onnx export to support older pytorch with example_outputs argument (P…

…roject-MONAI#6309)

Fixes # .
Project-MONAI#6297

### Description

PyTorch onnx exporter API has been changed since 1.10 to remove
example_outputs as a required input argument. Special handling is added
in this PR to support PyTorch version older than 1.10. Unit test is also
extended to covert this case.

### Types of changes
<!--- Put an `x` in all the boxes that apply, and remove the not
applicable items -->
- [x] Non-breaking change (fix or new feature that would not break
existing functionality).
- [ ] Breaking change (fix or new feature that would cause existing
functionality to change).
- [x] New tests added to cover the changes.
- [ ] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [ ] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.
- [ ] In-line docstrings updated.
- [ ] Documentation updated, tested `make html` command in the `docs/`
folder.

Signed-off-by: Liqun Fu <[email protected]>

monai/networks/utils.py

@@ -584,8 +584,10 @@ def convert_to_onnx(
         inputs: input sample data used by pytorch.onnx.export. It is also used in ONNX model verification.
         input_names: optional input names of the ONNX model.
         output_names: optional output names of the ONNX model.
-        opset_version: version of the (ai.onnx) opset to target. Must be >= 7 and <= 16, for more
-            details: https://github.com/onnx/onnx/blob/main/docs/Operators.md.
+        opset_version: version of the (ai.onnx) opset to target. Must be >= 7 and not exceed
+        the latest opset version supported by PyTorch, for more details:
+            https://github.com/onnx/onnx/blob/main/docs/Operators.md and
+            https://github.com/pytorch/pytorch/blob/master/torch/onnx/_constants.py
         dynamic_axes: specifies axes of tensors as dynamic (i.e. known only at run-time). If set to None,
             the exported model will have the shapes of all input and output tensors set to match given
             ones, for more details: https://pytorch.org/docs/stable/onnx.html#torch.onnx.export.
@@ -603,31 +605,45 @@ def convert_to_onnx(
     """
     model.eval()
     with torch.no_grad():
+        torch_versioned_kwargs = {}
         if use_trace:
-            script_module = torch.jit.trace(model, example_inputs=inputs)
+            # let torch.onnx.export to trace the model.
+            mode_to_export = model
         else:
-            script_module = torch.jit.script(model, **kwargs)
+            if not pytorch_after(1, 10):
+                if "example_outputs" not in kwargs:
+                    # https://github.com/pytorch/pytorch/blob/release/1.9/torch/onnx/__init__.py#L182
+                    raise TypeError(
+                        "example_outputs is required in scripting mode before PyTorch 1.10."
+                        "Please provide example outputs or use trace mode to export onnx model."
+                    )
+                torch_versioned_kwargs["example_outputs"] = kwargs["example_outputs"]
+                del kwargs["example_outputs"]
+            mode_to_export = torch.jit.script(model, **kwargs)
+
         if filename is None:
             f = io.BytesIO()
             torch.onnx.export(
-                script_module,
-                inputs,
+                mode_to_export,
+                tuple(inputs),
                 f=f,
                 input_names=input_names,
                 output_names=output_names,
                 dynamic_axes=dynamic_axes,
                 opset_version=opset_version,
+                **torch_versioned_kwargs,
             )
             onnx_model = onnx.load_model_from_string(f.getvalue())
         else:
             torch.onnx.export(
-                script_module,
-                inputs,
+                mode_to_export,
+                tuple(inputs),
                 f=filename,
                 input_names=input_names,
                 output_names=output_names,
                 dynamic_axes=dynamic_axes,
                 opset_version=opset_version,
+                **torch_versioned_kwargs,
             )
             onnx_model = onnx.load(filename)
 

tests/test_convert_to_onnx.py

@@ -19,42 +19,60 @@
 
 from monai.networks import convert_to_onnx
 from monai.networks.nets import SegResNet, UNet
+from monai.utils.module import pytorch_after
 from tests.utils import SkipIfBeforePyTorchVersion, SkipIfNoModule, optional_import
 
 if torch.cuda.is_available():
     TORCH_DEVICE_OPTIONS = ["cpu", "cuda"]
 else:
     TORCH_DEVICE_OPTIONS = ["cpu"]
-TESTS = list(itertools.product(TORCH_DEVICE_OPTIONS, [True, False]))
+TESTS = list(itertools.product(TORCH_DEVICE_OPTIONS, [True, False], [True, False]))
 TESTS_ORT = list(itertools.product(TORCH_DEVICE_OPTIONS, [True]))
 
 onnx, _ = optional_import("onnx")
 
 
 @SkipIfNoModule("onnx")
-@SkipIfBeforePyTorchVersion((1, 10))
+@SkipIfBeforePyTorchVersion((1, 9))
 class TestConvertToOnnx(unittest.TestCase):
     @parameterized.expand(TESTS)
-    def test_unet(self, device, use_ort):
+    def test_unet(self, device, use_trace, use_ort):
         if use_ort:
             _, has_onnxruntime = optional_import("onnxruntime")
             if not has_onnxruntime:
                 self.skipTest("onnxruntime is not installed probably due to python version >= 3.11.")
         model = UNet(
             spatial_dims=2, in_channels=1, out_channels=3, channels=(16, 32, 64), strides=(2, 2), num_res_units=0
         )
-        onnx_model = convert_to_onnx(
-            model=model,
-            inputs=[torch.randn((16, 1, 32, 32), requires_grad=False)],
-            input_names=["x"],
-            output_names=["y"],
-            verify=True,
-            device=device,
-            use_ort=use_ort,
-            use_trace=True,
-            rtol=1e-3,
-            atol=1e-4,
-        )
+        if pytorch_after(1, 10) or use_trace:
+            onnx_model = convert_to_onnx(
+                model=model,
+                inputs=[torch.randn((16, 1, 32, 32), requires_grad=False)],
+                input_names=["x"],
+                output_names=["y"],
+                verify=True,
+                device=device,
+                use_ort=use_ort,
+                use_trace=use_trace,
+                rtol=1e-3,
+                atol=1e-4,
+            )
+        else:
+            # https://github.com/pytorch/pytorch/blob/release/1.9/torch/onnx/__init__.py#L182
+            # example_outputs is required in scripting mode before PyTorch 3.10
+            onnx_model = convert_to_onnx(
+                model=model,
+                inputs=[torch.randn((16, 1, 32, 32), requires_grad=False)],
+                input_names=["x"],
+                output_names=["y"],
+                example_outputs=[torch.randn((16, 3, 32, 32), requires_grad=False)],
+                verify=True,
+                device=device,
+                use_ort=use_ort,
+                use_trace=use_trace,
+                rtol=1e-3,
+                atol=1e-4,
+            )
         self.assertTrue(isinstance(onnx_model, onnx.ModelProto))
 
     @parameterized.expand(TESTS_ORT)