zhongyu1998 / GTR

Official implementation of the TPAMI 2023 paper "Growing Like a Tree: Finding Trunks From Graph Skeleton Trees"

Graph Trunk Network (GTR)

The official implementation of Growing Like a Tree: Finding Trunks From Graph Skeleton Trees (TPAMI 2023).

1  Installation Instructions

Follow the steps below to set up the virtual environment.

Create and activate the environment:

conda create -n gtr python=3.6
conda activate gtr

Install dependencies in the listed order:

pip install rdkit-pypi==2021.3.5
pip install -r requirements.txt
pip install torch==1.4.0+cu100 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch_scatter==2.0.4 -f https://pytorch-geometric.com/whl/torch-1.4.0+cu100.html
pip install torch_sparse==0.6.1 -f https://pytorch-geometric.com/whl/torch-1.4.0+cu100.html
pip install torch-geometric==1.7.0
pip install ogb==1.2.6

2  Experiments

2.1  TUDataset

Reproduce the results reported in Table 3. Repeat the experiment 100 times using the same evaluation protocol as Karhadkar et al. to obtain the reported results.

MUTAG

python main_tu.py --name MUTAG --batch_size 16 --hidden_dim 64 --num_layers 4 --dropout 0.5 --lr 0.001 --lr_factor 0.5 --lr_limit 1e-5 --max_level 5

ENZYMES

python main_tu.py --name ENZYMES --batch_size 16 --hidden_dim 64 --num_layers 4 --dropout 0.5 --bidirectional --lr 0.001 --lr_factor 0.5 --lr_limit 1e-5 --max_level 10

PROTEINS

python main_tu.py --name PROTEINS --batch_size 16 --hidden_dim 64 --num_layers 4 --dropout 0.5 --bidirectional --lr 0.001 --lr_factor 0.1 --lr_limit 5e-6 --max_level 10

IMDB-BINARY

python main_tu.py --name IMDB-BINARY --batch_size 16 --hidden_dim 64 --num_layers 4 --dropout 0.5 --lr 0.001 --lr_factor 0.5 --lr_limit 2e-5 --max_level 2

COLLAB

python main_tu.py --name COLLAB --batch_size 32 --hidden_dim 64 --num_layers 4 --dropout 0.5 --lr 0.001 --lr_factor 0.5 --lr_limit 2e-5 --max_level 12

2.2  Graph Benchmarks

Reproduce the results reported in Table 4.

ogbg-molhiv

Repeat the experiment 10 times using the same evaluation protocol as Hu et al. to obtain the reported results.

python main_ogbhiv.py --name ogbg-molhiv --batch_size 128 --hidden_dim 128 --num_layers 4 --dropout 0.5 --lr 0.0003 --lr_factor 0.5 --lr_limit 5e-5 --max_level 4

Peptides-func

Repeat the experiment 4 times using the same evaluation protocol as Dwivedi et al. to obtain the reported results.

python main_pepfunc.py --name Peptides-func --batch_size 128 --hidden_dim 256 --num_layers 4 --dropout 0.5 --lr 0.0004 --lr_factor 0.5 --lr_limit 2e-5 --max_level 12

Peptides-struct

Repeat the experiment 4 times using the same evaluation protocol as Dwivedi et al. to obtain the reported results.

python main_pepstrc.py --name Peptides-struct --batch_size 128 --hidden_dim 256 --num_layers 3 --dropout 0.2 --lr 0.0004 --lr_factor 0.5 --lr_limit 2e-5 --max_level 12

2.3  Visualization

We provide the IPython Notebook file visualization.ipynb for reproducing the visualizations shown in Figure 3 (Section 5.6).

3  Citation

If you find this code or our GTR paper helpful for your research, please cite our paper using the following information:

@article{huang2023growing,
  title   = {Growing Like a Tree: Finding Trunks From Graph Skeleton Trees},
  author  = {Huang, Zhongyu and Wang, Yingheng and Li, Chaozhuo and He, Huiguang},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year    = {2023}
}