[Bug] [Relax][ONNX] NonMaxSuppression silently replaces a runtime output limit with zero

未关闭
  • 哪里不对？
#20,898 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
1-2 天
新手友好度
67/100
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
machine-learning

调研方向

Start in python/tvm/relax/frontend/onnx/onnx_frontend.py at the converter lines linked in the report, and trace how a runtime max_output_boxes_per_class Var is handled when it is absent from the initializer table. Compare the dynamic-input and initializer cases using the reproduction with scalar inputs and an explicit IoU threshold. Done means the runtime limit selects the same two rows as the constant limit, with relevant tests passing.

由索引模型根据 Issue 内容生成。

描述

needs-triage type: bug
Expected behavior

A runtime max_output_boxes_per_class=2 should select the same two disjoint boxes as an initializer containing the identical value.

Actual behavior

The constant-limit control matches ORT. Making only the limit a graph input silently changes the result from two rows to no selections:

constant limit (2, 3) [[0, 0, 0], [0, 0, 1]]
runtime limit (0, 3) []

The converter substitutes zero when a limit Var is absent from the initializer table. This case uses zero-dimensional scalars and an explicit IoU threshold, unlike #19693 and #20506.

Environment

TVM 0.28.dev0, LLVM 18.1.3, Linux x86_64; Python 3.12.3, NumPy 2.5.4, ONNX 1.22.0, ORT 1.29.0. The installed binary build revision is unknown; current main was inspected statically, not rebuilt.

Steps to reproduce

Run with an LLVM-enabled TVM installation:

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

boxes = np.array([[[0, 0, 1, 1], [2, 2, 3, 3]]], dtype="float32")
scores = np.array([[[0.9, 0.8]]], dtype="float32")
limit = np.array(2, dtype="int64")
expected = np.array([[0, 0, 0], [0, 0, 1]], dtype="int64")
for dynamic in (False, True):
    inputs = [h.make_tensor_value_info("boxes", T.FLOAT, [1, 2, 4]),
              h.make_tensor_value_info("scores", T.FLOAT, [1, 1, 2])]
    constants = [nh.from_array(np.array(0.5, dtype="float32"), "iou")]
    feeds = {"boxes": boxes, "scores": scores}
    if dynamic:
        inputs.append(h.make_tensor_value_info("limit", T.INT64, []))
        feeds["limit"] = limit
    else:
        constants.append(nh.from_array(limit, "limit"))
    model = h.make_model(h.make_graph([
        h.make_node("NonMaxSuppression", ["boxes", "scores", "limit", "iou"], ["y"])],
        "nms", inputs, [h.make_tensor_value_info("y", T.INT64, [None, 3])], constants),
        opset_imports=[h.make_opsetid("", 11)], ir_version=10)
    onnx.checker.check_model(model, full_check=True)
    options = ort.SessionOptions()
    options.intra_op_num_threads = options.inter_op_num_threads = 1
    options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
    source = ort.InferenceSession(model.SerializeToString(), options,
        providers=["CPUExecutionProvider"]).run(None, feeds)[0]
    np.testing.assert_array_equal(source, expected)
    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(v) for v in feeds.values()]).numpy()
    print("runtime limit" if dynamic else "constant limit", actual.shape, actual.tolist())
    if not dynamic:
        np.testing.assert_array_equal(actual, expected)
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。