def main():
#Import shared parameters and verify output dir exists
if not os.path.exists(temp_dir):
raise IOError
#==============================================================================
# Read in link data and update content files accordingly
#==============================================================================
#Get list of files containing link info and chop it up
linkfiles = glob.glob(temp_dir + '*' + extensions['links'])
linkchunks = listchopper(linkfiles)
linkfiles_read = 0
for linkchunk in linkchunks:
#Hash mapping each article to a set of articles linking to it
linkhash = {}
for filename in linkchunk:
with open(filename, 'r') as f:
newstuff = shared.load(f)
#Add link info to linkhash
for target, sources in newstuff.iteritems():
try:
linkhash[target].update(set(sources))
except KeyError:
linkhash[target] = set(sources)
#Log status
linkfiles_read += 1
log("Read " + filename + " - " +
str(100 * linkfiles_read / len(linkfiles))[:4] +
" % of link data.")
log("Chunk finished - updating content files")
#Update concept with newly read link data
contentfiles = glob.glob(temp_dir + '*' + extensions['content'])
contentfiles_read = 0
for filename in contentfiles:
#Read file. Content is like {'article title' : {'text' : blah}}
with open(filename, 'r') as f:
content = shared.load(f)
#Search linkhash for links going TO concept
for concept in content.keys():
try:
sources = linkhash[concept]
except KeyError:
sources = set([]) #Missing key => zero incoming links
#Update link info for concept
try:
content[concept]['links_in'] = set(
content[concept]['links_in'])
content[concept]['links_in'].update(sources)
except KeyError:
content[concept]['links_in'] = sources
#Save updated content
with open(filename, 'w') as f:
shared.dump(content, f)
contentfiles_read += 1
if contentfiles_read % 100 == 0:
log("Fixed " +
str(100 * contentfiles_read / len(contentfiles))[:4] +
"% of content files")
pass #Proceed to next link chunk
#==============================================================================
# Finished link processing
# Remove unworthy concepts and combine concept/word lists.
#==============================================================================
#What, you think memory grows on trees?
del linkhash
gc.collect()
#Set of all approved concepts
concept_list = set([])
#Purge inferior concepts (with insufficient incoming links)
for filename in contentfiles:
#Read in content file
with open(filename, 'r') as f:
content = shared.load(f)
for concept in content.keys():
entry = content[concept]
if 'links_in' in entry and len(entry['links_in']) >= min_links_in:
concept_list.add(concept)
else:
del content[concept]
with open(filename, 'w') as f:
shared.dump(content, f)
log("Links done - saving index files")
#Make sure output dir exists
if not os.path.exists(matrix_dir):
os.makedirs(matrix_dir)
#Generate and save a concept index map. Structure: {concept : index}
concept_indices = {n: m for m, n in enumerate(concept_list)}
with open(matrix_dir + 'concept2index.ind', 'w') as f:
shared.dump(concept_indices, f)
#Read in all wordlists and combine them.
words = set([])
for filename in glob.glob(temp_dir + '*' + extensions['words']):
with open(filename, 'r') as f:
words.update(shared.load(f))
#Generate and save a word index map. Structure: {word : index}
word_indices = {n: m for m, n in enumerate(words)}
with open(matrix_dir + 'word2index.ind', 'w') as f:
shared.dump(word_indices, f)
log("Wrapping up.")
#Attempt to notify that job is done
if shared.notify:
try:
shared.pushme(sys.argv[0] + ' completed.')
except:
log("Job's done. Push failed.")
logfile.close()
def main():
#Cleanup
for f in glob.glob(matrix_dir + '/*' + extensions['matrix']):
os.remove(f)
#Set pruning parameters
window_size = shared.window_size
cutoff = shared.cutoff
#Read in dicts mapping words and concepts to their respective indices
log("Reading in word/index data")
word2index = shared.load(open(matrix_dir + 'word2index.ind', 'r'))
concept2index = shared.load(open(matrix_dir + 'concept2index.ind', 'r'))
log("...Done!")
#==============================================================================
# Construct count matrix in small chunks
#==============================================================================
#Count words and concepts
n_words = len(word2index)
n_concepts = len(concept2index)
#Determine matrix dimensions
matrix_shape = (n_words, n_concepts)
#Allocate sparse matrix. Dict-of-keys should be faster for iterative
#construction. Convert to csr for fast row operations later.
mtx = sps.dok_matrix(matrix_shape, dtype=datatype)
def matrix_chopper(matrix, dim):
'''Generator to split a huge matrix into small submatrices, which can
then be stored in individual files.
This is handy both when constructing the matrix (building the whole
matrix without saving to files in the process takes about 50 gigs RAM),
and when applying it, as this allows one to load only the submatrix
relevant to a given word.'''
ind = 0
counter = 0
rows = matrix.get_shape()[0]
while ind < rows:
end = min(ind + dim, rows)
#Return pair of submatrix number and the submatrix itself
yield counter, sps.vstack([matrix.getrow(i)\
for i in xrange(ind, end)], format = 'csr')
counter += 1
ind += dim
def writeout():
'''Saves the matrix as small submatrrices in separate files.'''
for n, submatrix in matrix_chopper(mtx, row_chunk_size):
filename = matrix_dir + str(n) + extensions['matrix']
#Update submatrix if it's already partially calculated
log("Writing out chunk %s" % n)
try:
with open(filename, 'r') as f:
submatrix = submatrix + shared.mload(f)
#
except IOError:
pass #File doesn't exist yet, so no need to change mtx
#Dump the submatrix to file
with open(filename, 'w') as f:
shared.mdump(submatrix, f)
return None
log("Constructing matrix.")
filelist = glob.glob(temp_dir + '*' + extensions['content'])
files_read = 0
for filename in filelist:
with open(filename, 'r') as f:
content = shared.load(f)
#Loop over concepts (columns) as so we don't waste time with rare words
for concept, entry, in content.iteritems():
#This is the column index (concept w. index j)
j = concept2index[concept]
#Convert concept 'countmap' like so: {word : n}
wordmap = Counter(entry['text'].split()).iteritems()
#Add them all to the matrix
for word, count in wordmap:
#Find row index of the current word
i = word2index[word]
#Add the number of times word i occurs in concept j to the matrix
mtx[i, j] = count
#
#Update file count
files_read += 1
log("Processed content file no. %s of %s - %s" %
(files_read, len(filelist) - 1, percentof(files_read,
len(filelist))))
if files_read % column_chunk_size == 0:
mtx = mtx.tocsr()
writeout()
mtx = sps.dok_matrix(matrix_shape)
#
#Convert matrix to CSR format and write to files.
mtx = mtx.tocsr()
writeout()
#==============================================================================
# Count matrix/matrices constructed - computing TF-IDF
#==============================================================================
log("Done - computing TF-IDF")
#Grap list of matrix files (containing the submatrices from before)
matrixfiles = glob.glob(matrix_dir + "*" + extensions['matrix'])
words_processed = 0 #for logging purposes
for filename in matrixfiles:
with open(filename, 'r') as f:
mtx = shared.mload(f)
#Number of words in a submatrix
n_rows = mtx.get_shape()[0]
for w in xrange(n_rows):
#Grap non-zero elements from the row corresonding to word w
row = mtx.data[mtx.indptr[w]:mtx.indptr[w + 1]]
if len(row) == 0:
continue
#Make a vectorized function to convert a full row to TF-IDF
f = np.vectorize(lambda m_ij: (1 + np.log(m_ij)) * np.log(
n_concepts / len(row)))
#Map all elements to TF-IDF and update matrix
row = f(row)
#Normalize the row
assert row.dtype.kind == 'f' #Non floats round to zero w/o warning
normfact = 1.0 / np.linalg.norm(row)
row *= normfact
#Start inverted index pruning
if prune:
#Number of documents containing w
n_docs = len(row)
#Don't prune if the windows exceeds the array bounds (duh)
if window_size < n_docs:
#Obtain list of indices such that row[index] is sorted
indices = np.argsort(row)[::-1]
#Generate a sorted row
sorted_row = [row[index] for index in indices]
#Go through sorted row and truncate when pruning condition is met
for i in xrange(n_docs - window_size):
if sorted_row[i +
window_size] >= cutoff * sorted_row[i]:
#Truncate, i.e. set the remaining entries to zero
sorted_row[i:] = [0] * (n_docs - i)
break
else:
pass
#Unsort to original positions
for i in xrange(n_docs):
row[indices[i]] = sorted_row[i]
#Update matrix
mtx.data[mtx.indptr[w]:mtx.indptr[w + 1]] = row
#Log it
words_processed += 1
if words_processed % 10**3 == 0:
log("Processing word %s of %s - %s" %
(words_processed, n_words,
percentof(words_processed, n_words)))
#Keep it sparse - no need to store zeroes
mtx.eliminate_zeros()
with open(filename, 'w') as f:
shared.mdump(mtx, f)
log("Done!")
#Notify that the job is done
if shared.notify:
try:
shared.pushme(sys.argv[0] + ' completed.')
except:
log("Job's done. Push failed.")
logfile.close()
return None
self.flush_output_buffer()
return None
if __name__ == "__main__":
if len(sys.argv) == 2:
file_to_parse = sys.argv[1]
else:
file_to_parse = DEFAULT_FILENAME
#Create and configure content handler
test = WikiHandler()
test.verbose = True
#Create a parser and set handler
ATST = SAX.make_parser()
ATST.setContentHandler(test)
#Let the parser walk the file
log("Parsing started...")
ATST.parse(file_to_parse)
log("...Parsing done!")
#Attempt to send notification that job is done
if shared.notify:
try:
shared.pushme(sys.argv[0]+' completed.')
except:
log("Job's done. Push failed.")
logfile.close()
def main():
#Import shared parameters and verify output dir exists
if not os.path.exists(temp_dir):
raise IOError
#==============================================================================
# Read in link data and update content files accordingly
#==============================================================================
#Get list of files containing link info and chop it up
linkfiles = glob.glob(temp_dir + '*'+extensions['links'])
linkchunks = listchopper(linkfiles)
linkfiles_read = 0
for linkchunk in linkchunks:
#Hash mapping each article to a set of articles linking to it
linkhash = {}
for filename in linkchunk:
with open(filename, 'r') as f:
newstuff = shared.load(f)
#Add link info to linkhash
for target, sources in newstuff.iteritems():
try:
linkhash[target].update(set(sources))
except KeyError:
linkhash[target] = set(sources)
#Log status
linkfiles_read += 1
log("Read " + filename + " - " +
str(100*linkfiles_read/len(linkfiles))[:4] + " % of link data.")
log("Chunk finished - updating content files")
#Update concept with newly read link data
contentfiles = glob.glob(temp_dir + '*'+extensions['content'])
contentfiles_read = 0
for filename in contentfiles:
#Read file. Content is like {'article title' : {'text' : blah}}
with open(filename, 'r') as f:
content = shared.load(f)
#Search linkhash for links going TO concept
for concept in content.keys():
try:
sources = linkhash[concept]
except KeyError:
sources = set([]) #Missing key => zero incoming links
#Update link info for concept
try:
content[concept]['links_in'] = set(content[concept]['links_in'])
content[concept]['links_in'].update(sources)
except KeyError:
content[concept]['links_in'] = sources
#Save updated content
with open(filename, 'w') as f:
shared.dump(content, f)
contentfiles_read += 1
if contentfiles_read % 100 == 0:
log("Fixed " + str(100*contentfiles_read/len(contentfiles))[:4]
+ "% of content files")
pass #Proceed to next link chunk
#==============================================================================
# Finished link processing
# Remove unworthy concepts and combine concept/word lists.
#==============================================================================
#What, you think memory grows on trees?
del linkhash
gc.collect()
#Set of all approved concepts
concept_list = set([])
#Purge inferior concepts (with insufficient incoming links)
for filename in contentfiles:
#Read in content file
with open(filename, 'r') as f:
content = shared.load(f)
for concept in content.keys():
entry = content[concept]
if 'links_in' in entry and len(entry['links_in']) >= min_links_in:
concept_list.add(concept)
else:
del content[concept]
with open(filename, 'w') as f:
shared.dump(content, f)
log("Links done - saving index files")
#Make sure output dir exists
if not os.path.exists(matrix_dir):
os.makedirs(matrix_dir)
#Generate and save a concept index map. Structure: {concept : index}
concept_indices = {n: m for m,n in enumerate(concept_list)}
with open(matrix_dir+'concept2index.ind', 'w') as f:
shared.dump(concept_indices, f)
#Read in all wordlists and combine them.
words = set([])
for filename in glob.glob(temp_dir + '*'+extensions['words']):
with open(filename, 'r') as f:
words.update(shared.load(f))
#Generate and save a word index map. Structure: {word : index}
word_indices = {n: m for m,n in enumerate(words)}
with open(matrix_dir+'word2index.ind', 'w') as f:
shared.dump(word_indices, f)
log("Wrapping up.")
#Attempt to notify that job is done
if shared.notify:
try:
shared.pushme(sys.argv[0]+' completed.')
except:
log("Job's done. Push failed.")
logfile.close()
def main():
#Cleanup
for f in glob.glob(matrix_dir + '/*'+extensions['matrix']):
os.remove(f)
#Set pruning parameters
window_size = shared.window_size
cutoff = shared.cutoff
#Read in dicts mapping words and concepts to their respective indices
log("Reading in word/index data")
word2index = shared.load(open(matrix_dir+'word2index.ind', 'r'))
concept2index = shared.load(open(matrix_dir+'concept2index.ind', 'r'))
log("...Done!")
#==============================================================================
# Construct count matrix in small chunks
#==============================================================================
#Count words and concepts
n_words = len(word2index)
n_concepts = len(concept2index)
#Determine matrix dimensions
matrix_shape = (n_words, n_concepts)
#Allocate sparse matrix. Dict-of-keys should be faster for iterative
#construction. Convert to csr for fast row operations later.
mtx = sps.dok_matrix(matrix_shape, dtype = datatype)
def matrix_chopper(matrix, dim):
'''Generator to split a huge matrix into small submatrices, which can
then be stored in individual files.
This is handy both when constructing the matrix (building the whole
matrix without saving to files in the process takes about 50 gigs RAM),
and when applying it, as this allows one to load only the submatrix
relevant to a given word.'''
ind = 0
counter = 0
rows = matrix.get_shape()[0]
while ind < rows:
end = min(ind+dim, rows)
#Return pair of submatrix number and the submatrix itself
yield counter, sps.vstack([matrix.getrow(i)\
for i in xrange(ind, end)], format = 'csr')
counter += 1
ind += dim
def writeout():
'''Saves the matrix as small submatrrices in separate files.'''
for n, submatrix in matrix_chopper(mtx, row_chunk_size):
filename = matrix_dir+str(n)+extensions['matrix']
#Update submatrix if it's already partially calculated
log("Writing out chunk %s" % n)
try:
with open(filename, 'r') as f:
submatrix = submatrix + shared.mload(f)
#
except IOError:
pass #File doesn't exist yet, so no need to change mtx
#Dump the submatrix to file
with open(filename, 'w') as f:
shared.mdump(submatrix, f)
return None
log("Constructing matrix.")
filelist = glob.glob(temp_dir + '*'+extensions['content'])
files_read = 0
for filename in filelist:
with open(filename, 'r') as f:
content = shared.load(f)
#Loop over concepts (columns) as so we don't waste time with rare words
for concept, entry, in content.iteritems():
#This is the column index (concept w. index j)
j = concept2index[concept]
#Convert concept 'countmap' like so: {word : n}
wordmap = Counter(entry['text'].split()).iteritems()
#Add them all to the matrix
for word, count in wordmap:
#Find row index of the current word
i = word2index[word]
#Add the number of times word i occurs in concept j to the matrix
mtx[i,j] = count
#
#Update file count
files_read += 1
log("Processed content file no. %s of %s - %s"
% (files_read, len(filelist)-1, percentof(files_read, len(filelist))))
if files_read % column_chunk_size == 0:
mtx = mtx.tocsr()
writeout()
mtx = sps.dok_matrix(matrix_shape)
#
#Convert matrix to CSR format and write to files.
mtx = mtx.tocsr()
writeout()
#==============================================================================
# Count matrix/matrices constructed - computing TF-IDF
#==============================================================================
log("Done - computing TF-IDF")
#Grap list of matrix files (containing the submatrices from before)
matrixfiles = glob.glob(matrix_dir + "*" + extensions['matrix'])
words_processed = 0 #for logging purposes
for filename in matrixfiles:
with open(filename, 'r') as f:
mtx = shared.mload(f)
#Number of words in a submatrix
n_rows = mtx.get_shape()[0]
for w in xrange(n_rows):
#Grap non-zero elements from the row corresonding to word w
row = mtx.data[mtx.indptr[w] : mtx.indptr[w+1]]
if len(row) == 0:
continue
#Make a vectorized function to convert a full row to TF-IDF
f = np.vectorize(lambda m_ij: (1+np.log(m_ij))*
np.log(n_concepts/len(row)))
#Map all elements to TF-IDF and update matrix
row = f(row)
#Normalize the row
assert row.dtype.kind == 'f' #Non floats round to zero w/o warning
normfact = 1.0/np.linalg.norm(row)
row *= normfact
#Start inverted index pruning
if prune:
#Number of documents containing w
n_docs = len(row)
#Don't prune if the windows exceeds the array bounds (duh)
if window_size < n_docs:
#Obtain list of indices such that row[index] is sorted
indices = np.argsort(row)[::-1]
#Generate a sorted row
sorted_row = [row[index] for index in indices]
#Go through sorted row and truncate when pruning condition is met
for i in xrange(n_docs-window_size):
if sorted_row[i+window_size] >= cutoff*sorted_row[i]:
#Truncate, i.e. set the remaining entries to zero
sorted_row[i:] = [0]*(n_docs-i)
break
else:
pass
#Unsort to original positions
for i in xrange(n_docs):
row[indices[i]] = sorted_row[i]
#Update matrix
mtx.data[mtx.indptr[w] : mtx.indptr[w+1]] = row
#Log it
words_processed += 1
if words_processed % 10**3 == 0:
log("Processing word %s of %s - %s" %
(words_processed, n_words,
percentof(words_processed, n_words)))
#Keep it sparse - no need to store zeroes
mtx.eliminate_zeros()
with open(filename, 'w') as f:
shared.mdump(mtx, f)
log("Done!")
#Notify that the job is done
if shared.notify:
try:
shared.pushme(sys.argv[0]+' completed.')
except:
log("Job's done. Push failed.")
logfile.close()
return None
return None
if __name__ == "__main__":
if len(sys.argv) == 2:
file_to_parse = sys.argv[1]
else:
file_to_parse = DEFAULT_FILENAME
#Create and configure content handler
test = WikiHandler()
test.verbose = True
#Create a parser and set handler
ATST = SAX.make_parser()
ATST.setContentHandler(test)
#Let the parser walk the file
log("Parsing started...")
ATST.parse(file_to_parse)
log("...Parsing done!")
#Attempt to send notification that job is done
if shared.notify:
try:
shared.pushme(sys.argv[0] + ' completed.')
except:
log("Job's done. Push failed.")
logfile.close()
