def test_iaf():
# test for substitute logic for exposed methods `sample_posterior` and `get_transforms`
N, dim = 3000, 3
data = random.normal(random.PRNGKey(0), (N, dim))
true_coefs = jnp.arange(1.0, dim + 1.0)
logits = jnp.sum(true_coefs * data, axis=-1)
labels = dist.Bernoulli(logits=logits).sample(random.PRNGKey(1))
def model(data, labels):
coefs = numpyro.sample("coefs",
dist.Normal(jnp.zeros(dim), jnp.ones(dim)))
offset = numpyro.sample("offset", dist.Uniform(-1, 1))
logits = offset + jnp.sum(coefs * data, axis=-1)
return numpyro.sample("obs", dist.Bernoulli(logits=logits), obs=labels)
adam = optim.Adam(0.01)
rng_key_init = random.PRNGKey(1)
guide = AutoIAFNormal(model)
svi = SVI(model, guide, adam, Trace_ELBO())
svi_state = svi.init(rng_key_init, data, labels)
params = svi.get_params(svi_state)
x = random.normal(random.PRNGKey(0), (dim + 1, ))
rng_key = random.PRNGKey(1)
actual_sample = guide.sample_posterior(rng_key, params)
actual_output = guide._unpack_latent(guide.get_transform(params)(x))
flows = []
for i in range(guide.num_flows):
if i > 0:
flows.append(transforms.PermuteTransform(
jnp.arange(dim + 1)[::-1]))
arn_init, arn_apply = AutoregressiveNN(
dim + 1,
[dim + 1, dim + 1],
permutation=jnp.arange(dim + 1),
skip_connections=guide._skip_connections,
nonlinearity=guide._nonlinearity,
)
arn = partial(arn_apply, params["auto_arn__{}$params".format(i)])
flows.append(InverseAutoregressiveTransform(arn))
flows.append(guide._unpack_latent)
transform = transforms.ComposeTransform(flows)
_, rng_key_sample = random.split(rng_key)
expected_sample = transform(
dist.Normal(jnp.zeros(dim + 1), 1).sample(rng_key_sample))
expected_output = transform(x)
assert_allclose(actual_sample["coefs"], expected_sample["coefs"])
assert_allclose(
actual_sample["offset"],
transforms.biject_to(constraints.interval(-1, 1))(
expected_sample["offset"]),
)
check_eq(actual_output, expected_output)
def test_reparam_log_joint(model, kwargs):
guide = AutoIAFNormal(model)
svi = SVI(model, guide, Adam(1e-10), Trace_ELBO(), **kwargs)
svi_state = svi.init(random.PRNGKey(0))
params = svi.get_params(svi_state)
neutra = NeuTraReparam(guide, params)
reparam_model = neutra.reparam(model)
_, pe_fn, _, _ = initialize_model(random.PRNGKey(1), model, model_kwargs=kwargs)
init_params, pe_fn_neutra, _, _ = initialize_model(random.PRNGKey(2), reparam_model, model_kwargs=kwargs)
latent_x = list(init_params[0].values())[0]
pe_transformed = pe_fn_neutra(init_params[0])
latent_y = neutra.transform(latent_x)
log_det_jacobian = neutra.transform.log_abs_det_jacobian(latent_x, latent_y)
pe = pe_fn(guide._unpack_latent(latent_y))
assert_allclose(pe_transformed, pe - log_det_jacobian)
def test_neutra_reparam_unobserved_model():
model = dirichlet_categorical
data = jnp.ones(10, dtype=jnp.int32)
guide = AutoIAFNormal(model)
svi = SVI(model, guide, Adam(1e-3), Trace_ELBO())
svi_state = svi.init(random.PRNGKey(0), data)
params = svi.get_params(svi_state)
neutra = NeuTraReparam(guide, params)
reparam_model = neutra.reparam(model)
with handlers.seed(rng_seed=0):
reparam_model(data=None)
def test_neals_funnel_smoke():
dim = 10
guide = AutoIAFNormal(neals_funnel)
svi = SVI(neals_funnel, guide, Adam(1e-10), Trace_ELBO())
svi_state = svi.init(random.PRNGKey(0), dim)
def body_fn(i, val):
svi_state, loss = svi.update(val, dim)
return svi_state
svi_state = lax.fori_loop(0, 1000, body_fn, svi_state)
params = svi.get_params(svi_state)
neutra = NeuTraReparam(guide, params)
model = neutra.reparam(neals_funnel)
nuts = NUTS(model)
mcmc = MCMC(nuts, num_warmup=50, num_samples=50)
mcmc.run(random.PRNGKey(1), dim)
samples = mcmc.get_samples()
transformed_samples = neutra.transform_sample(samples['auto_shared_latent'])
assert 'x' in transformed_samples
assert 'y' in transformed_samples
