Codes for [Preprint] Bag of Tricks for Training Deeper Graph Neural Networks: A Comprehensive Benchmark Study

Tianlong Chen*, Kaixiong Zhou*, Keyu Duan, Wenqing Zheng, Peihao Wang, Xia Hu, Zhangyang Wang

This is the first fair and reproducible benchmark dedicated to assessing the "tricks" of training deep GNNs. We categorize existing approaches, investigate their hyperparameter sensitivity, and unify the basic configuration. Comprehensive evaluations are then conducted on tens of representative graph datasets including the recent large-scale Open Graph Benchmark (OGB), with diverse deep GNN backbones. Based on synergistic studies, we discover the transferable combo of superior training tricks, that lead us to attain the new state-of-the-art results for deep GCNs, across multiple representative graph datasets.

conda create -n deep_gcn_benchmark
conda activate deep_gcn_benchmark
pip install -r requirement.txt

What env configs that we tried that have succeeded: Mac/Linux + cuda driver 11.2 + Torch with cuda 11.1 + torch_geometric/torch sparse/etc with cuda 11.1.

What env configs that we tried but didn't work: Linux+Cuda 11.1/11.0/10.2 + whatever version of Torch.

In the above case when it did work, we adopted the following installation commands, and it automatically downloaded built wheels, and the installation completes within seconds.

In the case when it did not work, the installation appears to be very slow (ten minutes level for torch sparse/torch scatter). Then the installation did not produce any error, while when import torch_geometric in python code, it reports errors of different types.

Installation codes that we adopted on Linux cuda 11.2 that did work:

pip3 install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.9.0+cu111.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.9.0+cu111.html
pip install torch-geometric

.
├── Dataloader.py
├── main.py
├── trainer.py
├── models
│   ├── *.py
├── options
│   ├── base_options.py
│   └── configs
│       ├── *.yml
├── tricks
│   ├── tricks
│   │   ├── dropouts.py
│   │   ├── norms.py
│   │   ├── others.py
│   │   └── skipConnections.py
│   └── tricks_comb.py
└── utils.py

To train a deep GCN model <model> on dataset <dataset> as your baseline, run:

python main.py --compare_model=1 --cuda_num=0 --type_model=<model> --dataset=<dataset>
# <model>   in  [APPNP, DAGNN, GAT, GCN, GCNII, GPRGNN, JKNet, SGC]
# <dataset> in  [Cora, Citeseer, Pubmed, ogbn-arixv, CoauthorCS, CoauthorPhysics, AmazonComputers, AmazonPhoto, TEXAS, WISCONSIN, CORNELL, ACTOR]

we comprehensively explored the optimal hyperparameters for all models we implemented and train the models under the well-studied hyperparameter settings. For model-specific hyperparameter configs, please refer to options/configs/*.yml

To explore different trick combinations, we provide a tricks_comb model, which integrates different types of tricks as follows:

dropouts:        DropEdge, DropNode, FastGCN, LADIES
norms:           BatchNorm, PairNorm, NodeNorm, MeanNorm, GroupNorm, CombNorm
skipConnections: Residual, Initial, Jumping, Dense
others:          IdentityMapping

To train a tricks_comb model with specific tricks, run:

python main.py --compare_model=0 --cuda_num=0 --type_trick=<trick_1>+<trick_2>+...+<trick_n> --dataset=<dataset>

, where you can assign type_trick with any number of tricks. For instance, to train a trick_comb model with Initial, EdgeDrop, BatchNorm and IdentityMapping on Cora, run:

python main.py --compare_model=0 --cuda_num=0 --type_trick=Initial+EdgeDrop+BatchNorm+IdentityMapping --dataset=Cora

We provide two backbones --type_model=GCN and --type_tricks=SGC for trick combinations. Specifically, when --type_model=SGC and --type_trick=IdenityMapping co-occur, IdentityMapping has higher priority.

You are welcome to make any type of contributions. Here we provide a brief guidance to add your own deep GCN models and tricks.

Several simple steps to add your own deep GCN model <DeepGCN>.

1. Create a python file named <DeepGCN>.py
2. Implement your own model as a torch.nn.Module, where the class name is recommended to be consistent with your filename <DeepGCN>
3. Make sure the commonly-used hyperparameters is consistent with ours (listed as follows). To create any new hyperparameter, add it in options/base_options.py.

 --dim_hidden        # hidden dimension
 --num_layers        # number of GCN layers
 --dropout           # rate of dropout for GCN layers
 --lr:               # learning rate
 --weight_decay      # rate of l2 regularization

4. Register your model in models/__init__.py by add the following codes:

from <DeepGCN> import <DeepGCN>
__all__.append('<DeepGCN>')

5. You are recommend to use YAML to store your dataset-specific hyperparameter configuration. Create a YAML file <DeepGCN>.yml in options/configs and add the hyperparameters as the following style:

<dataset_1>
  <hyperparameter_1> : value_1
  <hyperparameter_2> : value_2

Now your own model <DeepGCN> should be added successfully into our benchmark framework. To test the performance of <DeepGCN> on <dataset>, run:

python main.py --compare_model=1 --type_model=<DeepGCN> --dataset=<dataset>

As all implemented tricks are coupled in tricks_comb.py tightly, we do not recommend integrating your own trick to trick_comb to avoid unexpected errors. However, you can use the interfaces we provided in tricks/tricks/ to combine your own trick with ours.

• Superior performance of our best combo with 32 layers deep GCNs

• Transferability of our best combo with 32 layers deep GCNs

• Takeaways of the best combo

if you find this repo is helpful, please cite

@misc{chen2021bag,
      title={Bag of Tricks for Training Deeper Graph Neural Networks: A Comprehensive Benchmark Study}, 
      author={Tianlong Chen and Kaixiong Zhou and Keyu Duan and Wenqing Zheng and Peihao Wang and Xia Hu and Zhangyang Wang},
      year={2021},
      eprint={2108.10521},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}