Docker image with requirements for "Strategies for Pre-training Graph Neural Networks"

[mdanb]

I am trying to run the code from the paper "Strategies for Pre-training Graph Neural Networks". The code requires a really old version of Pytorch: Pytorch 1.0.1
I'm trying to run the code from a Docker container. So I first pulled the pytorch/pytorch:1.0.1-cuda10.0-cudnn7-runtime docker image. I then installed all requirements, including torch-geometric and torch-scatter etc... and committed to create a new image with the necessary requirements
I then tried running the code in the container, but I get the error:
ModuleNotFoundError: No module named 'torch_scatter.scatter_cuda'
Was wondering how I can solve this?

[rusty1s]

Here is an old Dockerfile which might be helpful to you.

Nonetheless, I don't think there were any breaking changes from PyTorch 1.0.1 and PyTorch Geometric 1.0.3. It should be pretty straightforward to update the code to support later versions of them.

[mdanb]

Actually, I have a few questions about that since that's what I tried to do initially. First, do you mean Pytorch Geometric 1.3.0 instead of 1.0.3? Because the authors also mention 1.0.3 instead of 1.3.0 which confused me because I couldn't find 1.0.3
So I then decided just to search the 1.3.0 documentation. I was looking for a method that works on Data objects called cat_dim since the authors' code includes the line:
batch[key] = torch.cat(batch[key], dim=data_list[0].cat_dim(key, batch[key][0]))
and when I would run it I would get:
AttributeError: 'Data' object has no attribute 'cat_dim'
I couldn't find a cat_dim, but I found a __cat_dim__ instead (are they the same?). I then just replaced the former by the latter, which did make the error go away, but then I got another error saying:
ValueError: MessagePassing.propagate only supports torch.LongTensor of shape [2, num_messages] or torch_sparse.SparseTensor for argument edge_index.

Now, right after the line above they have:

batch.batch = torch.cat(batch.batch, dim=-1)
 return batch.contiguous()

and I added the line:
batch.edge_index = batch.edge_index.to_sparse()
between them
But then I got
RuntimeError: sparse tensors do not have is_contiguous

I then finally tried just doing return batch instead of return batch.contiguous()
Then, the line edge_index = add_self_loops(edge_index, num_nodes = x.size(0)) in the code was causing issues:

    edge_index = torch.cat([edge_index, loop_index], dim=1)
RuntimeError: Concatenating sparse tensors, but a dense tensor was found at position 1.

I stopped trying at that point haha because I figured the problem was maybe in the initial change from cat_dim to __cat_dim__ but I don't know I'm not sure

[rusty1s]

Indeed. cat_dim got renamed to __cat_dim__.

The second error relates to a change in the propagate API. self.aggr is now an argument passed to the constructor, i.e., super(GINConv, self).__init__(aggr=aggr) and the propagate call should look like:

return self.propagate(edge_index, x=x, edge_attr=edge_embeddings)