def batch_generator(corpus, k, V, thread_batch, lockedEta):
# ids
doc_id = 0
batch_id = 0
round_id = 0
# temp data
doc_buffer = []
voc_temp = set()
# process contral
nBatch = LockedSum(0, Lock())
for doc in corpus:
for vid, count in doc:
voc_temp.add(vid)
doc_buffer.append(doc)
if doc_id % thread_batch == thread_batch - 1:
eta_temp = lockedEta.get_eta(k, voc_temp)
etaSum = lockedEta.get_eta_sum(k, V)
alpha = _mea.get_alpha(k)
yield (doc_buffer, eta_temp, etaSum, alpha, batch_id)
# clear buffer
doc_buffer = []
voc_temp = set()
batch_id += 1
doc_id += 1
# some remain doc may not be processed
if len(doc_buffer) > 0:
eta_temp = lockedEta.get_eta(k, voc_temp)
etaSum = lockedEta.get_eta_sum(k, V)
alpha = _mea.get_alpha(k)
yield (doc_buffer, eta_temp, etaSum, alpha, batch_id)
batch_id += 1
def batch_generator(corpus,k,V,thread_batch,lockedEta):
# ids
doc_id = 0
batch_id = 0
round_id = 0
# temp data
doc_buffer = []
voc_temp = set()
# process contral
nBatch = LockedSum(0,Lock())
for doc in corpus:
for vid,count in doc:
voc_temp.add(vid)
doc_buffer.append(doc)
if doc_id % thread_batch == thread_batch - 1:
eta_temp = lockedEta.get_eta(k,voc_temp)
etaSum = lockedEta.get_eta_sum(k,V)
alpha = _mea.get_alpha(k)
yield (doc_buffer,eta_temp,etaSum,alpha,batch_id)
# clear buffer
doc_buffer = []
voc_temp = set()
batch_id += 1
doc_id += 1
# some remain doc may not be processed
if len(doc_buffer) > 0:
eta_temp = lockedEta.get_eta(k,voc_temp)
etaSum = lockedEta.get_eta_sum(k,V)
alpha = _mea.get_alpha(k)
yield (doc_buffer,eta_temp,etaSum,alpha,batch_id)
batch_id += 1
def test():
'''
$1 path to config file
'''
logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s',
level=logging.INFO)
# loading configs;
config = get_config(sys.argv[2])
k = config['k']
test_path = config['test_path']
test_train_path = config['test_train']
test_test_path = config['test_test']
eta_path = config['eta_path']
gensim = config['gensim']
print eta_path
corpus = _mCorpus.get_corpus(test_path)
V = corpus.num_terms
voc_set = set()
for doc in corpus:
for wid, count in doc:
voc_set.add(wid)
etaTest, etaSum = None, None
if gensim:
etaTest, etaSum = _mea.get_gensim_eta_etaSum(eta_path, voc_set)
else:
eta = _mea.load_eta(eta_path)
etaTest = _mea.get_eta(k, eta, voc_set)
etaSum = _mea.get_eta_sum(eta, k, V)
test_test = _mCorpus.get_corpus(test_test_path)
test_train = _mCorpus.get_corpus(test_train_path)
alpha = _mea.get_alpha(k)
perplexity = _mper.perplexity(test_train, test_test, alpha, etaTest,
etaSum)
print perplexity
def test():
'''
$1 path to config file
'''
logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)
# loading configs;
config = get_config(sys.argv[2])
k = config['k']
test_path = config['test_path']
test_train_path = config['test_train']
test_test_path = config['test_test']
eta_path = config['eta_path']
gensim= config['gensim']
print eta_path
corpus = _mCorpus.get_corpus(test_path)
V = corpus.num_terms
voc_set = set()
for doc in corpus:
for wid,count in doc:
voc_set.add(wid)
etaTest, etaSum = None,None
if gensim:
etaTest, etaSum = _mea.get_gensim_eta_etaSum(eta_path,voc_set)
else:
eta = _mea.load_eta(eta_path)
etaTest = _mea.get_eta(k,eta,voc_set)
etaSum = _mea.get_eta_sum(eta,k,V)
test_test = _mCorpus.get_corpus(test_test_path)
test_train = _mCorpus.get_corpus(test_train_path)
alpha = _mea.get_alpha(k)
perplexity = _mper.perplexity(test_train,test_test,alpha,etaTest,etaSum)
print perplexity
def asyn_framework(corpus,k,V,nthread,minibatch,var_path,record_eta = False):
# configs
thread_batch = minibatch/nthread
# ids
doc_id = 0
batch_id = 0
round_id = 0
# temp data
doc_buffer = []
voc_temp = set()
old_doc_seen = {}
old_doc_seen[0] = 0
# global data
lockedEta = LockedEta({},Lock())
# process contral
pool = Pool(processes = nthread)
nActPro = LockedSum(0,Lock())
nBatch = LockedSum(0,Lock())
results = []
for doc in corpus:
for vid,count in doc:
voc_temp.add(vid)
doc_buffer.append(doc)
if doc_id % thread_batch == thread_batch - 1:
eta_temp = lockedEta.get_eta(k,voc_temp)
etaSum = lockedEta.get_eta_sum(k,V)
alpha = _mea.get_alpha(k)
while True: # check for active processes amount
if nActPro.get_value() < nthread:
break
cb = lambda x: callback(x,lockedEta,nActPro,nBatch,var_path,nthread,thread_batch,old_doc_seen)
result = pool.apply_async(asyn_workder,(doc_buffer,eta_temp,etaSum,alpha),callback = cb)
results.append(result)
nActPro.add_value(1)
# clear buffer
doc_buffer = []
voc_temp = set()
batch_id += 1
doc_id += 1
# some remain doc may not be processed
if len(doc_buffer) > 0:
eta_temp = lockedEta.get_eta(k,voc_temp)
etaSum = lockedEta.get_eta_sum(k,V)
alpha = _mea.get_alpha(k)
while True: # check for active processes amount
if nActPro.get_value() < nthread:
break
cb = lambda x: callback(x,lockedEta,nActPro,nBatch,var_path,nthread,thread_batch,old_doc_seen)
result = pool.apply_async(asyn_workder,(doc_buffer,eta_temp,etaSum,alpha),callback = cb)
results.append(result)
nActPro.add_value(1)
batch_id += 1
for r in results:
r.wait()
if nBatch.get_value() % nthread != 0:
nBatch_value = nBatch.get_value()
fn = 'eta.{}.pickle'.format(nBatch_value/nthread)
path = os.path.join(var_path,fn)
lockedEta.write_eta(path)
return lockedEta.eta
def syn_framework(corpus,
k,
V,
nthread,
minibatch,
var_path,
record_eta=False):
# configs
thread_batch = minibatch / nthread
# ids
doc_id = 0
batch_id = 0
round_id = 0
# temp data
doc_buffer = []
batch_buffer = [] # [(docs,etas)]
voc_temp = set()
# global data
eta = {}
for doc in corpus:
for vid, count in doc:
voc_temp.add(vid)
doc_buffer.append(doc)
if doc_id % thread_batch == thread_batch - 1:
eta_temp = _mea.get_eta(k, eta, voc_temp)
etaSum = _mea.get_eta_sum(eta, k, V)
batch_buffer.append((doc_buffer, eta_temp, etaSum))
# clear doc buffer
doc_buffer = []
voc_temp = set()
if batch_id % nthread == nthread - 1:
# update eta
eta = syn_master(batch_buffer, k, nthread, eta,
_mea.get_alpha(k))
if record_eta:
fn = 'eta.{}.pickle'.format(round_id)
path = os.path.join(var_path, fn)
_mea.write_eta(eta, path)
# clear batch_buffer
batch_buffer = []
round_id += 1
logging.info('round:{}, batch:{}'.format(round_id, batch_id))
batch_id += 1
doc_id += 1
# process the docs in current doc_buffer
if len(doc_buffer) > 0:
# form a new batch
eta_temp = _mea.get_eta(k, eta, voc_temp)
etaSum = _mea.get_eta_sum(eta, k, V)
batch_buffer.append((doc_buffer, eta_temp, etaSum))
# form a new round
eta = syn_master(batch_buffer, k, len(batch_buffer), eta,
_mea.get_alpha(k))
if record_eta:
fn = 'eta.{}.pickle'.format(round_id)
path = os.path.join(var_path, fn)
_mea.write_eta(eta, path)
round_id += 1
batch_id += 1
return eta
def asyn_framework(corpus,
k,
V,
nthread,
minibatch,
var_path,
record_eta=False):
# configs
thread_batch = minibatch / nthread
# ids
doc_id = 0
batch_id = 0
round_id = 0
# temp data
doc_buffer = []
voc_temp = set()
old_doc_seen = {}
old_doc_seen[0] = 0
# global data
lockedEta = LockedEta({}, Lock())
# process contral
pool = Pool(processes=nthread)
nActPro = LockedSum(0, Lock())
nBatch = LockedSum(0, Lock())
results = []
for doc in corpus:
for vid, count in doc:
voc_temp.add(vid)
doc_buffer.append(doc)
if doc_id % thread_batch == thread_batch - 1:
eta_temp = lockedEta.get_eta(k, voc_temp)
etaSum = lockedEta.get_eta_sum(k, V)
alpha = _mea.get_alpha(k)
while True: # check for active processes amount
if nActPro.get_value() < nthread:
break
cb = lambda x: callback(x, lockedEta, nActPro, nBatch, var_path,
nthread, thread_batch, old_doc_seen)
result = pool.apply_async(asyn_workder,
(doc_buffer, eta_temp, etaSum, alpha),
callback=cb)
results.append(result)
nActPro.add_value(1)
# clear buffer
doc_buffer = []
voc_temp = set()
batch_id += 1
doc_id += 1
# some remain doc may not be processed
if len(doc_buffer) > 0:
eta_temp = lockedEta.get_eta(k, voc_temp)
etaSum = lockedEta.get_eta_sum(k, V)
alpha = _mea.get_alpha(k)
while True: # check for active processes amount
if nActPro.get_value() < nthread:
break
cb = lambda x: callback(x, lockedEta, nActPro, nBatch, var_path,
nthread, thread_batch, old_doc_seen)
result = pool.apply_async(asyn_workder,
(doc_buffer, eta_temp, etaSum, alpha),
callback=cb)
results.append(result)
nActPro.add_value(1)
batch_id += 1
for r in results:
r.wait()
if nBatch.get_value() % nthread != 0:
nBatch_value = nBatch.get_value()
fn = 'eta.{}.pickle'.format(nBatch_value / nthread)
path = os.path.join(var_path, fn)
lockedEta.write_eta(path)
return lockedEta.eta
def syn_framework(corpus,k,V,nthread,minibatch,var_path,record_eta = False):
# configs
thread_batch = minibatch/nthread
# ids
doc_id = 0
batch_id = 0
round_id = 0
# temp data
doc_buffer = []
batch_buffer = [] # [(docs,etas)]
voc_temp = set()
# global data
eta = {}
for doc in corpus:
for vid,count in doc:
voc_temp.add(vid)
doc_buffer.append(doc)
if doc_id % thread_batch == thread_batch - 1:
eta_temp = _mea.get_eta(k,eta, voc_temp)
etaSum = _mea.get_eta_sum(eta,k,V)
batch_buffer.append((doc_buffer,eta_temp,etaSum))
# clear doc buffer
doc_buffer = []
voc_temp = set()
if batch_id % nthread == nthread - 1:
# update eta
eta = syn_master(batch_buffer,k,nthread,eta,_mea.get_alpha(k))
if record_eta:
fn = 'eta.{}.pickle'.format(round_id)
path = os.path.join(var_path,fn)
_mea.write_eta(eta,path)
# clear batch_buffer
batch_buffer = []
round_id += 1
logging.info('round:{}, batch:{}'.format(round_id,batch_id))
batch_id += 1
doc_id += 1
# process the docs in current doc_buffer
if len(doc_buffer) > 0:
# form a new batch
eta_temp = _mea.get_eta(k,eta, voc_temp)
etaSum = _mea.get_eta_sum(eta,k,V)
batch_buffer.append((doc_buffer,eta_temp,etaSum))
# form a new round
eta = syn_master(batch_buffer,k,len(batch_buffer),eta,_mea.get_alpha(k))
if record_eta:
fn = 'eta.{}.pickle'.format(round_id)
path = os.path.join(var_path,fn)
_mea.write_eta(eta,path)
round_id +=1
batch_id +=1
return eta
