gentlezhu/shift-robust-gnns

"Shift-Robust GNNs: Overcoming the Limitations of Localized Graph Training Data" (NeurIPS 21')

shift-robust-gnns's Issues

Errors when running GAT and GraphSAGE

Hi,

I get the following errors when running main_gnn.py:

1. GAT:

Using backend: pytorch
number of classes 7
Using CUDA
Traceback (most recent call last):
  File "main_gnn.py", line 729, in <module>
    micro_f1, macro_f1, out_acc = main(args, [])
  File "main_gnn.py", line 463, in main
    args.aggregator_type
  File ".../Shift-Robust-GNNs/dgl_models.py", line 254, in __init__
    self.layers.append(GATConv(in_feats, n_hidden, num_heads=num_heads, feat_drop=dropout, activation=activation))
  File ".../python3.7/site-packages/dgl/nn/pytorch/conv/gatconv.py", line 160, in __init__
    self._in_src_feats, out_feats * num_heads, bias=False)
  File ".../python3.7/site-packages/torch/nn/modules/linear.py", line 81, in __init__
    self.weight = Parameter(torch.empty((out_features, in_features), **factory_kwargs))
TypeError: empty() received an invalid combination of arguments - got (tuple, dtype=NoneType, device=NoneType), but expected one of:
 * (tuple of ints size, *, tuple of names names, torch.memory_format memory_format, torch.dtype dtype, torch.layout layout, torch.device device, bool pin_memory, bool requires_grad)
 * (tuple of ints size, *, torch.memory_format memory_format, Tensor out, torch.dtype dtype, torch.layout layout, torch.device device, bool pin_memory, bool requires_grad)

and 2. GraphSAGE:

Using backend: pytorch
number of classes 7
Using CUDA
Traceback (most recent call last):
  File "main_gnn.py", line 729, in <module>
    micro_f1, macro_f1, out_acc = main(args, [])
  File "main_gnn.py", line 544, in main
    total_loss = loss + 1 * cmd(model.h[idx_train, :], model.h[iid_train, :])
  File ".../python3.7/site-packages/torch/nn/modules/module.py", line 1131, in __getattr__
    type(self).__name__, name))
AttributeError: 'GraphSAGE' object has no attribute 'h'

relevant packages form pip freeze:
dgl-cu102==0.6.1
torch==1.9.0+cu102

Thanks.

Suggest to loosen the dependency on networkx

Hi, your project Shift-Robust-GNNs requires "networkx==2.5" in its dependency. After analyzing the source code, we found that some other versions of networkx can also be suitable without affecting your project, i.e., networkx 2.5.1. Therefore, we suggest to loosen the dependency on networkx from "networkx==2.5" to "networkx>=2.5,<=2.5.1" to avoid any possible conflict for importing more packages or for downstream projects that may use Shift-Robust-GNNs.

May I pull a request to loosen the dependency on networkx?

By the way, could you please tell us whether such dependency analysis may be potentially helpful for maintaining dependencies easier during your development?

---

For your reference, here are details in our analysis.

Your project Shift-Robust-GNNs(commit id: 28ddefd) directly uses 4 APIs from package networkx.

networkx.classes.graph.Graph.__init__, networkx.convert.from_dict_of_lists, networkx.algorithms.centrality.betweenness._single_source_shortest_path_basic, networkx.linalg.graphmatrix.adjacency_matrix

From which, 2 functions are then indirectly called, including 1 networkx's internal APIs and 1 outsider APIs, as follows (neglecting some repeated function occurrences).

[/GentleZhu/Shift-Robust-GNNs]
+--networkx.classes.graph.Graph.__init__
|      +--networkx.convert.to_networkx_graph
|      |      +--networkx.convert.from_dict_of_dicts
|      |      +--networkx.convert.from_dict_of_lists
|      |      +--warnings.warn
|      |      +--networkx.convert.from_edgelist
+--networkx.convert.from_dict_of_lists
+--networkx.algorithms.centrality.betweenness._single_source_shortest_path_basic
+--networkx.linalg.graphmatrix.adjacency_matrix

We scan networkx's versions among [2.5.1] and 2.5, the changing functions (diffs being listed below) have none intersection with any function or API we mentioned above (either directly or indirectly called by this project).

diff: 2.5(original) 2.5.1
[](no clear difference between the source codes of two versions)

As for other packages, the APIs of @outside_package_name are called by networkx in the call graph and the dependencies on these packages also stay the same in our suggested versions, thus avoiding any outside conflict.

Therefore, we believe that it is quite safe to loose your dependency on networkx from "networkx==2.5" to "networkx>=2.5,<=2.5.1". This will improve the applicability of Shift-Robust-GNNs and reduce the possibility of any further dependency conflict with other projects/packages.

How to verify performance on datasets with unbiased training samples?

How to verify performance on datasets with unbiased training samples?

Inquiry About Figure 1 and Related Code

I would like to express my sincere appreciation for your contributions to this project. The work you've shared has been invaluable.

I'm currently delving deeper into your research and was particularly intrigued by Figure 1 presented in your paper. Could you please provide further details? If feasible, access to the related code would significantly enhance my understanding and ability to follow along with your work.

Thank you for considering my request.