This project is a PyTorch implementation of Unbiased Augmentation for Graph Learning (submitted to NeurIPS 2021).
Our implementation is based on Python 3.7 and PyTorch Geometric.
Please see the full list of packages required to run our codes in requirements.txt.
- Python 3.7
- PyTorch 1.4.0
- PyTorch Geometric 1.6.3
PyTorch Geometric requires a separate installation process from the other packages.
We included install.sh to guide the installation process of PyTorch Geometric based on the OS and CUDA version.
The code includes the cases for Linux + CUDA 10.0, Linux + CUDA 10.1, and MacOS + CPU.
We use seven molecular datasets in our work, which are not included in this repository due to their size but can be
downloaded easily by PyTorch Geometric.
You can run data.py in the src directory to download the datasets in the data/graphs directory.
Our split indices in data/splits are also based on these datasets.
| Name | Graphs | Nodes | Edges | Features | Labels |
|---|---|---|---|---|---|
| DD | 1178 | 284.3 | 1431.3 | 89 | 2 |
| ENZYMES | 600 | 32.6 | 124.3 | 3 | 6 |
| MUTAG | 188 | 17.9 | 39.6 | 7 | 2 |
| NCI1 | 4110 | 29.9 | 64.6 | 37 | 2 |
| NCI109 | 4127 | 29.7 | 64.3 | 38 | 2 |
| PROTEINS | 1113 | 39.1 | 145.6 | 3 | 2 |
| PTC_MR | 334 | 14.3 | 29.4 | 18 | 2 |
We included demo.sh, which reproduces the experimental results of our paper.
The code automatically downloads the datasets and trains a GIN classifier with all of our proposed approaches for graph
augmentation.
In other words, you just have to type the following command.
bash demo.sh
This demo script uses all of your GPUs by default and runs four workers for each GPU to reduce the running time.
You can change experimenal arguments such as the number of workers in run.py and the other hyperparameters such as
the number of epochs, batch size, or the initial learning rate in main.py.
Since run.py is a wrapper script for the parallel execution of main.py, all optional arguments given to run.py are passed also to main.py.