def saveCorpus(fname, corpus, id2word = None):
"""
Save a corpus in the List-of-words format.
"""
if id2word is None:
logging.info("no word id mapping provided; initializing from corpus")
id2word = utils.dictFromCorpus(corpus)
logging.info("storing corpus in List-Of-Words format: %s" % fname)
truncated = 0
fout = open(fname, 'w')
fout.write('%i\n' % len(corpus))
for doc in corpus:
words = []
for wordId, value in doc:
if abs(int(value) - value) > 1e-6:
truncated += 1
words.extend([str(id2word[wordId])] * int(value))
fout.write('%s\n' % ' '.join(words))
fout.close()
if truncated:
logging.warning("List-of-words format can only save vectors with \
integer entries; %i float entries were truncated to integer value" %
truncated)
def saveCorpus(fname, corpus, id2word=None):
"""
Save a corpus in the LDA-C format.
There are actually two files saved: `fname` and `fname.vocab`, where
`fname.vocab` is the vocabulary file.
This function is automatically called by `BleiCorpus.serialize`; don't
call it directly, call `serialize` instead.
"""
if id2word is None:
logging.info("no word id mapping provided; initializing from corpus")
id2word = utils.dictFromCorpus(corpus)
numTerms = len(id2word)
else:
numTerms = 1 + max([-1] + id2word.keys())
logging.info("storing corpus in Blei's LDA-C format: %s" % fname)
with open(fname, 'w') as fout:
offsets = []
for doc in corpus:
doc = list(doc)
offsets.append(fout.tell())
fout.write("%i %s\n" % (len(doc), ' '.join("%i:%s" % p for p in doc)))
fout.close()
# write out vocabulary, in a format compatible with Blei's topics.py script
fnameVocab = fname + '.vocab'
logging.info("saving vocabulary of %i words to %s" % (numTerms, fnameVocab))
fout = open(fnameVocab, 'w')
for featureId in xrange(numTerms):
fout.write("%s\n" % utils.toUtf8(id2word.get(featureId, '---')))
return offsets
def saveCorpus(fname, corpus, id2word=None):
"""
Save a corpus in the List-of-words format.
This function is automatically called by `LowCorpus.serialize`; don't
call it directly, call `serialize` instead.
"""
if id2word is None:
logging.info(
"no word id mapping provided; initializing from corpus")
id2word = utils.dictFromCorpus(corpus)
logging.info("storing corpus in List-Of-Words format: %s" % fname)
truncated = 0
offsets = []
with open(fname, 'w') as fout:
fout.write('%i\n' % len(corpus))
for doc in corpus:
words = []
for wordId, value in doc:
if abs(int(value) - value) > 1e-6:
truncated += 1
words.extend([str(id2word[wordId])] * int(value))
offsets.append(fout.tell())
fout.write('%s\n' % ' '.join(words))
if truncated:
logging.warning(
"List-of-words format can only save vectors with "
"integer elements; %i float entries were truncated to integer value"
% truncated)
return offsets
def saveCorpus(fname, corpus, id2word = None):
"""
Save a corpus in the Matrix Market format.
There are actually two files saved: fname and fname.vocab, where
fname.vocab is the vocabulary file.
"""
if id2word is None:
logging.info("no word id mapping provided; initializing from corpus")
id2word = utils.dictFromCorpus(corpus)
numTerms = len(id2word)
else:
numTerms = 1 + max([-1] + id2word.keys())
logging.info("storing corpus in Blei's LDA-C format: %s" % fname)
fout = open(fname, 'w')
for doc in corpus:
fout.write("%i %s\n" % (len(doc), ' '.join("%i:%f" % p for p in doc)))
fout.close()
# write out vocabulary, in a format compatible with Blei's topics.py script
fnameVocab = fname + '.vocab'
logging.info("saving vocabulary of %i words to %s" % (numTerms, fnameVocab))
fout = open(fnameVocab, 'w')
for featureId in xrange(numTerms):
fout.write("%s\n" % id2word.get(featureId, '---'))
fout.close()
def saveCorpus(fname, corpus, id2word = None):
"""
Save a corpus in the Matrix Market format.
There are actually two files saved: `fname` and `fname.vocab`, where
`fname.vocab` is the vocabulary file.
"""
if id2word is None:
logging.info("no word id mapping provided; initializing from corpus")
id2word = utils.dictFromCorpus(corpus)
numTerms = len(id2word)
else:
numTerms = 1 + max([-1] + id2word.keys())
logging.info("storing corpus in Blei's LDA-C format: %s" % fname)
fout = open(fname, 'w')
for doc in corpus:
doc = list(doc)
fout.write("%i %s\n" % (len(doc), ' '.join("%i:%s" % p for p in doc)))
fout.close()
# write out vocabulary, in a format compatible with Blei's topics.py script
fnameVocab = fname + '.vocab'
logging.info("saving vocabulary of %i words to %s" % (numTerms, fnameVocab))
fout = open(fnameVocab, 'w')
for featureId in xrange(numTerms):
fout.write("%s\n" % utils.toUtf8(id2word.get(featureId, '---')))
fout.close()
def initialize(self, corpus):
"""
Compute inverse document weights, which will be used to modify term
frequencies for documents.
"""
if self.id2word is None:
logging.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
logging.info("calculating IDF weights over %i documents" % len(corpus))
idfs = {}
numNnz = 0
for docNo, bow in enumerate(corpus):
if docNo % 5000 == 0:
logging.info("PROGRESS: processing document %i/%i" %
(docNo, len(corpus)))
numNnz += len(bow)
for termId, termCount in bow:
idfs[termId] = idfs.get(termId, 0) + 1
idfs = dict((termId, math.log(1.0 * (docNo + 1) / docFreq, 2)) # the IDF weight formula
for termId, docFreq in idfs.iteritems())
self.idfs = idfs
# keep some stats about the training corpus
self.numDocs = len(corpus)
self.numNnz = numNnz
def initialize(self, corpus, chunks = 100, keepDecomposition = False, dtype = numpy.float64):
"""
Run SVD decomposition on the corpus. This will define the latent space into
which terms and documents will be mapped.
The SVD is created incrementally, in blocks of `chunks` documents. In the
end, a `self.projection` matrix is constructed that can be used to transform
documents into the latent space. The `U, S, V` decomposition itself is
discarded, unless `keepDecomposition` is True, in which case it is stored
in `self.u`, `self.s` and `self.v`.
`dtype` dictates precision used for intermediate computations; the final
projection will however always be of type numpy.float32.
The algorithm is adapted from:
**M. Brand. 2006. Fast low-rank modifications of the thin singular value decomposition**
"""
if self.id2word is None:
logging.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
# initialize decomposition (zero documents so far)
self.u = numpy.matrix(numpy.zeros((self.numTerms, self.numTopics)), dtype = dtype)
self.s = numpy.matrix(numpy.zeros((self.numTopics, self.numTopics)), dtype = dtype)
#self.v = numpy.matrix(numpy.zeros((0, self.numTopics)), dtype = dtype)
self.v = None
# do the actual work -- perform iterative singular value decomposition.
# this is done by sequentially updating SVD with `chunks` new documents
chunker = itertools.groupby(enumerate(corpus), key = lambda val: val[0] / chunks)
for chunkNo, (key, group) in enumerate(chunker):
# convert the chunk of sparse documents to full vectors
docs = [matutils.sparse2full(doc, self.numTerms) for docNo, doc in group]
# self.svdAddCols(docs, reorth = chunkNo % 100 == 99) # reorthogonalize once in every "100*chunks" documents
self.svdAddCols(docs, reorth = False)
logging.info("processed documents up to #%s" % docNo)
# calculate projection needed to get document-topic matrix from term-document matrix.
#
# the way to represent a vector `x` in latent space is lsi[x] = v = self.s^-1 * self.u^-1 * x,
# so the projection is self.s^-1 * self.u^-1.
#
# the way to compare two documents `x1`, `x2` is to compute v1 * self.s^2 * v2.T, so
# we pre-multiply v * s (ie., scale axes by singular values), and return
# that directly as the representation of `x` in LSI space.
#
# this conveniently simplifies to lsi[x] = self.u.T * x, so the projection is
# just self.u.T
#
# note that neither `v` (the right singular vectors) nor `s` (the singular
# values) are used at all in the transformation
self.projection = self.u.T.astype(numpy.float32).copy('C')
if not keepDecomposition:
# once we have the projection stored in self, discard u*s*v decomposition to free up memory
del self.u, self.v
def saveCorpus(fname, corpus, id2word=None):
"""
Save a corpus in the List-of-words format.
This function is automatically called by `LowCorpus.serialize`; don't
call it directly, call `serialize` instead.
"""
if id2word is None:
logging.info("no word id mapping provided; initializing from corpus")
id2word = utils.dictFromCorpus(corpus)
logging.info("storing corpus in List-Of-Words format: %s" % fname)
truncated = 0
offsets = []
with open(fname, "w") as fout:
fout.write("%i\n" % len(corpus))
for doc in corpus:
words = []
for wordId, value in doc:
if abs(int(value) - value) > 1e-6:
truncated += 1
words.extend([str(id2word[wordId])] * int(value))
offsets.append(fout.tell())
fout.write("%s\n" % " ".join(words))
if truncated:
logging.warning(
"List-of-words format can only save vectors with "
"integer elements; %i float entries were truncated to integer value" % truncated
)
return offsets
def initialize(self, corpus, chunks=100, keepDecomposition=False, dtype=numpy.float64):
"""
Run SVD decomposition on the corpus. This will define the latent space into
which terms and documents will be mapped.
The SVD is created incrementally, in blocks of `chunks` documents. In the
end, a `self.projection` matrix is constructed that can be used to transform
documents into the latent space. The `U, S, V` decomposition itself is
discarded, unless `keepDecomposition` is True, in which case it is stored
in `self.u`, `self.s` and `self.v`.
`dtype` dictates precision used for intermediate computations; the final
projection will however always be of type numpy.float32.
The algorithm is adapted from:
**M. Brand. 2006. Fast low-rank modifications of the thin singular value decomposition**
"""
if self.id2word is None:
logging.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
# initialize decomposition (zero documents so far)
self.u = numpy.matrix(numpy.zeros((self.numTerms, self.numTopics)), dtype=dtype)
self.s = numpy.matrix(numpy.zeros((self.numTopics, self.numTopics)), dtype=dtype)
# self.v = numpy.matrix(numpy.zeros((0, self.numTopics)), dtype = dtype)
self.v = None
# do the actual work -- perform iterative singular value decomposition.
# this is done by sequentially updating SVD with `chunks` new documents
chunker = itertools.groupby(enumerate(corpus), key=lambda val: val[0] / chunks)
for chunkNo, (key, group) in enumerate(chunker):
# convert the chunk of sparse documents to full vectors
docs = [matutils.sparse2full(doc, self.numTerms) for docNo, doc in group]
# self.svdAddCols(docs, reorth = chunkNo % 100 == 99) # reorthogonalize once in every "100*chunks" documents
self.svdAddCols(docs, reorth=False)
logging.info("processed documents up to #%s" % docNo)
# calculate projection needed to get document-topic matrix from term-document matrix.
#
# the way to represent a vector `x` in latent space is lsi[x] = v = self.s^-1 * self.u^-1 * x,
# so the projection is self.s^-1 * self.u^-1.
#
# the way to compare two documents `x1`, `x2` is to compute v1 * self.s^2 * v2.T, so
# we pre-multiply v * s (ie., scale axes by singular values), and return
# that directly as the representation of `x` in LSI space.
#
# this conveniently simplifies to lsi[x] = self.u.T * x, so the projection is
# just self.u.T
#
# note that neither `v` (the right singular vectors) nor `s` (the singular
# values) are used at all in the transformation
self.projection = self.u.T.astype(numpy.float32).copy("C")
if not keepDecomposition:
# once we have the projection stored in self, discard u*s*v decomposition to free up memory
del self.u, self.v
def __init__(self, corpus = None, id2word = None, numTopics = 200, extraDims = 10,
chunks = 100, dtype = numpy.float64):
"""
`numTopics` is the number of requested factors (latent dimensions).
After the model has been trained, you can estimate topics for an
arbitrary, unseen document, using the ``topics = self[document]`` dictionary
notation. You can also add new training documents, with ``self.addDocuments``,
so that training can be stopped and resumed at any time, and the
LSI transformation is available at any point.
`extraDims` is the number of extra dimensions that will be internally
computed (ie. `numTopics + extraDims`) to improve numerical properties of
the SVD algorithm. These extra dimensions will be eventually chopped off
for the final projection. Set to 0 to save memory; set to ~10 to
2*numTopics for increased SVD precision.
If you specify a `corpus`, it will be used to train the model. See the
method `addDocuments` for a description of the `chunks` and `decay` parameters.
The algorithm is based on
**Brand, 2006: Fast low-rank modifications of the thin singular value decomposition**.
Example:
>>> lsi = LsiModel(corpus, numTopics = 10)
>>> print lsi[doc_tfidf]
>>> lsi.addDocuments(corpus2) # update LSI on additional documents
>>> print lsi[doc_tfidf]
"""
self.id2word = id2word
self.numTopics = numTopics # number of latent topics
self.extraDims = extraDims
self.dtype = dtype
if corpus is None and self.id2word is None:
raise ValueError('at least one of corpus/id2word must be specified, to establish input space dimensionality')
if self.id2word is None:
logging.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
self.projection = numpy.asmatrix(numpy.zeros((self.numTopics, self.numTerms), dtype = dtype))
self.u = None
self.s = numpy.asmatrix(numpy.zeros((self.numTopics + self.extraDims, self.numTopics + self.extraDims)), dtype = dtype)
self.v = None
if corpus is not None:
self.addDocuments(corpus, chunks = chunks, updateProjection = True)
def initialize(self, corpus):
"""
Initialize the random projection matrix.
"""
if self.id2word is None:
logging.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
# Now construct the projection matrix itself.
# Here i use a particular form, derived in "Achlioptas: Database-friendly random projection",
# and his (1) scenario of Theorem 1.1 in particular (all entries are +1/-1).
randmat = 1 - 2 * numpy.random.binomial(1, 0.5, (self.numTopics, self.numTerms)) # convert from 0/1 to +1/-1
self.projection = numpy.asmatrix(randmat, dtype = numpy.float32) # convert from int32 to floats, for faster multiplications
def initialize(self, corpus):
"""
Initialize the random projection matrix.
"""
if self.id2word is None:
logging.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
# Now construct the projection matrix itself.
#
# Here i use a particular form, derived in "Achlioptas: Database-friendly random projection",
# and his (1) scenario of Theorem 1.1 in particular (all entries are +1/-1).
tmp = numpy.random.binomial(1, 0.5, (self.numTopics, self.numTerms)) # FIXME temporary array unnecessarily big (int32 -> int8)
self.projection = numpy.asmatrix(1 - 2 * tmp.astype(numpy.int8)) # convert from 0/1 to +1/-1
def initialize(self, corpus):
"""
Initialize the random projection matrix.
"""
if self.id2word is None:
logging.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
shape = self.numTopics, self.numTerms
logger.info("constructing %s random matrix" % str(shape))
# Now construct the projection matrix itself.
# Here i use a particular form, derived in "Achlioptas: Database-friendly random projection",
# and his (1) scenario of Theorem 1.1 in particular (all entries are +1/-1).
randmat = 1 - 2 * numpy.random.binomial(1, 0.5, shape) # convert from 0/1 to +1/-1
self.projection = numpy.asfortranarray(randmat, dtype=numpy.float32) # convert from int32 to floats, for faster multiplications
def __init__(self, corpus, id2word = None, numTopics = 200, alpha = None, initMode = 'random'):
"""
Initialize the model based on corpus.
`id2word` is a mapping from word ids (integers) to words (strings). It is
used to determine the vocabulary size, as well as for debugging and topic
printing.
`numTopics` is the number of requested topics.
`alpha` is either None (to be estimated during training) or a number
between (0.0, 1.0).
"""
# store user-supplied parameters
self.id2word = id2word
if self.id2word is None:
logger.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
self.numTopics = numTopics # number of latent topics
# internal constants; can be manually changed after having called this constructor and before calling `initialize()`
self.ESTIMATE_ALPHA = alpha is None
if alpha is None: # no alpha supplied by user => get some initial estimate
alpha = 10.0 / numTopics # initial estimate is 50 / numTopics, as suggested in Steyvers&Griffiths: Probabilistic Topic Models
self.alpha = min(0.99999, max(0.00001, alpha)) # dirichlet prior; make sure it's within bounds
# set EM training constants
self.EM_MAX_ITER = 50 # maximum number of EM iterations; usually converges much earlier
self.EM_CONVERGED = 0.0001 # relative difference between two iterations; if lower than this, stop the EM training
self.VAR_MAX_ITER = 20 # maximum number of document inference iterations
self.VAR_CONVERGED = 0.000001 # relative difference between document inference iterations needed to stop sooner than VAR_MAX_ITER
if corpus is not None:
self.initialize(corpus, initMode)
def saveCorpus(fname, corpus, id2word=None):
"""
Save a corpus in the LDA-C format.
There are actually two files saved: `fname` and `fname.vocab`, where
`fname.vocab` is the vocabulary file.
This function is automatically called by `BleiCorpus.serialize`; don't
call it directly, call `serialize` instead.
"""
if id2word is None:
logging.info(
"no word id mapping provided; initializing from corpus")
id2word = utils.dictFromCorpus(corpus)
numTerms = len(id2word)
else:
numTerms = 1 + max([-1] + id2word.keys())
logging.info("storing corpus in Blei's LDA-C format: %s" % fname)
with open(fname, 'w') as fout:
offsets = []
for doc in corpus:
doc = list(doc)
offsets.append(fout.tell())
fout.write("%i %s\n" % (len(doc), ' '.join("%i:%s" % p
for p in doc)))
fout.close()
# write out vocabulary, in a format compatible with Blei's topics.py script
fnameVocab = fname + '.vocab'
logging.info("saving vocabulary of %i words to %s" %
(numTerms, fnameVocab))
fout = open(fnameVocab, 'w')
for featureId in xrange(numTerms):
fout.write("%s\n" %
utils.toUtf8(id2word.get(featureId, '---')))
return offsets
def initialize(self, corpus):
"""
Initialize the random projection matrix.
"""
if self.id2word is None:
logging.info(
"no word id mapping provided; initializing from corpus, assuming identity"
)
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
# Now construct the projection matrix itself.
#
# Here i use a particular form, derived in "Achlioptas: Database-friendly random projection",
# and his (1) scenario of Theorem 1.1 in particular (all entries are +1/-1).
tmp = numpy.random.binomial(
1, 0.5,
(self.numTopics, self.numTerms
)) # FIXME temporary array unnecessarily big (int32 -> int8)
self.projection = numpy.asmatrix(
1 - 2 * tmp.astype(numpy.int8)) # convert from 0/1 to +1/-1
def __init__(self,
corpus=None,
numTopics=200,
id2word=None,
chunks=20000,
decay=1.0,
distributed=False,
onepass=False):
"""
`numTopics` is the number of requested factors (latent dimensions).
After the model has been trained, you can estimate topics for an
arbitrary, unseen document, using the ``topics = self[document]`` dictionary
notation. You can also add new training documents, with ``self.addDocuments``,
so that training can be stopped and resumed at any time, and the
LSI transformation is available at any point.
If you specify a `corpus`, it will be used to train the model. See the
method `addDocuments` for a description of the `chunks` and `decay` parameters.
If your document stream is one-pass only (the stream cannot be repeated),
turn on `onepass` to force a single pass SVD algorithm (slower).
Turn on `distributed` to enforce distributed computing (only makes sense
if `onepass` is set at the same time, too).
Example:
>>> lsi = LsiModel(corpus, numTopics=10)
>>> print lsi[doc_tfidf] # project some document into LSI space
>>> lsi.addDocuments(corpus2) # update LSI on additional documents
>>> print lsi[doc_tfidf]
"""
self.id2word = id2word
self.numTopics = int(numTopics)
self.chunks = int(chunks)
self.decay = float(decay)
if distributed:
if not onepass:
logger.warning(
"forcing the one-pass algorithm for distributed LSA")
onepass = True
self.onepass = onepass
if corpus is None and self.id2word is None:
raise ValueError(
'at least one of corpus/id2word must be specified, to establish input space dimensionality'
)
if self.id2word is None:
logger.info(
"no word id mapping provided; initializing from corpus, assuming identity"
)
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
self.docs_processed = 0
self.projection = Projection(self.numTerms, self.numTopics)
if not distributed:
logger.info("using serial LSI version on this node")
self.dispatcher = None
else:
if not onepass:
raise NotImplementedError(
"distributed stochastic LSA not implemented yet; "
"run either distributed one-pass, or serial randomized.")
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy('PYRONAME:gensim.dispatcher@%s' %
ns._pyroUri.location)
logger.debug("looking for dispatcher at %s" %
str(dispatcher._pyroUri))
dispatcher.initialize(id2word=self.id2word,
numTopics=numTopics,
chunks=chunks,
decay=decay,
distributed=False,
onepass=onepass)
self.dispatcher = dispatcher
logger.info("using distributed version with %i workers" %
len(dispatcher.getworkers()))
except Exception, err:
# distributed version was specifically requested, so this is an error state
logger.error("failed to initialize distributed LSI (%s)" % err)
raise RuntimeError(
"failed to initialize distributed LSI (%s)" % err)
def __init__(self, corpus=None, numTopics=100, id2word=None, distributed=False,
chunks=10000, passes=1, update_every=1, alpha=None, eta=None, decay=0.5):
"""
`numTopics` is the number of requested latent topics to be extracted from
the training corpus.
`id2word` is a mapping from word ids (integers) to words (strings). It is
used to determine the vocabulary size, as well as for debugging and topic
printing.
`alpha` and `eta` are hyperparameters on document-topic (theta) and
topic-word (lambda) distributions. Both default to a symmetric 1.0/numTopics
(but can be set to a vector, for assymetric priors).
Turn on `distributed` to force distributed computing (see the web tutorial
on how to set up a cluster of machines for gensim).
Example:
>>> lda = LdaModel(corpus, numTopics=100)
>>> print lda[doc_bow] # get topic probability distribution for a document
>>> lda.update(corpus2) # update the LDA model with additional documents
>>> print lda[doc_bow]
"""
# store user-supplied parameters
self.id2word = id2word
if corpus is None and self.id2word is None:
raise ValueError('at least one of corpus/id2word must be specified, to establish input space dimensionality')
if self.id2word is None:
logger.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
if self.numTerms == 0:
raise ValueError("cannot compute LDA over an empty collection (no terms)")
self.distributed = bool(distributed)
self.numTopics = int(numTopics)
self.chunks = chunks
self.decay = decay
self.num_updates = 0
self.passes = passes
self.update_every = update_every
if alpha is None:
self.alpha = 1.0 / numTopics
else:
self.alpha = alpha
if eta is None:
self.eta = 1.0 / numTopics
else:
self.eta = eta
# VB constants
self.VAR_MAXITER = 50
self.VAR_THRESH = 0.001
# set up distributed environment if necessary
if not distributed:
logger.info("using serial LDA version on this node")
self.dispatcher = None
self.numworkers = 1
else:
# set up distributed version
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy('PYRONAME:gensim.lda_dispatcher@%s' % ns._pyroUri.location)
dispatcher._pyroOneway.add("exit")
logger.debug("looking for dispatcher at %s" % str(dispatcher._pyroUri))
dispatcher.initialize(id2word=id2word, numTopics=numTopics,
chunks=chunks, alpha=alpha, eta=eta, distributed=False)
self.dispatcher = dispatcher
self.numworkers = len(dispatcher.getworkers())
logger.info("using distributed version with %i workers" % self.numworkers)
except Exception, err:
logger.error("failed to initialize distributed LDA (%s)" % err)
raise RuntimeError("failed to initialize distributed LDA (%s)" % err)
def __init__(self,
corpus=None,
id2word=None,
numTopics=200,
extraDims=10,
chunks=100,
dtype=numpy.float64):
"""
`numTopics` is the number of requested factors (latent dimensions).
After the model has been trained, you can estimate topics for an
arbitrary, unseen document, using the ``topics = self[document]`` dictionary
notation. You can also add new training documents, with ``self.addDocuments``,
so that training can be stopped and resumed at any time, and the
LSI transformation is available at any point.
`extraDims` is the number of extra dimensions that will be internally
computed (ie. `numTopics + extraDims`) to improve numerical properties of
the SVD algorithm. These extra dimensions will be eventually chopped off
for the final projection. Set to 0 to save memory; set to ~10 to
2*numTopics for increased SVD precision.
If you specify a `corpus`, it will be used to train the model. See the
method `addDocuments` for a description of the `chunks` and `decay` parameters.
The algorithm is based on
**Brand, 2006: Fast low-rank modifications of the thin singular value decomposition**.
Example:
>>> lsi = LsiModel(corpus, numTopics = 10)
>>> print lsi[doc_tfidf]
>>> lsi.addDocuments(corpus2) # update LSI on additional documents
>>> print lsi[doc_tfidf]
"""
self.id2word = id2word
self.numTopics = numTopics # number of latent topics
self.extraDims = extraDims
self.dtype = dtype
if corpus is None and self.id2word is None:
raise ValueError(
'at least one of corpus/id2word must be specified, to establish input space dimensionality'
)
if self.id2word is None:
logging.info(
"no word id mapping provided; initializing from corpus, assuming identity"
)
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
self.projection = numpy.asmatrix(
numpy.zeros((self.numTopics, self.numTerms), dtype=dtype))
self.u = None
self.s = numpy.asmatrix(numpy.zeros((self.numTopics + self.extraDims,
self.numTopics + self.extraDims)),
dtype=dtype)
self.v = None
if corpus is not None:
self.addDocuments(corpus, chunks=chunks, updateProjection=True)
def __init__(self, corpus=None, numTopics=200, id2word=None, chunks=20000,
decay=1.0, distributed=False, onepass=False):
"""
`numTopics` is the number of requested factors (latent dimensions).
After the model has been trained, you can estimate topics for an
arbitrary, unseen document, using the ``topics = self[document]`` dictionary
notation. You can also add new training documents, with ``self.addDocuments``,
so that training can be stopped and resumed at any time, and the
LSI transformation is available at any point.
If you specify a `corpus`, it will be used to train the model. See the
method `addDocuments` for a description of the `chunks` and `decay` parameters.
If your document stream is one-pass only (the stream cannot be repeated),
turn on `onepass` to force a single pass SVD algorithm (slower).
Turn on `distributed` to force distributed computing.
Example:
>>> lsi = LsiModel(corpus, numTopics=10)
>>> print lsi[doc_tfidf] # project some document into LSI space
>>> lsi.addDocuments(corpus2) # update LSI on additional documents
>>> print lsi[doc_tfidf]
"""
self.id2word = id2word
self.numTopics = int(numTopics)
self.chunks = int(chunks)
self.decay = float(decay)
if distributed:
if not onepass:
logger.warning("forcing the one-pass algorithm for distributed LSA")
onepass = True
self.onepass = onepass
if corpus is None and self.id2word is None:
raise ValueError('at least one of corpus/id2word must be specified, to establish input space dimensionality')
if self.id2word is None:
logger.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
self.docs_processed = 0
self.projection = Projection(self.numTerms, self.numTopics)
if not distributed:
logger.info("using serial LSI version on this node")
self.dispatcher = None
else:
if not onepass:
raise NotImplementedError("distributed stochastic LSA not implemented yet; "
"run either distributed one-pass, or serial randomized.")
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy('PYRONAME:gensim.lsi_dispatcher@%s' % ns._pyroUri.location)
dispatcher._pyroOneway.add("exit")
logger.debug("looking for dispatcher at %s" % str(dispatcher._pyroUri))
dispatcher.initialize(id2word = self.id2word, numTopics = numTopics,
chunks = chunks, decay = decay,
distributed = False, onepass = onepass)
self.dispatcher = dispatcher
logger.info("using distributed version with %i workers" % len(dispatcher.getworkers()))
except Exception, err:
# distributed version was specifically requested, so this is an error state
logger.error("failed to initialize distributed LSI (%s)" % err)
raise RuntimeError("failed to initialize distributed LSI (%s)" % err)
def __init__(self, corpus=None, numTopics=100, id2word=None, distributed=False,
chunks=1000, passes=1, update_every=1, alpha=None, eta=None, decay=0.5):
"""
`numTopics` is the number of requested latent topics to be extracted from
the training corpus.
`id2word` is a mapping from word ids (integers) to words (strings). It is
used to determine the vocabulary size, as well as for debugging and topic
printing.
`alpha` and `eta` are hyperparameters on document-topic (theta) and
topic-word (lambda) distributions. Both default to a symmetric 1.0/numTopics
(but can be set to a vector, for assymetric priors).
Turn on `distributed` to force distributed computing (see the web tutorial
on how to set up a cluster of machines for gensim).
Example:
>>> lda = LdaModel(corpus, numTopics=100)
>>> print lda[doc_bow] # get topic probability distribution for a document
>>> lda.update(corpus2) # update the LDA model with additional documents
>>> print lda[doc_bow]
"""
# store user-supplied parameters
self.id2word = id2word
if corpus is None and self.id2word is None:
raise ValueError('at least one of corpus/id2word must be specified, to establish input space dimensionality')
if self.id2word is None:
logger.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
if self.numTerms == 0:
raise ValueError("cannot compute LDA over an empty collection (no terms)")
self.distributed = bool(distributed)
self.numTopics = int(numTopics)
self.chunks = chunks
self.decay = decay
self.num_updates = 0
self.passes = passes
self.update_every = update_every
if alpha is None:
self.alpha = 1.0 / numTopics
else:
self.alpha = alpha
if eta is None:
self.eta = 1.0 / numTopics
else:
self.eta = eta
# VB constants
self.VAR_MAXITER = 50
self.VAR_THRESH = 0.001
# set up distributed environment if necessary
if not distributed:
logger.info("using serial LDA version on this node")
self.dispatcher = None
self.numworkers = 1
else:
# set up distributed version
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy('PYRONAME:gensim.lda_dispatcher@%s' % ns._pyroUri.location)
dispatcher._pyroOneway.add("exit")
logger.debug("looking for dispatcher at %s" % str(dispatcher._pyroUri))
dispatcher.initialize(id2word=id2word, numTopics=numTopics,
chunks=chunks, alpha=alpha, eta=eta, distributed=False)
self.dispatcher = dispatcher
self.numworkers = len(dispatcher.getworkers())
logger.info("using distributed version with %i workers" % self.numworkers)
except Exception, err:
logger.error("failed to initialize distributed LDA (%s)" % err)
raise RuntimeError("failed to initialize distributed LDA (%s)" % err)
def __init__(self,
corpus=None,
id2word=None,
numTopics=200,
chunks=None,
decay=1.0,
serial_only=None):
"""
`numTopics` is the number of requested factors (latent dimensions).
After the model has been trained, you can estimate topics for an
arbitrary, unseen document, using the ``topics = self[document]`` dictionary
notation. You can also add new training documents, with ``self.addDocuments``,
so that training can be stopped and resumed at any time, and the
LSI transformation is available at any point.
If you specify a `corpus`, it will be used to train the model. See the
method `addDocuments` for a description of the `chunks` and `decay` parameters.
The algorithm will automatically try to find active nodes on other computers
and run in a distributed manner; if this fails, it falls back to serial mode
(single core). To suppress distributed computing, set the `serial_only`
constructor parameter to True.
Example:
>>> lsi = LsiModel(corpus, numTopics = 10)
>>> print lsi[doc_tfidf]
>>> lsi.addDocuments(corpus2) # update LSI on additional documents
>>> print lsi[doc_tfidf]
"""
self.id2word = id2word
self.numTopics = numTopics # number of latent topics
if chunks is None:
# by default, proceed in chunks as big as number of topics, to improve accuracy
self.chunks = max(numTopics, 100)
else:
self.chunks = chunks
self.decay = decay
if corpus is None and self.id2word is None:
raise ValueError(
'at least one of corpus/id2word must be specified, to establish input space dimensionality'
)
if self.id2word is None:
logger.info(
"no word id mapping provided; initializing from corpus, assuming identity"
)
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
self.docs_processed = 0
self.projection = Projection(self.numTerms, self.numTopics)
if serial_only:
logger.info("using slave LSI version on this node")
self.dispatcher = None
else:
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy('PYRONAME:gensim.dispatcher@%s' %
ns._pyroUri.location)
logger.debug("looking for dispatcher at %s" %
str(dispatcher._pyroUri))
dispatcher.initialize(id2word=self.id2word,
numTopics=numTopics,
chunks=chunks,
decay=decay,
serial_only=True)
self.dispatcher = dispatcher
logger.info("using distributed version with %i workers" %
len(dispatcher.getworkers()))
except Exception, err:
if serial_only is not None:
# distributed version was specifically requested, so this is an error state
logger.error("failed to initialize distributed LSI (%s)" %
err)
raise RuntimeError(
"failed to initialize distributed LSI (%s)" % err)
else:
# user didn't request distributed specifically; just let him know we're running in serial
logger.info(
"distributed LSI not available, running LSI in serial mode (%s)"
% err)
self.dispatcher = None
def __init__(self, corpus = None, id2word = None, numTopics = 200,
chunks = 10000, decay = 1.0, serial_only = None):
"""
`numTopics` is the number of requested factors (latent dimensions).
After the model has been trained, you can estimate topics for an
arbitrary, unseen document, using the ``topics = self[document]`` dictionary
notation. You can also add new training documents, with ``self.addDocuments``,
so that training can be stopped and resumed at any time, and the
LSI transformation is available at any point.
If you specify a `corpus`, it will be used to train the model. See the
method `addDocuments` for a description of the `chunks` and `decay` parameters.
The algorithm will automatically try to find active nodes on other computers
and run in a distributed manner; if this fails, it falls back to serial mode
(single core). To suppress distributed computing, set the `serial_only`
constructor parameter to True.
Example:
>>> lsi = LsiModel(corpus, numTopics = 10)
>>> print lsi[doc_tfidf]
>>> lsi.addDocuments(corpus2) # update LSI on additional documents
>>> print lsi[doc_tfidf]
"""
self.id2word = id2word
self.numTopics = numTopics # number of latent topics
self.chunks = chunks
self.decay = decay
if corpus is None and self.id2word is None:
raise ValueError('at least one of corpus/id2word must be specified, to establish input space dimensionality')
if self.id2word is None:
logger.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
self.docs_processed = 0
self.projection = Projection(self.numTerms, self.numTopics)
if serial_only:
logger.info("using slave LSI version on this node")
self.dispatcher = None
else:
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy('PYRONAME:gensim.dispatcher@%s' % ns._pyroUri.location)
logger.debug("looking for dispatcher at %s" % str(dispatcher._pyroUri))
dispatcher.initialize(id2word = self.id2word, numTopics = numTopics,
chunks = chunks, decay = decay,
serial_only = True)
self.dispatcher = dispatcher
logger.info("using distributed version with %i workers" % len(dispatcher.getworkers()))
except Exception, err:
if serial_only is not None:
# distributed version was specifically requested, so this is an error state
logger.error("failed to initialize distributed LSI (%s)" % err)
raise RuntimeError("failed to initialize distributed LSI (%s)" % err)
else:
# user didn't request distributed specifically; just let him know we're running in serial
logger.info("distributed LSI not available, running LSI in serial mode (%s)" % err)
self.dispatcher = None
def __init__(self,
corpus=None,
numTopics=200,
id2word=None,
distributed=False,
chunks=None,
alpha=None,
initMode='random',
dtype=numpy.float64):
"""
`numTopics` is the number of requested latent topics to be extracted from
the training corpus.
`id2word` is a mapping from word ids (integers) to words (strings). It is
used to determine the vocabulary size, as well as for debugging and topic
printing.
`initMode` can be either 'random', for a fast random initialization of
the model parameters, or 'seeded', for an initialization based on a handful
of real documents. The 'seeded' mode requires an extra sweep over the entire
input corpus, and is thus much slower.
`alpha` is either None (to be estimated during training) or a number
between (0.0, 1.0).
Turn on `distributed` to force distributed computing (see the web tutorial
on how to set up a cluster).
Example:
>>> lda = LdaModel(corpus, numTopics=100)
>>> print lda[doc_tfidf] # get topic probability distribution for a documents
>>> lda.addDocuments(corpus2) # update LDA with additional documents
>>> print lda[doc_tfidf]
"""
# store user-supplied parameters
self.id2word = id2word
if self.id2word is None:
logger.info(
"no word id mapping provided; initializing from corpus, assuming identity"
)
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
if self.numTerms == 0:
raise ValueError(
"cannot compute LDA over an empty collection (no terms)")
self.distributed = bool(distributed)
self.numTopics = int(numTopics)
self.state = LdaState()
self.chunks = chunks
# initialize wordtype/topic counts
if initMode == 'seeded': # init from corpus (slow)
self.state.classWord = self.countsFromCorpus(corpus, numInitDocs=2)
elif initMode == 'random': # init with 1/k+noise
self.state.classWord = 1.0 / self.numTerms + numpy.random.rand(
self.numTopics, self.numTerms) # add noise from <0, 1)
else:
raise NotImplementedError(
"LDA initialization mode '%s' not supported" % str(initMode))
self.state.classWord = self.state.classWord.astype(dtype)
# internal algorithm constants
self.estimate_alpha = alpha is None
if self.estimate_alpha: # no alpha supplied by user => get some initial estimate
alpha = 10.0 / numTopics # n / numTopics, as suggested in Steyvers&Griffiths: Probabilistic Topic Models
self.alpha = min(0.99999, max(
0.00001, alpha)) # dirichlet prior; make sure it's within bounds
# EM training constants
self.EM_MAX_ITER = 50 # maximum number of EM iterations; usually converges earlier
self.EM_CONVERGED = 1e-4 # relative difference between two iterations; if lower than this, stop the EM training
self.VAR_MAX_ITER = 20 # maximum number of document inference iterations
self.VAR_CONVERGED = 1e-6 # relative difference between document inference iterations needed to stop sooner than VAR_MAX_ITER
if not distributed:
logger.info("using serial LDA version on this node")
self.dispatcher = None
else:
# set up distributed version
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy(
'PYRONAME:gensim.lda_dispatcher@%s' % ns._pyroUri.location)
dispatcher._pyroOneway.add("exit")
logger.debug("looking for dispatcher at %s" %
str(dispatcher._pyroUri))
dispatcher.initialize(id2word=self.id2word,
numTopics=numTopics,
chunks=chunks,
alpha=alpha,
distributed=False)
self.dispatcher = dispatcher
logger.info("using distributed version with %i workers" %
len(dispatcher.getworkers()))
except Exception, err:
logger.error("failed to initialize distributed LDA (%s)" % err)
raise RuntimeError(
"failed to initialize distributed LDA (%s)" % err)
def __init__(self, corpus=None, numTopics=200, id2word=None, distributed=False,
chunks=None, alpha=None, initMode='random', dtype=numpy.float64):
"""
`numTopics` is the number of requested latent topics to be extracted from
the training corpus.
`id2word` is a mapping from word ids (integers) to words (strings). It is
used to determine the vocabulary size, as well as for debugging and topic
printing.
`initMode` can be either 'random', for a fast random initialization of
the model parameters, or 'seeded', for an initialization based on a handful
of real documents. The 'seeded' mode requires an extra sweep over the entire
input corpus, and is thus much slower.
`alpha` is either None (to be estimated during training) or a number
between (0.0, 1.0).
Turn on `distributed` to force distributed computing (see the web tutorial
on how to set up a cluster).
Example:
>>> lda = LdaModel(corpus, numTopics=100)
>>> print lda[doc_tfidf] # get topic probability distribution for a documents
>>> lda.addDocuments(corpus2) # update LDA with additional documents
>>> print lda[doc_tfidf]
"""
# store user-supplied parameters
self.id2word = id2word
if self.id2word is None:
logger.info("no word id mapping provided; initializing from corpus, assuming identity")
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
if self.numTerms == 0:
raise ValueError("cannot compute LDA over an empty collection (no terms)")
self.distributed = bool(distributed)
self.numTopics = int(numTopics)
self.state = LdaState()
self.chunks = chunks
# initialize wordtype/topic counts
if initMode == 'seeded': # init from corpus (slow)
self.state.classWord = self.countsFromCorpus(corpus, numInitDocs=2)
elif initMode == 'random': # init with 1/k+noise
self.state.classWord = 1.0 / self.numTerms + numpy.random.rand(self.numTopics, self.numTerms) # add noise from <0, 1)
else:
raise NotImplementedError("LDA initialization mode '%s' not supported" % str(initMode))
self.state.classWord = self.state.classWord.astype(dtype)
# internal algorithm constants
self.estimate_alpha = alpha is None
if self.estimate_alpha: # no alpha supplied by user => get some initial estimate
alpha = 10.0 / numTopics # n / numTopics, as suggested in Steyvers&Griffiths: Probabilistic Topic Models
self.alpha = min(0.99999, max(0.00001, alpha)) # dirichlet prior; make sure it's within bounds
# EM training constants
self.EM_MAX_ITER = 50 # maximum number of EM iterations; usually converges earlier
self.EM_CONVERGED = 1e-4 # relative difference between two iterations; if lower than this, stop the EM training
self.VAR_MAX_ITER = 20 # maximum number of document inference iterations
self.VAR_CONVERGED = 1e-6 # relative difference between document inference iterations needed to stop sooner than VAR_MAX_ITER
if not distributed:
logger.info("using serial LDA version on this node")
self.dispatcher = None
else:
# set up distributed version
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy('PYRONAME:gensim.lda_dispatcher@%s' % ns._pyroUri.location)
dispatcher._pyroOneway.add("exit")
logger.debug("looking for dispatcher at %s" % str(dispatcher._pyroUri))
dispatcher.initialize(id2word=self.id2word, numTopics=numTopics,
chunks=chunks, alpha=alpha, distributed=False)
self.dispatcher = dispatcher
logger.info("using distributed version with %i workers" % len(dispatcher.getworkers()))
except Exception, err:
logger.error("failed to initialize distributed LDA (%s)" % err)
raise RuntimeError("failed to initialize distributed LDA (%s)" % err)
def __init__(self,
corpus=None,
numTopics=200,
id2word=None,
chunks=20000,
decay=1.0,
distributed=False,
onepass=True,
power_iters=P2_EXTRA_ITERS,
extra_samples=P2_EXTRA_DIMS):
"""
`numTopics` is the number of requested factors (latent dimensions).
After the model has been trained, you can estimate topics for an
arbitrary, unseen document, using the ``topics = self[document]`` dictionary
notation. You can also add new training documents, with ``self.addDocuments``,
so that training can be stopped and resumed at any time, and the
LSI transformation is available at any point.
If you specify a `corpus`, it will be used to train the model. See the
method `addDocuments` for a description of the `chunks` and `decay` parameters.
Turn `onepass` off to force a multi-pass stochastic algorithm.
`power_iters` and `extra_samples` affect the accuracy of the stochastic
multi-pass algorithm, which is used either internally (`onepass=True`) or
as the front-end algorithm (`onepass=False`). Increasing the number of
power iterations improves accuracy, but lowers performance. See [2]_ for
some hard numbers.
Turn on `distributed` to enable distributed computing.
Example:
>>> lsi = LsiModel(corpus, numTopics=10)
>>> print lsi[doc_tfidf] # project some document into LSI space
>>> lsi.addDocuments(corpus2) # update LSI on additional documents
>>> print lsi[doc_tfidf]
.. [2] http://nlp.fi.muni.cz/~xrehurek/nips/rehurek_nips.pdf
"""
self.id2word = id2word
self.numTopics = int(numTopics)
self.chunks = int(chunks)
self.decay = float(decay)
if distributed:
if not onepass:
logger.warning(
"forcing the one-pass algorithm for distributed LSA")
onepass = True
self.onepass = onepass
self.extra_samples, self.power_iters = extra_samples, power_iters
if corpus is None and self.id2word is None:
raise ValueError(
'at least one of corpus/id2word must be specified, to establish input space dimensionality'
)
if self.id2word is None:
logger.info(
"no word id mapping provided; initializing from corpus, assuming identity"
)
self.id2word = utils.dictFromCorpus(corpus)
self.numTerms = len(self.id2word)
else:
self.numTerms = 1 + max([-1] + self.id2word.keys())
self.docs_processed = 0
self.projection = Projection(self.numTerms, self.numTopics)
if not distributed:
logger.info("using serial LSI version on this node")
self.dispatcher = None
else:
if not onepass:
raise NotImplementedError(
"distributed stochastic LSA not implemented yet; "
"run either distributed one-pass, or serial randomized.")
try:
import Pyro
ns = Pyro.naming.locateNS()
dispatcher = Pyro.core.Proxy(
'PYRONAME:gensim.lsi_dispatcher@%s' % ns._pyroUri.location)
dispatcher._pyroOneway.add("exit")
logger.debug("looking for dispatcher at %s" %
str(dispatcher._pyroUri))
dispatcher.initialize(id2word=self.id2word,
numTopics=numTopics,
chunks=chunks,
decay=decay,
distributed=False,
onepass=onepass)
self.dispatcher = dispatcher
logger.info("using distributed version with %i workers" %
len(dispatcher.getworkers()))
except Exception, err:
# distributed version was specifically requested, so this is an error state
logger.error("failed to initialize distributed LSI (%s)" % err)
raise RuntimeError(
"failed to initialize distributed LSI (%s)" % err)
def stochasticSvd(corpus, rank, num_terms=None, chunks=20000, extra_dims=None, dtype=numpy.float64, eps=1e-6):
"""
Return U, S -- the left singular vectors and the singular values of the streamed
input corpus.
This may actually return less than the requested number of top `rank` factors,
in case the input is of lower rank. Also note that the decomposition is unique
up the the sign of the left singular vectors (columns of U).
This is a streamed, two-pass algorithm, without power-iterations. In case you can
only afford a single pass over the input corpus, set `onepass=True` in LsiModel and
avoid using this algorithm.
The decomposition algorithm is based on
**Halko, Martinsson, Tropp. Finding structure with randomness, 2009.**
"""
rank = int(rank)
if extra_dims is None:
samples = max(10, 2 * rank) # use more samples than requested factors, to improve accuracy
else:
samples = rank + int(extra_dims)
if num_terms is None:
logger.warning("number of terms not provided; will scan the corpus (ONE EXTRA PASS, MAY BE SLOW) to determine it")
num_terms = len(utils.dictFromCorpus(corpus))
logger.info("found %i terms" % num_terms)
else:
num_terms = int(num_terms)
eps = max(float(eps), 1e-9) # must ignore near-zero eigenvalues (probably numerical error); the associated eigenvectors are typically unstable/garbage
# first pass: construct the orthonormal action matrix Q = orth(Y) = orth(A * O)
# build Y in blocks of `chunks` documents (much faster than going one-by-one
# and more memory friendly than processing all documents at once)
y = numpy.zeros(dtype = dtype, shape = (num_terms, samples))
logger.info("1st pass: constructing %s action matrix" % str(y.shape))
chunker = itertools.groupby(enumerate(corpus), key = lambda (docno, doc): docno / chunks)
for chunk_no, (key, group) in enumerate(chunker):
logger.info('PROGRESS: at document #%i' % (chunk_no * chunks))
# construct the chunk as a sparse matrix, to minimize memory overhead
# definitely avoid materializing it as a dense (num_terms x chunks) matrix!
chunk = matutils.corpus2csc((doc for _, doc in group), num_terms=num_terms, dtype=dtype) # documents = columns of sparse CSC
m, n = chunk.shape
assert m == num_terms
assert n <= chunks # the very last chunk of A may be smaller
logger.debug("multiplying chunk * gauss")
o = numpy.random.normal(0.0, 1.0, (n, samples)).astype(dtype) # draw a random gaussian matrix
sparsetools.csc_matvecs(num_terms, n, samples, chunk.indptr, # y = y + chunk * o
chunk.indices, chunk.data, o.ravel(), y.ravel())
del chunk, o
logger.info("orthonormalizing %s action matrix" % str(y.shape))
q, r = numpy.linalg.qr(y) # orthonormalize the range
del y # Y not needed anymore, free up mem
samples = clipSpectrum(numpy.diag(r), samples, discard = eps)
qt = q[:, :samples].T.copy() # discard bogus columns, in case Y was rank-deficient
del q
# second pass: construct the covariance matrix X = B * B.T, where B = Q.T * A
# again, construct X incrementally, in chunks of `chunks` documents from the streaming
# input corpus A, to avoid using O(number of documents) memory
x = numpy.zeros(shape = (samples, samples), dtype = dtype)
logger.info("2nd pass: constructing %s covariance matrix" % str(x.shape))
chunker = itertools.groupby(enumerate(corpus), key = lambda (docno, doc): docno / chunks)
for chunk_no, (key, group) in enumerate(chunker):
logger.info('PROGRESS: at document #%i' % (chunk_no * chunks))
chunk = matutils.corpus2csc((doc for _, doc in group), num_terms=num_terms, dtype=dtype)
b = qt * chunk # dense * sparse matrix multiply
x += numpy.dot(b, b.T) # TODO should call the BLAS routine SYRK, but there is no SYRK wrapper in scipy :(
del chunk, b
# now we're ready to compute decomposition of the small matrix X
logger.info("computing decomposition of the %s covariance matrix" % str(x.shape))
u, s, vt = numpy.linalg.svd(x) # could use linalg.eigh, but who cares... and svd returns the factors already sorted :)
keep = clipSpectrum(s, rank, discard = eps)
logger.info("computing the final decomposition")
s = numpy.sqrt(s[:keep]) # sqrt to go back from singular values of X to singular values of B = singular values of the corpus
u = numpy.dot(qt.T, u[:, :keep]) # go back from left singular vectors of B to left singular vectors of the corpus
return u.astype(dtype), s.astype(dtype)
def stochasticSvd(corpus,
rank,
num_terms=None,
chunks=20000,
extra_dims=None,
dtype=numpy.float64,
eps=1e-6):
"""
Return U, S -- the left singular vectors and the singular values of the streamed
input corpus.
This may actually return less than the requested number of top `rank` factors,
in case the input is of lower rank. Also note that the decomposition is unique
up the the sign of the left singular vectors (columns of U).
This is a streamed, two-pass algorithm, without power-iterations. In case you can
only afford a single pass over the input corpus, set `onepass=True` in LsiModel and
avoid using this algorithm.
The decomposition algorithm is based on
**Halko, Martinsson, Tropp. Finding structure with randomness, 2009.**
"""
rank = int(rank)
if extra_dims is None:
samples = max(
10, 2 * rank
) # use more samples than requested factors, to improve accuracy
else:
samples = rank + int(extra_dims)
if num_terms is None:
logger.warning(
"number of terms not provided; will scan the corpus (ONE EXTRA PASS, MAY BE SLOW) to determine it"
)
num_terms = len(utils.dictFromCorpus(corpus))
logger.info("found %i terms" % num_terms)
else:
num_terms = int(num_terms)
eps = max(
float(eps), 1e-9
) # must ignore near-zero eigenvalues (probably numerical error); the associated eigenvectors are typically unstable/garbage
# first pass: construct the orthonormal action matrix Q = orth(Y) = orth(A * O)
# build Y in blocks of `chunks` documents (much faster than going one-by-one
# and more memory friendly than processing all documents at once)
y = numpy.zeros(dtype=dtype, shape=(num_terms, samples))
logger.info("1st pass: constructing %s action matrix" % str(y.shape))
chunker = itertools.groupby(enumerate(corpus),
key=lambda (docno, doc): docno / chunks)
for chunk_no, (key, group) in enumerate(chunker):
logger.info('PROGRESS: at document #%i' % (chunk_no * chunks))
# construct the chunk as a sparse matrix, to minimize memory overhead
# definitely avoid materializing it as a dense (num_terms x chunks) matrix!
chunk = matutils.corpus2csc(
(doc for _, doc in group), num_terms=num_terms,
dtype=dtype) # documents = columns of sparse CSC
m, n = chunk.shape
assert m == num_terms
assert n <= chunks # the very last chunk of A may be smaller
logger.debug("multiplying chunk * gauss")
o = numpy.random.normal(0.0, 1.0, (n, samples)).astype(
dtype) # draw a random gaussian matrix
sparsetools.csc_matvecs(
num_terms,
n,
samples,
chunk.indptr, # y = y + chunk * o
chunk.indices,
chunk.data,
o.ravel(),
y.ravel())
del chunk, o
logger.info("orthonormalizing %s action matrix" % str(y.shape))
q, r = numpy.linalg.qr(y) # orthonormalize the range
del y # Y not needed anymore, free up mem
samples = clipSpectrum(numpy.diag(r), samples, discard=eps)
qt = q[:, :samples].T.copy(
) # discard bogus columns, in case Y was rank-deficient
del q
# second pass: construct the covariance matrix X = B * B.T, where B = Q.T * A
# again, construct X incrementally, in chunks of `chunks` documents from the streaming
# input corpus A, to avoid using O(number of documents) memory
x = numpy.zeros(shape=(samples, samples), dtype=dtype)
logger.info("2nd pass: constructing %s covariance matrix" % str(x.shape))
chunker = itertools.groupby(enumerate(corpus),
key=lambda (docno, doc): docno / chunks)
for chunk_no, (key, group) in enumerate(chunker):
logger.info('PROGRESS: at document #%i' % (chunk_no * chunks))
chunk = matutils.corpus2csc((doc for _, doc in group),
num_terms=num_terms,
dtype=dtype)
b = qt * chunk # dense * sparse matrix multiply
x += numpy.dot(
b, b.T
) # TODO should call the BLAS routine SYRK, but there is no SYRK wrapper in scipy :(
del chunk, b
# now we're ready to compute decomposition of the small matrix X
logger.info("computing decomposition of the %s covariance matrix" %
str(x.shape))
u, s, vt = numpy.linalg.svd(
x
) # could use linalg.eigh, but who cares... and svd returns the factors already sorted :)
keep = clipSpectrum(s, rank, discard=eps)
logger.info("computing the final decomposition")
s = numpy.sqrt(
s[:keep]
) # sqrt to go back from singular values of X to singular values of B = singular values of the corpus
u = numpy.dot(
qt.T, u[:, :keep]
) # go back from left singular vectors of B to left singular vectors of the corpus
return u.astype(dtype), s.astype(dtype)
