Allow rebuilding a graph in toposort_replace

Author: ricardoV94

pymc_experimental/model_transform/conditioning.py

@@ -14,9 +14,8 @@
     model_from_fgraph,
     model_named,
     model_observed_rv,
-    toposort_replace,
 )
-from pymc_experimental.utils.pytensorf import rvs_in_graph
+from pymc_experimental.utils.pytensorf import rvs_in_graph, toposort_replace
 
 
 def observe(model: Model, vars_to_observations: Dict[Union["str", TensorVariable], Any]) -> Model:

pymc_experimental/tests/utils/test_pytensorf.py

@@ -0,0 +1,71 @@
+import pytensor.tensor as pt
+import pytest
+from pytensor.graph import FunctionGraph
+from pytensor.graph.basic import equal_computations
+
+from pymc_experimental.utils.pytensorf import toposort_replace
+
+
+class TestToposortReplace:
+    @pytest.mark.parametrize("compatible_type", (True, False))
+    @pytest.mark.parametrize("num_replacements", (1, 2))
+    @pytest.mark.parametrize("rebuild", (True, False))
+    def test_horizontal_dependency(self, compatible_type, num_replacements, rebuild):
+        x = pt.vector("x", shape=(5,))
+        y = pt.vector("y", shape=(5,))
+
+        out1 = pt.exp(x + y) + pt.log(x + y)
+        out2 = pt.cos(out1)
+
+        new_shape = (5,) if compatible_type else (10,)
+        new_x = pt.vector("new_x", shape=new_shape)
+        new_y = pt.vector("new_y", shape=new_shape)
+        if num_replacements == 1:
+            replacements = [(y, new_y)]
+        else:
+            replacements = [(x, new_x), (y, new_y)]
+
+        fg = FunctionGraph([x, y], [out1, out2], clone=False)
+
+        # If types are incompatible, and we don't rebuild or only replace one of the variables,
+        # The function should fail
+        if not compatible_type and (not rebuild or num_replacements == 1):
+            with pytest.raises((TypeError, ValueError)):
+                toposort_replace(fg, replacements, rebuild=rebuild)
+            return
+        toposort_replace(fg, replacements, rebuild=rebuild)
+
+        if num_replacements == 1:
+            expected_out1 = pt.exp(x + new_y) + pt.log(x + new_y)
+        else:
+            expected_out1 = pt.exp(new_x + new_y) + pt.log(new_x + new_y)
+        expected_out2 = pt.cos(expected_out1)
+        assert equal_computations(fg.outputs, [expected_out1, expected_out2])
+
+    @pytest.mark.parametrize("compatible_type", (True, False))
+    @pytest.mark.parametrize("num_replacements", (2, 3))
+    @pytest.mark.parametrize("rebuild", (True, False))
+    def test_vertical_dependency(self, compatible_type, num_replacements, rebuild):
+        x = pt.vector("x", shape=(5,))
+        a1 = pt.exp(x)
+        a2 = pt.log(a1)
+        out = a1 + a2
+
+        new_x = pt.vector("new_x", shape=(5 if compatible_type else 10,))
+        if num_replacements == 2:
+            replacements = [(x, new_x), (a1, pt.cos(a1)), (a2, pt.sin(a2 + 5))]
+        else:
+            replacements = [(a1, pt.cos(pt.exp(new_x))), (a2, pt.sin(a2 + 5))]
+
+        fg = FunctionGraph([x], [out], clone=False)
+
+        if not compatible_type and not rebuild:
+            with pytest.raises(TypeError):
+                toposort_replace(fg, replacements, rebuild=rebuild)
+            return
+        toposort_replace(fg, replacements, rebuild=rebuild)
+
+        expected_a1 = pt.cos(pt.exp(new_x))
+        expected_a2 = pt.sin(pt.log(expected_a1) + 5)
+        expected_out = expected_a1 + expected_a2
+        assert equal_computations(fg.outputs, [expected_out])

pymc_experimental/utils/model_fgraph.py

@@ -1,4 +1,4 @@
-from typing import Dict, Optional, Sequence, Tuple
+from typing import Dict, Optional, Tuple
 
 import pytensor
 from pymc.logprob.transforms import RVTransform
@@ -10,7 +10,7 @@
 from pytensor.scalar import Identity
 from pytensor.tensor.elemwise import Elemwise
 
-from pymc_experimental.utils.pytensorf import StringType
+from pymc_experimental.utils.pytensorf import StringType, toposort_replace
 
 
 class ModelVar(Op):
@@ -89,19 +89,6 @@ def model_free_rv(rv, value, transform, *dims):
 model_named = ModelNamed()
 
 
-def toposort_replace(
-    fgraph: FunctionGraph, replacements: Sequence[Tuple[Variable, Variable]], reverse: bool = False
-) -> None:
-    """Replace multiple variables in topological order."""
-    toposort = fgraph.toposort()
-    sorted_replacements = sorted(
-        replacements,
-        key=lambda pair: toposort.index(pair[0].owner) if pair[0].owner else -1,
-        reverse=reverse,
-    )
-    fgraph.replace_all(sorted_replacements, import_missing=True)
-
-
 @node_rewriter([Elemwise])
 def local_remove_identity(fgraph, node):
     if isinstance(node.op.scalar_op, Identity):

pymc_experimental/utils/pytensorf.py

@@ -1,10 +1,12 @@
-from typing import Sequence
+from collections import deque
+from itertools import chain
+from typing import Iterable, Sequence, Set, Tuple
 
 import pytensor
 from pymc import SymbolicRandomVariable
 from pytensor import Variable
-from pytensor.graph import Constant, Type
-from pytensor.graph.basic import walk
+from pytensor.graph import Constant, FunctionGraph, Type
+from pytensor.graph.basic import Apply, walk
 from pytensor.graph.op import HasInnerGraph
 from pytensor.tensor.random.op import RandomVariable
 
@@ -58,3 +60,110 @@ def expand(r):
         for node in walk(vars, expand, False)
         if node.owner and isinstance(node.owner.op, (RandomVariable, SymbolicRandomVariable))
     )
+
+
+def _replace_rebuild_all(
+    fgraph: FunctionGraph, replacements: Iterable[Tuple[Variable, Variable]], **kwargs
+) -> FunctionGraph:
+    """Replace variables and rebuild dependent graph if needed.
+
+    Rebuilding allows for replacements that change the semantics of the graph
+    (different types), which may not be possible for all Ops.
+    """
+
+    def get_client_nodes(vars) -> Set[Apply]:
+        nodes = set()
+        d = deque(
+            chain.from_iterable(fgraph.clients[var] for var in vars if var in fgraph.variables)
+        )
+        while d:
+            node, _ = d.pop()
+            if node in nodes or node == "output":
+                continue
+            nodes.add(node)
+            d.extend(chain.from_iterable(fgraph.clients[out] for out in node.outputs))
+        return nodes
+
+    repl_dict = {old: new for old, new in replacements}
+    root_nodes = {var.owner for var in repl_dict.keys()}
+
+    # Build sorted queue with all nodes that depend on replaced variables
+    topo_order = {node: order for order, node in enumerate(fgraph.toposort())}
+    client_nodes = get_client_nodes(repl_dict.keys())
+    d = deque(sorted(client_nodes, key=lambda node: topo_order[node]))
+    while d:
+        node = d.popleft()
+        if node in root_nodes:
+            continue
+
+        new_inputs = [repl_dict.get(i, i) for i in node.inputs]
+        if new_inputs == node.inputs:
+            continue
+
+        # Either remake the node or do a simple inplace replacement
+        # This property is not yet present in PyTensor
+        if getattr(node.op, "_output_type_depends_on_input_value", False):
+            remake_node = True
+        else:
+            remake_node = any(
+                not inp.type == new_inp.type for inp, new_inp in zip(node.inputs, new_inputs)
+            )
+
+        if remake_node:
+            new_node = node.clone_with_new_inputs(new_inputs, strict=False)
+            fgraph.import_node(new_node, import_missing=True)
+            for out, new_out in zip(node.outputs, new_node.outputs):
+                repl_dict[out] = new_out
+        else:
+            replace = list(zip(node.inputs, new_inputs))
+            fgraph.replace_all(replace, import_missing=True)
+
+    # We need special logic for the cases where we had to rebuild the output nodes
+    for i, (new_output, old_output) in enumerate(
+        zip(
+            (repl_dict.get(out, out) for out in fgraph.outputs),
+            fgraph.outputs,
+        )
+    ):
+        if new_output is old_output:
+            continue
+        fgraph.outputs[i] = new_output
+        fgraph.import_var(new_output, import_missing=True)
+        client = ("output", i)
+        fgraph.add_client(new_output, client)
+        fgraph.remove_client(old_output, client)
+        fgraph.execute_callbacks("on_change_input", "output", i, old_output, new_output)
+
+
+def toposort_replace(
+    fgraph: FunctionGraph,
+    replacements: Sequence[Tuple[Variable, Variable]],
+    reverse: bool = False,
+    rebuild: bool = False,
+) -> None:
+    """Replace multiple variables in topological order."""
+    if rebuild and reverse:
+        raise NotImplementedError("reverse rebuild not supported")
+
+    toposort = fgraph.toposort()
+    sorted_replacements = sorted(
+        replacements,
+        key=lambda pair: toposort.index(pair[0].owner) if pair[0].owner else -1,
+        reverse=reverse,
+    )
+
+    if rebuild:
+        if len(replacements) > 1:
+            # In this case we need to introduce the replacements inside each other
+            # To avoid undoing previous changes
+            sorted_replacements = [list(pairs) for pairs in sorted_replacements]
+            for i in range(1, len(replacements)):
+                # Replace-rebuild each successive replacement with the previous replacements (in topological order)
+                temp_fgraph = FunctionGraph(
+                    outputs=[repl for _, repl in sorted_replacements[i:]], clone=False
+                )
+                _replace_rebuild_all(temp_fgraph, replacements=sorted_replacements[:i])
+                sorted_replacements[i][1] = temp_fgraph.outputs[0]
+        _replace_rebuild_all(fgraph, sorted_replacements, import_missing=True)
+    else:
+        fgraph.replace_all(sorted_replacements, import_missing=True)