Run GNNExplainer

[paulamartingonzalez]

I am trying to run the GNN Explainer code. Although I have used the install.batch to generate the conda environment, I am having issues with the packages. Namely with tap (in the line "from tap import Tap") . I tried installing it myself and I can import tap but doesn't seem to find anything called Tap on it. Could you guide me here please?

Also, I was going through the example and I was wondering which settings should I change to run it on the graph level predictions. Is there an example for that? And, is it possible to get the feature importances and the important subgraphs? I didn't see any outputs in the explain functions.

Thanks in advance!

[Oceanusity]

Hi, thank you for your issue.

First, the Tap package can be installed by the command pip install typed-argument-parser==1.5.4.

Second, if you want to explain the graph level predictions, you can simply set the explain_graph attribute to True explain_graph=True. In addition, GNNExplainer returns the importance score for each edge, and you can find it as the return value edge_masks in the forward function of GNNExplainer. By the way, edge_masks is a list, where edge_masks[0] is the edge_mask for the class 0, and edge_mask[1] is the edge_mask for class 1.

[Oceanusity]

One more thing, there will be a default sigmoid function for the edge_masks value in the propagation method in the torch_geometric.nn.MessagePassing class. Thus, the edge_mask value will be in the range 0-1 after the sigmoid function, and the return value edge_masks in the forward function doesn't in the range 0-1.

[paulamartingonzalez]

Many thanks for your help! I have everything installed now! Nevertheless, I am encountering a few issues when running the code:

• if I run python -m benchmark.kernel.pipeline --task explain --model_name [GCN_2l/GCN_3l/GIN_2l/GIN_3l] --dataset_name [ba_shape/ba_lrp/tox21/clintox] --target_idx [0/2] --explainer GNNExplainer --sparsity [0.5/...] from the console, I get the error: -m: error: argument --target_idx: invalid int value: '[0/2]'
• If I run python -m benchmark.kernel.pipeline --task explain --model_name GCN_3l --dataset_name clintox --target_idx 0 --explainer GNNExplainer --sparsity 0.5 --debug --vis --nolabel also on the console, I get the error: `TypeError: draw_networkx_nodes() got an unexpected keyword argument 'num_nodes'
• And, when trying to run the script /benchmark/pipeline.py on spyder, I get the error AssertionError: Explaining on multi tasks is meaningless. that comes from benchmark/data/dataset.py

Any hint on how to make any of these work?

I'd like to explore the examples before I get to try my own data, so I might follow-up on the explanation bit and edge masks once I manage to get the code running 😃 . Thanks again!
`

[Oceanusity]

Hi, I have the following suggestions for the issues.

• Please select one from the optional selections for parameters.
• Please follow issue TypeError: draw_networkx_nodes() got an unexpected keyword argument 'num_nodes' #12.
• I am not sure about this problem. If it happens again with the others fixed, feel free to give me more details.

[paulamartingonzalez]

Many thanks again! Following that issue, the second point works well.

Now, when I try to run: python -m benchmark.kernel.pipeline --task explain --model_name GCN_3l --dataset_name ba_shape --target_idx 2 --explainer GNNExplainer --sparsity 0.5 I get

File "/home/martin09/DIG/dig/xgraph/GNNExplainer/benchmark/models/model_manager.py", line 45, in config_model
    print(f'#E#Checkpoint not found at {os.path.abspath(args.test_ckpt)}')
This overload of nonzero is deprecated:
	nonzero()
Consider using one of the following signatures instead:
	nonzero(*, bool as_tuple) (Triggered internally at  /opt/conda/conda-bld/pytorch_1595629395347/work/torch/csrc/utils/python_arg_parser.cpp:766.)

I tried reinstalling everything from scratch to check if that was the issue, but it persisted. Do you have any hint of what may be happening?

[paulamartingonzalez]

Small update: using python -m benchmark.kernel.pipeline --task explain --model_name GCN_3l --dataset_name tox21 --target_idx 2 --explainer GNNExplainer --sparsity 0.3 --vis it worked, which makes me think that perhaps the model for the combination above did not exist.

I think I will close the issue, I will be exploring the code the next couple of days and I might reopen it with more questions. 😃 Thanks!

[paulamartingonzalez]

I am exploring the code on TOX21 dataset and I would like to understand a bit better the outputs (both the edge masks and the features masks). For the edge masks, is it the masks variable in sample_explain the one I should be looking at? For the feature masks, I don't really know where to find them. Would you mind pointing me? Is there a general feature mask for all the cohort? or is it one per graph?

Ultimately, I am looking to get something as Figure 5 in GNNExplainer paper: https://arxiv.org/pdf/1903.03894.pdf

[Oceanusity]

For the GNNExplainer class, you can extract the node mask and edge mask from the class attribute self.node_feat_mask and self.edge_mask. Please take mask_features=True to explain the importance of node features.

For the figure, GNNExplainer provides the visualization as shown in README.md. It's not the same as the GNNExplainer paper, but I think it can help you with the figure to some extend.