def test_variational_complete():
# construct objects
train_inputs, train_outputs, test_inputs, num_train, inducing_inputs = construct_input(
)
input_dim = 1
output_dim = 1
args = dict(num_samples=5000000,
num_components=1,
optimize_inducing=False,
use_loo=True,
diag_post=False,
sn=1.0,
length_scale=0.5,
sf=1.0,
iso=False,
cov='SquaredExponential')
vi = inference.Variational(args, 'LikelihoodGaussian', output_dim,
num_train, inducing_inputs)
vi.build((num_train, input_dim))
# compute losses and predictions
losses = vi.inference(train_inputs, train_outputs, True)
nelbo = losses['NELBO']
loo = losses['LOO_VARIATIONAL']
pred_mean, pred_var = vi.prediction(test_inputs)
# check results
np.testing.assert_allclose(nelbo.numpy(), 3.844, rtol=1e-3)
np.testing.assert_allclose(
loo.numpy(), 4.45, rtol=1e-2) # test with a relative tolerance of 1%
np.testing.assert_allclose(pred_mean.numpy(), 0.0, RTOL)
np.testing.assert_allclose(tf.squeeze(pred_var).numpy(), 2.0, 1e-2)
def construct_simple_full():
input_dim = 1
output_dim = 1
num_train = 1
inf = inference.Variational(PARAMS, 'LikelihoodGaussian', output_dim,
num_train, num_train)
inf.build((num_train, input_dim))
return inf
def construct_simple_full():
likelihood = lik.LikelihoodGaussian({'sn': 1.0})
kernel = [
cov.SquaredExponential(input_dim=1,
args=dict(length_scale=1.0, sf=1.0, iso=False))
]
# In most of our unit test, we will replace this value with something else.
return inference.Variational(kernel, likelihood, 1, 1, PARAMS)
def construct_simple_diag():
input_dim = 1
output_dim = 1
num_train = 1
inf = inference.Variational({
**PARAMS, 'diag_post': True
}, 'LikelihoodGaussian', output_dim, num_train, num_train)
inf.build((num_train, input_dim))
return inf
def construct_simple_diag():
likelihood = lik.LikelihoodGaussian({'sn': 1.0})
kernel = [
cov.SquaredExponential(input_dim=1,
args=dict(length_scale=1.0, sf=1.0, iso=False))
]
return inference.Variational(kernel, likelihood, 1, 1, {
**PARAMS, 'diag_post': True
})
def construct_multi_full():
likelihood = lik.LikelihoodSoftmax({'num_samples_pred': 100})
kernels = [
cov.SquaredExponential(input_dim=2,
args=dict(length_scale=1.0, sf=1.0, iso=False))
for _ in range(2)
]
return inference.Variational(kernels, likelihood, 1, 1, {
**PARAMS, 'num_components': 2
})
def construct_multi_full():
input_dim = 2
output_dim = 2
num_train = 1
inf = inference.Variational(
{
**PARAMS, 'num_samples_pred': 100,
'num_components': 2
}, 'LikelihoodSoftmax', output_dim, num_train, num_train)
inf.build((num_train, input_dim))
return inf
def test_variational_complete():
# construct objects
train_inputs, train_outputs, test_inputs, num_train, inducing_inputs = construct_input()
likelihood = lik.LikelihoodGaussian({'sn': 1.0})
kernel = [cov.SquaredExponential(input_dim=1, args=dict(length_scale=0.5, sf=1.0, iso=False))]
vi = inference.Variational(kernel, likelihood, num_train, inducing_inputs,
{'num_samples': 5000000, 'num_components': 1, 'optimize_inducing': False,
'use_loo': True, 'diag_post': False})
# compute losses and predictions
losses, _ = vi.inference(train_inputs, train_outputs, True)
nelbo = losses['NELBO']
loo = losses['LOO_VARIATIONAL']
pred_mean, pred_var = vi.predict(test_inputs)
# check results
np.testing.assert_allclose(nelbo.numpy(), 3.844, rtol=1e-3)
np.testing.assert_allclose(loo.numpy(), 4.45, rtol=1e-2) # test with a relative tolerance of 1%
np.testing.assert_allclose(pred_mean.numpy(), 0.0, RTOL)
np.testing.assert_allclose(tf.squeeze(pred_var).numpy(), 2.0, 1e-2)