def make_training_step_method(classifier,
num_positive_regions,
positive_fraction,
lr=1e-3,
l2=0,
method='GE-binomial',
pi=0,
slack=-1,
autoencoder=0):
import topaz.methods as methods
criteria = nn.BCEWithLogitsLoss()
optim = torch.optim.Adam
# pi sets the expected fraction of positives
# but during training, we iterate over unlabeled data with labeled positives removed
# therefore, we expected the fraction of positives in the unlabeled data
# to be pi - fraction of labeled positives
# if we are using the 'GE-KL' or 'GE-binomial' loss functions
p_observed = positive_fraction
if pi <= p_observed and method in ['GE-KL', 'GE-binomial']:
# if pi <= p_observed, then we think the unlabeled data is all negatives
# report this to the user and switch method to 'PN' if it isn't already
print(
'WARNING: pi={} but the observed fraction of positives is {} and method is set to {}.'
.format(pi, p_observed, method),
file=sys.stderr)
print('WARNING: setting method to PN with pi={} instead.'.format(
p_observed),
file=sys.stderr)
print('WARNING: if you meant to use {}, please set pi > {}.'.format(
method, p_observed),
file=sys.stderr)
pi = p_observed
method = 'PN'
elif method in ['GE-KL', 'GE-binomial']:
pi = pi - p_observed
split = 'pn'
if method == 'PN':
optim = optim(classifier.parameters(), lr=lr)
trainer = methods.PN(classifier,
optim,
criteria,
pi=pi,
l2=l2,
autoencoder=autoencoder)
elif method == 'GE-KL':
if slack < 0:
slack = 10
optim = optim(classifier.parameters(), lr=lr)
trainer = methods.GE_KL(classifier,
optim,
criteria,
pi,
l2=l2,
slack=slack)
elif method == 'GE-binomial':
if slack < 0:
slack = 1
optim = optim(classifier.parameters(), lr=lr)
trainer = methods.GE_binomial(classifier,
optim,
criteria,
pi,
l2=l2,
slack=slack,
autoencoder=autoencoder)
elif method == 'PU':
split = 'pu'
optim = optim(classifier.parameters(), lr=lr)
trainer = methods.PU(classifier,
optim,
criteria,
pi,
l2=l2,
autoencoder=autoencoder)
else:
raise Exception('Invalid method: ' + method)
return trainer, criteria, split
def make_training_step_method(classifier, num_positive_regions,
positive_fraction, args):
import topaz.methods as methods
criteria = nn.BCEWithLogitsLoss()
optim = torch.optim.Adam
lr = args.learning_rate
l2 = args.l2
pi = args.pi
slack = args.slack
split = 'pn'
if args.method == 'PN':
optim = optim(classifier.parameters(), lr=lr)
trainer = methods.PN(classifier,
optim,
criteria,
pi=pi,
l2=l2,
autoencoder=args.autoencoder)
elif args.method == 'GE-KL':
#split = 'pu'
if slack < 0:
slack = 10 * num_positive_regions
assert positive_fraction <= pi
pi = pi - positive_fraction
optim = optim(classifier.parameters(), lr=lr)
trainer = methods.GE_KL(classifier,
optim,
criteria,
pi,
l2=l2,
slack=slack)
elif args.method == 'GE-binomial':
#split = 'pu'
if slack < 0:
slack = 1
assert positive_fraction <= pi
pi = pi - positive_fraction
optim = optim(classifier.parameters(), lr=lr)
trainer = methods.GE_binomial(classifier,
optim,
criteria,
pi,
l2=l2,
slack=slack,
autoencoder=args.autoencoder)
elif args.method == 'PU':
split = 'pu'
optim = optim(classifier.parameters(), lr=lr)
trainer = methods.PU(classifier,
optim,
criteria,
pi,
l2=l2,
autoencoder=args.autoencoder)
else:
raise Exception('Invalid method: ' + args.method)
return trainer, criteria, split