Summary

Fixes #1266.

In ONNX, optional inputs that are not provided are represented by an empty string (""). For example, a Resize node can have inputs like : ['Quant_0_out0', '', 'Resize_0_param0']

where the second entry corresponds to the optional ROI input.

At the moment, get_input_shape() attempts to resolve the shape for every input, including the empty placeholder. Since an empty string does not correspond to a valid tensor, this results in,

RuntimeError: Could not find the shape for input

This change simply skips empty input placeholders while keeping the existing behavior unchanged for valid input names. If a non empty input is missing its shape, the same error is still raised.

Changes

• Ignore empty input placeholders in get_input_shape().
• Update the docstring to clarify that shapes are returned only for non empty inputs.
• Add a regression test for a Resize node with the optional ROI input omitted.

Testing

pytest test/pytest/test_onnx_to_hls.py