Closed JackTemaki closed 1 year ago
In case we want to use the model in pipelines for ONNX export, we can not use padding="same" for now: https://github.com/pytorch/pytorch/issues/68880
padding="same"
While workarounds would be possible, I would strongly advise to not introduce export hacks already for our first model. Also while export might work, it is unclear if the ONNX runtime does.
I think it should be fine to only use odd kernels. This is anw what other frameworks use (e.g Fairseq and ESPNet)
As there is no further comments I merge for now, I want to test #8
In case we want to use the model in pipelines for ONNX export, we can not use
padding="same"
for now: https://github.com/pytorch/pytorch/issues/68880While workarounds would be possible, I would strongly advise to not introduce export hacks already for our first model. Also while export might work, it is unclear if the ONNX runtime does.