dutZ1855 opened a new issue, #18608: URL: https://github.com/apache/tvm/issues/18608
### Expected behavior TVM should be able to import and compile an ONNX `Resize` model with **non-4D** input tensors (e.g. 3D), matching ONNX Runtime behavior (and also OpenVINO, which can run this model). Per the ONNX `Resize` operator spec, `Resize` supports resizing **N-D** tensors (not restricted to 4D): - Spec: https://onnx.ai/onnx/operators/onnx__Resize.html ### Actual behavior For the following model, <img width="145" height="284" alt="Image" src="https://github.com/user-attachments/assets/49af8147-d1f0-429d-b2f2-3e5f65fa71af"; /> <img width="493" height="681" alt="Image" src="https://github.com/user-attachments/assets/00e596f3-c025-45e7-877c-1d2acbbe6177"; /> When importing the attached model with TVM Relax (`tvm.relax.frontend.onnx.from_onnx`), TVM fails during ONNX conversion with: > `AssertionError: Only resize2d is currently supported.` The failure is from TVM's ONNX `Resize` converter which asserts `ndims == 4`. ``` Error converting operator Resize, with inputs: [x, None, None, sizes] Traceback (most recent call last): File "DLCompilers/bug/tvm/resize_only_resize2d/run_repro.py", line 66, in <module> test(onnx_model) ^^^^^^^^^^^^^^^^ File "DLCompilers/bug/tvm/resize_only_resize2d/run_repro.py", line 51, in test tvm_model = from_onnx(model, opset=18, keep_params_in_input=True) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 4260, in from_onnx return g.from_onnx(graph, opset) ^^^^^^^^^^^^^^^^^^^^^^^^^ File "DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 3890, in from_onnx self._construct_nodes(graph) File "DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 4071, in _construct_nodes raise err File "DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 4066, in _construct_nodes op = self._convert_operator(op_name, inputs, attr, self.opset) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 4166, in _convert_operator sym = op_function(self.bb, inputs, attrs, [self._nodes, self._params]) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 2221, in _impl_v18 assert ndims == 4, "Only resize2d is currently supported." ^^^^^^^^^^ AssertionError: Only resize2d is currently supported. ``` ONNX Runtime can execute the same model successfully. ``` ORT run finished [[[0.1257]]] ``` ### Environment Operating System：Ubuntu 22.04.4 LTS TVM version：0.23.0dev pytorch version:2.9.1 ort version:1.23.2 onnx version: 1.20.0 python:3.11.14 ### Steps to reproduce Download the model and run the following code to obtain the results. [model.zip](https://github.com/user-attachments/files/24328158/model.zip) ``` from typing import Dict, List, Literal, Optional import sys import os from pathlib import Path import numpy as np import onnx import onnxruntime from onnx import ModelProto, TensorProto, helper _REPO_ROOT = Path(__file__).resolve().parents[3] _TVM_PYTHON = _REPO_ROOT / "tvm" / "python" _TVM_BUILD = _REPO_ROOT / "tvm" / "build" if _TVM_PYTHON.exists(): sys.path.insert(0, _TVM_PYTHON.as_posix()) os.environ.setdefault("TVM_LIBRARY_PATH", _TVM_BUILD.as_posix()) import tvm import tvm.testing from tvm import relax from tvm.relax.frontend.onnx import from_onnx import argparse import pickle def test(model: ModelProto,) -> None: model.ir_version = 8 model.opset_import[0].version = 18 with open("oracle.pkl", 'rb') as fp: inputs = pickle.load(fp) try: ort_session = onnxruntime.InferenceSession( model.SerializeToString(), providers=["CPUExecutionProvider"] ) feed = inputs.get("input", inputs) if isinstance(inputs, dict) else inputs ort_output = ort_session.run([], feed) print("ORT run finished") for idx, tensor in enumerate(ort_output): print(tensor) print("-" * 40) except Exception as e: print("This model cannot be executed by onnxruntime!") sys.exit(1) tvm_model = from_onnx(model, opset=18, keep_params_in_input=True) tvm_model = relax.transform.DecomposeOpsForInference()(tvm_model) tvm_model = relax.transform.LegalizeOps()(tvm_model) tvm_model, params = relax.frontend.detach_params(tvm_model) with tvm.transform.PassContext(opt_level=3): ex = tvm.compile(tvm_model, target="llvm") if __name__ == "__main__": onnx_model = onnx.load("model.onnx") test(onnx_model) ``` ### Triage * needs-triage -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
