to obtain our external ip (from the pyro nameserver's point of view) This tries to sidestep the issue of bogus `/etc/hosts` entries and other local misconfigurations, which often mess up hostname resolution. If all else fails, fall back to simple `socket.gethostbyname()` lookup. """ import socket try: import Pyro4 # we know the nameserver must exist, so use it as our anchor point ns = Pyro4.naming.locateNS() s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM) s.connect((, ns._pyroUri.port)) result, port = s.getsockname() except: try: # see what ifconfig says about our default interface import commands result = commands.getoutput("ifconfig").split("\n")[1].split()[1][5:] if len(result.split('.')) != 4: raise Exception() except: # give up, leave the resolution to gethostbyname result = socket.gethostbyname(socket.gethostname()) return result
[docs]class RepeatCorpus(SaveLoad): """ Used in the tutorial on distributed computing and likely not useful anywhere else. """
[docs] def __init__(self, corpus, reps): """ Wrap a `corpus` as another corpus of length `reps`. This is achieved by repeating documents from `corpus` over and over again, until the requested length `len(result)==reps` is reached. Repetition is done on-the-fly=efficiently, via `itertools`. >>> corpus = [[(1, 0.5)], []] # 2 documents >>> list(RepeatCorpus(corpus, 5)) # repeat 2.5 times to get 5 documents [[(1, 0.5)], [], [(1, 0.5)], [], [(1, 0.5)]] """ self.corpus = corpus self.reps = reps
def __iter__(self): return itertools.islice(itertools.cycle(self.corpus), self.reps)
[docs]class RepeatCorpusNTimes(SaveLoad):
[docs] def __init__(self, corpus, n): """ Repeat a `corpus` `n` times. >>> corpus = [[(1, 0.5)], []] >>> list(RepeatCorpusNTimes(corpus, 3)) # repeat 3 times [[(1, 0.5)], [], [(1, 0.5)], [], [(1, 0.5)], []] """ self.corpus = corpus self.n = n
def __iter__(self): for _ in xrange(self.n): for document in self.corpus: yield document
[docs]class ClippedCorpus(SaveLoad):
[docs] def __init__(self, corpus, max_docs=None): """ Return a corpus that is the "head" of input iterable `corpus`. Any documents after `max_docs` are ignored. This effectively limits the length of the returned corpus to <= `max_docs`. Set `max_docs=None` for "no limit", effectively wrapping the entire input corpus. """ self.corpus = corpus self.max_docs = max_docs
def __iter__(self): return itertools.islice(self.corpus, self.max_docs) def __len__(self): return min(self.max_docs, len(self.corpus))
[docs]class SlicedCorpus(SaveLoad):
[docs] def __init__(self, corpus, slice_): """ Return a corpus that is the slice of input iterable `corpus`. Negative slicing can only be used if the corpus is indexable. Otherwise, the corpus will be iterated over. Slice can also be a numpy.ndarray to support fancy indexing. NOTE: calculating the size of a SlicedCorpus is expensive when using a slice as the corpus has to be iterated over once. Using a list or numpy.ndarray does not have this drawback, but consumes more memory. """ self.corpus = corpus self.slice_ = slice_ self.length = None
def __iter__(self): if hasattr(self.corpus, 'index') and len(self.corpus.index) > 0: return (self.corpus.docbyoffset(i) for i in self.corpus.index[self.slice_]) else: return itertools.islice(self.corpus, self.slice_.start, self.slice_.stop, self.slice_.step) def __len__(self): # check cached length, calculate if needed if self.length is None: if isinstance(self.slice_, (list, numpy.ndarray)): self.length = len(self.slice_) else: self.length = sum(1 for x in self) return self.length
[docs]def safe_unichr(intval): try: return unichr(intval) except ValueError: # ValueError: unichr() arg not in range(0x10000) (narrow Python build) s = "\\U%08x" % intval # return UTF16 surrogate pair return s.decode('unicode-escape')
[docs]def decode_htmlentities(text): """ Decode HTML entities in text, coded as hex, decimal or named. Adapted from >>> u = u'E tu vivrai nel terrore - L&#x27;aldil&#xE0; (1981)' >>> print(decode_htmlentities(u).encode('UTF-8')) E tu vivrai nel terrore - L'aldilà (1981) >>> print(decode_htmlentities("l&#39;eau")) l'eau >>> print(decode_htmlentities("foo &lt; bar")) foo < bar """ def substitute_entity(match): try: ent = if == "#": # decoding by number if == '': # number is in decimal return safe_unichr(int(ent)) elif in ['x', 'X']: # number is in hex return safe_unichr(int(ent, 16)) else: # they were using a name cp = n2cp.get(ent) if cp: return safe_unichr(cp) else: return except: # in case of errors, return original input return return RE_HTML_ENTITY.sub(substitute_entity, text)
[docs]def chunkize_serial(iterable, chunksize, as_numpy=False): """ Return elements from the iterable in `chunksize`-ed lists. The last returned element may be smaller (if length of collection is not divisible by `chunksize`). >>> print(list(grouper(range(10), 3))) [[0, 1, 2], [3, 4, 5], [6, 7, 8], [9]] """ import numpy it = iter(iterable) while True: if as_numpy: # convert each document to a 2d numpy array (~6x faster when transmitting # chunk data over the wire, in Pyro) wrapped_chunk = [[numpy.array(doc) for doc in itertools.islice(it, int(chunksize))]] else: wrapped_chunk = [list(itertools.islice(it, int(chunksize)))] if not wrapped_chunk[0]: break # memory opt: wrap the chunk and then pop(), to avoid leaving behind a dangling reference yield wrapped_chunk.pop()
grouper = chunkize_serial
[docs]class InputQueue(multiprocessing.Process):
[docs] def __init__(self, q, corpus, chunksize, maxsize, as_numpy): super(InputQueue, self).__init__() self.q = q self.maxsize = maxsize self.corpus = corpus self.chunksize = chunksize self.as_numpy = as_numpy
[docs] def run(self): if self.as_numpy: import numpy # don't clutter the global namespace with a dependency on numpy it = iter(self.corpus) while True: chunk = itertools.islice(it, self.chunksize) if self.as_numpy: # HACK XXX convert documents to numpy arrays, to save memory. # This also gives a scipy warning at runtime: # "UserWarning: indices array has non-integer dtype (float64)" wrapped_chunk = [[numpy.asarray(doc) for doc in chunk]] else: wrapped_chunk = [list(chunk)] if not wrapped_chunk[0]: self.q.put(None, block=True) break try: qsize = self.q.qsize() except NotImplementedError: qsize = '?' logger.debug("prepared another chunk of %i documents (qsize=%s)" % (len(wrapped_chunk[0]), qsize)) self.q.put(wrapped_chunk.pop(), block=True)
#endclass InputQueue if == 'nt':"detected Windows; aliasing chunkize to chunkize_serial") def chunkize(corpus, chunksize, maxsize=0, as_numpy=False): for chunk in chunkize_serial(corpus, chunksize, as_numpy=as_numpy): yield chunk else:
[docs] def chunkize(corpus, chunksize, maxsize=0, as_numpy=False): """ Split a stream of values into smaller chunks. Each chunk is of length `chunksize`, except the last one which may be smaller. A once-only input stream (`corpus` from a generator) is ok, chunking is done efficiently via itertools. If `maxsize > 1`, don't wait idly in between successive chunk `yields`, but rather keep filling a short queue (of size at most `maxsize`) with forthcoming chunks in advance. This is realized by starting a separate process, and is meant to reduce I/O delays, which can be significant when `corpus` comes from a slow medium (like harddisk). If `maxsize==0`, don't fool around with parallelism and simply yield the chunksize via `chunkize_serial()` (no I/O optimizations). >>> for chunk in chunkize(range(10), 4): print(chunk) [0, 1, 2, 3] [4, 5, 6, 7] [8, 9] """ assert chunksize > 0 if maxsize > 0: q = multiprocessing.Queue(maxsize=maxsize) worker = InputQueue(q, corpus, chunksize, maxsize=maxsize, as_numpy=as_numpy) worker.daemon = True worker.start() while True: chunk = [q.get(block=True)] if chunk[0] is None: break yield chunk.pop() else: for chunk in chunkize_serial(corpus, chunksize, as_numpy=as_numpy): yield chunk
[docs]def smart_extension(fname, ext): fname, oext = os.path.splitext(fname) if oext.endswith('.bz2'): fname = fname + oext[:-4] + ext + '.bz2' elif oext.endswith('.gz'): fname = fname + oext[:-3] + ext + '.gz' else: fname = fname + oext + ext return fname
[docs]def pickle(obj, fname, protocol=2): """Pickle object `obj` to file `fname`. `protocol` defaults to 2 so pickled objects are compatible across Python 2.x and 3.x. """ with smart_open(fname, 'wb') as fout: # 'b' for binary, needed on Windows _pickle.dump(obj, fout, protocol=protocol)
[docs]def unpickle(fname): """Load pickled object from `fname`""" with smart_open(fname) as f: # Because of loading from S3 load can't be used (missing readline in smart_open) return _pickle.loads(
[docs]def revdict(d): """ Reverse a dictionary mapping. When two keys map to the same value, only one of them will be kept in the result (which one is kept is arbitrary). """ return dict((v, k) for (k, v) in iteritems(d))
[docs]def toptexts(query, texts, index, n=10): """ Debug fnc to help inspect the top `n` most similar documents (according to a similarity index `index`), to see if they are actually related to the query. `texts` is any object that can return something insightful for each document via `texts[docid]`, such as its fulltext or snippet. Return a list of 3-tuples (docid, doc's similarity to the query, texts[docid]). """ sims = index[query] # perform a similarity query against the corpus sims = sorted(enumerate(sims), key=lambda item: -item[1]) result = [] for topid, topcosine in sims[:n]: # only consider top-n most similar docs result.append((topid, topcosine, texts[topid])) return result
[docs]def randfname(prefix='gensim'): randpart = hex(random.randint(0, 0xffffff))[2:] return os.path.join(tempfile.gettempdir(), prefix + randpart)
[docs]def upload_chunked(server, docs, chunksize=1000, preprocess=None): """ Memory-friendly upload of documents to a SimServer (or Pyro SimServer proxy). Use this function to train or index large collections -- avoid sending the entire corpus over the wire as a single Pyro in-memory object. The documents will be sent in smaller chunks, of `chunksize` documents each. """ start = 0 for chunk in grouper(docs, chunksize): end = start + len(chunk)"uploading documents %i-%i" % (start, end - 1)) if preprocess is not None: pchunk = [] for doc in chunk: doc['tokens'] = preprocess(doc['text']) del doc['text'] pchunk.append(doc) chunk = pchunk server.buffer(chunk) start = end
[docs]def getNS(host=None, port=None, broadcast=True, hmac_key=None): """ Return a Pyro name server proxy. """ import Pyro4 try: return Pyro4.locateNS(host, port, broadcast, hmac_key) except Pyro4.errors.NamingError: raise RuntimeError("Pyro name server not found")
[docs]def pyro_daemon(name, obj, random_suffix=False, ip=None, port=None, ns_conf={}): """ Register object with name server (starting the name server if not running yet) and block until the daemon is terminated. The object is registered under `name`, or `name`+ some random suffix if `random_suffix` is set. """ if random_suffix: name += '.' + hex(random.randint(0, 0xffffff))[2:] import Pyro4 with getNS(**ns_conf) as ns: with Pyro4.Daemon(ip or get_my_ip(), port or 0) as daemon: # register server for remote access uri = daemon.register(obj, name) ns.remove(name) ns.register(name, uri)"%s registered with nameserver (URI '%s')" % (name, uri)) daemon.requestLoop()
[docs]def has_pattern(): """ Function to check if there is installed pattern library """ pattern = False try: from pattern.en import parse pattern = True except ImportError:"Pattern library is not installed, lemmatization won't be available.") return pattern
[docs]def lemmatize(content, allowed_tags=re.compile('(NN|VB|JJ|RB)'), light=False, stopwords=frozenset(), min_length=2, max_length=15): """ This function is only available when the optional 'pattern' package is installed. Use the English lemmatizer from `pattern` to extract UTF8-encoded tokens in their base form=lemma, e.g. "are, is, being" -> "be" etc. This is a smarter version of stemming, taking word context into account. Only considers nouns, verbs, adjectives and adverbs by default (=all other lemmas are discarded). >>> lemmatize('Hello World! How is it going?! Nonexistentword, 21') ['world/NN', 'be/VB', 'go/VB', 'nonexistentword/NN'] >>> lemmatize('The study ranks high.') ['study/NN', 'rank/VB', 'high/JJ'] >>> lemmatize('The ranks study hard.') ['rank/NN', 'study/VB', 'hard/RB'] """ if not has_pattern(): raise ImportError("Pattern library is not installed. Pattern library is needed in order \ to use lemmatize function") from pattern.en import parse if light: import warnings warnings.warn("The light flag is no longer supported by pattern.") # tokenization in `pattern` is weird; it gets thrown off by non-letters, # producing '==relate/VBN' or '**/NN'... try to preprocess the text a little # FIXME this throws away all fancy parsing cues, including sentence structure, # abbreviations etc. content = u(' ').join(tokenize(content, lower=True, errors='ignore')) parsed = parse(content, lemmata=True, collapse=False) result = [] for sentence in parsed: for token, tag, _, _, lemma in sentence: if min_length <= len(lemma) <= max_length and not lemma.startswith('_') and lemma not in stopwords: if allowed_tags.match(tag): lemma += "/" + tag[:2</