How to get input tensor shape in onnxruntime-web?

[wirhabenzeit]

Given an ort session, say

session = await ort.InferenceSession.create(
  "model.onnx", { executionProviders: ["wasm"] }
)

how can I get the expected input shape?

To get the input names, I can do session.inputNames but the docs don't explain how to get the input shapes.

[hariharans29]

You can inspect the model by opening it on Netron and checking what the model expcts its input shape to be

[wirhabenzeit]

My question was about the Javascript-runtime specifically. My use case is that I load one of several model.onnx file and depending on which one I downscale the data in Javascript before feeding it into onnxruntime. Now I can of course somewhere store the information on the model input shape but it seems superfluous since this information is stored in the model itself. For instance the python API exposes such information using session.get_inputs() but I take it the Javascript runtime does not?

[hariharans29]

I see. @fs-eire / @guschmue - Does onnxruntime-web allow querying the model metadata like input shape ?

[fs-eire]

we don't support this feature so far.

The major reason is that for most of the use cases we expect users to know the input type and shape requirement, so we lack motivation to implement this feature.

There are also other trivial reasons -

• some customers want to optimize for artifact size, exporting metadata increases that for some and cannot be reduced using tree-shaking (because it crosses JS/WASM boundrary).
• implementation is non-trivial as the JS API also for node.js binding and react-native.

If you are interested in it, contribution is welcome.