ONNX uses protobuf (3.11.3) to define its model graph. Actually I am not sure it's a protobuf issue or it's something we can handle from ONNX. I also tried a solution ...

When reading a file using h5py==2.80, float values read from the file as a numpy.float64 do not cast correctly back to a python.float i.e. read_float = h5py.File ...