predictor_type='Linear',
respects_MC=False)
accuracy_s_rf[i] = evaluations.evaluate_private_representations(
network.encoder,
trainset,
testset,
device,
verbose=True,
predictor_type='RandomForest',
respects_MC=False)
X, Y = testset.data, testset.hidden
Z, Z_mean = network.encoder(
torch.FloatTensor(X).to(device),
torch.FloatTensor(Y).to(device))
mine_network = mine.MINE(1,
representation_dim,
hidden_size=100,
moving_average_rate=0.1).to(device)
print('MINE calculations...')
IYZ[i] = mine_network.train(Y,
Z.detach().cpu().numpy(),
batch_size=batch_size_mine,
n_iterations=int(5e4),
n_verbose=-1,
n_window=100,
save_progress=-1)
print(f'I(Y;Z): {IYZ[i]}')
np.save(logsdir + 'betas', betas)
np.save(logsdir + 'IYZ', IYZ)
np.save(logsdir + 'accuracy_s_lin', accuracy_s_lin)
np.save(logsdir + 'accuracy_s_rf', accuracy_s_rf)
rhos = np.linspace(-0.95, 0.95, n_rhos)
for i, rho in enumerate(rhos):
print(f'Computing MI for rho {rho:.2f}')
# Compute the real MI
cov = np.eye(2 * d)
cov[d:2 * d, 0:d] = rho * np.eye(d)
cov[0:d, d:2 * d] = rho * np.eye(d)
MI_real[i] = -0.5 * np.log(np.linalg.det(cov)) / np.log(2)
# Compute the estimation of the MI
dataset = utils.BiVariateGaussianDatasetForMI(d, rho, 5000)
mine_network = mine.MINE(d, d, hidden_size=100).to(device)
MI_mine[i] = mine_network.train(dataset,
batch_size=args.batch_size,
n_iterations=args.n_iterations,
n_verbose=args.n_verbose,
n_window=args.n_window,
save_progress=args.save_progress)
print(
f'Rho {rho:.2f}: Real ({MI_real[i]:.3f}) - Estimated ({MI_mine[i]:.3f})'
)
plt.plot(rhos, MI_real, 'black', label='Real MI')
plt.plot(rhos, MI_mine, 'orange', label='Estimated MI')
plt.legend()
plt.savefig(f'../figures/{d}-dimensional_gaussian_mi.png')