[IR] Create a shape inference pass using onnx shape inference

onnxscript/ir/_io.py

@@ -78,7 +78,7 @@ def save(
 
         # Store the original initializer values so they can be restored if modify_model=False
         initializer_values = tuple(model.graph.initializers.values())
-        tensors = [v.const_value for v in model.graph.initializers.values()]
+        tensors = [v.const_value for v in initializer_values]
 
         try:
             model = _external_data.unload_from_model(

onnxscript/ir/passes/common/shape_inference.py

@@ -0,0 +1,77 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+"""Shape inference pass using onnx.shape_inference."""
+
+from __future__ import annotations
+
+import logging
+
+import onnx
+
+from onnxscript import ir
+
+logger = logging.getLogger(__name__)
+
+
+class ShapeInferencePass(ir.passes.PassBase):
+    """This pass performs shape inference on the graph."""
+
+    def __init__(
+        self, check_type: bool = True, strict_mode: bool = True, data_prop: bool = True
+    ) -> None:
+        """Initialize the shape inference pass.
+
+        Args:
+            check_type: If True, check the types of the inputs and outputs.
+            strict_mode: If True, use strict mode for shape inference.
+            data_prop: If True, use data propagation for shape inference.
+        """
+        super().__init__()
+        self.check_type = check_type
+        self.strict_mode = strict_mode
+        self.data_prop = data_prop
+
+    in_place = False
+
+    def call(self, model: ir.Model) -> ir.passes.PassResult:
+        # Store the original initializer values so they can be restored
+        initializer_values = tuple(model.graph.initializers.values())
+        tensors = {v.name: v.const_value for v in initializer_values}
+        original_inputs_len = len(model.graph.inputs)
+        initializer_names = {v.name for v in initializer_values}
+
+        # Turn the initializers into inputs and clear the initializers
+        # to limit the model size
+        for initializer in initializer_values:
+            if initializer not in model.graph.inputs:
+                model.graph.inputs.append(initializer)
+            initializer.const_value = None
+        model.graph.initializers.clear()
+
+        # Perform shape inference
+        try:
+            proto = ir.serde.serialize_model(model)
+            proto = onnx.shape_inference.infer_shapes(
+                proto,
+                check_type=self.check_type,
+                strict_mode=self.strict_mode,
+                data_prop=self.data_prop,
+            )
+            model = ir.serde.deserialize_model(proto)
+        except Exception:
+            logger.warning("Shape inference failed. The model is not modified", exc_info=True)
+        finally:
+            # Restore the original initializer values so the model is unchanged
+            for new_input in model.graph.inputs:
+                # Assign the tensors back to the initializers
+                if new_input.name in initializer_names:
+                    model.graph.register_initializer(new_input)
+                    new_input.const_value = tensors[new_input.name]
+            # Remove the inputs that were added
+            inputs = model.graph.inputs[:original_inputs_len]
+            model.graph.inputs.clear()
+            model.graph.inputs.extend(inputs)
+
+        # Even though modified, we know the pass will not change the model if we ran it again.
+        # So set modified to False
+        return ir.passes.PassResult(model, modified=False)

onnxscript/ir/passes/common/shape_inference_test.py

@@ -0,0 +1,48 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+from __future__ import annotations
+
+import unittest
+
+from onnxscript import ir
+from onnxscript.ir.passes.common import shape_inference
+
+
+class TestShapeInference(unittest.TestCase):
+    def test_shape_inference(self):
+        # Create a simple ONNX model with shape inference
+        # Define the model
+        inputs = [
+            ir.Value(
+                name="input_a", type=ir.TensorType(ir.DataType.FLOAT), shape=ir.Shape((1, 2))
+            ),
+            ir.Value(
+                name="input_b", type=ir.TensorType(ir.DataType.FLOAT), shape=ir.Shape((1, 2))
+            ),
+        ]
+
+        add_node = ir.Node("", "Add", inputs=inputs)
+
+        model = ir.Model(
+            ir.Graph(
+                inputs=inputs,
+                outputs=add_node.outputs,
+                nodes=[add_node],
+                opset_imports={"": 20},
+            ),
+            ir_version=10,
+        )
+        self.assertIsNone(add_node.outputs[0].shape)
+        self.assertIsNone(add_node.outputs[0].dtype)
+
+        # Perform shape inference
+        result = shape_inference.ShapeInferencePass()(model)
+        self.assertFalse(result.modified)
+        self.assertEqual(result.model.graph.node(0).outputs[0].shape, ir.Shape((1, 2)))
+        self.assertEqual(result.model.graph.node(0).outputs[0].dtype, ir.DataType.FLOAT)
+        self.assertEqual(result.model.graph.outputs[0].shape, ir.Shape((1, 2)))
+        self.assertEqual(result.model.graph.outputs[0].dtype, ir.DataType.FLOAT)
+
+
+if __name__ == "__main__":
+    unittest.main()