Sam VS
Sam VS
I found that torch now has it's own `torch.scatter_reduce()`, so we probably could probably just add support for that and forget about the `torch_scatter` package.
Also I think it should be possible to simplify the `ScatterElements` implementation using [numpy.put_along_axis](https://numpy.org/doc/stable/reference/generated/numpy.put_along_axis.html#numpy.put_along_axis).
Hi @ramkrishna2910, thanks. Do you know who is working on this and when it might be ready?
Thanks @yaox12, upon downgrading to PT 1.12.0 the issue is resolved.
I had a look at converting PyG models using `torch.onnx.export` and the primary problem (at least for GATConv layers) seems to be the use of `torch_scatter` in the PyG [`softmax`](https://github.com/pyg-team/pytorch_geometric/blob/master/torch_geometric/utils/softmax.py#L43)...
Thanks, that's good to know. We still would need to support `torch.scatter_reduce()` in ONNX but I am hoping this will not be too difficult.
Hi @Joecerniglia! I tried to contact you on academia.edu but this is probably a better place. This contribution to Turing Data Stories is really very exciting! Thank you so much...
Hi Joe! Oh dear sorry about that, can you please try [this link](https://join.slack.com/t/turingdatastories/shared_invite/zt-1717gg1v7-vhd_CHDfx79GZcMZw4EGOw) instead? Hopefully that works a little better :). We also tend to meet on Wednesdays at around...
hackmd: https://hackmd.io/RzaozlRNSAqHU1CXn4AYKA
Some more ideas in the notes from David and I https://hackmd.io/vixz2MUmQk-sxX0950mXKA https://hackmd.io/@David-Beavan/BJggYRrDP