Create a new node type for inputs that are initialized (#29)

<img width="823" alt="image"
src="https://github.com/justinchuby/model-explorer-onnx/assets/11205048/7be1e179-c5b2-4540-8973-cbe2a70a211e">

src/model_explorer_onnx/main.py

@@ -359,6 +359,13 @@ def add_graph_io(
             id=get_graph_io_node_name(value),
             label=type_,
         )
+        set_attr(
+            node,
+            "explanation",
+            f"This is a graph {type_.lower()}. ONNX {type_.lower()}s are values (edges), not nodes. "
+            f"A node is displayed here for visualization.",
+        )
+
         producer = value.producer()
         if producer is not None:
             assert producer.name, "Bug: Node name is required"
@@ -437,20 +444,31 @@ def add_initializers(
                 "Initializer does not have a name. Skipping: %s", initializer
             )
             continue
-        initializer_node_name = get_initializer_node_name(initializer)
-        if initializer_node_name in all_nodes:
-            # The initializer is also a graph input. Fill in the missing metadata.
-            node = all_nodes[initializer_node_name]
-            metadata = node.outputsMetadata[0]
+        input_node_name = get_graph_io_node_name(initializer)
+        if input_node_name in all_nodes:
+            # The initializer is also a graph input.
+            # Convert it into an InitializedInput and fill in the missing metadata
+            node = all_nodes[input_node_name]
+            node.label = "InitializedInput"
+            set_attr(
+                node,
+                "explanation",
+                "This is a graph input that is initialized by an initializer. "
+                "ONNX inputs are values (edges), not nodes. A node is displayed here for visualization.",
+            )
+            # Display the constant value
             if can_display_tensor_json(initializer.const_value, settings=settings):
                 assert initializer.const_value is not None
                 set_attr(
                     node,
                     "__value",
                     display_tensor_json(initializer.const_value, settings=settings),
                 )
+            # Set output metadata
+            metadata = node.outputsMetadata[0]
             set_attr(metadata, "value", display_tensor_repr(initializer.const_value))
             continue
+        initializer_node_name = get_initializer_node_name(initializer)
         node = graph_builder.GraphNode(
             id=initializer_node_name,
             label="Initializer",

testdata/initializer_node_tests.textproto

@@ -0,0 +1,95 @@
+ir_version: 8
+graph {
+  node {
+    input: "input"
+    output: "output"
+    name: "node_Identity_0"
+    op_type: "Identity"
+    doc_string: "An identity node 1"
+  }
+  node {
+    input: "input_initialized"
+    output: "output_initialized"
+    name: "node_Identity_1"
+    op_type: "Identity"
+    doc_string: "An identity node 2"
+  }
+  node {
+    input: "normal_initializer"
+    output: "val_0"
+    name: "node_Identity_2"
+    op_type: "Identity"
+    doc_string: "An identity node 3"
+  }
+  name: "initializer_node_tests"
+  initializer {
+    dims: 3
+    data_type: 1
+    name: "input_initialized"
+    raw_data: "\000\000\000\000\000\000\200?\000\000\000@"
+  }
+  initializer {
+    dims: 1
+    data_type: 7
+    name: "normal_initializer"
+    raw_data: "*\000\000\000\000\000\000\000"
+  }
+  input {
+    name: "input"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 42
+          }
+        }
+      }
+    }
+    doc_string: "An input"
+  }
+  input {
+    name: "input_initialized"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 3
+          }
+        }
+      }
+    }
+    doc_string: "An initialized input"
+  }
+  output {
+    name: "output"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 42
+          }
+        }
+      }
+    }
+  }
+  output {
+    name: "output_initialized"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 3
+          }
+        }
+      }
+    }
+  }
+}
+opset_import {
+  domain: ""
+  version: 20
+}

tools/test_generation/initializers.py

@@ -0,0 +1,70 @@
+import pathlib
+
+import onnx
+from onnxscript import ir
+import numpy as np
+
+
+def main():
+    input_ = ir.Input(
+        "input", ir.Shape([42]), ir.TensorType(ir.DataType.FLOAT), "An input"
+    )
+    input_initialized = ir.Input(
+        "input_initialized",
+        ir.Shape([3]),
+        ir.TensorType(ir.DataType.FLOAT),
+        "An initialized input",
+    )
+    input_initializer = ir.Tensor(
+        np.array([0.0, 1.0, 2.0], dtype=np.float32), name="input_initialized"
+    )
+    input_initialized.const_value = input_initializer
+    normal_value = ir.Value(name="normal_initializer")
+    normal_initializer = ir.Tensor(
+        np.array([42], dtype=np.int64), name="normal_initializer"
+    )
+    normal_value.const_value = normal_initializer
+    output = ir.Value(
+        shape=ir.Shape([42]), type=ir.TensorType(ir.DataType.FLOAT), name="output"
+    )
+    output_initialized = ir.Value(
+        shape=ir.Shape([3]),
+        type=ir.TensorType(ir.DataType.FLOAT),
+        name="output_initialized",
+    )
+    graph = ir.Graph(
+        [input_, input_initialized],
+        [output, output_initialized],
+        nodes=[
+            ir.Node(
+                "",
+                "Identity",
+                [input_],
+                outputs=[output],
+                doc_string="An identity node 1",
+            ),
+            ir.Node(
+                "",
+                "Identity",
+                [input_initialized],
+                outputs=[output_initialized],
+                doc_string="An identity node 2",
+            ),
+            ir.Node("", "Identity", [normal_value], doc_string="An identity node 3"),
+        ],
+        initializers=[input_initialized, normal_value],
+        opset_imports={"": 20},
+        name="initializer_node_tests",
+    )
+    model = ir.Model(graph, ir_version=8)
+    model_proto = ir.serde.serialize_model(model)
+    onnx.save(
+        model_proto,
+        pathlib.Path(__file__).parent.parent.parent
+        / "testdata"
+        / "initializer_node_tests.textproto",
+    )
+
+
+if __name__ == "__main__":
+    main()