ONNX export uses hardcoded feature shape
In to_onnx.py, the number of jet features and number of tracks per jet are hardcoded:
jets, tracks = inputs_sep_no_pad(1, 40, 2, self.in_dims[0] - 2)
It would be handy if this could somehow be configured automatically when loading the model. I am using a different number of features and encountered issues when trying to use the ONNX export in athena:
Decor_PrimaryVertices ERROR Got invalid dimensions for input: jet_features for the following indices
index: 1 Got: 7 Expected: 2
Please fix either the inputs or the model.
Happy to try and implement this, but maybe it would be a very simple fix by an expert.
Edited by Jackson Carl Burzynski