Question: how to handle input model not containing tensor shapes?

[radu-matei]

When using the ONNX C API through the Rust bindings, in more than one place we rely on the input model containing the shape for input and output tensors - this is fine in set_input, where the guest provides the shape of the input tensor (if the model doesn't contain it, just use the one supplied by the guest) - but is prone to failing in compute and get_output, when we have to store the output tensors and then return them to the guest into a unidimensional buffer, without any information about its shape.

Currently, in compute, we have to unwrap on the output's dimensions(), which could very well be None.

let shape = outputs
    .get(index)
    .unwrap()
    .dimensions()
    .map(|d| d.unwrap())
    .collect::<Vec<usize>>();

[radu-matei]

Update: we can handle this gracefully within the Tract runtime, still need some more experimentation on how to do it properly when using the native ONNX runtime.

[radu-matei]

Update: I think I managed to fix this for the native ONNX runtime as well, PR incoming.