def active_learning(filename,
query='',
stop='true',
stopat=0.95,
error='none',
interval=100000,
seed=0):
stopat = float(stopat)
thres = 0
starting = 1
counter = 0
pos_last = 0
np.random.seed(seed)
read = MAR()
read = read.create(filename)
read.interval = interval
read.BM25(query.strip().split('_'))
num2 = read.get_allpos()
target = int(num2 * stopat)
print("number of target, true/close here:", target)
if stop == 'est':
read.enable_est = True
else:
read.enable_est = False
while True:
pos, neg, total = read.get_numbers()
try:
print("%d, %d, %d" % (pos, pos + neg, read.est_num))
except:
print("%d, %d" % (pos, pos + neg))
if pos + neg >= total:
if stop == 'knee' and error == 'random':
coded = np.where(
np.array(read.body['code']) != "undetermined")[0]
seq = coded[np.argsort(read.body['time'][coded])]
part1 = set(seq[:read.kneepoint * read.step]) & set(
np.where(np.array(read.body['code']) == "no")[0])
part2 = set(seq[read.kneepoint * read.step:]) & set(
np.where(np.array(read.body['code']) == "yes")[0])
for id in part1 | part2:
read.code_error(id, error=error)
break
if pos < starting or pos + neg < thres:
for id in read.BM25_get():
read.code_error(id, error=error)
else:
a, b, c, d = read.train(weighting=True, pne=True)
if stop == 'est':
if stopat * read.est_num <= pos:
break
elif stop == 'soft':
if pos >= 10 and pos_last == pos:
counter = counter + 1
else:
counter = 0
pos_last = pos
if counter >= 5:
break
elif stop == 'knee':
if pos >= 10:
if read.knee():
if error == 'random':
coded = np.where(
np.array(read.body['code']) != "undetermined"
)[0]
seq = coded[np.argsort(
np.array(read.body['time'])[coded])]
part1 = set(
seq[:read.kneepoint * read.step]) & set(
np.where(
np.array(read.body['code']) == "no")
[0])
part2 = set(
seq[read.kneepoint * read.step:]) & set(
np.where(
np.array(read.body['code']) == "yes")
[0])
for id in part1 | part2:
read.code_error(id, error=error)
break
else:
if pos >= target:
break
if pos < 10:
for id in a:
read.code_error(id, error=error)
else:
for id in c:
read.code_error(id, error=error)
set_trace()
return read
def active_learning(filename,
query='',
stop='true',
stopat=1.00,
error='none',
interval=100000,
seed=0):
stopat = float(stopat)
thres = 0
starting = 1
counter = 0
pos_last = 0
np.random.seed(seed)
read = MAR()
read = read.create(filename)
# random sampling or by querying similar documents
# self.bm is provided with a list or a view of a dict's value which is not sorted
read.BM25(query.strip().split('_'))
# get the rest #pos documents
num2 = read.get_allpos()
target = int(
num2 * stopat
) # stopat is 1. Is it the minum num of pos to activate svm training ?
if stop == 'est': # stop = 'true'
read.enable_est = True
else:
read.enable_est = False # will excute this line
while True:
pos, neg, total = read.get_numbers()
try:
print("%d, %d, %d" %
(pos, pos + neg, read.est_num)) # what is est_num ?
except:
print("%d, %d" % (pos, pos + neg)) # execute this line
if pos + neg >= total: # do not go inside
if stop == 'knee' and error == 'random':
coded = np.where(
np.array(read.body['code']) != "undetermined")[0]
seq = coded[np.argsort(read.body['time'][coded])]
part1 = set(seq[:read.kneepoint * read.step]) & set(
np.where(np.array(read.body['code']) == "no")[0])
part2 = set(seq[read.kneepoint * read.step:]) & set(
np.where(np.array(read.body['code']) == "yes")[0])
for id in part1 | part2:
read.code_error(id, error=error)
break
if pos < starting or pos + neg < thres: # the second condition doesn't work
for id in read.BM25_get(
): # select a set of candidates from self.pool
read.code_error(
id, error=error
) # simulate human labeling error, default is no error
else:
a, b, c, d = read.train(weighting=True, pne=True)
if stop == 'est':
if stopat * read.est_num <= pos:
break
elif stop == 'soft':
if pos >= 10 and pos_last == pos:
counter = counter + 1
else:
counter = 0
pos_last = pos
if counter >= 5:
break
elif stop == 'knee':
if pos >= 10:
if read.knee():
if error == 'random':
coded = np.where(
np.array(read.body['code']) != "undetermined"
)[0]
seq = coded[np.argsort(
np.array(read.body['time'])[coded])]
part1 = set(
seq[:read.kneepoint * read.step]) & set(
np.where(
np.array(read.body['code']) == "no")
[0])
part2 = set(
seq[read.kneepoint * read.step:]) & set(
np.where(
np.array(read.body['code']) == "yes")
[0])
for id in part1 | part2:
read.code_error(id, error=error)
break
else:
if pos >= target:
break
if pos < 10:
for id in a:
read.code_error(id, error=error)
else:
for id in c:
read.code_error(id, error=error)
return read
root = os.getcwd().split("src")[0] + "src/src/util"
sys.path.append(root)
from mar import MAR
from pdb import set_trace
if __name__ == "__main__":
data_path = "Hall.csv"
target_recall = 0.95
thres = 10
query = "defect prediction"
read = MAR()
read = read.create(data_path)
read.enable_est = True
if query:
read.BM25(query.split())
while True:
pos, neg, total = read.get_numbers()
try:
print("%d, %d, %d" % (pos, pos + neg, read.est_num))
except:
print("%d, %d" % (pos, pos + neg))
if pos + neg >= total:
break
if pos < 1:
if query:
ids, scores = read.BM25_get()
for id in ids:
read.code(id, read.body["label"][id])
