huggingface/candle · error

Unsupported coordinate_transformation_mode for resize: {}

Error message

Unsupported coordinate_transformation_mode for resize: {}

What it means

candle-onnx's Resize implementation only supports coordinate_transformation_mode="asymmetric". ONNX defines several modes (half_pixel, align_corners, pytorch_half_pixel, tf_half_pixel_for_nn, etc.) and the default is half_pixel, so most exported models trigger this error even when mode/nearest_mode are supported.

Source

Thrown at candle-onnx/src/eval.rs:2358

                let coordinate_transformation_mode =
                    get_attr_opt::<str>(node, "coordinate_transformation_mode")?
                        .unwrap_or("half_pixel");
                // Interpolation mode: nearest, linear, or cubic.
                let mode = get_attr_opt::<str>(node, "mode")?.unwrap_or("nearest");
                // How to determine the "nearest" pixel in nearest interpolation mode.
                let nearest_mode =
                    get_attr_opt::<str>(node, "nearest_mode")?.unwrap_or("round_prefer_floor");

                if mode != "nearest" {
                    bail!("Unsupported resize mode: {}", mode);
                }

                if nearest_mode != "floor" {
                    bail!("Unsupported nearest_mode for resize: {}", nearest_mode);
                }

                if coordinate_transformation_mode != "asymmetric" {
                    bail!(
                        "Unsupported coordinate_transformation_mode for resize: {}",
                        coordinate_transformation_mode
                    );
                }

                let h = output_dims[2];
                let w = output_dims[3];
                let output = input.upsample_nearest2d(h, w)?;

                values.insert(node.output[0].clone(), output);
            }
            "Trilu" => {
                let input = get(&node.input[0])?;

                // Get the diagonal offset 'k' from the second input if provided
                let k = if node.input.len() > 1 && !node.input[1].is_empty() {
                    to_vec0_flexible::<i64>(get(&node.input[1])?)?
                } else {

View on GitHub (pinned to d5fee525bf)

Solutions

1. Rewrite the Resize node attributes to coordinate_transformation_mode="asymmetric" via onnx-surgement/graph surgery (verify numerics)
2. Add half_pixel support to candle-onnx eval.rs
3. Compute the resize outside the ONNX graph in candle code
4. Use onnxruntime for this model

Example fix

// before
// coordinate_transformation_mode = "half_pixel" (default)
// after
node.attribute.push(onnx_attr("coordinate_transformation_mode", "asymmetric"));

if node.op_type == "Resize" {
    let ctm = get_attr_opt::<String>(node, "coordinate_transformation_mode")?.unwrap_or_else(|| "half_pixel".into());
    if ctm != "asymmetric" { return Err(format!("ctm {} unsupported", ctm)); }
}

fn uses_asymmetric_ctm(node: &Node) -> bool {
    get_attr_opt::<String>(node, "coordinate_transformation_mode").ok().flatten().map_or(false, |m| m == "asymmetric")
}

match eval(...) {
    Err(e) if e.contains("coordinate_transformation_mode") => fallback_to_onnxruntime(model),
    other => other,
}

When it happens

Trigger: Evaluating a Resize node whose coordinate_transformation_mode attribute is anything other than "asymmetric", including the spec default "half_pixel".

Common situations: PyTorch exports using align_corners=False produce half_pixel; align_corners=True exports produce align_corners; TF exports produce tf_half_pixel_for_nn — all rejected.

Related errors

• Unsupported resize mode: {}
• Unsupported nearest_mode for resize: {}
• Either scales or sizes should be present
• attribute {} was of type TENSOR, but no tensor was found
• attribute {} of type TENSOR was an invalid data_type number

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/006cfd4b13e8519f. Report an issue: GitHub.