Faster infer_static_shape

Author: ricardoV94

pytensor/tensor/basic.py

@@ -22,10 +22,11 @@
 from pytensor import compile, config, printing
 from pytensor import scalar as aes
 from pytensor.gradient import DisconnectedType, grad_undefined
+from pytensor.graph import RewriteDatabaseQuery
 from pytensor.graph.basic import Apply, Constant, Variable
 from pytensor.graph.fg import FunctionGraph
 from pytensor.graph.op import Op
-from pytensor.graph.rewriting.utils import rewrite_graph
+from pytensor.graph.rewriting.db import EquilibriumDB
 from pytensor.graph.type import HasShape, Type
 from pytensor.link.c.op import COp
 from pytensor.link.c.params_type import ParamsType
@@ -1356,6 +1357,45 @@ def identity_like(x, dtype: Optional[Union[str, np.generic, np.dtype]] = None):
     return eye(_x.shape[0], _x.shape[1], k=0, dtype=dtype)
 
 
+class CachedEquilibrimDB(EquilibriumDB):
+    """A subclass of EquilibriumDB that allows caching of a default query for faster reuse."""
+
+    def __init__(self, default_query):
+        super().__init__()
+        self._default_query = default_query
+        self._cached_default_query = None
+
+    def register(self, *args, **kwargs):
+        # If new rewrites are registered, the default cached query is void
+        self.cached_default_query = None
+        super().register(*args, **kwargs)
+
+    @property
+    def default_query(self):
+        if self._cached_default_query is None:
+            self._cached_default_query = self.query(self._default_query)
+        return self._cached_default_query
+
+
+infer_shape_db = CachedEquilibrimDB(
+    default_query=RewriteDatabaseQuery(include=("infer_shape",))
+)
+
+
+def register_infer_shape(rewrite, *tags, **kwargs):
+    if isinstance(rewrite, str):
+
+        def register(inner_lopt):
+            return register_infer_shape(inner_lopt, rewrite, *tags, **kwargs)
+
+        return register
+    else:
+        name = kwargs.pop("name", None) or rewrite.__name__
+
+        infer_shape_db.register(name, rewrite, *tags, "infer_shape", **kwargs)
+        return rewrite
+
+
 def infer_static_shape(
     shape: Union[Variable, Sequence[Union[Variable, int]]]
 ) -> tuple[Sequence["TensorLike"], Sequence[Optional[int]]]:
@@ -1390,14 +1430,16 @@ def check_type(s):
 
         raise TypeError(f"Shapes must be scalar integers; got {s_as_str}")
 
-    sh = [check_type(as_tensor_variable(s, ndim=0)) for s in shape]
+    sh = folded_shape = [check_type(as_tensor_variable(s, ndim=0)) for s in shape]
+
+    if not all(isinstance(s, Constant) for s in folded_shape):
+        shape_fg = FunctionGraph(outputs=sh, features=[ShapeFeature()], clone=True)
+        with config.change_flags(optdb__max_use_ratio=10, cxx=""):
+            infer_shape_db.default_query.rewrite(shape_fg)
+            if not all(isinstance(s, Constant) for s in shape_fg.outputs):
+                topo_constant_folding.rewrite(shape_fg)
+        folded_shape = shape_fg.outputs
 
-    shape_fg = FunctionGraph(
-        outputs=sh,
-        features=[ShapeFeature()],
-        clone=True,
-    )
-    folded_shape = rewrite_graph(shape_fg, custom_rewrite=topo_constant_folding).outputs
     static_shape = tuple(
         s.data.item() if isinstance(s, Constant) else None for s in folded_shape
     )

pytensor/tensor/rewriting/basic.py

@@ -58,6 +58,7 @@
     get_underlying_scalar_constant_value,
     join,
     ones_like,
+    register_infer_shape,
     switch,
     tensor_copy,
     zeros,
@@ -420,6 +421,7 @@ def local_fill_to_alloc(fgraph, node):
 )
 
 
+@register_infer_shape
 @register_canonicalize("fast_compile", "shape_unsafe")
 @register_useless("shape_unsafe")
 @node_rewriter([fill])
@@ -441,6 +443,7 @@ def local_useless_fill(fgraph, node):
         return [v]
 
 
+@register_infer_shape
 @register_specialize("shape_unsafe")
 @register_stabilize("shape_unsafe")
 @register_canonicalize("shape_unsafe")
@@ -530,6 +533,7 @@ def local_alloc_empty_to_zeros(fgraph, node):
 )
 
 
+@register_infer_shape
 @register_useless
 @register_canonicalize("fast_compile")
 @register_specialize
@@ -806,6 +810,7 @@ def local_remove_all_assert(fgraph, node):
 )
 
 
+@register_infer_shape
 @register_specialize
 @register_canonicalize
 @register_useless
@@ -826,6 +831,7 @@ def local_join_1(fgraph, node):
 
 
 # TODO: merge in local_useless_join
+@register_infer_shape
 @register_useless
 @register_specialize
 @register_canonicalize
@@ -1066,6 +1072,7 @@ def local_merge_switch_same_cond(fgraph, node):
     ]
 
 
+@register_infer_shape
 @register_useless
 @register_canonicalize
 @register_specialize
@@ -1149,6 +1156,7 @@ def constant_folding(fgraph, node):
 register_specialize(topo_constant_folding, "fast_compile", final_rewriter=True)
 
 
+@register_infer_shape
 @register_canonicalize("fast_compile")
 @register_useless("fast_compile")
 @node_rewriter(None)
@@ -1157,6 +1165,7 @@ def local_view_op(fgraph, node):
         return node.inputs
 
 
+@register_infer_shape
 @register_useless
 @register_canonicalize
 @register_stabilize

pytensor/tensor/rewriting/math.py

@@ -32,6 +32,7 @@
     extract_constant,
     get_underlying_scalar_constant_value,
     ones_like,
+    register_infer_shape,
     switch,
     zeros_like,
 )
@@ -1745,6 +1746,7 @@ def local_reduce_join(fgraph, node):
         return [ret]
 
 
+@register_infer_shape
 @register_canonicalize("fast_compile", "local_cut_useless_reduce")
 @register_useless("local_cut_useless_reduce")
 @node_rewriter(ALL_REDUCE)

pytensor/tensor/rewriting/shape.py

@@ -25,6 +25,7 @@
     constant,
     extract_constant,
     get_underlying_scalar_constant_value,
+    register_infer_shape,
     stack,
 )
 from pytensor.tensor.elemwise import DimShuffle, Elemwise
@@ -964,6 +965,7 @@ def local_reshape_lift(fgraph, node):
         return [e]
 
 
+@register_infer_shape
 @register_useless
 @register_canonicalize
 @node_rewriter([SpecifyShape])
@@ -990,6 +992,17 @@ def local_merge_consecutive_specify_shape(fgraph, node):
     return [specify_shape(inner_obj, shape)]
 
 
+@register_infer_shape
+@node_rewriter([Shape])
+def local_shape_ground(fgraph, node):
+    """Rewrite shape(x) -> make_vector(x.type.shape) when this is constant."""
+    [x] = node.inputs
+    static_shape = x.type.shape
+    if not any(dim is None for dim in static_shape):
+        return [stack([constant(dim, dtype="int64") for dim in static_shape])]
+
+
+@register_infer_shape
 @register_useless
 @register_canonicalize
 @node_rewriter([Shape])
@@ -1014,6 +1027,7 @@ def local_Shape_of_SpecifyShape(fgraph, node):
     return [stack(shape).astype(np.int64)]
 
 
+@register_infer_shape
 @register_canonicalize
 @register_specialize
 @node_rewriter([SpecifyShape])
@@ -1060,6 +1074,7 @@ def local_specify_shape_lift(fgraph, node):
         return new_out
 
 
+@register_infer_shape
 @register_useless
 @register_canonicalize
 @node_rewriter([Shape_i])
@@ -1079,6 +1094,7 @@ def local_Shape_i_ground(fgraph, node):
         return [as_tensor_variable(s_val, dtype=np.int64)]
 
 
+@register_infer_shape
 @register_specialize
 @register_canonicalize
 @node_rewriter([Shape])

pytensor/tensor/rewriting/subtensor.py

@@ -26,6 +26,7 @@
     concatenate,
     extract_constant,
     get_underlying_scalar_constant_value,
+    register_infer_shape,
     switch,
 )
 from pytensor.tensor.elemwise import Elemwise
@@ -328,6 +329,7 @@ def local_subtensor_of_dot(fgraph, node):
     return [r]
 
 
+@register_infer_shape
 @register_useless
 @register_canonicalize
 @register_specialize
@@ -599,6 +601,7 @@ def local_subtensor_remove_broadcastable_index(fgraph, node):
         return [node.inputs[0].dimshuffle(tuple(remain_dim))]
 
 
+@register_infer_shape
 @register_useless
 @register_canonicalize
 @register_specialize
@@ -707,6 +710,7 @@ def local_subtensor_inc_subtensor(fgraph, node):
             return
 
 
+@register_infer_shape
 @register_specialize
 @register_canonicalize("fast_compile")
 @register_useless
@@ -785,6 +789,7 @@ def local_subtensor_make_vector(fgraph, node):
             pass
 
 
+@register_infer_shape
 @register_useless
 @register_canonicalize
 @register_specialize
@@ -1461,6 +1466,7 @@ def local_adv_sub1_adv_inc_sub1(fgraph, node):
     return [r2]
 
 
+@register_infer_shape
 @register_specialize
 @register_stabilize
 @register_canonicalize

pytensor/tensor/shape.py

@@ -1,12 +1,13 @@
 import warnings
 from numbers import Number
 from textwrap import dedent
-from typing import Union
+from typing import Union, cast
 
 import numpy as np
 
 import pytensor
 from pytensor.gradient import DisconnectedType
+from pytensor.graph import Op
 from pytensor.graph.basic import Apply, Variable
 from pytensor.graph.replace import _vectorize_node
 from pytensor.graph.type import HasShape
@@ -145,14 +146,14 @@ def c_code_cache_version(self):
 def shape(x: Union[np.ndarray, Number, Variable]) -> Variable:
     """Return the shape of `x`."""
     if not isinstance(x, Variable):
-        x = at.as_tensor_variable(x)
+        x = at.as_tensor_variable(x)  # type: ignore
 
-    return _shape(x)
+    return cast(Variable, _shape(x))
 
 
-@_get_vector_length.register(Shape)
-def _get_vector_length_Shape(op, var):
-    return var.owner.inputs[0].type.ndim
+@_get_vector_length.register(Shape)  # type: ignore
+def _get_vector_length_Shape(op: Op, var: TensorVariable) -> int:
+    return cast(int, var.owner.inputs[0].type.ndim)
 
 
 @_vectorize_node.register(Shape)
@@ -181,7 +182,7 @@ def shape_tuple(x: TensorVariable) -> tuple[Variable, ...]:
         # We assume/call it a scalar
         return ()
 
-    res = ()
+    res: tuple[Variable, ...] = ()
     symbolic_shape = shape(x)
     static_shape = x.type.shape
     for i in range(x.type.ndim):
@@ -191,7 +192,7 @@ def shape_tuple(x: TensorVariable) -> tuple[Variable, ...]:
             # TODO: Why not use uint64?
             res += (pytensor.scalar.ScalarConstant(pytensor.scalar.int64, shape_val),)
         else:
-            res += (symbolic_shape[i],)
+            res += (symbolic_shape[i],)  # type: ignore
 
     return res
 
@@ -366,7 +367,7 @@ def shape_i_op(i):
     return shape_i_op.cache[key]
 
 
-shape_i_op.cache = {}
+shape_i_op.cache = {}  # type: ignore
 
 
 def register_shape_i_c_code(typ, code, check_input, version=()):
@@ -578,7 +579,7 @@ def specify_shape(
 
     # If the specified shape is already encoded in the input static shape, do nothing
     # This ignores PyTensor constants in shape
-    x = at.as_tensor_variable(x)
+    x = at.as_tensor_variable(x)  # type: ignore
     new_shape_info = any(
         s != xts for (s, xts) in zip(shape, x.type.shape) if s is not None
     )
@@ -589,10 +590,10 @@ def specify_shape(
     return _specify_shape(x, *shape)
 
 
-@_get_vector_length.register(SpecifyShape)
-def _get_vector_length_SpecifyShape(op, var):
+@_get_vector_length.register(SpecifyShape)  # type: ignore
+def _get_vector_length_SpecifyShape(op: Op, var: TensorVariable) -> int:
     try:
-        return at.get_underlying_scalar_constant_value(var.owner.inputs[1]).item()
+        return int(at.get_underlying_scalar_constant_value(var.owner.inputs[1]).item())
     except NotScalarConstantError:
         raise ValueError(f"Length of {var} cannot be determined")
 
@@ -1104,4 +1105,4 @@ def _vectorize_unbroadcast(op: Unbroadcast, node: Apply, x: TensorVariable) -> A
     batched_ndims = x.type.ndim - node.inputs[0].type.ndim
     old_axes = op.axes
     new_axes = (old_axis + batched_ndims for old_axis in old_axes)
-    return unbroadcast(x, *new_axes).owner
+    return cast(Apply, unbroadcast(x, *new_axes).owner)

scripts/mypy-failing.txt

@@ -27,7 +27,6 @@ pytensor/tensor/random/basic.py
 pytensor/tensor/random/op.py
 pytensor/tensor/random/utils.py
 pytensor/tensor/rewriting/basic.py
-pytensor/tensor/shape.py
 pytensor/tensor/slinalg.py
 pytensor/tensor/subtensor.py
 pytensor/tensor/type.py