def test_generate(self):
path_checkpoint='.'
prefix_checkpoint='test'
n_epochs=10
cor = Corgan()
# dummy dataset
n_gen = 500
n = 1000
m = 7
x = np.random.randint(low=0, high=2, size=(n,m))
model = cor.train(x=x,
n_epochs_pretrain=10,
n_epochs=10,
path_checkpoint=path_checkpoint,
prefix_checkpoint=prefix_checkpoint)
x_synth = cor.generate(model = model, n_gen=n_gen)
# clean up
file_ckpt=os.path.join(path_checkpoint, prefix_checkpoint + ".model_epoch_%d.pth" % n_epochs)
os.remove(file_ckpt)
assert len(x_synth) == n_gen
def test_train(self):
path_checkpoint='.'
prefix_checkpoint='test'
n_epochs=10
cor = Corgan()
# dummy dataset
n = 1000
m = 7
x = np.random.randint(low=0, high=2, size=(n,m))
model = cor.train(x=x,
n_epochs_pretrain=10,
n_epochs=10,
batch_size=512,
path_checkpoint=path_checkpoint,
prefix_checkpoint=prefix_checkpoint)
file_ckpt=os.path.join(path_checkpoint, prefix_checkpoint + ".model_epoch_%d.pth" % n_epochs)
res = os.path.isfile(file_ckpt)
os.remove(file_ckpt)
assert res
def membership_inference_hayes(self, r_trn, r_tst, s_all, n_cpu):
"""Membership inference scenario as in Hayes et al. 2018.
Parameters
----------
r_trn : TYPE
DESCRIPTION.
r_tst : TYPE
DESCRIPTION.
s_all : TYPE
DESCRIPTION.
n_cpu : TYPE
DESCRIPTION.
Returns
-------
dict
DESCRIPTION.
"""
cor = Corgan()
# evaluation set
x_all = np.row_stack((r_tst, r_trn))
y_all = np.append(np.zeros(len(r_tst)), np.ones(len(r_trn)))
# train shadow GAN
gan_shadow = cor.train(x=s_all, n_cpu=n_cpu, debug=True)
# load shadow discriminator
minibatch_averaging = gan_shadow['parameter_dict'][
'minibatch_averaging']
feature_size = gan_shadow['parameter_dict']['feature_size']
d_shadow = Discriminator(minibatch_averaging=minibatch_averaging,
feature_size=feature_size)
d_shadow.load_state_dict(gan_shadow['Discriminator_state_dict'])
d_shadow.eval()
# calculate probabilities from shadow discriminator
p_all = d_shadow(x_all)
roc = metrics.roc_curve(y_true=y_all, y_score=p_all)
auc = metrics.roc_auc_score(y_true=y_all, y_score=p_all)
return {
'prob': p_all,
'label': y_all,
'roc': roc,
'auc': auc,
'analysis': 'membership_inference'
}
def test_save_and_load(self):
path_checkpoint='.'
prefix_checkpoint='test'
n_epochs=10
cor = Corgan()
# dummy dataset
n_gen = 500
n = 1000
m = 7
x = np.random.randint(low=0, high=2, size=(n,m))
model_saved = cor.train(x=x,
n_epochs_pretrain=10,
n_epochs=10,
path_checkpoint=path_checkpoint,
prefix_checkpoint=prefix_checkpoint)
file = 'test.pkl'
cor.save_obj(obj=model_saved, file_name=file)
model_loaded = cor.load_obj(file)
x_synth = cor.generate(model = model_loaded, n_gen=n_gen)
# clean up
file_ckpt=os.path.join(path_checkpoint, prefix_checkpoint + ".model_epoch_%d.pth" % n_epochs)
os.remove(file_ckpt)
os.remove(file)
assert len(x_synth) == n_gen
# preprocess
meta = pre.get_metadata(arr=arr, header=header)
obj_d = pre.get_discretized_matrix(arr=arr, meta=meta, header=header)
# split data
r_all = obj_d['x']
n_subset_r = round(len(r_all) * args.frac_train)
idx_trn = np.random.choice(len(r_all), n_subset_r, replace=False)
idx_tst = np.setdiff1d(range(len(r_all)), idx_trn)
r_trn = r_all[idx_trn, :]
r_tst = r_all[idx_tst, :]
# train and save model
if args.train_type == 'corgan':
syn = Corgan(debug=debug, n_cpu=args.n_cpu_train)
elif args.train_type == 'ppgan':
syn = Ppgan(debug=debug, n_cpu=args.n_cpu_train)
model = syn.train(x=r_trn, n_epochs=args.n_epoch)
model['m'] = meta
model['header'] = obj_d['header']
syn.save_obj(model, outfile)
elif args.task == 'generate':
pre = Preprocessor(missing_value=args.missing_value_generate)
outfile = args.outprefix_generate + '.csv'
syn = Corgan()
model = syn.load_obj(args.file_model)
if model['parameter_dict']['model'] == 'corgan':