def blur_no_infogain(blue_c_imb, data_dir, task, pos_class):
# assumed b_imb == 0.5
blue_dir = data_dir.replace('Red', 'Blue')
d_red = setup.get_label_dict_knowing(data_dir, task, pos_class)
d_blue = setup.get_label_dict_knowing(blue_dir, task, pos_class)
red_c_imb = float(len(
d_red['Default'])) / (len(d_red[task]) + len(d_red['Default']))
blue_c_imb = float(blue_c_imb)
if red_c_imb >= blue_c_imb:
# can't add all negatives
num_pos = len(d_red[task])
num_neg = num_pos * (blue_c_imb / float(1 - blue_c_imb))
# num_neg = num_pos * (len(d_blue['Default'])/len(d_blue[task]))
else:
# can't add all positives
k = ((1 - blue_c_imb) / float(blue_c_imb))
print k
for key in d_red.keys():
try:
print len(d_red[key]), key
except:
pass
num_neg = len(d_red['Default'])
num_pos = num_neg * k
# num_pos = num_neg * (len(d_blue[task])/len(d_blue['Default']))
return int(num_pos), int(num_neg)
def blur_no_infogain(blue_c_imb, data_dir, task, pos_class):
# assumed b_imb == 0.5
blue_dir = data_dir.replace('Red','Blue')
d_red = setup.get_label_dict_knowing(data_dir, task, pos_class)
d_blue = setup.get_label_dict_knowing(blue_dir, task, pos_class)
red_c_imb=float(len(d_red['Default']))/(len(d_red[task])+len(d_red['Default']))
blue_c_imb = float(blue_c_imb)
if red_c_imb >= blue_c_imb:
# can't add all negatives
num_pos = len(d_red[task])
num_neg = num_pos * (blue_c_imb/float(1-blue_c_imb))
# num_neg = num_pos * (len(d_blue['Default'])/len(d_blue[task]))
else:
# can't add all positives
k = ((1-blue_c_imb)/float(blue_c_imb))
print k
for key in d_red.keys():
try:
print len(d_red[key]), key
except: pass
num_neg = len(d_red['Default'])
num_pos = num_neg * k
# num_pos = num_neg * (len(d_blue[task])/len(d_blue['Default']))
return int(num_pos), int(num_neg)
def what_redbox_numbers(c_imb, b_imb, data_dir, task, pos_class, b_pos, b_neg):
# big prob: after redbox sampling, imbalance has changed.
# so actually, redbox sampling and undersampling both need to be
# determined before either takes place.
# other prob: given b_imb, compute num_neg num_pos
# maybe easier is given info gain, compute num_neg num_pos
d = setup.get_label_dict_knowing(data_dir, task, pos_class)
red_c_imb = float(len(d['Default'])) / (len(d[task]) + len(d['Default']))
if red_c_imb >= c_imb:
r_pos = len(d[task])
r_neg = r_pos * (b_imb / (1 - b_imb)) * (r_pos / (r_pos))
return r_pos, r_neg
# if red_c_imb lower, would have to
elif red_c_imb <= c_imb:
print "class imbalance going to decrease! :D"
return len(d[task]), len(d['Default']) * (c_imb / red_c_imb)
def what_redbox_numbers(c_imb, b_imb, data_dir, task, pos_class,
b_pos, b_neg):
# big prob: after redbox sampling, imbalance has changed.
# so actually, redbox sampling and undersampling both need to be
# determined before either takes place.
# other prob: given b_imb, compute num_neg num_pos
# maybe easier is given info gain, compute num_neg num_pos
d = setup.get_label_dict_knowing(data_dir, task, pos_class)
red_c_imb=float(len(d['Default']))/(len(d[task])+len(d['Default']))
if red_c_imb >= c_imb:
r_pos = len(d[task])
r_neg = r_pos * (b_imb/(1-b_imb)) * (r_pos/(r_pos))
return r_pos, r_neg
# if red_c_imb lower, would have to
elif red_c_imb <= c_imb:
print "class imbalance going to decrease! :D"
return len(d[task]), len(d['Default'])*(c_imb/red_c_imb)
def same_amount_as_bluebox(data_dir, task, pos_class):
d = setup.get_label_dict_knowing(data_dir, task, pos_class)
# ASSUMING MODEL LEARNS P(label|data) !
return len(d[task]), len(d[task])
def same_amount_as_bluebox(data_dir, task, pos_class): d = setup.get_label_dict_knowing(data_dir, task, pos_class) # ASSUMING MODEL LEARNS P(label|data) ! return len(d[task]), len(d[task])
