A reproducible CPU diagnostic

The red image arrived at the model as blue.

This synthetic ONNX example isolates one common pipeline mismatch: BGR input is passed to a model expecting RGB. The input shape, data type and normalization stay the same.

2 × 2 synthetic red imageNCHW · float32ONNX Runtime CPU

A small demonstration of one cause. It does not validate a production model or an entire vision pipeline.

Same model · same input · channel order changed

The channel means show the mismatch.

The model averages each input channel. With incorrect order, red is in the third channel. Reversing BGR to RGB puts red in the first.

Observed channel means after normalizing pixel values by 255
Input to modelChannel means [R, G, B]Interpretation
BGR passed as RGB[0.0, 0.0, 1.0]Order mismatch
BGR converted to RGB[1.0, 0.0, 0.0]Red in first channel

One issue at a time

A bounded diagnosis for a reproducible pipeline bug.

The proposed pilot covers one symptom in one Python script or module. The output is a written reproduction and diagnostic. A local fix or diff is included only when the cause is confirmed and fits the agreed scope.

What a useful intake includes

  1. The smallest command and code needed to reproduce the symptom.
  2. The expected output and the output observed.
  3. Python, dependency and operating system versions.
  4. CPU or GPU provider and relevant input shape, dtype, channel order and normalization.
  5. A sanitized sample the requester has permission to share.

Scope, timing and price must be agreed with the service owner after reviewing the reproduction. This page has no upload or payment flow.

The example is synthetic. This static page contains no form, tracker or upload. See data handling.