Matthias Fey

I personally prefer to move this to `torch-sparse` (by providing a `spmm_coo` operator similar to [here](https://github.com/rusty1s/pytorch_sparse/blob/master/csrc/spmm.cpp#L22-L25)): ```python spmm_coo(Tensor row, Tensor col, Tensor mat, int dim_size, string reduce) ``` If you...

Super, thank you!

We query the last element of `indptr` to determine the output size of the `out` tensor. I don't think there is any real workaround besides passing in an `out` tensor...

I've never seen this error before, I'm sorry. Any chance you can post the full log of the installation?

I think we could overload `Reducer::write` such that it supports `arg_out_data` or not. We can then pass in `arg_out_data` in case `reduce == min | max`, and otherwise not.

Thanks for reporting. Interesting finding! For `scatter_add`, we make use of PyTorch's internal `torch.scatter_add_` functionality, so there is not much we can do to improve performance on our end despite...

I think this is supported by [`torch_geometric.utils.to_dense_batch`](https://pytorch-geometric.readthedocs.io/en/latest/modules/utils.html#torch_geometric.utils.to_dense_batch). Let me know if that works for you!

Thanks @jeanfeydy for this insightful thread. I will need to take a closer look at this. It looks like PyTorch C++ backward linkage is indeed different from Python backward linkage,...

The wheels for PyTorch 1.7.0 will work for PyTorch 1.7.1 as well (that's why we just link to them). Your issue can be resolved by specifying versioning numbers according to...

The only way to make this installable without version numbers if via `--no-index` option (which will not look at PyPi for newer versions): ``` pip install --no-index torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.1+cu101.html...