Troubleshooting

Messages you might see when uploading, and how to fix them.

MessageFix
Input "x" is in the model but missing from peekml.yamlAdd it under inputs: properties: with exactly the same name. Names are case-sensitive.
Input "x" is in peekml.yaml but the model has no such inputCheck the spelling against the model's input names (see step 3 of the Quick start).
"X" holds 4 values in one input…Your model takes all features in one input. Re-export with one input per feature: scikit-learn, PyTorch.
"input" looks like a fixed-size image input…PeekML passes pictures at their own size. Build the resizing and scaling into the model so it takes uint8 pixels shaped [1, height, width, 3]: see Image models.
The test run failed with a 100 × 80 picture…The model only works at one picture size. Export it with flexible height and width (dynamic_shapes), as in Image models.
"output_probability" is a list of {class: probability} pairs…Export with options={…: {"zipmap": False}} to show probabilities. Without it, only the predicted class is shown.
The model file could not be loadedRe-export with target_opset 17, as in the export guide. For PyTorch, also use external_data=False.
probabilities: the model returns 3 values but x-classes lists 2 classesList one class name per probability column in x-classes, in the model's order.
an array output needs x-classesAdd x-classes, or remove type: array to show the raw values instead.
a slider needs minimum, maximum and multipleOf (the step)Add minimum, maximum and multipleOf, or remove x-widget: slider.
Not enough space: you're using … of 500 MBDelete models you no longer need, or make the model smaller (int8 quantization usually shrinks it about 4×).

On the model page

  • "Fill in the inputs and press Submit": the form opens empty. Add examples to peekml.yaml so visitors can try the model with one click.
  • "default" is no longer used (a warning when publishing): there are no starting values any more. Remove default and add an example instead.
  • The prediction looks wrong: check that dropdown options match the values the model was trained on, and that inputs are scaled the same way as in training.