I'm cross-referencing these issue https://github.com/pytorch/pytorch/issues/26893. I think the underlying issue there is that einsum is simply not supported by ONNX. Could this be added?
@matthewfeickert @kratsg
lukasheinrich
All 3 comments
Does this require its own ONNX op? From my understanding einsum could be implemented as a sequence of already existing ONNX ops.
Does anyone know if this is the case, or is there functionality not currently covered?
conjam
on 19 Oct 2019
@conjam I think most frameworks have einsum as a first-class op. Maybe @mattjj can comment on what the process was to add it to jax (see source code here:
https://github.com/google/jax/blob/master/jax/numpy/lax_numpy.py#L2017
https://github.com/google/jax/issues/37
lukasheinrich
on 21 Oct 2019
From our SIG meeting, einsum should be added as OP to ONNX. @conjam what you are saying is true, but every exporter or converter need to understand how to parse the einsum expression and generate the corresponding ONNX sub-graph. Furthermore, the generated sub-graph will be slower to run.
ebarsoum
on 7 Nov 2019
Related issues
ma-hei
路
4Comments
AnaRhisT94
路
3Comments
PenghuiCheng
路
3Comments
radikalliberal
路
3Comments
fdwr
路
3Comments
Most helpful comment
From our SIG meeting, einsum should be added as OP to ONNX. @conjam what you are saying is true, but every exporter or converter need to understand how to parse the einsum expression and generate the corresponding ONNX sub-graph. Furthermore, the generated sub-graph will be slower to run.