def test_multivariate_gaussian() -> None:
num_samples = 2000
dim = 2
mu = np.arange(0, dim) / float(dim)
L_diag = np.ones((dim, ))
L_low = 0.1 * np.ones((dim, dim)) * np.tri(dim, k=-1)
L = np.diag(L_diag) + L_low
Sigma = L.dot(L.transpose())
distr = MultivariateGaussian(mu=mx.nd.array(mu), L=mx.nd.array(L))
samples = distr.sample(num_samples)
mu_hat, L_hat = maximum_likelihood_estimate_sgd(
MultivariateGaussianOutput(dim=dim),
samples,
init_biases=
None, # todo we would need to rework biases a bit to use it in the multivariate case
hybridize=False,
learning_rate=PositiveFloat(0.01),
num_epochs=PositiveInt(10),
)
distr = MultivariateGaussian(mu=mx.nd.array([mu_hat]),
L=mx.nd.array([L_hat]))
Sigma_hat = distr.variance[0].asnumpy()
assert np.allclose(
mu_hat, mu, atol=0.1,
rtol=0.1), f"mu did not match: mu = {mu}, mu_hat = {mu_hat}"
assert np.allclose(
Sigma_hat, Sigma, atol=0.1, rtol=0.1
), f"Sigma did not match: sigma = {Sigma}, sigma_hat = {Sigma_hat}"
(
StudentTOutput(),
mx.nd.random.normal(shape=(3, 4, 5, 6)),
[None, mx.nd.ones(shape=(3, 4, 5))],
(3, 4, 5),
(),
),
(
GammaOutput(),
mx.nd.random.gamma(shape=(3, 4, 5, 6)),
[None, mx.nd.ones(shape=(3, 4, 5))],
(3, 4, 5),
(),
),
(
MultivariateGaussianOutput(dim=5),
mx.nd.random.normal(shape=(3, 4, 10)),
[None, mx.nd.ones(shape=(3, 4, 5))],
(3, 4),
(5, ),
),
(
LowrankMultivariateGaussianOutput(dim=5, rank=4),
mx.nd.random.normal(shape=(3, 4, 10)),
[None, mx.nd.ones(shape=(3, 4, 5))],
(3, 4),
(5, ),
),
(
DirichletOutput(dim=5),
mx.nd.random.gamma(shape=(3, 4, 5)),
< 0.05
)
# can only calculated cdf for gaussians currently
if isinstance(distr1, Gaussian) and isinstance(distr2, Gaussian):
emp_cdf, edges = empirical_cdf(samples_mix.asnumpy())
calc_cdf = mixture.cdf(mx.nd.array(edges)).asnumpy()
assert np.allclose(calc_cdf[1:, :], emp_cdf, atol=1e-2)
@pytest.mark.parametrize(
"distribution_outputs",
[
((GaussianOutput(), GaussianOutput()),),
((GaussianOutput(), StudentTOutput(), LaplaceOutput()),),
((MultivariateGaussianOutput(3), MultivariateGaussianOutput(3)),),
],
)
def test_mixture_output(distribution_outputs) -> None:
mdo = MixtureDistributionOutput(*distribution_outputs)
args_proj = mdo.get_args_proj()
args_proj.initialize()
input = mx.nd.ones(shape=(512, 30))
distr_args = args_proj(input)
d = mdo.distribution(distr_args)
samples = d.sample(num_samples=NUM_SAMPLES)
LowrankMultivariateGaussianOutput(dim=target_dim, rank=2),
10,
estimator,
False,
False,
),
(
LowrankMultivariateGaussianOutput(dim=target_dim, rank=2),
10,
estimator,
True,
False,
),
(None, 10, estimator, True, True),
(
MultivariateGaussianOutput(dim=target_dim),
10,
estimator,
False,
True,
),
(
MultivariateGaussianOutput(dim=target_dim),
10,
estimator,
True,
True,
),
],
)
def test_deepvar(
plt.plot(custom_datasetx[0])
plt.show()
start = pd.Timestamp("01-01-2019", freq=freq)
train_ds = [{
'target': x,
'start': start
} for x in custom_datasetx[:, :, :-prediction_length]]
test_ds = [{'target': x, 'start': start} for x in custom_datasetx[:, :, :]]
# Trainer parameters
epochs = 1
learning_rate = 1E-3
batch_size = 1
num_batches_per_epoch = 2
# create estimator
estimator = DeepAREstimator(
prediction_length=prediction_length,
context_length=prediction_length,
freq=freq,
# trainer=Trainer(ctx="gpu", epochs=epochs, learning_rate=learning_rate, hybridize=True,
# batch_size=batch_size, num_batches_per_epoch=num_batches_per_epoch,),
distr_output=MultivariateGaussianOutput(dim=2),
)
predictor = estimator.train(train_ds)