def test_categorical_likelihood(num_cats: int, cat_probs: np.ndarray,
hybridize: bool):
"""
Test to check that maximizing the likelihood recovers the parameters
"""
cat_prob = mx.nd.array(cat_probs)
cat_probs = mx.nd.zeros((NUM_SAMPLES, num_cats)) + cat_prob
distr = Categorical(cat_probs.log())
samples = distr.sample()
cat_prob_init = mx.nd.random_uniform(1 - TOL, 1 + TOL, num_cats) * cat_prob
cat_prob_init = cat_prob_init / cat_prob_init.sum()
init_biases = [cat_prob_init]
cat_log_prob_hat = maximum_likelihood_estimate_sgd(
CategoricalOutput(num_cats),
samples,
init_biases=init_biases,
hybridize=hybridize,
learning_rate=PositiveFloat(0.05),
num_epochs=PositiveInt(25),
)
cat_prob_hat = np.exp(cat_log_prob_hat)
prob_deviation = np.abs(cat_prob_hat - cat_prob.asnumpy()).flatten()
tolerance = (TOL * cat_prob.asnumpy()).flatten()
assert np.all(
np.less(prob_deviation, tolerance)
), f"cat_prob did not match: cat_prob = {cat_prob}, cat_prob_hat = {cat_prob_hat}"
def __init__(
self,
num_marks: int,
interval_length: float,
time_distr_output: TPPDistributionOutput = distribution.WeibullOutput(), # noqa: E501
embedding_dim: int = 5,
num_hidden_dimensions: int = 10,
output_scale: Optional[Tensor] = None,
apply_log_to_rnn_inputs: bool = True,
**kwargs,
) -> None:
super().__init__(**kwargs)
self.num_marks = num_marks
self.interval_length = interval_length
self.rnn_hidden_size = num_hidden_dimensions
self.output_scale = output_scale
self.apply_log_to_rnn_inputs = apply_log_to_rnn_inputs
with self.name_scope():
self.embedding = mx.gluon.nn.Embedding(
input_dim=num_marks, output_dim=embedding_dim
)
self.rnn = mx.gluon.rnn.GRU(
num_hidden_dimensions,
input_size=embedding_dim + 1,
layout="NTC",
)
# Conditional distribution over the inter-arrival times
self.time_distr_output = time_distr_output
self.time_distr_args_proj = self.time_distr_output.get_args_proj()
# Conditional distribution over the marks
if num_marks > 1:
self.mark_distr_output = CategoricalOutput(num_marks)
self.mark_distr_args_proj = (
self.mark_distr_output.get_args_proj()
)
def test_nanmixture_categorical_inference() -> None:
nmdo = NanMixtureOutput(CategoricalOutput(3))
args_proj = nmdo.get_args_proj()
args_proj.initialize()
args_proj.hybridize()
input = mx.nd.ones((NUM_SAMPLES))
trainer = mx.gluon.Trainer(
args_proj.collect_params(), "sgd", {"learning_rate": 0.000002}
)
mixture_samples = mx.nd.array(cat_samples)
N = 3000
t = tqdm(list(range(N)))
for _ in t:
with mx.autograd.record():
distr_args = args_proj(input)
d = nmdo.distribution(distr_args)
loss = d.loss(mixture_samples)
loss.backward()
loss_value = loss.mean().asnumpy()
t.set_postfix({"loss": loss_value})
trainer.step(NUM_SAMPLES)
distr_args = args_proj(input)
d = nmdo.distribution(distr_args)
cat_probs_hat = d.distribution.probs.asnumpy()
nan_prob_hat = d.nan_prob.asnumpy()
assert np.allclose(
cat_probs, cat_probs_hat, atol=TOL
), f"categorical dist: cat_probs did not match: cat_probs = {cat_probs}, cat_probs_hat = {cat_probs_hat}"
assert (
np.abs(nan_prob - nan_prob_hat) < TOL
), f"categorical dist: nan_prob did not match: nan_prob = {nan_prob}, nan_prob_hat = {nan_prob_hat}"
), f"categorical dist: nan_prob did not match: nan_prob = {nan_prob}, nan_prob_hat = {nan_prob_hat}"
n_cat = 3
cat_probs = np.array([0.2, 0.3, 0.5])
cat_samples = np.random.choice(
list(range(n_cat)), p=cat_probs, size=NUM_SAMPLES
)
cat_samples = np.where(
np.random.uniform(size=(NUM_SAMPLES)) > nan_prob, cat_samples, np.nan
)
@pytest.mark.parametrize(
"distribution_output",
[GaussianOutput(), StudentTOutput(), CategoricalOutput(num_cats=2),],
)
@pytest.mark.parametrize("serialize_fn", serialize_fn_list)
def test_nanmixture_output(distribution_output, serialize_fn) -> None:
nmdo = NanMixtureOutput(distribution_output)
args_proj = nmdo.get_args_proj()
args_proj.initialize()
input = mx.nd.ones(shape=(3, 2))
distr_args = args_proj(input)
d = nmdo.distribution(distr_args)
d = serialize_fn(d)
class DeepTPPNetworkBase(mx.gluon.HybridBlock):
"""
Temporal point process model based on a recurrent neural network.
Parameters
----------
num_marks
Number of discrete marks (correlated processes), that are available
in the data.
interval_length
The length of the total time interval that is in the prediction
range. Note that in contrast to discrete-time models in the rest
of GluonTS, the network is trained to predict an interval, in
continuous time.
time_distr_output
Output distribution for the inter-arrival times. Available
distributions can be found in gluonts.model.tpp.distribution.
embedding_dim
Dimension of vector embeddings of marks (used only as input).
num_hidden_dimensions
Number of hidden units in the RNN.
output_scale
Positive scaling applied to the inter-event times. You should provide
this argument if the average inter-arrival time is much larger than 1.
apply_log_to_rnn_inputs
Apply logarithm to inter-event times that are fed into the RNN.
"""
@validated()
def __init__(
self,
num_marks: int,
interval_length: float,
time_distr_output: TPPDistributionOutput = distribution.WeibullOutput(), # noqa: E501
embedding_dim: int = 5,
num_hidden_dimensions: int = 10,
output_scale: Optional[Tensor] = None,
apply_log_to_rnn_inputs: bool = True,
**kwargs,
) -> None:
super().__init__(**kwargs)
self.num_marks = num_marks
self.interval_length = interval_length
self.rnn_hidden_size = num_hidden_dimensions
self.output_scale = output_scale
self.apply_log_to_rnn_inputs = apply_log_to_rnn_inputs
with self.name_scope():
self.embedding = mx.gluon.nn.Embedding(
input_dim=num_marks, output_dim=embedding_dim
)
self.rnn = mx.gluon.rnn.GRU(
num_hidden_dimensions,
input_size=embedding_dim + 1,
layout="NTC",
)
# Conditional distribution over the inter-arrival times
self.time_distr_output = time_distr_output
self.time_distr_args_proj = self.time_distr_output.get_args_proj()
# Conditional distribution over the marks
if num_marks > 1:
self.mark_distr_output = CategoricalOutput(num_marks)
self.mark_distr_args_proj = (
self.mark_distr_output.get_args_proj()
)
def hybridize(self, active=True, **kwargs):
if active:
raise NotImplementedError(
"DeepTPP blocks do not support hybridization"
)
PiecewiseLinearOutput,
PoissonOutput,
StudentTOutput,
UniformOutput,
ZeroAndOneInflatedBetaOutput,
ZeroInflatedBetaOutput,
ZeroInflatedNegativeBinomialOutput,
ZeroInflatedPoissonOutput,
)
@pytest.mark.parametrize(
"distr_output",
[
BetaOutput(),
CategoricalOutput(num_cats=3),
DeterministicOutput(value=42.0),
DirichletMultinomialOutput(dim=3, n_trials=5),
DirichletOutput(dim=4),
EmpiricalDistributionOutput(num_samples=10,
distr_output=GaussianOutput()),
GammaOutput(),
GaussianOutput(),
GenParetoOutput(),
LaplaceOutput(),
LogitNormalOutput(),
LoglogisticOutput(),
LowrankMultivariateGaussianOutput(dim=5, rank=2),
MultivariateGaussianOutput(dim=4),
NegativeBinomialOutput(),
OneInflatedBetaOutput(),
