| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.Resize |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 19 |
|
|
| ## Description |
|
|
| Resizes the input tensor by sampling neighboring input values. Output dimensions are determined by per-axis scale factors or explicit target sizes. Supports `nearest`, `linear`, and `cubic` interpolation with configurable coordinate transformation modes. |
|
|
| See the [ONNX `Resize` spec](https://onnx.ai/onnx/operators/onnx__Resize.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `x` | `X` | `T` | — | — | N-D input tensor to be resized. | required | |
|
|
| ## Outputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `y` | `Y` | `T` | same as `x` | — | N-D output tensor after resizing. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `antialias` | `0` | When set to 1, stretches the resampling filter during downscaling so that more input pixels contribute to each output pixel, reducing aliasing. | |
| | `axes` | — | Optional axes to resize. When omitted, the scale values or requested output size apply to every input axis. | |
| | `coordinate_transformation_mode` | `"half_pixel"` | How to map a coordinate in the output tensor back to the input tensor; supported modes include `"half_pixel"`, `"pytorch_half_pixel"`, `"align_corners"`, `"asymmetric"`, and `"tf_crop_and_resize"`. | |
| | `cubic_coeff_a` | `-0.75` | Coefficient `a` in the cubic interpolation filter; the default is -0.75. Valid only when `mode` is `"cubic"`. | |
| | `exclude_outside` | `0` | When set to 1, assigns zero weight to sampling locations outside the input tensor and renormalizes the remaining weights to sum to 1. | |
| | `extrapolation_value` | `0` | Fill value used for out-of-bounds samples when `coordinate_transformation_mode` is `"tf_crop_and_resize"`. | |
| | `keep_aspect_ratio_policy` | `"stretch"` | How explicit sizes preserve aspect ratio. This package supports only the default `"stretch"` policy; `"not_larger"` and `"not_smaller"` are unsupported. | |
| | `mode` | `"nearest"` | Interpolation mode: `"nearest"` (default), `"linear"` (bilinear/N-linear), or `"cubic"` (bicubic/N-cubic). | |
| | `nearest_mode` | `"round_prefer_floor"` | Rounding strategy used when `mode` is `"nearest"`: `"round_prefer_floor"` (default), `"round_prefer_ceil"`, `"floor"`, or `"ceil"`. | |
| | `roi` | `[]` | Values of the optional `roi` tensor, supplied as `[starts..., ends...]` with one pair per input axis or per `axes` entry; used only by `"tf_crop_and_resize"`. | |
| | `scales` | `[]` | Values of the optional `scales` tensor. Supply one positive value per input axis, or one per `axes` entry when that attribute is present; omit it when the output shape represents the exact `sizes` input. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16`, `uint8`, `int8` | |
|
|
| ## Files |
|
|
| - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, provenance) |
| - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth) |
| - [`test.json`](build/webgpu/test.json) — correctness cases |
| - [`bench.json`](build/webgpu/bench.json) — benchmark + tuning cases |
| - [`resize-antialias.wgsl.jinja`](build/webgpu/resize-antialias.wgsl.jinja) |
| - [`resize-coord-transform-5d.wgsl.jinja`](build/webgpu/resize-coord-transform-5d.wgsl.jinja) |
| - [`resize-coord-transform.wgsl.jinja`](build/webgpu/resize-coord-transform.wgsl.jinja) |
| - [`resize-cubic.wgsl.jinja`](build/webgpu/resize-cubic.wgsl.jinja) |
| - [`resize-generic.wgsl.jinja`](build/webgpu/resize-generic.wgsl.jinja) |
| - [`resize-linear-2x-stencil-x8.wgsl.jinja`](build/webgpu/resize-linear-2x-stencil-x8.wgsl.jinja) |
| - [`resize-linear-2x-stencil.wgsl.jinja`](build/webgpu/resize-linear-2x-stencil.wgsl.jinja) |
| - [`resize-nearest-integer-scale.wgsl.jinja`](build/webgpu/resize-nearest-integer-scale.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| npm install --save-exact @huggingface/kernels@0.0.1-preview.2 |
| ``` |
|
|
| Outputs with inferable metadata are allocated automatically. Explicit `outputs` entries request optional results or provide metadata that cannot be inferred from the supplied inputs and attributes. |
|
|
| This example supplies explicit metadata for: |
|
|
| - `y` |
|
|
| The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version. |
| It follows the `v1` branch as fixes land. To pin exact artifact bytes, pass a 40-character commit `revision` instead of `version`. |
|
|
| Replace each `*Data` placeholder with a typed array containing the corresponding input data. |
|
|
| ```js |
| import { getKernel } from "@huggingface/kernels"; |
| |
| const kernel = await getKernel("webgpu-kernels/ai.onnx.Resize", { version: 1 }); |
| // Explicit destinations request optional results or supply metadata that cannot be inferred. |
| const { y } = await kernel({ x: { data: xData, shape: [1, 1, 2, 4] } }, { |
| outputs: { y: { shape: [1, 1, 1, 2], dtype: "float32" } }, |
| }); |
| ``` |
|
|
