is there any ways for me to change the name of input variables of traced module? I have looked into the documentation of torch.jit.trace and it could not be changed beforehand as well, e.g. with dictionary.
basically I’m trying to mimic the behavior of input_names argument in torch.onnx.export
I don’t believe we expose a way to do this. Names are not semantically meaningful in a JIT graph anyway, they are only for human readability
thanks for the reply.
what I want to do is so that I could use keyword arguments as an input to forward.
in my traced module, there are 3 arguments, the name of the first two arguments have already been properly inferred by tracing, but the last one is not.
my question now becomes how does the tracer infer the argument name for this?
The tracer looks in the python interpreter frame state’s local variables (f_local) for names (see here. So if, for a given traced value, it finds a corresponding local, it will add that as the name in the graph.
Thank you, I will look into that