Preserve the identities of valued/observed variables

Author: brandonwillard

aeppl/joint_logprob.py

@@ -141,18 +141,16 @@ def conditional_logprob(
     # maps to the logprob graphs and value variables before returning them.
     rv_values = {**original_rv_values, **realized}
 
-    fgraph, _, memo = construct_ir_fgraph(rv_values, ir_rewriter=ir_rewriter)
-
-    if extra_rewrites is not None:
-        extra_rewrites.add_requirements(fgraph, rv_values, memo)
-        extra_rewrites.apply(fgraph)
+    fgraph, new_rv_values = construct_ir_fgraph(
+        rv_values, ir_rewriter=ir_rewriter, extra_rewrites=extra_rewrites
+    )
 
     # We assign log-densities on a per-node basis, and not per-output/variable.
     realized_vars = set()
     new_to_old_rvs = {}
     nodes_to_vals: Dict["Apply", List[Tuple["Variable", "Variable"]]] = {}
 
-    for bnd_var, (old_mvar, old_val) in zip(fgraph.outputs, rv_values.items()):
+    for bnd_var, (old_mvar, val) in zip(fgraph.outputs, new_rv_values.items()):
         mnode = bnd_var.owner
         assert mnode and isinstance(mnode.op, ValuedVariable)
 
@@ -165,11 +163,7 @@ def conditional_logprob(
         if old_mvar in realized:
             realized_vars.add(rv_var)
 
-        # Do this just in case a value variable was changed.  (Some transforms
-        # do this.)
-        new_val = memo[old_val]
-
-        nodes_to_vals.setdefault(rv_node, []).append((val_var, new_val))
+        nodes_to_vals.setdefault(rv_node, []).append((val_var, val))
 
         new_to_old_rvs[rv_var] = old_mvar
 

aeppl/rewriting.py

@@ -1,11 +1,16 @@
-from typing import Dict, Optional, Tuple
+from typing import Dict, Optional, Tuple, Union
 
 import aesara.tensor as at
 from aesara.compile.mode import optdb
 from aesara.graph.basic import Apply, Variable
 from aesara.graph.features import Feature
 from aesara.graph.fg import FunctionGraph
-from aesara.graph.rewriting.basic import GraphRewriter, in2out, node_rewriter
+from aesara.graph.rewriting.basic import (
+    GraphRewriter,
+    NodeRewriter,
+    in2out,
+    node_rewriter,
+)
 from aesara.graph.rewriting.db import EquilibriumDB, RewriteDatabaseQuery, SequenceDB
 from aesara.tensor.elemwise import DimShuffle, Elemwise
 from aesara.tensor.extra_ops import BroadcastTo
@@ -180,9 +185,10 @@ def incsubtensor_rv_replace(fgraph, node):
 
 
 def construct_ir_fgraph(
-    rv_values: Dict[Variable, Variable],
+    rvs_to_values: Dict[Variable, Variable],
     ir_rewriter: Optional[GraphRewriter] = None,
-) -> Tuple[FunctionGraph, Dict[Variable, Variable], Dict[Variable, Variable]]:
+    extra_rewrites: Optional[Union[GraphRewriter, NodeRewriter]] = None,
+) -> Tuple[FunctionGraph, Dict[Variable, Variable]]:
     r"""Construct a `FunctionGraph` in measurable IR form for the keys in `rv_values`.
 
     A custom IR rewriter can be specified. By default,
@@ -215,9 +221,8 @@ def construct_ir_fgraph(
     Returns
     -------
     A `FunctionGraph` of the measurable IR, a copy of `rv_values` containing
-    the new, cloned versions of the original variables in `rv_values`, and
-    a ``dict`` mapping all the original variables to their cloned values in
-    the `FunctionGraph`.
+    the new, cloned versions of the original variables in `rv_values`.
+
     """
 
     # We're going to create a `FunctionGraph` that effectively represents the
@@ -233,16 +238,20 @@ def construct_ir_fgraph(
     # so that they're distinct nodes in the graph.  This allows us to replace
     # all instances of the original random variables with their value
     # variables, while leaving the output clones untouched.
-    rv_value_clones = {}
+    rv_clone_to_value_clone = {}
+    rv_to_value_clone = {}
+    value_clone_to_value = {}
     measured_outputs = {}
-    memo = {}
-    for rv, val in rv_values.items():
+    memo: Dict[Variable, Variable] = {}
+    for rv, val in rvs_to_values.items():
         rv_node_clone = rv.owner.clone()
         rv_clone = rv_node_clone.outputs[rv.owner.outputs.index(rv)]
-        rv_value_clones[rv_clone] = val
-        measured_outputs[rv] = valued_variable(rv_clone, val)
-        # Prevent value variables from being cloned
-        memo[val] = val
+        val_clone = val.clone()
+        val_clone.name = "val_clone"
+        rv_clone_to_value_clone[rv_clone] = val_clone
+        rv_to_value_clone[rv] = val_clone
+        value_clone_to_value[val_clone] = val
+        measured_outputs[rv] = valued_variable(rv_clone, val_clone)
 
     # We add `ShapeFeature` because it will get rid of references to the old
     # `RandomVariable`s that have been lifted; otherwise, it will be difficult
@@ -257,9 +266,6 @@ def construct_ir_fgraph(
         copy_inputs=False,
     )
 
-    # Update `rv_values` so that it uses the new cloned variables
-    rv_value_clones = {memo[k]: v for k, v in rv_value_clones.items()}
-
     # Replace valued non-output variables with their values
     fgraph.replace_all(
         [(memo[rv], val) for rv, val in measured_outputs.items() if rv in memo],
@@ -272,11 +278,22 @@ def construct_ir_fgraph(
 
     ir_rewriter.rewrite(fgraph)
 
+    if extra_rewrites is not None:
+        # Expect `value_clone_to_value` to be updated in-place
+        extra_rewrites.add_requirements(fgraph, rv_to_value_clone, value_clone_to_value)
+        extra_rewrites.apply(fgraph)
+
     # Undo un-valued measurable IR rewrites
     new_to_old = tuple((v, k) for k, v in fgraph.measurable_conversions.items())
-    fgraph.replace_all(new_to_old, reason="undo-unvalued-measurables")
+    # and add the original value variables back in
+    new_to_old += tuple(value_clone_to_value.items())
+    fgraph.replace_all(
+        new_to_old, reason="undo-unvalued-measurables", import_missing=True
+    )
+
+    new_rvs_to_values = dict(zip(rvs_to_values.keys(), value_clone_to_value.values()))
 
-    return fgraph, rv_value_clones, memo
+    return fgraph, new_rvs_to_values
 
 
 @register_useless

aeppl/transforms.py

@@ -161,14 +161,10 @@ def transform_values(fgraph: FunctionGraph, node: Apply):
     4. Replace the old `ValuedVariable` with a new one containing a
     `TransformedVariable` value.
 
-    Step 3. is currently accomplished by updating the `memo` dictionary
-    associated with the `FunctionGraph`.  Our main entry-point,
+    Step 3. is currently accomplished by updating the `rvs_to_values`
+    dictionary associated with the `FunctionGraph`.  Our main entry-point,
     `conditional_logprob`, checks this dictionary for value variable changes.
 
-    TODO: This approach is less than ideal, because it puts awkward demands on
-    users/callers of this rewrite to check with `memo`; let's see if we can do
-    something better.
-
     The new value variable mentioned in Step 2. may be of a different `Type`
     (e.g. extra/fewer dimensions) than the original value variable; this is why
     we must replace the corresponding original value variables before we
@@ -235,8 +231,8 @@ def transform_values(fgraph: FunctionGraph, node: Apply):
 
     # This effectively lets the caller know that a value variable has been
     # replaced (i.e. they should filter all their old value variables through
-    # the memo/replacements map).
-    fgraph.memo[value_var] = trans_value_var
+    # the replacements map).
+    fgraph.value_clone_to_value[value_var] = trans_value_var
 
     trans_var = trans_node.outputs[rv_var_out_idx]
     new_var = valued_variable(trans_var, untrans_value_var)
@@ -252,7 +248,7 @@ class TransformValuesMapping(Feature):
 
     """
 
-    def __init__(self, values_to_transforms, memo):
+    def __init__(self, values_to_transforms, value_clone_to_value):
         """
         Parameters
         ==========
@@ -261,20 +257,19 @@ def __init__(self, values_to_transforms, memo):
             value variable can be assigned one of `RVTransform`,
             `DEFAULT_TRANSFORM`, or ``None``. Random variables with no
             transform specified remain unchanged.
-        memo
-            Mapping from variables to their clones.  This is updated
-            in-place whenever a value variable is transformed.
-
+        value_clone_to_value
+            Mapping between random variable value clones and their original
+            value variables.
         """
         self.values_to_transforms = values_to_transforms
-        self.memo = memo
+        self.value_clone_to_value = value_clone_to_value
 
     def on_attach(self, fgraph):
         if hasattr(fgraph, "values_to_transforms"):
             raise AlreadyThere()
 
         fgraph.values_to_transforms = self.values_to_transforms
-        fgraph.memo = self.memo
+        fgraph.value_clone_to_value = self.value_clone_to_value
 
 
 class TransformValuesRewrite(GraphRewriter):
@@ -322,6 +317,7 @@ def __init__(
             measurable variable can be assigned an `RVTransform` instance,
             `DEFAULT_TRANSFORM`, or ``None``. Measurable variables with no
             transform specified remain unchanged.
+        rvs_to_values
 
         """
 
@@ -330,14 +326,16 @@ def __init__(
     def add_requirements(
         self,
         fgraph,
-        rv_to_values: Dict[TensorVariable, TensorVariable],
-        memo: Dict[TensorVariable, TensorVariable],
+        rvs_to_values: Dict[TensorVariable, TensorVariable],
+        value_clone_to_value: Dict[TensorVariable, TensorVariable],
     ):
         values_to_transforms = {
-            rv_to_values[rv]: transform
+            rvs_to_values[rv]: transform
             for rv, transform in self.rvs_to_transforms.items()
         }
-        values_transforms_feature = TransformValuesMapping(values_to_transforms, memo)
+        values_transforms_feature = TransformValuesMapping(
+            values_to_transforms, value_clone_to_value
+        )
         fgraph.attach_feature(values_transforms_feature)
 
     def apply(self, fgraph: FunctionGraph):

tests/test_composite_logprob.py

@@ -79,25 +79,20 @@ def test_unvalued_ir_reversion():
     """Make sure that un-valued IR rewrites are reverted."""
     srng = at.random.RandomStream(0)
 
-    x_rv = srng.normal()
+    x_rv = srng.normal(name="X")
     y_rv = at.clip(x_rv, 0, 1)
-    z_rv = srng.normal(y_rv, 1, name="z")
+    y_rv.name = "Y"
+    z_rv = srng.normal(y_rv, 1, name="Z")
     z_vv = z_rv.clone()
+    z_vv.name = "z"
 
     # Only the `z_rv` is "valued", so `y_rv` doesn't need to be converted into
     # measurable IR.
     rv_values = {z_rv: z_vv}
 
-    z_fgraph, _, memo = construct_ir_fgraph(rv_values)
+    z_fgraph, new_rvs_to_values = construct_ir_fgraph(rv_values)
 
-    assert memo[y_rv] in z_fgraph.measurable_conversions
-
-    measurable_y_rv = z_fgraph.measurable_conversions[memo[y_rv]]
-    assert isinstance(measurable_y_rv.owner.op, MeasurableClip)
-
-    # `construct_ir_fgraph` should've reverted the un-valued measurable IR
-    # change
-    assert measurable_y_rv not in z_fgraph
+    assert not any(isinstance(node.op, MeasurableClip) for node in z_fgraph.apply_nodes)
 
 
 def test_shifted_cumsum():

tests/test_convolutions.py

@@ -74,7 +74,7 @@ def test_add_independent_normals(mu_x, mu_y, sigma_x, sigma_y, x_shape, y_shape,
     Z_rv.name = "Z"
     z_vv = Z_rv.clone()
 
-    fgraph, _, _ = construct_ir_fgraph({Z_rv: z_vv})
+    fgraph, *_ = construct_ir_fgraph({Z_rv: z_vv})
 
     (valued_var_out_node) = fgraph.outputs[0].owner
     # The convolution should be applied, and not the transform
@@ -108,7 +108,7 @@ def test_normal_add_input_valued():
     Z_rv.name = "Z"
     z_vv = Z_rv.clone()
 
-    fgraph, _, _ = construct_ir_fgraph({Z_rv: z_vv, X_rv: x_vv})
+    fgraph, *_ = construct_ir_fgraph({Z_rv: z_vv, X_rv: x_vv})
 
     valued_var_out_node = fgraph.outputs[0].owner
     # We should not expect the convolution to be applied; instead, the
@@ -136,7 +136,7 @@ def test_normal_add_three_inputs():
     Z_rv.name = "Z"
     z_vv = Z_rv.clone()
 
-    fgraph, _, _ = construct_ir_fgraph({Z_rv: z_vv})
+    fgraph, *_ = construct_ir_fgraph({Z_rv: z_vv})
 
     valued_var_out_node = fgraph.outputs[0].owner
     # The convolution should be applied, and not the transform

tests/test_mixture.py

@@ -685,7 +685,7 @@ def test_switch_mixture():
     z_vv = Z1_rv.clone()
     z_vv.name = "z1"
 
-    fgraph, _, _ = construct_ir_fgraph({Z1_rv: z_vv, I_rv: i_vv})
+    fgraph, *_ = construct_ir_fgraph({Z1_rv: z_vv, I_rv: i_vv})
 
     out_rv = fgraph.outputs[0].owner.inputs[0]
     assert isinstance(out_rv.owner.op, MixtureRV)
@@ -696,7 +696,7 @@ def test_switch_mixture():
 
     Z1_rv.name = "Z1"
 
-    fgraph, _, _ = construct_ir_fgraph({Z1_rv: z_vv, I_rv: i_vv})
+    fgraph, *_ = construct_ir_fgraph({Z1_rv: z_vv, I_rv: i_vv})
 
     out_rv = fgraph.outputs[0].owner.inputs[0]
     assert out_rv.name == "Z1-mixture"
@@ -705,7 +705,7 @@ def test_switch_mixture():
 
     Z2_rv = at.stack((X_rv, Y_rv))[I_rv]
 
-    fgraph2, _, _ = construct_ir_fgraph({Z2_rv: z_vv, I_rv: i_vv})
+    fgraph2, *_ = construct_ir_fgraph({Z2_rv: z_vv, I_rv: i_vv})
 
     assert equal_computations(fgraph.outputs, fgraph2.outputs)