How to concatenate strings in ONNX?

Ask a Question

Question

We have a use case where we would like to concatenate string tensor pairwise. Here is some pseudo code of what I’m trying to accomplish:

tensor1 = ["a", "b", "c"]
tensor2 = ["d", "e", "f"]

concat(tensor1, tensor2)
> ["ad", "be", "cf"]

# Should also support broadcasting
concat(tensor1, "g")
> ["ag", "bg", "cg"]

The existing concat is not helping me, as it only concats the tensor into a tensor of a higher rank instead of concatenating the strings.

Is it possible to accomplish this with the current ONNX operator set?

Further information

• Relevant Area (e.g. model usage, backend, best practices, converters, shape_inference, version_converter, training, test, operators):

• Is this issue related to a specific model? No.

[AlexandreEichenberger]

FYI this feature is one that is being requested in the ONNX Roadmap discussion. https://onnx.ai/news.html#new60

It will be discussed Sept 8th at 9pm EST Sept 8th (5:30 to 6pm – PST)

1. Takuya Nakaike (IBM) – New operators for data processing to cover ML pipeline (eg: Stringconcatenator, Stringsplitter, Date)
2. Adam Pocock (Oracle labs) – C API for C++ components of ONNX (to assist in wrapper for model checker functionality)
3. Adam Pocock (Oracle labs) – Better support for emitting ONNX models from other languages beyond Python

Check the link above on how to join.

[janjagusch]

https://www.youtube.com/watch?v=RJKXRgfKjkQ

Thanks! I will close this issue for now.