def test_gan_train_match(self):
rea = Realism()
n = 1000
m_2 = 3
threshold = 0.05
max_beta = 10
n_epoch = 100
beta = np.append(np.random.randint(low=-max_beta,high=0,size=(m_2,1)),
np.random.randint(low=0,high=max_beta,size=(m_2,1)))
x_real = np.random.randint(low=0, high=2, size=(n,m_2*2))
x_for_e = np.reshape(np.matmul(x_real, beta), (n,1)) + 0.5 * np.random.random(size=(n,1))
y_real = np.reshape(np.round(1.0 / (1.0 + np.exp(-x_for_e))), (n,))
res_real = rea.gan_train(x_synth=x_real, y_synth=y_real,
x_real=x_real, y_real=y_real, n_epoch=n_epoch)
res_gan_train1 = rea.gan_train(x_synth=x_real, y_synth=y_real,
x_real=x_real, y_real=y_real, n_epoch=n_epoch)
assert (abs(res_real['auc'] - res_gan_train1['auc']) < threshold)
def test_gan_test_mismatch(self):
rea = Realism()
n = 1000
m_2 = 3
threshold = 0.05
max_beta = 10
n_epoch = 100
beta = np.append(np.random.randint(low=-max_beta,high=0,size=(m_2,1)),
np.random.randint(low=0,high=max_beta,size=(m_2,1)))
x_real = np.random.randint(low=0, high=2, size=(n,m_2*2))
x_for_e = np.reshape(np.matmul(x_real, beta), (n,1)) + 0.5 * np.random.random(size=(n,1))
y_real = np.reshape(np.round(1.0 / (1.0 + np.exp(-x_for_e))), (n,))
# flip label to ensure AUCs are very different
x_synth = x_real
y_synth = 1 - y_real
res_real = rea.gan_train(x_synth=x_real, y_synth=y_real,
x_real=x_real, y_real=y_real, n_epoch=n_epoch)
res_gan_test2 = rea.gan_test(x_synth, y_synth, x_real, y_real, n_epoch=n_epoch)
assert (abs(res_real['auc'] - res_gan_test2['auc']) > threshold)