Associate transforms with random variables

Author: rlouf

aeppl/joint_logprob.py

@@ -110,8 +110,16 @@ def conditional_logprob(
 
     fgraph, rv_values, _ = construct_ir_fgraph(rv_values, ir_rewriter=ir_rewriter)
 
+    # The interface for transformations assumes that the value variables are in
+    # the transformed space. To get the correct `shape` and `dtype` for the
+    # value variables we return we need to apply the forward transformation to
+    # our RV copies, and return the type of the resulting variable as a value
+    # variable.
+    vv_remapper = {}
     if extra_rewrites is not None:
-        extra_rewrites.rewrite(fgraph)
+        extra_rewrites.add_requirements(fgraph, {**original_rv_values, **realized})
+        extra_rewrites.apply(fgraph)
+        vv_remapper = fgraph.values_to_untransformed
 
     rv_remapper = fgraph.preserve_rv_mappings
 
@@ -145,6 +153,7 @@ def conditional_logprob(
     q = deque(fgraph.toposort())
 
     logprob_vars = {}
+    value_variables = {}
 
     while q:
         node = q.popleft()
@@ -201,6 +210,9 @@ def conditional_logprob(
 
             logprob_vars[q_rv] = q_logprob_var
 
+            q_value_var = vv_remapper.get(q_value_var, q_value_var)
+            value_variables[q_rv] = q_value_var
+
         # Recompute test values for the changes introduced by the
         # replacements above.
         if config.compute_test_value != "off":
@@ -213,7 +225,7 @@ def conditional_logprob(
             f"The logprob terms of the following random variables could not be derived: {missing_value_terms}"
         )
 
-    return logprob_vars, list(original_rv_values.values())
+    return logprob_vars, [value_variables[rv] for rv in original_rv_values.keys()]
 
 
 def joint_logprob(

aeppl/transforms.py

@@ -122,9 +122,14 @@ def transform_values(fgraph: FunctionGraph, node: Node) -> Optional[List[Node]]:
     """
 
     rv_map_feature = getattr(fgraph, "preserve_rv_mappings", None)
+    values_to_untransformed = getattr(fgraph, "values_to_untransformed", None)
     values_to_transforms = getattr(fgraph, "values_to_transforms", None)
 
-    if rv_map_feature is None or values_to_transforms is None:
+    if (
+        rv_map_feature is None
+        or values_to_transforms is None
+        or values_to_untransformed is None
+    ):
         return None  # pragma: no cover
 
     try:
@@ -133,6 +138,7 @@ def transform_values(fgraph: FunctionGraph, node: Node) -> Optional[List[Node]]:
     except ValueError:
         return None
 
+    value_var: TensorVariable
     value_var = rv_map_feature.rv_values.get(rv_var, None)
     if value_var is None:
         return None
@@ -154,10 +160,21 @@ def transform_values(fgraph: FunctionGraph, node: Node) -> Optional[List[Node]]:
         trans_node.outputs[rv_var_out_idx].name = node.outputs[rv_var_out_idx].name
 
     # We now assume that the old value variable represents the *transformed space*.
-    # This means that we need to replace all instance of the old value variable
+
+    # Since we initialize value variables as copies of the random variables,
+    # thus in the untransformed space, we need to apply the forward
+    # transformation to get value variables in the transformed space.
+    old_value_var: TensorVariable = transform.forward(
+        value_var, *trans_node.inputs
+    ).type()
+    if value_var.name:
+        old_value_var.name = value_var.name
+    values_to_untransformed[value_var] = old_value_var
+
+    # We need to replace all instance of the old value variable
     # with "inversely/un-" transformed versions of itself.
     new_value_var = transformed_variable(
-        transform.backward(value_var, *trans_node.inputs), value_var
+        transform.backward(old_value_var, *trans_node.inputs), old_value_var
     )
     if value_var.name and getattr(transform, "name", None):
         new_value_var.name = f"{value_var.name}_{transform.name}"
@@ -170,16 +187,24 @@ def transform_values(fgraph: FunctionGraph, node: Node) -> Optional[List[Node]]:
 
 
 class TransformValuesMapping(Feature):
-    r"""A `Feature` that maintains a map between value variables and their transforms."""
+    r"""A `Feature` that maintains a map between value variables and their transforms as
+    well as between value variables and their transformed counterpart.
+
+    """
 
     def __init__(self, values_to_transforms):
         self.values_to_transforms = values_to_transforms
+        self.values_to_untransformed: Dict[TensorVariable, TensorVariable] = {}
 
     def on_attach(self, fgraph):
         if hasattr(fgraph, "values_to_transforms"):
             raise AlreadyThere()
 
         fgraph.values_to_transforms = self.values_to_transforms
+        fgraph.values_to_untransformed = self.values_to_untransformed
+
+    def update_untransformed_value(self, value, untransformed_value):
+        self.values_to_untransformed[value] = untransformed_value
 
 
 class TransformValuesRewrite(GraphRewriter):
@@ -189,25 +214,31 @@ class TransformValuesRewrite(GraphRewriter):
 
     def __init__(
         self,
-        values_to_transforms: Dict[
+        rvs_to_transforms: Dict[
             TensorVariable, Union[RVTransform, DefaultTransformSentinel, None]
         ],
     ):
         """
         Parameters
         ==========
         values_to_transforms
-            Mapping between value variables and their transformations.  Each
-            value variable can be assigned one of `RVTransform`,
-            ``DEFAULT_TRANSFORM``, or ``None``. If a transform is not specified
-            for a specific value variable it will not be transformed.
+            Mapping between random variables and their transformations. Each
+            random variable can be assigned one of `RVTransform`,
+            ``DEFAULT_TRANSFORM``, or ``None``. Random variables with no
+            transform specified remain unchanged.
 
         """
 
-        self.values_to_transforms = values_to_transforms
+        self.rvs_to_transforms = rvs_to_transforms
 
-    def add_requirements(self, fgraph):
-        values_transforms_feature = TransformValuesMapping(self.values_to_transforms)
+    def add_requirements(
+        self, fgraph, rv_to_values: Dict[TensorVariable, TensorVariable]
+    ):
+        values_to_transforms = {
+            rv_to_values[rv]: transform
+            for rv, transform in self.rvs_to_transforms.items()
+        }
+        values_transforms_feature = TransformValuesMapping(values_to_transforms)
         fgraph.attach_feature(values_transforms_feature)
 
     def apply(self, fgraph: FunctionGraph):

tests/test_censoring.py

@@ -154,10 +154,8 @@ def test_clip_transform():
     x_rv = at.random.normal(0.5, 1)
     cens_x_rv = at.clip(x_rv, 0, x_rv)
 
-    cens_x_vv = cens_x_rv.clone()
-
-    transform = TransformValuesRewrite({cens_x_vv: LogTransform()})
-    logp, _ = joint_logprob(realized={cens_x_rv: cens_x_vv}, extra_rewrites=transform)
+    transform = TransformValuesRewrite({cens_x_rv: LogTransform()})
+    logp, (cens_x_vv,) = joint_logprob(cens_x_rv, extra_rewrites=transform)
 
     cens_x_vv_testval = -1
     obs_logp = logp.eval({cens_x_vv: cens_x_vv_testval})

tests/test_joint_logprob.py

@@ -247,6 +247,6 @@ def test_multiple_rvs_to_same_value_raises():
     x = x_rv1.type()
     x.name = "x"
 
-    msg = "More than one logprob factor was assigned to the value variable x"
+    msg = "More than one logprob factor was assigned to the random variable x"
     with pytest.raises(ValueError, match=msg):
         joint_logprob(realized={x_rv1: x, x_rv2: x})

tests/test_mixture.py

@@ -41,7 +41,10 @@ def create_mix_model(size, axis):
     I_rv = env["I_rv"]
     M_rv = env["M_rv"]
 
-    with pytest.raises(RuntimeError, match="could not be derived: {m}"):
+    with pytest.raises(
+        RuntimeError,
+        match="The logprob terms of the following random variables could not be derived: {M}",
+    ):
         conditional_logprob(M_rv, I_rv, X_rv)
 
     with pytest.raises(NotImplementedError):

tests/test_transforms.py

@@ -4,7 +4,6 @@
 import pytest
 import scipy as sp
 import scipy.special
-from aesara.graph.basic import equal_computations
 from aesara.graph.fg import FunctionGraph
 from numdifftools import Jacobian
 
@@ -22,7 +21,6 @@
     TransformValuesMapping,
     TransformValuesRewrite,
     _default_transformed_rv,
-    transformed_variable,
 )
 from tests.utils import assert_no_rvs
 
@@ -176,15 +174,13 @@ def test_transformed_logprob(at_dist, dist_params, sp_dist, size):
 
     a = at_dist(*dist_params, size=size)
     a.name = "a"
-    a_value_var = at.tensor(a.dtype, shape=(None,) * a.ndim)
-    a_value_var.name = "a_value"
 
     b = at.random.normal(a, 1.0)
     b.name = "b"
 
-    transform_rewrite = TransformValuesRewrite({a_value_var: DEFAULT_TRANSFORM})
-    res, (b_value_var,) = joint_logprob(
-        b, realized={a: a_value_var}, extra_rewrites=transform_rewrite
+    transform_rewrite = TransformValuesRewrite({a: DEFAULT_TRANSFORM})
+    res, (b_value_var, a_value_var) = joint_logprob(
+        b, a, extra_rewrites=transform_rewrite
     )
 
     test_val_rng = np.random.RandomState(3238)
@@ -268,12 +264,10 @@ def a_backward_fn_(x):
 @pytest.mark.parametrize("use_jacobian", [True, False])
 def test_simple_transformed_logprob_nojac(use_jacobian):
     X_rv = at.random.halfnormal(0, 3, name="X")
-    x_vv = X_rv.clone()
-    x_vv.name = "x"
 
-    transform_rewrite = TransformValuesRewrite({x_vv: DEFAULT_TRANSFORM})
-    tr_logp, _ = joint_logprob(
-        realized={X_rv: x_vv},
+    transform_rewrite = TransformValuesRewrite({X_rv: DEFAULT_TRANSFORM})
+    tr_logp, (x_vv,) = joint_logprob(
+        X_rv,
         extra_rewrites=transform_rewrite,
         use_jacobian=use_jacobian,
     )
@@ -321,19 +315,17 @@ def test_hierarchical_uniform_transform():
     upper_rv = at.random.uniform(9, 10, name="upper")
     x_rv = at.random.uniform(lower_rv, upper_rv, name="x")
 
-    lower = lower_rv.clone()
-    upper = upper_rv.clone()
-    x = x_rv.clone()
-
     transform_rewrite = TransformValuesRewrite(
         {
-            lower: DEFAULT_TRANSFORM,
-            upper: DEFAULT_TRANSFORM,
-            x: DEFAULT_TRANSFORM,
+            lower_rv: DEFAULT_TRANSFORM,
+            upper_rv: DEFAULT_TRANSFORM,
+            x_rv: DEFAULT_TRANSFORM,
         }
     )
-    logp, _ = joint_logprob(
-        realized={lower_rv: lower, upper_rv: upper, x_rv: x},
+    logp, (lower, upper, x) = joint_logprob(
+        lower_rv,
+        upper_rv,
+        x_rv,
         extra_rewrites=transform_rewrite,
     )
 
@@ -346,20 +338,18 @@ def test_nondefault_transforms():
     scale_rv = at.random.uniform(-1, 1, name="scale")
     x_rv = at.random.normal(loc_rv, scale_rv, name="x")
 
-    loc = loc_rv.clone()
-    scale = scale_rv.clone()
-    x = x_rv.clone()
-
     transform_rewrite = TransformValuesRewrite(
         {
-            loc: None,
-            scale: LogOddsTransform(),
-            x: LogTransform(),
+            loc_rv: None,
+            scale_rv: LogOddsTransform(),
+            x_rv: LogTransform(),
         }
     )
 
-    logp, _ = joint_logprob(
-        realized={loc_rv: loc, scale_rv: scale, x_rv: x},
+    logp, (loc, scale, x) = joint_logprob(
+        loc_rv,
+        scale_rv,
+        x_rv,
         extra_rewrites=transform_rewrite,
     )
 
@@ -391,12 +381,11 @@ def test_default_transform_multiout():
     # multiple outputs and no default output.
     sd = at.linalg.svd(at.eye(1))[1][0]
     x_rv = at.random.normal(0, sd, name="x")
-    x = x_rv.clone()
 
-    transform_rewrite = TransformValuesRewrite({x: DEFAULT_TRANSFORM})
+    transform_rewrite = TransformValuesRewrite({x_rv: DEFAULT_TRANSFORM})
 
-    logp, _ = joint_logprob(
-        realized={x_rv: x},
+    logp, (x,) = joint_logprob(
+        x_rv,
         extra_rewrites=transform_rewrite,
     )
 
@@ -412,12 +401,11 @@ def test_nonexistent_default_transform():
     transform does not fail
     """
     x_rv = at.random.normal(name="x")
-    x = x_rv.clone()
 
-    transform_rewrite = TransformValuesRewrite({x: DEFAULT_TRANSFORM})
+    transform_rewrite = TransformValuesRewrite({x_rv: DEFAULT_TRANSFORM})
 
-    logp, _ = joint_logprob(
-        realized={x_rv: x},
+    logp, (x,) = joint_logprob(
+        x_rv,
         extra_rewrites=transform_rewrite,
     )
 
@@ -446,9 +434,8 @@ def test_original_values_output_dict():
     the logprob factor
     """
     p_rv = at.random.beta(1, 1, name="p")
-    p_vv = p_rv.clone()
 
-    tr = TransformValuesRewrite({p_vv: DEFAULT_TRANSFORM})
+    tr = TransformValuesRewrite({p_rv: DEFAULT_TRANSFORM})
     logp_dict, _ = conditional_logprob(p_rv, extra_rewrites=tr)
 
     assert p_rv in logp_dict
@@ -469,29 +456,28 @@ def test_mixture_transform():
     Y_rv = at.stack([Y_1_rv, Y_2_rv])[I_rv]
     Y_rv.name = "Y"
 
-    logp_no_trans, (y_vv, i_vv) = joint_logprob(Y_rv, I_rv)
+    logp, (y_vv, i_vv) = joint_logprob(
+        Y_rv,
+        I_rv,
+    )
 
-    transform_rewrite = TransformValuesRewrite({y_vv: LogTransform()})
+    transform_rewrite = TransformValuesRewrite({Y_rv: LogOddsTransform()})
 
     with pytest.warns(None) as record:
         # This shouldn't raise any warnings
-        logp_trans, _ = joint_logprob(
-            realized={Y_rv: y_vv, I_rv: i_vv},
+        logp_trans, (y_vv_trans, i_vv_trans) = joint_logprob(
+            Y_rv,
+            I_rv,
             extra_rewrites=transform_rewrite,
             use_jacobian=False,
         )
 
     assert not record.list
 
-    # The untransformed graph should be the same as the transformed graph after
-    # replacing the `Y_rv` value variable with a transformed version of itself
-    logp_nt_fg = FunctionGraph(outputs=[logp_no_trans], clone=False)
-    y_trans = transformed_variable(at.exp(y_vv), y_vv)
-    y_trans.name = "y_log"
-    logp_nt_fg.replace(y_vv, y_trans)
-    logp_nt = logp_nt_fg.outputs[0]
-
-    assert equal_computations([logp_nt], [logp_trans])
+    logp_fn = aesara.function((i_vv, y_vv), logp)
+    logp_trans_fn = aesara.function((i_vv_trans, y_vv_trans), logp_trans)
+    np.isclose(logp_trans_fn(0, np.log(0.1 / 0.9)), logp_fn(0, 0.1))
+    np.isclose(logp_trans_fn(1, np.log(0.1 / 0.9)), logp_fn(1, 0.1))
 
 
 def test_invalid_interval_transform():
@@ -642,11 +628,10 @@ def test_scale_transform_rv(rv_size, scale_type):
 def test_transformed_rv_and_value():
     y_rv = at.random.halfnormal(-1, 1, name="base_rv") + 1
     y_rv.name = "y"
-    y_vv = y_rv.clone()
 
-    transform_rewrite = TransformValuesRewrite({y_vv: LogTransform()})
+    transform_rewrite = TransformValuesRewrite({y_rv: LogTransform()})
 
-    logp, _ = joint_logprob(realized={y_rv: y_vv}, extra_rewrites=transform_rewrite)
+    logp, (y_vv,) = joint_logprob(y_rv, extra_rewrites=transform_rewrite)
     assert_no_rvs(logp)
     logp_fn = aesara.function([y_vv], logp)