def test_serialize_and_load(self, shape):
def make_model(context_):
alpha = context_.named_parameter(
"alpha", inf.Prior(Normal, loc=0.0, scale=1.0))
beta = context_.named_parameter(
"beta", inf.Prior(LogNormal, loc=0.0, scale=1.0))
return ts.models.OrnsteinUhlenbeck(alpha, beta, beta)
with inf.make_context() as context:
make_model(context)
context.initialize_parameters(shape)
as_state = context.state_dict()
with inf.make_context() as new_context:
model = make_model(new_context)
new_context.load_state_dict(as_state)
assert (context.stack_parameters() ==
new_context.stack_parameters()).all()
for p1, p2 in zip(model.functional_parameters(),
new_context.parameters.values()):
assert p1 is p2
def test_exchange_context(self):
with inf.make_context() as context:
alpha = context.named_parameter(
"alpha", inf.Prior(Normal, loc=0.0, scale=1.0))
beta = context.named_parameter(
"beta", inf.Prior(LogNormal, loc=0.0, scale=1.0))
context.initialize_parameters(batch_shape)
with context.make_new() as sub_context:
alpha_sub = sub_context.named_parameter(
"alpha", inf.Prior(Normal, loc=0.0, scale=1.0))
beta_sub = sub_context.named_parameter(
"beta", inf.Prior(LogNormal, loc=0.0, scale=1.0))
assert alpha_sub is not alpha and beta_sub is not beta
sub_context.initialize_parameters(batch_shape)
mask: torch.BoolTensor = (torch.empty(
batch_shape[0]).normal_() > 0.0).bool()
context.exchange(sub_context, mask)
assert (alpha[mask]
== alpha_sub[mask]).all() and (beta[mask]
== beta_sub[mask]).all()
def test_verify_not_batched(self):
with inf.make_context() as context:
with pytest.raises(AssertionError):
beta = context.named_parameter(
"beta",
inf.Prior(LogNormal,
loc=torch.zeros(1),
scale=torch.ones(1)))
def test_apply_fun(self, shape):
with inf.make_context() as context:
beta = context.named_parameter(
"beta", inf.Prior(LogNormal, loc=0.0, scale=1.0))
context.initialize_parameters(shape)
sub_context = context.apply_fun(lambda u: u.mean())
for p, v in sub_context.parameters.items():
assert v.shape == torch.Size([])
def test_sample_parameters(self):
with inf.make_context() as context:
alpha = context.named_parameter(
"alpha", inf.Prior(Normal, loc=0.0, scale=1.0))
beta = context.named_parameter(
"beta", inf.Prior(LogNormal, loc=0.0, scale=1.0))
assert isinstance(context.get_parameter(alpha._name), inf.PriorBoundParameter) and \
isinstance(context.get_parameter(beta._name), inf.PriorBoundParameter)
context.initialize_parameters(batch_shape)
assert alpha.shape == batch_shape and beta.shape == batch_shape
def test_assert_sampling_multiple_same(self):
with inf.make_context() as context:
alpha = context.named_parameter(
"alpha", inf.Prior(Normal, loc=0.0, scale=1.0))
beta = context.named_parameter(
"beta", inf.Prior(LogNormal, loc=0.0, scale=1.0))
alpha2 = context.named_parameter(
"alpha", inf.Prior(Normal, loc=0.0, scale=1.0))
with pytest.raises(AssertionError):
alpha = context.named_parameter(
"alpha", inf.Prior(Normal, loc=0.0, scale=2.0))
def test_check_inactive_context_raises(self, models, filter_):
model, _ = models
from pyfilter.inference import make_context
context = make_context()
def model_builder(context_):
return model
# with pytest.raises()
with pytest.raises(Exception):
f = filter_(model_builder)
def test_algorithms(self, models, algorithm, callback):
torch.manual_seed(123)
true_model, build_model = models
_, y = true_model.sample_states(100).get_paths()
with inf.make_context() as context:
filter_ = filts.APF(build_model, 200)
alg = algorithm(filter_)
alg.register_callback(callback)
# TODO: Add something to test
result = alg.fit(y)
def test_multidimensional_parameters(self, shape):
with inf.make_context() as context:
alpha = context.named_parameter(
"alpha",
inf.Prior(Normal,
loc=torch.zeros((2, 2)),
scale=torch.ones((2, 2)),
reinterpreted_batch_ndims=2))
beta = context.named_parameter(
"beta", inf.Prior(LogNormal, loc=0.0, scale=1.0))
context.initialize_parameters(shape)
stacked = context.stack_parameters()
assert stacked.shape == torch.Size([shape.numel(), 5])
def test_algorithms_serialize(self, models, algorithm, callback):
torch.manual_seed(123)
true_model, build_model = models
_, y = true_model.sample_states(100).get_paths()
train_split = y.shape[0] // 2
particles = 250
with inf.make_context() as context:
filter_ = filts.APF(build_model, particles)
alg = algorithm(filter_)
alg.register_callback(callback)
result = alg.fit(y[:train_split])
algorithm_state = result.state_dict()
context_state = context.state_dict()
with inf.make_context() as new_context:
new_filter = filts.APF(build_model, particles)
new_context.load_state_dict(context_state)
new_alg = algorithm(new_filter)
new_result = new_alg.initialize()
new_result.load_state_dict(algorithm_state)
assert ((new_result.ess == result.ess).all()
and (new_result.w == result.w).all())
for yt in y[train_split:]:
new_result = new_alg.step(yt, new_result)
assert ((new_result.ess.shape[0] == y.shape[0] + 1)
and (new_result.filter_state.latest_state.x.time_index
== y.shape[0]).all())
def test_pmcmc(self, models, kernel_and_record_states):
true_model, build_model = models
_, y = true_model.sample_states(50).get_paths()
with inf.make_context() as context:
kernel, record_states = kernel_and_record_states
filter_ = filts.APF(lambda u: build_model(u, use_cuda=False), 150, record_states=record_states)
# TODO: Just make sure it runs
pmcmc = inf.batch.mcmc.PMMH(filter_, 100, initializer="mean", proposal=kernel)
result = pmcmc.fit(y)
# TODO: Add something to test
print()
def test_make_model_and_resample(self):
with inf.make_context() as context:
model = build_model(context)
context.initialize_parameters(batch_shape)
old_dict = {k: v.clone() for k, v in context.parameters.items()}
indices: torch.IntTensor = torch.randint(low=0,
high=batch_shape[0],
size=batch_shape[:1])
context.resample(indices)
for p_model, (n, p) in zip(model.hidden.functional_parameters(),
context.get_parameters()):
assert (p == old_dict[n][indices]).all() and (p_model is p)
def test_resample_context(self):
with inf.make_context() as context:
alpha = context.named_parameter(
"alpha", inf.Prior(Normal, loc=0.0, scale=1.0))
beta = context.named_parameter(
"beta", inf.Prior(LogNormal, loc=0.0, scale=1.0))
context.initialize_parameters(batch_shape)
indices: torch.IntTensor = torch.randint(low=0,
high=batch_shape[0],
size=batch_shape[:1])
alpha_clone = alpha.clone()
beta_clone = beta.clone()
context.resample(indices)
assert (alpha == alpha_clone[indices]).all() and (
beta == beta_clone[indices]).all()
def test_initialize_parameter(self, batch_shape):
with inf.make_context() as cntxt:
prior = inf.Prior(Normal, loc=0.0, scale=1.0)
parameter = cntxt.named_parameter("kappa", prior)
sts = StructuralStochasticProcess((parameter, ), None)
assert (next(sts.parameters()) is
parameter) and (cntxt.get_parameter("kappa") is parameter)
for p in sts.parameters():
p.sample_(batch_shape)
assert (p is parameter) and (p.shape == batch_shape)
if not torch.cuda.is_available():
return
sts = sts.cuda()
assert (next(sts.parameters()) is
parameter) and (cntxt.get_parameter("kappa") is parameter)
for p in sts.parameters():
p.sample_(batch_shape)
assert (p is parameter) and (p.shape == batch_shape)
def test_assert_fails_register_inactive(self):
context = inf.make_context()
with pytest.raises(AssertionError):
a = context.named_parameter("a",
inf.Prior(Normal, loc=0.0, scale=1.0))
