Repository navigation
Expand file tree
/
Copy pathmatutils.py
More file actions
623 lines (509 loc) · 22.9 KB
/
Copy pathmatutils.py
File metadata and controls
623 lines (509 loc) · 22.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
#!/usr/bin/env python
# -*- coding: utf-8 -*-
#
# Copyright (C) 2011 Radim Rehurek <radimrehurek@seznam.cz>
# Licensed under the GNU LGPL v2.1 - http://www.gnu.org/licenses/lgpl.html
"""
This module contains math helper functions.
"""
from __future__ import with_statement
import logging
import math
import os
import itertools
import numpy
import scipy.sparse
import scipy.linalg
from scipy.linalg.lapack import get_lapack_funcs, find_best_lapack_type
# scipy is not a stable package yet, locations change, so try to work
# around differences (currently only concerns location of 'triu' in scipy 0.7 vs. 0.8)
try:
from scipy.linalg.basic import triu
except ImportError:
from scipy.linalg.special_matrices import triu
try:
from numpy import triu_indices
except ImportError:
# numpy < 1.4
def triu_indices(n, k=0):
m = numpy.ones((n, n), int)
a = triu(m, k)
return numpy.where(a != 0)
blas = lambda name, ndarray: scipy.linalg.get_blas_funcs((name,), (ndarray,))[0]
try:
# with bottleneck installed, we can use faster partial sorting
import bottleneck
def argsort(x, topn=None):
"""Return indices of the `topn` greatest elements in numpy array `x`, in order."""
if topn is None:
topn = x.size
if topn <= 0:
return []
if topn >= x.size:
return numpy.argsort(x)[::-1]
biggest = bottleneck.argpartsort(x, x.size - topn)[-topn:]
# the indices in `biggest` are not sorted by magnitude => sort & return
return biggest.take(numpy.argsort(x.take(biggest))[::-1])
except ImportError:
# no bottleneck => fall back to numpy
def argsort(x, topn=None):
if topn is None:
topn = x.size
return numpy.argsort(x)[::-1][:topn]
logger = logging.getLogger("gensim.matutils")
def corpus2csc(corpus, num_terms=None, dtype=numpy.float64, num_docs=None, num_nnz=None, printprogress=0):
"""
Convert corpus into a sparse matrix, in scipy.sparse.csc_matrix format,
with documents as columns.
If the number of terms, documents and non-zero elements is known, you can pass
them here as parameters and a more memory efficient code path will be taken.
"""
try:
# if the input corpus has the `num_nnz`, `num_docs` and `num_terms` attributes
# (as is the case with MmCorpus for example), we can use a more efficient code path
if num_terms is None:
num_terms = corpus.num_terms
if num_docs is None:
num_docs = corpus.num_docs
if num_nnz is None:
num_nnz = corpus.num_nnz
except AttributeError, e:
pass # not a MmCorpus...
if printprogress:
logger.info("creating sparse matrix from corpus")
if num_terms is not None and num_docs is not None and num_nnz is not None:
# faster and much more memory-friendly version of creating the sparse csc
posnow, indptr = 0, [0]
indices = numpy.empty((num_nnz,), dtype=numpy.int32) # HACK assume feature ids fit in 32bit integer
data = numpy.empty((num_nnz,), dtype=dtype)
for docno, doc in enumerate(corpus):
if printprogress and docno % printprogress == 0:
logger.info("PROGRESS: at document #%i/%i" % (docno, num_docs))
posnext = posnow + len(doc)
indices[posnow: posnext] = [feature_id for feature_id, _ in doc]
data[posnow: posnext] = [feature_weight for _, feature_weight in doc]
indptr.append(posnext)
posnow = posnext
assert posnow == num_nnz, "mismatch between supplied and computed number of non-zeros"
result = scipy.sparse.csc_matrix((data, indices, indptr), shape=(num_terms, num_docs), dtype=dtype)
else:
# slower version; determine the sparse matrix parameters during iteration
num_nnz, data, indices, indptr = 0, [], [], [0]
for docno, doc in enumerate(corpus):
if printprogress and docno % printprogress == 0:
logger.info("PROGRESS: at document #%i" % (docno))
indices.extend([feature_id for feature_id, _ in doc])
data.extend([feature_weight for _, feature_weight in doc])
num_nnz += len(doc)
indptr.append(num_nnz)
if num_terms is None:
num_terms = max(indices) + 1 if indices else 0
num_docs = len(indptr) - 1
# now num_docs, num_terms and num_nnz contain the correct values
data = numpy.asarray(data, dtype=dtype)
indices = numpy.asarray(indices)
result = scipy.sparse.csc_matrix((data, indices, indptr), shape=(num_terms, num_docs), dtype=dtype)
return result
def pad(mat, padrow, padcol):
"""
Add additional rows/columns to a numpy.matrix `mat`. The new rows/columns
will be initialized with zeros.
"""
if padrow < 0:
padrow = 0
if padcol < 0:
padcol = 0
rows, cols = mat.shape
return numpy.bmat([[mat, numpy.matrix(numpy.zeros((rows, padcol)))],
[numpy.matrix(numpy.zeros((padrow, cols + padcol)))]])
def zeros_aligned(shape, dtype, order='C', align=128):
"""Like `numpy.zeros()`, but the array will be aligned at `align` byte boundary."""
nbytes = numpy.prod(shape) * numpy.dtype(dtype).itemsize
buffer = numpy.zeros(nbytes + align, dtype=numpy.uint8)
start_index = -buffer.ctypes.data % align
return buffer[start_index : start_index + nbytes].view(dtype).reshape(shape, order=order)
def ismatrix(m):
return isinstance(m, numpy.ndarray) and m.ndim == 2 or scipy.sparse.issparse(m)
def any2sparse(vec, eps=1e-9):
"""Convert a numpy/scipy vector into gensim format (list of 2-tuples)."""
if isinstance(vec, numpy.ndarray):
return dense2vec(vec, eps)
if scipy.sparse.issparse(vec):
return scipy2sparse(vec, eps)
return [(int(fid), float(fw)) for fid, fw in vec if numpy.abs(fw) > eps]
def scipy2sparse(vec, eps=1e-9):
"""Convert a scipy.sparse vector to gensim format (list of 2-tuples)."""
vec = vec.tocsr()
assert vec.shape[0] == 1
return [(int(pos), float(val)) for pos, val in zip(vec.indices, vec.data) if numpy.abs(val) > eps]
class Scipy2Corpus(object):
def __init__(self, vecs):
"""Convert a sequence of dense/sparse vector to a gensim corpus object."""
self.vecs = vecs
def __iter__(self):
for vec in self.vecs:
if isinstance(vec, numpy.ndarray):
yield full2sparse(vec)
else:
yield scipy2sparse(vec)
8000
def __len__(self):
return len(self.vecs)
def sparse2full(doc, length):
"""
Convert a document in sparse corpus format (sequence of 2-tuples) into a dense
numpy array (of size `length`).
"""
result = numpy.zeros(length, dtype=numpy.float32) # fill with zeroes (default value)
doc = dict(doc)
result[doc.keys()] = doc.values() # overwrite some of the zeroes with explicit values
return result
def full2sparse(vec, eps=1e-9):
"""
Convert a dense numpy array into the sparse corpus format (sequence of 2-tuples).
Values of magnitude < `eps` are treated as zero (ignored).
"""
vec = numpy.asarray(vec, dtype=float)
nnz = numpy.nonzero(abs(vec) > eps)[0]
return zip(nnz, vec.take(nnz))
dense2vec = full2sparse
def full2sparse_clipped(vec, topn, eps=1e-9):
"""
Like `full2sparse`, but only return the `topn` greatest elements (not all).
"""
# use numpy.argsort and only form tuples that are actually returned.
# this is about 40x faster than explicitly forming all 2-tuples to run sort() or heapq.nlargest() on.
if topn <= 0:
return []
vec = numpy.asarray(vec, dtype=float)
nnz = numpy.nonzero(abs(vec) > eps)[0]
biggest = nnz.take(argsort(vec.take(nnz), topn))
return zip(biggest, vec.take(biggest))
def corpus2dense(corpus, num_terms):
"""
Convert corpus into a dense numpy array (documents will be columns).
"""
return numpy.column_stack(sparse2full(doc, num_terms) for doc in corpus)
class Dense2Corpus(object):
"""
Treat dense numpy array as a sparse gensim corpus.
No data copy is made (changes to the underlying matrix imply changes in the
corpus).
"""
def __init__(self, dense, documents_columns=True):
if documents_columns:
self.dense = dense.T
else:
self.dense = dense
def __iter__(self):
for doc in self.dense:
yield full2sparse(doc.flat)
def __len__(self):
return len(self.dense)
#endclass DenseCorpus
class Sparse2Corpus(object):
"""
Convert a matrix in scipy.sparse format into a streaming gensim corpus.
"""
def __init__(self, sparse, documents_columns=True):
if documents_columns:
self.sparse = sparse.tocsc()
else:
self.sparse = sparse.tocsr().T # make sure shape[1]=number of docs (needed in len())
def __iter__(self):
for indprev, indnow in itertools.izip(self.sparse.indptr, self.sparse.indptr[1:]):
yield zip(self.sparse.indices[indprev:indnow], self.sparse.data[indprev:indnow])
def __len__(self):
return self.sparse.shape[1]
#endclass Sparse2Corpus
def veclen(vec):
if len(vec) == 0:
return 0.0
length = 1.0 * math.sqrt(sum(val**2 for _, val in vec))
assert length > 0.0, "sparse documents must not contain any explicit zero entries"
return length
blas_nrm2 = blas('nrm2', numpy.array([], dtype=float))
blas_scal = blas('scal', numpy.array([], dtype=float))
def unitvec(vec):
"""
Scale a vector to unit length. The only exception is the zero vector, which
is returned back unchanged.
Output will be in the same format as input (i.e., gensim vector=>gensim vector,
or numpy array=>numpy array, scipy.sparse=>scipy.sparse).
"""
if scipy.sparse.issparse(vec): # convert scipy.sparse to standard numpy array
vec = vec.tocsr()
veclen = numpy.sqrt(numpy.sum(vec.data ** 2))
if veclen > 0.0:
return vec / veclen
else:
return vec
if isinstance(vec, numpy.ndarray):
vec = numpy.asarray(vec, dtype=float)
veclen = blas_nrm2(vec)
if veclen > 0.0:
return blas_scal(1.0 / veclen, vec)
else:
return vec
try:
first = iter(vec).next() # is there at least one element?
except:
return vec
if isinstance(first, (tuple, list)) and len(first) == 2: # gensim sparse format?
length = 1.0 * math.sqrt(sum(val ** 2 for _, val in vec))
assert length > 0.0, "sparse documents must not contain any explicit zero entries"
if length != 1.0:
return [(termid, val / length) for termid, val in vec]
else:
return list(vec)
else:
raise ValueError("unknown input type")
def cossim(vec1, vec2):
vec1, vec2 = dict(vec1), dict(vec2)
if not vec1 or not vec2:
return 0.0
vec1len = 1.0 * math.sqrt(sum(val * val for val in vec1.itervalues()))
vec2len = 1.0 * math.sqrt(sum(val * val for val in vec2.itervalues()))
assert vec1len > 0.0 and vec2len > 0.0, "sparse documents must not contain any explicit zero entries"
if len(vec2) < len(vec1):
vec1, vec2 = vec2, vec1 # swap references so that we iterate over the shorter vector
result = sum(value * vec2.get(index, 0.0) for index, value in vec1.iteritems())
result /= vec1len * vec2len # rescale by vector lengths
return result
def qr_destroy(la):
"""
Return QR decomposition of `la[0]`. Content of `la` gets destroyed in the process.
Using this function should be less memory intense than calling `scipy.linalg.qr(la[0])`,
because the memory used in `la[0]` is reclaimed earlier.
"""
a = numpy.asfortranarray(la[0])
del la[0], la # now `a` is the only reference to the input matrix
m, n = a.shape
# perform q, r = QR(a); code hacked out of scipy.linalg.qr
logger.debug("computing QR of %s dense matrix" % str(a.shape))
geqrf, = get_lapack_funcs(('geqrf',), (a,))
qr, tau, work, info = geqrf(a, lwork=-1, overwrite_a=True)
qr, tau, work, info = geqrf(a, lwork=work[0], overwrite_a=True)
del a # free up mem
assert info >= 0
r = triu(qr[:n, :n])
if m < n: # rare case, #features < #topics
qr = qr[:, :m] # retains fortran order
gorgqr, = get_lapack_funcs(('orgqr',), (qr,))
q, work, info = gorgqr(qr, tau, lwork=-1, overwrite_a=True)
q, work, info = gorgqr(qr, tau, lwork=work[0], overwrite_a=True)
assert info >= 0, "qr failed"
assert q.flags.f_contiguous
return q, r
class MmWriter(object):
"""
Store a corpus in Matrix Market format.
Note that the output is written one document at a time, not the whole
matrix at once (unlike scipy.io.mmread). This allows us to process corpora
which are larger than the available RAM.
NOTE: the output file is created in a single pass through the input corpus, so
that the input can be a once-only stream (iterator).
To achieve this, a fake MM header is written first, statistics are collected
during the pass (shape of the matrix, number of non-zeroes), followed by a seek
back to the beginning of the file, rewriting the fake header with proper values.
"""
HEADER_LINE = '%%MatrixMarket matrix coordinate real general\n' # the only supported MM format
def __init__(self, fname):
self.fname = fname
tmp = open(self.fname, 'w') # reset/create the target file
tmp.close()
self.fout = open(self.fname, 'rb+') # open for both reading and writing
self.headers_written = False
def write_headers(self, num_docs, num_terms, num_nnz):
self.fout.write(MmWriter.HEADER_LINE)
if num_nnz < 0:
# we don't know the matrix shape/density yet, so only log a general line
logger.info("saving sparse matrix to %s" % self.fname)
self.fout.write(' ' * 50 + '\n') # 48 digits must be enough for everybody
else:
logger.info("saving sparse %sx%s matrix with %i non-zero entries to %s" %
(num_docs, num_terms, num_nnz, self.fname))
self.fout.write('%s %s %s\n' % (num_docs, num_terms, num_nnz))
self.last_docno = -1
self.headers_written = True
def fake_headers(self, num_docs, num_terms, num_nnz):
stats = '%i %i %i' % (num_docs, num_terms, num_nnz)
if len(stats) > 50:
raise ValueError('Invalid stats: matrix too large!')
self.fout.seek(len(MmWriter.HEADER_LINE))
self.fout.write(stats)
def write_vector(self, docno, vector):
"""
Write a single sparse vector to the file.
Sparse vector is any iterable yielding (field id, field value) pairs.
"""
assert self.headers_written, "must write Matrix Market file headers before writing data!"
assert self.last_docno < docno, "documents %i and %i not in sequential order!" % (self.last_docno, docno)
vector = sorted((i, w) for i, w in vector if abs(w) > 1e-12) # ignore near-zero entries
for termid, weight in vector: # write term ids in sorted order
self.fout.write("%i %i %s\n" % (docno + 1, termid + 1, weight)) # +1 because MM format starts counting from 1
self.last_docno = docno
return (vector[-1][0], len(vector)) if vector else (-1, 0)
@staticmethod
def write_corpus(fname, corpus, progress_cnt=1000, index=False, num_terms=None):
"""
Save the vector space representation of an entire corpus to disk.
Note that the documents are processed one at a time, so the whole corpus
is allowed to be larger than the available RAM.
"""
mw = MmWriter(fname)
# write empty headers to the file (with enough space to be overwritten later)
mw.write_headers(-1, -1, -1) # will print 50 spaces followed by newline on the stats line
# calculate necessary header info (nnz elements, num terms, num docs) while writing out vectors
_num_terms, num_nnz = 0, 0
docno, poslast = -1, -1
offsets = []
for docno, bow in enumerate(corpus):
if docno % progress_cnt == 0:
logger.info("PROGRESS: saving document #%i" % docno)
if index:
posnow = mw.fout.tell()
if posnow == poslast:
offsets[-1] = -1
offsets.append(posnow)
poslast = posnow
max_id, veclen = mw.write_vector(docno, bow)
_num_terms = max(_num_terms, 1 + max_id)
num_nnz += veclen
num_docs = docno + 1
num_terms = num_terms or _num_terms
if num_docs * num_terms != 0:
logger.info("saved %ix%i matrix, density=%.3f%% (%i/%i)" % (
num_docs, num_terms,
100.0 * num_nnz / (num_docs * num_terms),
num_nnz,
num_docs * num_terms))
# now write proper headers, by seeking and overwriting the spaces written earlier
mw.fake_headers(num_docs, num_terms, num_nnz)
mw.close()
if index:
return offsets
def __del__(self):
"""
Automatic destructor which closes the underlying file.
There must be no circular references contained in the object for __del__
to work! Closing the file explicitly via the close() method is preferred
and safer.
"""
self.close() # does nothing if called twice (on an already closed file), so no worries
def close(self):
logger.debug("closing %s" % self.fname)
self.fout.close()
#endclass MmWriter
class MmReader(object):
"""
Wrap a term-document matrix on disk (in matrix-market format), and present it
as an object which supports iteration over the rows (~documents).
Note that the file is read into memory one document at a time, not the whole
matrix at once (unlike scipy.io.mmread). This allows us to process corpora
which are larger than the available RAM.
"""
def __init__(self, input, transposed=True):
"""
Initialize the matrix reader.
The `input` refers to a file on local filesystem, which is expected to
be in the sparse (coordinate) Matrix Market format. Documents are assumed
to be rows of the matrix (and document features are columns).
`input` is either a string (file path) or a file-like object that supports
`seek()` (e.g. gzip.GzipFile, bz2.BZ2File).
"""
logger.info("initializing corpus reader from %s" % input)
self.input, self.transposed = input, transposed
if isinstance(input, basestring):
input = open(input)
header = input.next().strip()
if not header.lower().startswith('%%matrixmarket matrix coordinate real general'):
raise ValueError("File %s not in Matrix Market format with coordinate real general; instead found: \n%s" %
(self.input, header))
self.num_docs = self.num_terms = self.num_nnz = 0
for lineno, line in enumerate(input):
if not line.startswith('%'):
self.num_docs, self.num_terms, self.num_nnz = map(int, line.split())
if not self.transposed:
self.num_docs, self.num_terms = self.num_terms, self.num_docs
break
logger.info("accepted corpus with %i documents, %i features, %i non-zero entries" %
(self.num_docs, self.num_terms, self.num_nnz))
def __len__(self):
return self.num_docs
def __str__(self):
return ("MmCorpus(%i documents, %i features, %i non-zero entries)" %
(self.num_docs, self.num_terms, self.num_nnz))
def skip_headers(self, input_file):
"""
Skip file headers that appear before the first document.
"""
for line in input_file:
if line.startswith('%'):
continue
break
def __iter__(self):
"""
Iteratively yield vectors from the underlying file, in the format (row_no, vector),
where vector is a list of (col_no, value) 2-tuples.
Note that the total number of vectors returned is always equal to the
number of rows specified in the header; empty documents are inserted and
yielded where appropriate, even if they are not explicitly stored in the
Matrix Market file.
"""
if isinstance(self.input, basestring):
fin = open(self.input)
else:
fin = self.input
fin.seek(0)
self.skip_headers(fin)
previd = -1
for line in fin:
docid, termid, val = line.split()
if not self.transposed:
termid, docid = docid, termid
docid, termid, val = int(docid) - 1, int(termid) - 1, float(val) # -1 because matrix market indexes are 1-based => convert to 0-based
assert previd <= docid, "matrix columns must come in ascending order"
if docid != previd:
# change of document: return the document read so far (its id is prevId)
if previd >= 0:
yield previd, document
# return implicit (empty) documents between previous id and new id
# too, to keep consistent document numbering and corpus length
for previd in xrange(previd + 1, docid):
yield previd, []
# from now on start adding fields to a new document, with a new id
previd = docid
document = []
document.append((termid, val,)) # add another field to the current document
# handle the last document, as a special case
if previd >= 0:
yield previd, document
# return empty documents between the last explicit document and the number
# of documents as specified in the header
for previd in xrange(previd + 1, self.num_docs):
yield previd, []
def docbyoffset(self, offset):
"""Return document at file offset `offset` (in bytes)"""
# empty documents are not stored explicitly in MM format, so the index marks
# them with a special offset, -1.
if offset == -1:
return []
if isinstance(self.input, basestring):
fin = open(self.input)
else:
fin = self.input
fin.seek(offset) # works for gzip/bz2 input, too
previd, document = -1, []
for line in fin:
docid, termid, val = line.split()
if not self.transposed:
termid, docid = docid, termid
docid, termid, val = int(docid) - 1, int(termid) - 1, float(val) # -1 because matrix market indexes are 1-based => convert to 0-based
assert previd <= docid, "matrix columns must come in ascending order"
if docid != previd:
if previd >= 0:
return document
previd = docid
document.append((termid, val,)) # add another field to the current document
return document
#endclass MmReader