def train(folder):
'''
Function to train autoencoder.
'''
t = time.time()
lr_decay = 0.95
pa = ProbArray()
# Frequency to filter out low freq words
freq = {}
filepaths = list(map(lambda x: folder + "/" + x, os.listdir(folder)))
rgx = re.compile("([\w][\w']*\w)")
# Another iterator to count frequency of words
print("Pre-processing (clearning garbage words)")
for filepath in filepaths:
text = open(filepath).read().lower()
tokens = re.findall(rgx, text)
N = len(tokens)
for i in range(0, N):
if tokens[i] in freq:
freq[tokens[i]] += 1
else:
freq[tokens[i]] = 1
# Sort the frequencies storing in (freq, token) pairs and prune words with freq < min_count
tokenFreq = sorted(freq.items(), key=lambda x: x[1])
garbageWords = []
for item in tokenFreq:
if item[1] < min_count:
garbageWords.append(item[0])
print("Generating co-occurence matrix")
doc_text = ""
for filepath in filepaths:
text = open(filepath).read().lower()
words = re.findall(rgx, text)
N = len(words)
temp = [' '] * (N + contextSize)
temp[contextSize:(contextSize + N)] = words
words = temp
for i in range(contextSize, (contextSize + N)):
# Filter out garbage words"
#if words[i] not in garbageWords:
# Include context size specified by user
for j in range(i - contextSize, i):
if words[i] != ' ' and words[j] != ' ':
pa.addcontext(words[j], words[i])
pa.addcontext(words[i], words[j])
print("Co-occurence matrix generated")
print("Starting training")
tm = TrainModel(maxnum, newdims, lr)
pa.freeze()
tm.trainonone(pa, ntimes)
#lr /=float(1+k*lr_decay)
wordembeddings = {}
for numwordvec in pa.getallwordvecs():
(num, wordvec) = numwordvec
word = pa.wordnumrelation.getWord(num)
if backend == 'tf':
embedding = tm.getoutput(wordvec, './embedding.chk')
else:
embedding = tm.getoutput(wordvec)
wordembeddings[word] = embedding
print("Training proces done, dumping embedding into persistant storage!")
with open(r'./embeddings.pickle', "wb") as outfile:
pickle.dump(wordembeddings, outfile)
print("Training completed! Embedding done.")
print("time is %f" % (time.time() - t))
def train(folder,contextSize=5,min_count=100, newdims=100, ntimes=2,
maxnum=10000, lr=0.4):
'''
Function to train autoencoder.
'''
t = time.time()
lr_decay = 0.95
pa = ProbArray()
# Frequency to filter out low freq words
freq = {}
filepaths = map(lambda x: folder + "/" + x,os.listdir(folder))
rgx = re.compile("([\w][\w']*\w)")
# Another iterator to count frequency of words
print "Pre-processing (clearning garbage words)"
for filepath in filepaths:
text = open(filepath).read().lower()
tokens = re.findall(rgx, text)
N = len(tokens)
for i in xrange(0,N):
if tokens[i] in freq:
freq[tokens[i]] += 1
else:
freq[tokens[i]] = 1
# Sort the frequencies storing in (freq, token) pairs and prune words with freq < min_count
tokenFreq = sorted(freq.items(), key = lambda x: x[1])
garbageWords = []
for item in tokenFreq:
if item[1] < min_count:
garbageWords.append(item[0])
print "Generating co-occurence matrix"
doc_text = ""
for filepath in filepaths:
text = open(filepath).read().lower()
words = re.findall(rgx, text)
N = len(words)
temp = [' '] * (N + contextSize)
temp[contextSize : (contextSize + N)] = words
words = temp
for i in xrange(contextSize, (contextSize + N)):
# Filter out garbage words"
#if words[i] not in garbageWords:
# Include context size specified by user
for j in xrange(i-contextSize, i):
if words[i] != ' ' and words[j] != ' ':
pa.addcontext(words[j], words[i])
pa.addcontext(words[i], words[j])
print "Co-occurence matrix generated"
print "Starting training"
tm = TrainModel(maxnum, newdims)
pa.freeze()
for k in xrange(ntimes):
for numwordvec in pa.getallwordvecs():
tm.trainonone(numwordvec[1])
lr /=float(1+k*lr_decay)
wordembeddings = {}
for numwordvec in pa.getallwordvecs():
(num, wordvec) = numwordvec
word = pa.wordnumrelation.getWord(num)
embedding = tm.getoutput(wordvec)
wordembeddings[word] = embedding
print "Training proces done, dumping embedding into persistant storage!"
outfile = open("./embeddings.pickle", "w")
pickle.dump(wordembeddings, outfile)
outfile.close()
print "Training completed! Embedding done."
print "time is %f" % (time.time()-t)
from probarray import ProbArray
from backprop import TrainModel
import os,re
import cPickle as pickle
pa = ProbArray()
folder = "ie_data"
filepaths = map(lambda x: folder+"/"+x,os.listdir(folder))
rgx = re.compile("([\w][\w']*\w)")
for filepath in filepaths:
alltext = open(filepath).read().lower()
words = filter(lambda x: len(x)<15 ,re.findall(rgx,alltext))
for i in xrange(1,len(words)):
pa.addcontext(words[i-1],words[i])
pa.addcontext(words[i],words[i-1])
maxnum = 20000 #pa.wordnumrelation.maxnum
newdims = 100 #suppose 100 dimensions afterall
tm = TrainModel(maxnum,newdims)
ntimes = 2
pa.freeze()
for k in xrange(ntimes):
for numwordvec in pa.getallwordvecs():
tm.trainonone(numwordvec[1])
wordembeddings = {}
for numwordvec in pa.getallwordvecs():
(num,wordvec) = numwordvec
word = pa.wordnumrelation.getWord(num)
embedding = tm.getoutput(wordvec)
wordembeddings[word] = embedding
outfile = open("embeddings.pickle","w")
