Remove logic around Initializer-s in adapt_node (#191)

CHANGELOG.rst

@@ -18,6 +18,10 @@ Change log
 
 - Value propagation of string tensors no longer raises an erroneous ``ValueError`` in some instances.
 
+**Other changes**
+
+- The adaption logic is not using inferred values as initializers anymore.
+
 
 0.12.1 (2024-06-18)
 -------------------

src/spox/_adapt.py

@@ -4,7 +4,6 @@
 import warnings
 from typing import Optional
 
-import numpy as np
 import onnx
 import onnx.version_converter
 
@@ -14,7 +13,6 @@
 from ._node import Node
 from ._schemas import SCHEMAS
 from ._scope import Scope
-from ._utils import from_array
 from ._var import Var
 
 
@@ -43,11 +41,6 @@ def adapt_node(
             )
             for key, var in node.outputs.get_vars().items()
         ]
-        initializers = [
-            from_array(var._value, name)  # type: ignore
-            for name, var in node.inputs.get_vars().items()
-            if isinstance(var._value, np.ndarray)
-        ]
     except ValueError:
         return None
 
@@ -57,7 +50,6 @@ def adapt_node(
             "spox__singleton_adapter_graph",
             list(input_info.values()),
             output_info,
-            initializers,
         ),
         opset_imports=[onnx.helper.make_operatorsetid("", source_version)],
     )

tests/test_adapt.py

@@ -15,6 +15,7 @@
 from spox import Tensor, Var, argument, build, inline
 from spox._attributes import AttrInt64s
 from spox._fields import BaseAttributes, BaseInputs, BaseOutputs
+from spox._future import initializer
 from spox._graph import arguments, results
 from spox._node import OpType
 from spox._standard import StandardNode
@@ -157,6 +158,18 @@ def test_adapt_node_with_repeating_input_names():
     build({"a": a}, {"b": b, "c": c})
 
 
+def test_adapt_node_initializer():
+    init_data = [1.0, 2.0, 3.0]
+
+    a = argument(Tensor(np.float32, ("N",)))
+    b = initializer(init_data, np.float32)
+    c = op18.equal(a, b)
+    d = op19.identity(a)
+
+    model = build({"a": a}, {"b": b, "c": c, "d": d})
+    np.testing.assert_allclose(model.graph.initializer[0].float_data, init_data)
+
+
 def test_inline_model_custom_node_only():
     """Inline a model which only consists of a custom node.