def test_fast_computation(self):
from edge.utils import device, cuda
if device != cuda:
self.assertTrue(
False, "This test should be run with device=cuda. "
"See edge.utils.device.py to set this.")
x = np.linspace(0, 1, 9000, dtype=np.float32).reshape((-1, 1))
y = np.exp(-x**2).reshape(-1)
gp = MaternGP(x,
y,
noise_constraint=(0, 1e-3),
value_structure_discount_factor=0.5)
x_ = np.linspace(0, 1, 20, dtype=np.float32).reshape((-1, 1))
y_ = np.exp(-x_**2).reshape(-1)
#with gpytorch.settings.fast_computations(solves=False):
try:
pred = gp.predict(x_).mean.cpu().numpy()
except Exception as e:
if DEBUG:
try:
import pudb
pudb.post_mortem()
except ImportError:
pass
self.assertTrue(
False, f'Prediction failed with the following error: {str(e)}')
self.assertTrue(True)
def test_neighbor_erasing_dataset_1(self):
x = np.linspace(0, 1, 100, dtype=np.float32).reshape((-1, 1))
y = np.exp(-x**2).reshape(-1)
r = 0.0125
gp = MaternGP(x,
y,
noise_constraint=(0, 1e-3),
dataset_type='neighborerasing',
dataset_params={'radius': r})
x_ = np.linspace(0, 1, 20, dtype=np.float32).reshape((-1, 1))
y_ = np.exp(-x_**2).reshape(-1)
gp.append_data(x_, y_)
distances = np.abs(gp.train_x.cpu().numpy().reshape((-1, 1)) -
x_.squeeze())
self.assertTrue(np.all(distances[:-len(x_), :] >= r))
self.assertTrue((gp.train_x.cpu().numpy()[-len(x_):, :] == x_).all())
# Set to True to plot
if False:
import matplotlib.pyplot as plt
plt.figure()
plt.scatter(
gp.train_x.cpu().numpy()[-len(x_):],
gp.train_x.cpu().numpy()[-len(x_):],
color='r',
)
plt.scatter(gp.train_x.cpu().numpy()[:-len(x_)],
gp.train_x.cpu().numpy()[:-len(x_)],
color='b')
def test_hyper_optimization_1(self):
warnings.simplefilter('ignore', gpytorch.utils.warnings.GPInputWarning)
tol = 5e-3
lengthscale = 1.5
outputscale = 2
x = np.linspace(-1, 1, 501).reshape((-1, 1))
# Eventhough the data is generated by a Matern function, the lengthscale and outputscale learned by the Matern
# kernel GP do NOT need to coincide with the ones used to generate the data: the learned ones correspond to the
# influence of nearby points, not to the global structure of the data
y = self._matern_52(x,
lengthscale=lengthscale,
outputscale=outputscale).reshape(-1)
x_train = x[::2]
y_train = y[::2]
x_test = x[1::2]
y_test = y[1::2]
gp = MaternGP(x_train,
y_train,
nu=5 / 2,
noise_constraint=(0, 1e-5),
hyperparameters_initialization={
'covar_module.base_kernel.lengthscale': 1,
'covar_module.outputscale': 1
})
gp.optimizer = torch.optim.Adam
gp.optimize_hyperparameters(epochs=50, lr=0.01)
gp.optimize_hyperparameters(epochs=50, lr=0.001)
gp.optimize_hyperparameters(epochs=30, lr=0.0001)
predictions = gp.predict(x_test)
y_pred = predictions.mean.cpu().numpy()
passed = np.all(np.abs(y_test - y_pred) < tol)
if passed:
self.assertTrue(True)
else:
print(f'Max(diff): {np.abs(y_test - y_pred).max()}')
x_train = x_train.reshape(-1)
x_test = x_test.reshape(-1)
f, ax = plt.subplots(1, 1, figsize=(8, 8))
lower, upper = predictions.confidence_region()
ax.plot(x_train, y_train, 'k*')
ax.plot(x_test, y_pred, 'g-*')
ax.plot(x_test, y_test, 'r*')
ax.fill_between(x_test,
lower.cpu().numpy(),
upper.cpu().numpy(),
alpha=0.5)
ax.legend([
'Observed Data', 'Prediction', 'Hidden data', 'Confidence',
'Noise'
])
ax.grid(True)
#plt.show()
self.assertTrue(False)
def test_data_manipulation(self):
tol = 1e-1
x = np.linspace(0, 1, 101).reshape((-1, 1))
y = np.exp(-x**2).reshape(-1)
x_ = np.linspace(1.5, 2, 51).reshape((-1, 1))
y_ = 1 + np.exp(-x_**2).reshape(-1)
gp = MaternGP(x, y, noise_prior=(0.1, 0.1))
tmp = gp.empty_data()
self.assertEqual(tmp, gp)
self.assertTrue(tuple(gp.train_x.shape), (0, 1))
# GPyTorch fails when predicting with an empty dataset, so the following line fails if uncommented
# gp.predict(x)
gp.set_data(x, y)
self.assertEqual(tuple(gp.train_x.shape), (len(x), 1))
gp.optimize_hyperparameters(epochs=10)
gp_pred = gp.predict(x_).mean.cpu().numpy()
self.assertFalse(np.all(np.abs(gp_pred - y_) < tol))
tmp = gp.append_data(x_, y_)
# self.assertTrue(gp != tmp)
# self.assertEqual(tuple(gp.train_x.shape), (len(x), 1))
self.assertEqual(tuple(tmp.train_x.shape), (len(x) + len(x_), 1))
tmp.optimize_hyperparameters(epochs=10)
tmp_pred = tmp.predict(x_).mean.cpu().numpy()
self.assertTrue(np.all(np.abs(tmp_pred - y_) < tol))
def test_multi_dim_input(self):
tol = 0.1
lin = np.linspace(0, 1, 101)
x = lin.reshape((-1, 1))
x = x + x.T
y = np.exp(-lin**2)
gp = MaternGP(x, y, noise_prior=(0.1, 0.1))
gp.optimize_hyperparameters(epochs=10)
x_query = np.linspace(0.25, 0.75, 27)
y_ = np.exp(-x_query**2)
x_query = x_query.reshape((-1, 1)) + lin.reshape((1, -1))
pred = gp.predict(x_query).mean.cpu().numpy()
self.assertEqual(pred.shape, (27, ))
self.assertTrue(np.all(np.abs(pred - y_) < tol))
def test_large_ds_matern(self):
x = np.linspace(0, 1, 9000, dtype=np.float32).reshape((-1, 1))
y = np.exp(-x ** 2).reshape(-1)
gp = MaternGP(
x, y, noise_constraint=(0, 1e-3)
)
x_ = np.linspace(0, 1, 20, dtype=np.float32).reshape((-1, 1))
y_ = np.exp(-x_ ** 2).reshape(-1)
try:
pred = gp.predict(x_).mean.cpu().numpy()
except Exception as e:
if DEBUG:
import pudb
pudb.post_mortem()
self.assertTrue(False, f'Prediction failed with the following error: {str(e)}')
self.assertTrue(True)
def get_gp(x_train, y_train, noise):
return MaternGP(
x_train.reshape((-1, 1)),
y_train,
noise_prior=(noise**2, 1e-3),
lengthscale_prior=(1, 1e-3),
outputscale_prior=(1, 1e-3)
)
def test_load_save(self):
x = np.linspace(0, 1, 11)
y = np.sin(2 * np.pi * x) + np.random.randn(len(x)) * 0.2
model = MaternGP(x,
y,
noise_prior=(1, 0.1),
noise_constraint=0.5,
outputscale_constraint=(2.7, np.pi))
save_file = tempfile.NamedTemporaryFile(suffix='.pth').name
model.save(save_file)
self.assertTrue(os.path.isfile(save_file))
loaded = MaternGP.load(save_file, x, y)
self.assertEqual(model.covar_module.outputscale,
loaded.covar_module.outputscale)
def test_hyper_optimization_0(self):
warnings.simplefilter('ignore', gpytorch.utils.warnings.GPInputWarning)
tol = 1e-3
x = np.linspace(0, 1, 101).reshape((-1, 1))
y = np.exp(-x**2).reshape(-1)
gp = MaternGP(x, y, noise_constraint=(0, 1e-4))
gp.optimize_hyperparameters(epochs=20)
gp.optimize_hyperparameters(epochs=20, lr=0.01)
predictions = gp.predict(x)
y_pred = predictions.mean.cpu().numpy()
passed = np.all(np.abs(y - y_pred) < tol)
if passed:
self.assertTrue(True)
else:
x = x.reshape(-1)
f, ax = plt.subplots(1, 1, figsize=(4, 3))
lower, upper = predictions.confidence_region()
ax.plot(x, y, 'k*')
ax.plot(x, y_pred, 'b')
ax.fill_between(x,
lower.cpu().numpy(),
upper.cpu().numpy(),
alpha=0.5)
ax.legend(['Observed Data', 'Mean', 'Confidence', 'Noise'])
ax.grid(True)
#plt.show()
self.assertTrue(False) # This test is necessarily failed
def test_multivariate_normal_prior(self):
x = np.linspace(0, 1, 100, dtype=np.float32).reshape((-1, 2))
y = np.exp(-x @ x.T).reshape(-1)
gp = MaternGP(x, y, lengthscale_prior=((1, 0.01), (10, 1)))
kernel = gp.covar_module.base_kernel
prior = kernel.lengthscale_prior
self.assertIsInstance(gp.covar_module.base_kernel.lengthscale_prior,
gpytorch.priors.MultivariateNormalPrior)
self.assertTrue((kernel.lengthscale.squeeze() == prior.mean).all())
def get_gp(self, actions=None, rewards=None, gamma=0.5, n=11):
if actions is None or rewards is None:
actions, rewards = self.get_examples(n)
gp = MaternGP(
train_x=actions,
train_y=rewards,
lengthscale_prior=(0.1, 0.01),
noise_prior=(self.noise + 1e-6, 0.01),
noise_constraint=(self.noise / 10, self.noise * 10 + 1e-4),
value_structure_discount_factor=gamma,
)
return gp
def test_neighbor_erasing_dataset_2(self):
x = np.linspace(0, 1, 100, dtype=np.float32).reshape((-1, 1))
y = np.exp(-x**2).reshape(-1)
r = 0.0125
gp = MaternGP(x,
y,
noise_constraint=(0, 1e-3),
dataset_type='neighborerasing',
dataset_params={'radius': r})
x_ = np.linspace(0, 1, 20, dtype=np.float32).reshape((-1, 1))
y_ = np.exp(-x_**2).reshape(-1)
forgettable = [False] * len(y_)
gp.append_data(x_, y_, forgettable=forgettable)
x__ = np.linspace(0, 1, 21, dtype=np.float32).reshape((-1, 1))
y__ = np.exp(-x_**2).reshape(-1)
gp.append_data(x__, y__)
all_present_x_ = all(
x_ex.tolist() in gp.train_x.cpu().numpy().tolist() for x_ex in x_)
all_present_x__ = all(
x__ex.tolist() in gp.train_x.cpu().numpy().tolist()
for x__ex in x__)
self.assertTrue(all_present_x_ and all_present_x__)
def test_timeforgetting_dataset(self):
x = np.linspace(0, 1, 100, dtype=np.float32).reshape((-1, 1))
y = np.exp(-x**2).reshape(-1)
gp = MaternGP(x,
y,
noise_constraint=(0, 1e-3),
dataset_type='timeforgetting',
dataset_params={'keep': 50})
self.assertTrue((gp.train_x.cpu().numpy() == x[-50:]).all())
self.assertTrue((gp.train_y.cpu().numpy() == y[-50:]).all())
gp.append_data(x[:10], y[:10])
self.assertTrue((gp.train_x.cpu().numpy() == np.vstack(
(x[-40:], x[:10]))).all())
self.assertTrue((gp.train_y.cpu().numpy() == np.hstack(
(y[-40:], y[:10]))).all())
gp.set_data(x[:75], y[:75])
self.assertTrue((gp.train_x.cpu().numpy() == x[25:75]).all())
self.assertTrue((gp.train_y.cpu().numpy() == y[25:75]).all())
def test_initialization(self):
x = np.arange(18).reshape((6, 3))
y = np.arange(6)
gp = MaternGP(x, y)
self.assertTrue(torch.is_tensor(gp.train_x))
self.assertTrue(torch.is_tensor(gp.train_y))