[Bug] parse the output number of MaxPool incorrectly #14614

Activity

1. jikechao

   added

   needs-triagePRs or issues that need to be investigated by maintainers to find the right assignees to address itPRs or issues that need to be investigated by maintainers to find the right assignees to address it

   type: bug

   on Apr 13, 2023
2. jikechao commented on Apr 14, 2023

   

   jikechao

   on Apr 14, 2023

   MemberAuthor

   More actions

   @masahi @jwfromm @vvchernov Could you help me check this crash? Thanks a lot!
3. vvchernov commented on Apr 14, 2023

   

   vvchernov

   on Apr 14, 2023

   Contributor

   More actions

   Hello @jikechao! It is simply not supported on ONNX front-end of TVM. The implementation of MaxPool (AveragePool as well) was done for onnx opset = 1, but optional second output (Indices) appeared only since opset = 8. It needs to extend implementation for the op, if you do it, it will be good
4. jikechao commented on Apr 15, 2023

   

   jikechao

   on Apr 15, 2023

   Last edited by jikechao

   MemberAuthor

   More actions

   @vvchernov Thanks for your comments! Indeed, this crash was caused by Indices appearing since opset=8. However, this feature existing about 5 years. Thus, Supporting Indices to avoid this crash may not be trivial. Thus, I'll try to fix it.
5. vvchernov commented on Apr 18, 2023

   

   vvchernov

   on Apr 18, 2023

   Contributor

   More actions

   Hello @jikechao! Some nits may be help you to fix. There can be problems with optional output and indices calculation, in both cases you can see how it was implemented for scatter_elements (ScatterElements class)
6. tqchen

   closed this as not plannedon Feb 7, 2025
7. sunni13830 commented on Nov 7, 2026

   

   sunni13830

   on Nov 7, 2026

   More actions

   This also affects the Relax ONNX frontend in TVM 0.28.dev0. A one-output MaxPool executes correctly, but requesting the legal optional indices output fails with Missing outputs during conversion. Expected 2 but Got 1 in MaxPool. I am adding a self-contained Relax reproducer here rather than opening a duplicate report.

   Tested on Linux x86_64, LLVM CPU, TVM 0.28.dev0 (installed binary revision unknown); no current-main rebuild.

   import numpy as np
   import onnx
   import tvm
   from onnx import TensorProto as T, helper as h
   from onnx.reference import ReferenceEvaluator
   from tvm import relax
   from tvm.relax.frontend.onnx import from_onnx
   
   x = np.arange(16, dtype="float32").reshape(1, 1, 4, 4)
   for indices in [False, True]:
       outputs = [h.make_tensor_value_info("y", T.FLOAT, [1, 1, 2, 2])]
       if indices:
           outputs.append(h.make_tensor_value_info("i", T.INT64, [1, 1, 2, 2]))
       model = h.make_model(h.make_graph([
           h.make_node("MaxPool", ["x"], [v.name for v in outputs], kernel_shape=[2, 2], strides=[2, 2])],
           "pool", [h.make_tensor_value_info("x", T.FLOAT, x.shape)], outputs),
           opset_imports=[h.make_opsetid("", 13)], ir_version=10)
       onnx.checker.check_model(model, full_check=True)
       expected = ReferenceEvaluator(model).run(None, {"x": x})
       np.testing.assert_array_equal(expected[0], [[[[5, 7], [13, 15]]]])
       if indices:
           np.testing.assert_array_equal(expected[1], [[[[5, 7], [13, 15]]]])
       try:
           mod = from_onnx(model, keep_params_in_input=False)
           vm = relax.VirtualMachine(relax.build(mod, target="llvm"), tvm.cpu())
           actual = vm["main"](tvm.runtime.tensor(x))
           if indices:
               raise AssertionError("Expected the optional-output failure")
           np.testing.assert_array_equal(actual.numpy(), expected[0])
           print("single-output control passed")
       except Exception as error:
           if not indices:
               raise
           assert "MaxPool" in str(error) and ("output" in str(error).lower() or "Output" in str(error))
           print("indices failure:", type(error).__name__, str(error).splitlines()[-1])