ai.onnx.DeformConv

ai.onnx · standard ONNX operator · ONNX opset ≥ 22

Description

Performs deformable convolution by sampling input at spatially offset locations specified per output position, enabling the kernel to adapt its receptive field shape. The offset tensor provides (y, x) offsets for each kernel point and output location; fractional offsets are resolved via bilinear interpolation and out-of-bounds locations contribute zero. An optional mask tensor modulates each sampling point's contribution.

See the ONNX DeformConv spec for the reference semantics.

Inputs

Name	Bind key	Logical dtype	Rank	Shape	Description	Presence
`X`	`x`	`T`	`4`	—	Input data tensor of shape `(N, C, H, W)` for 2D or `(N, C, D1, ..., Dn)` in general.	required
`W`	`w`	`T`	`4`	—	Convolution weight tensor of shape `(oC, C/group, kH, kW)`.	required
`offset`	`offset`	`T`	`4`	—	Per-output-position sampling offsets of shape `(N, offset_group * kH * kW * 2, oH, oW)` for 2D data.	required
`B`	`bias`	`T`	`1`	—	Optional 1D bias of length oC added to the convolution output.	optional
`mask`	`mask`	`T`	`4`	—	Optional modulation mask of shape `(N, offset_group * kH * kW, oH, oW)` scaling each sampled kernel point; defaults to ones.	optional

Outputs

Name	Bind key	Logical dtype	Rank	Shape	Description	Presence
`Y`	`y`	`T`	`4`	derived; see description	Output tensor of shape `(N, oC, oH, oW)` containing the deformable convolution result.	required

Attributes

Attributes and default values (overridable per request):

Attribute	Default	Description
`group`	`1`	Number of groups the input channels C and output channels oC are divided into; both must be divisible by group.
`offset_group`	`1`	Number of offset groups; input channels `C` must be divisible by `offset_group`.
`dilations`	—	Dilation factors for the spatial axes in height-width order; defaults to `[1, 1]`.
`kernel_shape`	—	Kernel dimensions in height-width order. When omitted, they are inferred from the spatial dimensions of `W`.
`pads`	—	Padding at the beginning and end of each spatial axis in `[top, left, bottom, right]` order; defaults to zeros.
`strides`	—	Stride factors for the spatial axes in height-width order; defaults to `[1, 1]`.

Type constraints

Variable	Allowed dtypes
`T`	`float32`

Files

metadata.json — kernel metadata (id, digests, provenance)
manifest.json — the op contract (source of truth)
test.json — correctness cases
bench.json — benchmark + tuning cases
deform-conv2d-aligned-groups.wgsl.jinja
deform-conv2d-nchw.wgsl.jinja

Use with `@huggingface/kernels`

The loader derives every required output's shape and logical dtype from the manifest contract and this call. It then allocates the result tensors automatically.

The version: 1 option selects the published kernel contract; it is independent of any operator opset, contrib since_version, or model version.

Replace each *Data placeholder with a typed array containing the corresponding input data.

import { getKernel } from "@huggingface/kernels";

const kernel = await getKernel("webgpu-kernels/ai.onnx.DeformConv", { version: 1 });
const { y } = await kernel({
  x: { data: xData, shape: [1, 1, 1, 1] },
  w: { data: wData, shape: [1, 1, 1, 1] },
  offset: { data: offsetData, shape: [1, 2, 1, 1] },
});

Downloads last month: -

kernel

webgpu

wgsl

apache-2.0

WebGPU

Requires WebGPU support. See the compatibility table.