[Bug] [Relax][ONNX] Resize drops axes-scoped sizes before constructing resize2d

未关闭
  • 哪里不对？
#20,887 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
3/5
预计耗时
半天
新手友好度
65/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python
领域
compilers

调研方向

Start from the Resize converter in the Relax ONNX frontend, where sizes[2:] is sliced without applying the axes mapping. Run the provided repro.py to confirm the ValueError for axes=[2, 3] while full-rank sizes still import. Done when the axes-scoped case yields a 2-dim size and the output matches the full-rank result [1, 1, 4, 4].

由索引模型根据 Issue 内容生成。

描述

needs-triage type: bug
Expected behavior

For Resize-18, axes=[2, 3], sizes=[4, 4] on a [1, 1, 2, 2] tensor should produce the same [1, 1, 4, 4] result as full-rank sizes=[1, 1, 4, 4].

Actual behavior

Full-rank sizes import successfully. Axes-scoped sizes pass ONNX full checking and the ONNX reference evaluation, but fail during Relax import:

ValueError: Resize2D expects the given output image size to be a 2-dim shape, while the given one has ndim 0

The converter slices sizes[2:] without expanding the axes mapping. This uses the default stretch policy; the aspect-ratio change in #20846 does not cover it.

Environment

TVM 0.28.dev0 (local build; build commit unavailable), Python 3.12.3, Linux x86_64; ONNX 1.22.0, NumPy 2.5.4.

Steps to reproduce

Run with python repro.py:

import numpy as np
import onnx
from onnx import TensorProto as T, helper as h, numpy_helper as nh
from onnx.reference import ReferenceEvaluator
from tvm.relax.frontend.onnx import from_onnx

x = np.arange(4, dtype="float32").reshape(1, 1, 2, 2)
expected = x.repeat(2, axis=2).repeat(2, axis=3)
for axes, sizes in [(None, [1, 1, 4, 4]), ([2, 3], [4, 4])]:
    attrs = {} if axes is None else {"axes": axes}
    node = h.make_node("Resize", ["x", "", "", "sizes"], ["y"],
                       mode="nearest", coordinate_transformation_mode="asymmetric",
                       nearest_mode="floor", **attrs)
    model = h.make_model(h.make_graph([node], "resize",
        [h.make_tensor_value_info("x", T.FLOAT, [1, 1, 2, 2])],
        [h.make_tensor_value_info("y", T.FLOAT, [1, 1, 4, 4])],
        [nh.from_array(np.array(sizes, dtype="int64"), "sizes")]),
        opset_imports=[h.make_opsetid("", 18)], ir_version=10)
    onnx.checker.check_model(model, full_check=True)
    np.testing.assert_array_equal(ReferenceEvaluator(model).run(None, {"x": x})[0], expected)
    try:
        from_onnx(model, keep_params_in_input=False)
        print("axes", axes, "imported")
    except Exception as error:
        print("axes", axes, type(error).__name__, str(error).splitlines()[-1])
Triage
  • needs-triage
主要语言
Python
星标
13.9k
派生
4k
平均合并
8 小时 5 分钟
30 天内合并 PR
170

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

apache/tvm 的其他 Issue

查看 apache/tvm 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。