Make Predictive work with the SplitReparam reparameterizer [bugfix] (#…

…3388)

pyro/infer/predictive.py

@@ -10,6 +10,7 @@
 
 import pyro
 import pyro.poutine as poutine
+from pyro.infer.autoguide.initialization import InitMessenger, init_to_sample
 from pyro.infer.importance import LogWeightsMixin
 from pyro.infer.util import CloneMixin, plate_log_prob_sum
 from pyro.poutine.trace_struct import Trace
@@ -86,12 +87,15 @@ def _predictive(
     mask=True,
 ):
     model = torch.no_grad()(poutine.mask(model, mask=False) if mask else model)
-    max_plate_nesting = _guess_max_plate_nesting(model, model_args, model_kwargs)
+    initailized_model = InitMessenger(init_to_sample)(model)
+    max_plate_nesting = _guess_max_plate_nesting(
+        initailized_model, model_args, model_kwargs
+    )
     vectorize = pyro.plate(
         _predictive_vectorize_plate_name, num_samples, dim=-max_plate_nesting - 1
     )
     model_trace = prune_subsample_sites(
-        poutine.trace(model).get_trace(*model_args, **model_kwargs)
+        poutine.trace(initailized_model).get_trace(*model_args, **model_kwargs)
     )
     reshaped_samples = {}
 

tests/infer/reparam/test_split.py

@@ -13,8 +13,7 @@
 
 from .util import check_init_reparam
 
-
-@pytest.mark.parametrize(
+event_shape_splits_dim = pytest.mark.parametrize(
     "event_shape,splits,dim",
     [
         ((6,), [2, 1, 3], -1),
@@ -31,7 +30,13 @@
     ],
     ids=str,
 )
-@pytest.mark.parametrize("batch_shape", [(), (4,), (3, 2)], ids=str)
+
+
+batch_shape = pytest.mark.parametrize("batch_shape", [(), (4,), (3, 2)], ids=str)
+
+
+@event_shape_splits_dim
+@batch_shape
 def test_normal(batch_shape, event_shape, splits, dim):
     shape = batch_shape + event_shape
     loc = torch.empty(shape).uniform_(-1.0, 1.0).requires_grad_()
@@ -72,24 +77,8 @@ def model():
     assert_close(actual_grads, expected_grads)
 
 
-@pytest.mark.parametrize(
-    "event_shape,splits,dim",
-    [
-        ((6,), [2, 1, 3], -1),
-        (
-            (
-                2,
-                5,
-            ),
-            [2, 3],
-            -1,
-        ),
-        ((4, 2), [1, 3], -2),
-        ((2, 3, 1), [1, 2], -2),
-    ],
-    ids=str,
-)
-@pytest.mark.parametrize("batch_shape", [(), (4,), (3, 2)], ids=str)
+@event_shape_splits_dim
+@batch_shape
 def test_init(batch_shape, event_shape, splits, dim):
     shape = batch_shape + event_shape
     loc = torch.empty(shape).uniform_(-1.0, 1.0)
@@ -100,3 +89,49 @@ def model():
             return pyro.sample("x", dist.Normal(loc, scale).to_event(len(event_shape)))
 
     check_init_reparam(model, SplitReparam(splits, dim))
+
+
+@event_shape_splits_dim
+@batch_shape
+def test_predictive(batch_shape, event_shape, splits, dim):
+    shape = batch_shape + event_shape
+    loc = torch.empty(shape).uniform_(-1.0, 1.0)
+    scale = torch.empty(shape).uniform_(0.5, 1.5)
+
+    def model():
+        with pyro.plate_stack("plates", batch_shape):
+            pyro.sample("x", dist.Normal(loc, scale).to_event(len(event_shape)))
+
+    # Reparametrize model
+    rep = SplitReparam(splits, dim)
+    reparam_model = poutine.reparam(model, {"x": rep})
+
+    # Fit guide to reparametrized model
+    guide = pyro.infer.autoguide.guides.AutoMultivariateNormal(reparam_model)
+    optimizer = pyro.optim.Adam(dict(lr=0.01))
+    loss = pyro.infer.JitTrace_ELBO(
+        num_particles=20, vectorize_particles=True, ignore_jit_warnings=True
+    )
+    svi = pyro.infer.SVI(reparam_model, guide, optimizer, loss)
+    for count in range(1001):
+        loss = svi.step()
+        if count % 100 == 0:
+            print(f"iteration {count} loss = {loss}")
+
+    # Sample from model using the guide
+    num_samples = 100000
+    parallel = True
+    sites = ["x_split_{}".format(i) for i in range(len(splits))]
+    values = pyro.infer.Predictive(
+        reparam_model,
+        guide=guide,
+        num_samples=num_samples,
+        parallel=parallel,
+        return_sites=sites,
+    )()
+
+    # Verify sampling
+    mean = torch.cat([values[site].mean(0) for site in sites], dim=dim)
+    std = torch.cat([values[site].std(0) for site in sites], dim=dim)
+    assert_close(mean, loc, atol=0.1)
+    assert_close(std, scale, rtol=0.1)