What causes "Unexpected det output dims: [${dims.join(", ")}]" in PaddlePaddle/PaddleOCR?

Use the detection model assets shipped or referenced by this paddleocr-js version

Unexpected det output dims: [${dims.join(", ")}]

Error message

Unexpected det output dims: [${dims.join(", ")}]

What it means

Thrown by getDetMap() when slicing the detection model's output tensor: the map interpretation supports only 4D [N,C,H,W] or 3D [C,H,W]/[N,H,W] layouts. Any other rank (e.g. a 2D tensor) means the ONNX model's output shape is not what the postprocessing code was written for — typically a wrong or re-exported detection model.

Source

Thrown at paddleocr-js/packages/core/src/models/det.ts:345

  const chw = toBgrFloatCHWFromBgr(bgr.data, dstW, dstH, config.normalize);
  resized.delete();
  bgr.delete();

  return {
    tensor: new ort.Tensor("float32", chw, [1, 3, dstH, dstW]),
    srcW,
    srcH,
    dstW,
    dstH
  };
}

function getDetMap(outputTensor: Tensor): { data: Float32Array; h: number; w: number } {
  const dims = outputTensor.dims;
  const data = outputTensor.data as Float32Array;
  if (dims.length === 4) return { data, h: dims[2], w: dims[3] };
  if (dims.length === 3) return { data, h: dims[1], w: dims[2] };
  throw new Error(`Unexpected det output dims: [${dims.join(", ")}]`);
}

function createBatchDetTensor(
  ort: OrtModule,
  preps: DetPreprocessResult[],
  maxH: number,
  maxW: number
): Tensor {
  const batch = preps.length;
  const plane = 3 * maxH * maxW;
  const out = new Float32Array(batch * plane);
  for (let i = 0; i < batch; i += 1) {
    const prep = preps[i];
    const chw = prep.tensor.data as Float32Array;
    const { dstH, dstW } = prep;
    const base = i * plane;
    for (let c = 0; c < 3; c += 1) {
      const srcChannelBase = c * dstH * dstW;

View on GitHub (pinned to 2661c7c0ef)

Solutions

Use the detection model assets shipped or referenced by this paddleocr-js version
If exporting your own model, ensure the output tensor retains a 3D or 4D layout ([N,C,H,W] preferred)
Log the actual dims from the caught error and compare against the expected [batch,1,H,W]
Verify you loaded the det model into the det slot and not a rec/cls model

Defensive patterns

Strategy: try-catch

Type guard

function isDetMapDims(dims: readonly number[]): dims is [number, number, number] | [number, number, number, number] {
  return dims.length === 3 || dims.length === 4;
}

Try / catch

try { await detModel.predict(cv, mats); } catch (e) {
  if (e instanceof Error && /Unexpected det output dims/.test(e.message)) {
    // model/layout mismatch: switch back to the bundled det ONNX asset
  } else throw e;
}

Prevention

Do not substitute det ONNX files from other releases without checking output layout [N,C,H,W]
Keep model assets and package version locked as a pair
Run a one-image smoke test after any model swap to catch shape errors early

When it happens

Trigger: Substituting a custom or newer DBNet detection ONNX file whose graph output was squeezed/reshaped differently; loading a classifier or recognition model into the det slot; an ONNX Runtime version change altering dynamic-dim squeezing on the output.

Common situations: User swaps model_dir to a self-exported PaddleOCR det model without matching the expected output layout; mixed model versions (det model from one release, code from another); export with keepdims off producing rank-2 output.

Related errors

AI-assisted analysis of PaddlePaddle/PaddleOCR@2661c7c0ef (2026-08-14). Data as JSON: /api/errors/3f2124d131729cbc. Report an issue: GitHub.