Add backward conversions from 18->17 for reduce ops (#5606)

### Description backward conversion is needed for reduce ops that have axes attribute changed to input ### Motivation and Context version converter --------- Signed-off-by: Liqun Fu <liqfu@microsoft.com> Signed-off-by: liqun Fu <liqun.fu@microsoft.com> Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com> Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>
onnx · Sep 26, 2023 · b5111b8 · b5111b8
1 parent d14f721
commit b5111b8
Show file tree

Hide file tree

Showing 5 changed files with 201 additions and 98 deletions.
diff --git a/docs/AddNewOp.md b/docs/AddNewOp.md
@@ -74,7 +74,10 @@ Once the criteria of proposing new operator/function has been satisfied, you wil
     1. The testing examples will be extracted to the doc.
     2. We also generate binary data for it.
     3. Example: [onnx/backend/test/case/node/abs.py](/onnx/backend/test/case/node/abs.py)
-5. Add at least one automatic upgrade test for your operator in [onnx/test/automatic_upgrade_test.py](/onnx/test/automatic_upgrade_test.py) using `_test_op_upgrade`. These tests create a given operator at a given opset version (usually the version the operator was introduced in) and test that the version converter is able to convert them to the highest available version. So for a new operator `_test_op_upgrade` will not test anything, but as soon as the operator gets updated in a future opset the test will automatically become nontrivial.
+5. Write upgrade and downgrade tests:
+    1. Add at least one automatic upgrade test for your operator in [onnx/test/automatic_upgrade_test.py](/onnx/test/automatic_upgrade_test.py) using `_test_op_upgrade`. These tests create a given operator at a given opset version (usually the version the operator was introduced in) and test that the version converter is able to convert them to the highest available version. So for a new operator `_test_op_upgrade` will not test anything, but as soon as the operator gets updated in a future opset the test will automatically become nontrivial.
+    2. Similarly add at least one automatic downgrade test for your operator in [onnx/test/automatic_downgrade_test.py](/onnx/test/automatic_downgrade_test.py) using `_test_op_downgrade`. Specifying the current version so that once the op is updated at a higher opset version the test will ensure downward conversion is validated.
+
 6. Update the documentation and generate the test data.
     1. Running [the script](/tools/update_doc.sh). If you have files under `onnx/backend/test/data/node` which cannot be generated by the scripts from `onnx/backend/test/case/node`, please further use `python onnx/backend/test/cmd_tools.py generate-data --clean` to cleanup the directory and only preserve needed test data.
 to update the doc and generate the test data.

diff --git a/onnx/test/version_converter/automatic_conversion_test_base.py b/onnx/test/version_converter/automatic_conversion_test_base.py
@@ -0,0 +1,125 @@
+# Copyright (c) ONNX Project Contributors
+
+# SPDX-License-Identifier: Apache-2.0
+import string
+import unittest
+from typing import Any, Dict, List, Optional, Sequence, Union, cast
+
+import onnx
+from onnx import TensorProto, ValueInfoProto, helper, shape_inference, version_converter
+
+LATEST_OPSET = onnx.defs.onnx_opset_version()
+
+
+class TestAutomaticConversion(unittest.TestCase):
+    def _test_op_conversion(
+        self,
+        op: str,
+        from_opset: int,
+        input_shapes: Sequence[Union[Sequence[Optional[int]], str]] = ((3, 4, 5),),
+        output_shapes: Sequence[Sequence[Optional[int]]] = ((3, 4, 5),),
+        input_types: Optional[Sequence[Any]] = None,
+        output_types: Optional[Sequence[Any]] = None,
+        initializer: Sequence[Any] = (),
+        attrs: Optional[Dict[str, Any]] = None,
+        seq_inputs: Sequence[int] = (),
+        seq_outputs: Sequence[int] = (),
+        optional_inputs: Sequence[int] = (),
+        optional_outputs: Sequence[int] = (),
+        is_upgrade: bool = True,
+    ) -> None:
+        """
+        Args:
+            op: A string representing the name of the operator to test.
+            from_opset: An integer representing the lowest opset version to convert.
+            input_shapes: A sequence of tuples or strings representing the shapes of the input tensors.
+                The default value is ((3, 4, 5),).
+            output_shapes: A sequence of tuples representing the shapes of the output tensors.
+                The default value is ((3, 4, 5),).
+            input_types: An optional sequence of types representing the data types of the input tensors.
+            output_types: An optional sequence of types representing the data types of the output tensors.
+            initializer: A sequence of values representing the initial values of the input tensors.
+            attrs: An optional dictionary of attributes for the operator.
+            seq_inputs: A sequence of integers representing the indices of the input tensors that are sequences.
+            seq_outputs: A sequence of integers representing the indices of the output tensors that are sequences.
+            optional_inputs: A sequence of integers representing the indices of the input tensors that are optional.
+            optional_outputs: A sequence of integers representing the indices of the output tensors that are optional.
+            is_upgrade: A boolean value indicating whether to run the version converter from from_opset to
+                the most recent opset version (True) or from the most recent opset version to from_opset (False).
+                The default value is True. In both cases, runs checker and shape inference on the final model.
+        """
+        if attrs is None:
+            attrs = {}
+
+        n_inputs = len(input_shapes)
+        letters = list(string.ascii_lowercase)[:n_inputs]
+        input_names = [
+            letter if shape != "" else ""
+            for (letter, shape) in zip(letters, input_shapes)
+        ]
+        if input_types is None:
+            input_types = [TensorProto.FLOAT] * n_inputs
+        is_sequence = [0 if id not in seq_inputs else 1 for id in range(n_inputs)]
+        is_optional = [0 if id not in optional_inputs else 1 for id in range(n_inputs)]
+        # turn empty strings into [0] to ease type analysis, even though those entries
+        # will be ignored
+        input_shapes_cast = cast(
+            List[List[int]],
+            [[0] if isinstance(shape, str) else shape for shape in input_shapes],
+        )
+        inputs: List[ValueInfoProto] = []
+        for name, ttype, shape, is_seq, is_opt in zip(
+            input_names, input_types, input_shapes_cast, is_sequence, is_optional
+        ):
+            if name != "":
+                if is_seq:
+                    inputs += [
+                        helper.make_tensor_sequence_value_info(name, ttype, shape)
+                    ]
+                elif is_opt:
+                    type_proto = helper.make_tensor_type_proto(ttype, shape)
+                    optional_type_proto = helper.make_optional_type_proto(type_proto)
+                    inputs += [helper.make_value_info(name, optional_type_proto)]
+                else:
+                    inputs += [helper.make_tensor_value_info(name, ttype, shape)]
+
+        n_outputs = len(output_shapes)
+        output_names = list(string.ascii_lowercase)[n_inputs : n_inputs + n_outputs]
+        if output_types is None:
+            output_types = [TensorProto.FLOAT] * n_outputs
+        is_sequence = [0 if id not in seq_outputs else 1 for id in range(n_outputs)]
+        is_optional = [
+            0 if id not in optional_outputs else 1 for id in range(n_outputs)
+        ]
+        output_shapes_cast = cast(
+            List[List[int]],
+            [[0] if isinstance(shape, str) else shape for shape in output_shapes],
+        )
+        outputs: List[ValueInfoProto] = []
+        for name, ttype, shape, is_seq, is_opt in zip(
+            output_names, output_types, output_shapes_cast, is_sequence, is_optional
+        ):
+            if is_seq:
+                outputs += [helper.make_tensor_sequence_value_info(name, ttype, shape)]
+            elif is_opt:
+                type_proto = helper.make_tensor_type_proto(ttype, shape)
+                optional_type_proto = helper.make_optional_type_proto(type_proto)
+                outputs += [helper.make_value_info(name, optional_type_proto)]
+            else:
+                outputs += [helper.make_tensor_value_info(name, ttype, shape)]
+
+        node = helper.make_node(op, input_names, output_names, **attrs)
+        graph = helper.make_graph([node], op, inputs, outputs, initializer)
+        start_opset = from_opset if is_upgrade else LATEST_OPSET
+        end_opset = LATEST_OPSET if is_upgrade else from_opset
+        original = helper.make_model(
+            graph,
+            producer_name="test",
+            opset_imports=[helper.make_opsetid("", start_opset)],
+        )
+        onnx.checker.check_model(original)
+        shape_inference.infer_shapes(original, strict_mode=True)
+
+        converted = version_converter.convert_version(original, end_opset)
+        onnx.checker.check_model(converted)
+        shape_inference.infer_shapes(converted, strict_mode=True)
diff --git a/onnx/test/version_converter/automatic_downgrade_test.py b/onnx/test/version_converter/automatic_downgrade_test.py
@@ -0,0 +1,54 @@
+# Copyright (c) ONNX Project Contributors
+
+# SPDX-License-Identifier: Apache-2.0
+import unittest
+
+import automatic_conversion_test_base
+import numpy as np
+import parameterized
+
+import onnx
+from onnx import helper
+
+#####################################################################################
+# Every test calls _test_op_conversion to downgrade a model from the most recent opset version
+# to a early version and runs checker + shape inference on the downgraded model.
+####################################################################################
+
+
+class TestAutomaticDowngrade(automatic_conversion_test_base.TestAutomaticConversion):
+    def _test_op_downgrade(self, op: str, *args, **kwargs):
+        self._test_op_conversion(op, *args, **kwargs, is_upgrade=False)
+
+    @parameterized.parameterized.expand(
+        [
+            "ReduceL1",
+            "ReduceL2",
+            "ReduceLogSum",
+            "ReduceLogSumExp",
+            "ReduceMean",
+            "ReduceMax",
+            "ReduceMin",
+            "ReduceProd",
+            "ReduceSum",
+            "ReduceSumSquare",
+        ]
+    )
+    def test_reduce_ops(self, op) -> None:
+        # TODO: need to add test cases for missing axes input which depends on this pr:
+        # https://github.com/onnx/onnx/pull/5613
+        axes = helper.make_tensor(
+            "b", onnx.TensorProto.INT64, dims=[3], vals=np.array([0, 1, 2])
+        )
+        self._test_op_downgrade(
+            op,
+            from_opset=13,
+            input_shapes=[[3, 4, 5], [3]],
+            output_shapes=[[1, 1, 1]],
+            input_types=[onnx.TensorProto.FLOAT, onnx.TensorProto.INT64],
+            initializer=[axes],
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/onnx/test/automatic_upgrade_test.py → ...rsion_converter/automatic_upgrade_test.py b/onnx/test/automatic_upgrade_test.py → ...rsion_converter/automatic_upgrade_test.py
@@ -1,117 +1,27 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-import string
 import unittest
-from typing import Any, Dict, List, Optional, Sequence, Union, cast
 
+import automatic_conversion_test_base
 import numpy as np
 import pytest
 
 import onnx
-from onnx import TensorProto, ValueInfoProto, helper, shape_inference, version_converter
+from onnx import TensorProto, helper
 
 #####################################################################################
-# Every test creates a model containing a single operator from the lowest possible
-# opset version, upgrades it to the most recent opset version and then runs checker +
-# shape inference on the upgraded model.
+# Every test calls _test_op_conversion to upgrade a model from an initial opset version
+# to the most recent version and runs checker and shape inference on the final upgraded model.
 ####################################################################################
 
-LATEST_OPSET = onnx.defs.onnx_opset_version()
 tested_ops = []
 
 
-class TestAutomaticUpgrade(unittest.TestCase):
-    def _test_op_upgrade(
-        self,
-        op: str,
-        from_opset: int,
-        input_shapes: Sequence[Union[Sequence[Optional[int]], str]] = ((3, 4, 5),),
-        output_shapes: Sequence[Sequence[Optional[int]]] = ((3, 4, 5),),
-        input_types: Optional[Sequence[Any]] = None,
-        output_types: Optional[Sequence[Any]] = None,
-        initializer: Sequence[Any] = (),
-        attrs: Optional[Dict[str, Any]] = None,
-        seq_inputs: Sequence[int] = (),
-        seq_outputs: Sequence[int] = (),
-        optional_inputs: Sequence[int] = (),
-        optional_outputs: Sequence[int] = (),
-    ) -> None:
-        if attrs is None:
-            attrs = {}
-
+class TestAutomaticUpgrade(automatic_conversion_test_base.TestAutomaticConversion):
+    def _test_op_upgrade(self, op, *args, **kwargs):
         tested_ops.append(op)
-
-        n_inputs = len(input_shapes)
-        letters = list(string.ascii_lowercase)[:n_inputs]
-        input_names = [
-            letter if shape != "" else ""
-            for (letter, shape) in zip(letters, input_shapes)
-        ]
-        if input_types is None:
-            input_types = [TensorProto.FLOAT] * n_inputs
-        is_sequence = [0 if id not in seq_inputs else 1 for id in range(n_inputs)]
-        is_optional = [0 if id not in optional_inputs else 1 for id in range(n_inputs)]
-        # turn empty strings into [0] to ease type analysis, even though those entries
-        # will be ignored
-        input_shapes_cast = cast(
-            List[List[int]],
-            [[0] if isinstance(shape, str) else shape for shape in input_shapes],
-        )
-        inputs: List[ValueInfoProto] = []
-        for name, ttype, shape, is_seq, is_opt in zip(
-            input_names, input_types, input_shapes_cast, is_sequence, is_optional
-        ):
-            if name != "":
-                if is_seq:
-                    inputs += [
-                        helper.make_tensor_sequence_value_info(name, ttype, shape)
-                    ]
-                elif is_opt:
-                    type_proto = helper.make_tensor_type_proto(ttype, shape)
-                    optional_type_proto = helper.make_optional_type_proto(type_proto)
-                    inputs += [helper.make_value_info(name, optional_type_proto)]
-                else:
-                    inputs += [helper.make_tensor_value_info(name, ttype, shape)]
-
-        n_outputs = len(output_shapes)
-        output_names = list(string.ascii_lowercase)[n_inputs : n_inputs + n_outputs]
-        if output_types is None:
-            output_types = [TensorProto.FLOAT] * n_outputs
-        is_sequence = [0 if id not in seq_outputs else 1 for id in range(n_outputs)]
-        is_optional = [
-            0 if id not in optional_outputs else 1 for id in range(n_outputs)
-        ]
-        output_shapes_cast = cast(
-            List[List[int]],
-            [[0] if isinstance(shape, str) else shape for shape in output_shapes],
-        )
-        outputs: List[ValueInfoProto] = []
-        for name, ttype, shape, is_seq, is_opt in zip(
-            output_names, output_types, output_shapes_cast, is_sequence, is_optional
-        ):
-            if is_seq:
-                outputs += [helper.make_tensor_sequence_value_info(name, ttype, shape)]
-            elif is_opt:
-                type_proto = helper.make_tensor_type_proto(ttype, shape)
-                optional_type_proto = helper.make_optional_type_proto(type_proto)
-                outputs += [helper.make_value_info(name, optional_type_proto)]
-            else:
-                outputs += [helper.make_tensor_value_info(name, ttype, shape)]
-
-        node = helper.make_node(op, input_names, output_names, **attrs)
-        graph = helper.make_graph([node], op, inputs, outputs, initializer)
-        original = helper.make_model(
-            graph,
-            producer_name="test",
-            opset_imports=[helper.make_opsetid("", from_opset)],
-        )
-        onnx.checker.check_model(original)
-        shape_inference.infer_shapes(original, strict_mode=True)
-
-        converted = version_converter.convert_version(original, LATEST_OPSET)
-        onnx.checker.check_model(converted)
-        shape_inference.infer_shapes(converted, strict_mode=True)
+        self._test_op_conversion(op, *args, **kwargs, is_upgrade=True)
 
     def test_Abs(self) -> None:
         self._test_op_upgrade("Abs", 1, attrs={"consumed_inputs": [0]})

diff --git a/onnx/version_converter/convert.h b/onnx/version_converter/convert.h
@@ -545,6 +545,17 @@ class DefaultVersionConverter : public BaseVersionConverter {
     registerAdapter(std::make_unique<AxesAttributeToInput>("ReduceProd", OpSetID(17), OpSetID(18)));
     registerAdapter(std::make_unique<AxesAttributeToInput>("ReduceSumSquare", OpSetID(17), OpSetID(18)));
 
+    /******** 18 -> 17 ********/
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceL1", OpSetID(18), OpSetID(17)));
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceL2", OpSetID(18), OpSetID(17)));
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceLogSum", OpSetID(18), OpSetID(17)));
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceLogSumExp", OpSetID(18), OpSetID(17)));
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceMax", OpSetID(18), OpSetID(17)));
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceMean", OpSetID(18), OpSetID(17)));
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceMin", OpSetID(18), OpSetID(17)));
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceProd", OpSetID(18), OpSetID(17)));
+    registerAdapter(std::make_unique<AxesInputToAttribute>("ReduceSumSquare", OpSetID(18), OpSetID(17)));
+
     /******** 18 -> 19 ********/
     registerAdapter(std::make_unique<CompatibleAdapter>("Equal", OpSetID(18), OpSetID(19)));
     registerAdapter(std::make_unique<CompatibleAdapter>("AveragePool", OpSetID(18), OpSetID(19)));