SubTree Attention Graph Neural Network Source Code

The official implementation for "Tailoring Self-Attention for Graph via Rooted Subtrees" which is accepted to NeurIPS2023 as a poster.

Related materials: [arxiv]

In this repository, you'll find the requisite code to reproduce the empirical results in our research paper. There are two methods available to operate this code:

Utilizing Weights & Biases (WandB), which is our recommended approach.
Executing directly through provided scripts.

Dependencies

The following list presents the versions of the packages used in our experiments.

python==3.7.12
pytorch==1.8.0
torch_geometric==2.0.1
torch_sparse==0.6.11
torch_scatter==2.1.1
torch_cluster==1.6.1
torch_spline_conv==1.2.2
wandb==0.12.16

Preparing data

Please unzip data.zip first and put the unzipped data folder in the parent directory of this folder.

parent directory
├── data
│   ├── Amazon
│   └── Citation_Full
│   └── ...
└── SubTree-Attention
    ├── best_params_yamls
    └── scripts
    └── ...

Option 1: Using Provided Scripts We also provide script files that can be run directly. For CiteSeer, Cora, Deezer-Europe and Film, please use scripts/exp_setting_1.sh. For Computers, CoraFull, CS, Photo, Physics and Pubmed, please use best_params_yamls/scripts/exp_setting_2.sh.

Option 2: Using Weights & Biases

Setup: Initially, create the configs and remote folders within the root directory. These will be used to store wandb files.
Initiating Sweep: We provide the hyperparameters for each dataset within the best_params_yamls folder. You can use them to create a sweep. Make sure to specify your wandb username and project name when doing so.

For CiteSeer, Cora, Deezer-Europe and Film, please use yamls in best_params_yamls/setting_1. For Computers, CoraFull, CS, Photo, Physics and Pubmed, please use yamls in best_params_yamls/setting_2.

Here's an example command:
```
python sweep.py --entity=$YOUR_WANDB_ENTITY$ --project=$YOUR_WANDB_PROJECT$ --source=file --info=best_params_yamls/setting_1/citeseer.yaml
```
Please replace $YOUR_WANDB_ENTITY$ and $YOUR_WANDB_PROJECT$ with your wandb username and project name respectively.

Initiating Agent: Once you run the above command, you will receive a sweep ID $SWEEP_ID$ and sweep URL $SWEEP_URL$ , similar to the example shown below:

Create sweep with ID: $SWEEP_ID$
Sweep URL: $SWEEP_URL$

You can now choose to run the program in single process mode or in parallel.

For single process execution, use the following command:

python agents.py --entity=$YOUR_WANDB_ENTITY$ --project=$YOUR_WANDB_PROJECT$ --sweep_id=$SWEEP_ID$ --gpu_allocate=$INDEX_GPU$:1 --wandb_base=remote --mode=one-by-one --save_model=False

For parallel execution, use the command below:

python agents.py --entity=$YOUR_WANDB_ENTITY$ --project=$YOUR_WANDB_PROJECT$ --sweep_id=$SWEEP_ID$ --gpu_allocate=$INDEX_GPU$:$PARALLEL_RUNS$ --wandb_base=temp --mode=parallel --save_model=False

Here, the parameter $INDEX_GPU$:$PARALLEL_RUNS$ specifies that $PARALLEL_RUNS$ will run concurrently on GPU $INDEX_GPU$ . In multi-process mode, you can link multiple GPUs to distribute tasks concurrently using -.

For instance:

python agents.py --entity=$YOUR_WANDB_ENTITY$ --project=$YOUR_WANDB_PROJECT$ --sweep_id=$SWEEP_ID$ --gpu_allocate=$INDEX_GPU_1$:$PARALLEL_RUNS$-$INDEX_GPU_2$:$PARALLEL_RUNS$ --wandb_base=temp --mode=parallel --save_model=False

Results Evaluation: The outcomes of your experiment can be viewed at the $SWEEP_URL$ , which is a webpage hosted on wandb.ai.

LUMIA-Group/SubTree-Attention

SubTree Attention Graph Neural Network Source Code