PyTorch is all you need!
GraphGallery is a gallery for benchmarking Graph Neural Networks (GNNs) based on pure PyTorch backend. Alteratively, Pytorch Geometric (PyG) and Deep Graph Library (DGL) backend are also available in GraphGallery to facilitate your implementations.
- November 20, 2021: We now no longer support
TensorFlowbackend. - November 20, 2021: The module
graphgallery.attackis deprecated, users may refer to GraphWar for more information.
Please make sure you have installed PyTorch. Also, Pytorch Geometric (PyG) and Deep Graph Library (DGL) are alternative choices.
Install from source:
# Recommended git clone https://github.com/EdisonLeeeee/GraphGallery.git && cd GraphGallery pip install -e . --verbose
where -e means "editable" mode so you don't have to reinstall every time you make changes.
NOTE: GraphGallery is a frequently updated package and DO NOT install GraphGallery with pip, we're currently working on releasing a binary distribution on PyPI, stay tuned!
In detail, the following methods are currently implemented:
The graph purification methods are universal for all models, just specify:
graph_transform="purification_method"
so, here we only give the examples of GCN with purification methods, other models should work.
| Method | Author | Paper |
|---|---|---|
| GCN-Jaccard | Huijun Wu et al. | Adversarial Examples on Graph Data: Deep Insights into Attack and Defense (IJCAI'19) |
| GCN-SVD | Negin Entezari et al. | All You Need Is Low (Rank): Defending Against Adversarial Attacks on Graphs (WSDM'20) |
| Method | Author | Paper | PyTorch | PyG | DGL |
|---|---|---|---|---|---|
| GAE, VGAE | Thomas N. Kipf et al. | Variational Graph Auto-Encoders (NeuIPS'16) | βοΈ | βοΈ |
The following methods are framework-agnostic.
| Method | Author | Paper |
|---|---|---|
| Deepwalk | Bryan Perozzi et al. | DeepWalk: Online Learning of Social Representations (KDD'14) |
| Node2vec | Aditya Grover and Jure Leskovec | node2vec: Scalable Feature Learning for Networks (KDD'16) |
| Node2vec+ | Renming Liu et al. | Accurately Modeling Biased Random Walks on Weighted Graphs Using Node2vec+ |
| BANE | Hong Yang et al. | Binarized attributed network embedding (ICDM'18) |
- Planetoid: a collection of widely used benchmark datasets in graph learning tasks, including 'cora', 'citeseerr', 'pubmed' and 'nell' datasets.
- NPZDataset: a collection of graph datasets stored with numpy
.npzformat.
you can simply run dataset.available_datasets() to see the available datasets, e.g.,:
from graphgallery.datasets import Planetoid print(Planetoid.available_datasets())
more details please refer to GraphData.
It takes just a few lines of code.
import torch import graphgallery from graphgallery.datasets import Planetoid from graphgallery.gallery import callbacks data = Planetoid('cora', root="~/GraphData/datasets/", verbose=True) graph = data.graph splits = data.split_nodes() device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') graphgallery.set_backend("torch") from graphgallery.gallery.nodeclas import GCN trainer = GCN(device=device, seed=123).setup_graph(graph, feat_transform="normalize_feat").build() cb = callbacks.ModelCheckpoint('model.pth', monitor='val_accuracy') trainer.fit(splits.train_nodes, splits.val_nodes, verbose=1, callbacks=[cb]) results = trainer.evaluate(splits.test_nodes) print(f'Test loss {results.loss:.5}, Test accuracy {results.accuracy:.2%}')
import torch import graphgallery from graphgallery.gallery import callbacks from graphgallery.datasets import Planetoid data = Planetoid('cora', root="~/GraphData/datasets/", verbose=True) graph = data.graph splits = data.split_edges(random_state=15) device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') graphgallery.set_backend("torch") from graphgallery.gallery.linkpred import GAE trainer = GAE(device=device, seed=123).setup_graph(graph).build() cb = callbacks.ModelCheckpoint('model.pth', monitor='val_ap') trainer.fit(splits.train_pos_edge_index, val_data=(splits.val_pos_edge_index, splits.val_neg_edge_index), verbose=1, callbacks=[cb]) results = trainer.evaluate((splits.test_pos_edge_index, splits.test_neg_edge_index)) print(results)
If you have any troubles, you can simply run trainer.help() for more information.
>>> import graphgallery # Default: PyTorch backend >>> graphgallery.backend() PyTorch 1.9.0+cu111 Backend # Switch to PyTorch Geometric backend >>> graphgallery.set_backend("pyg") # Switch to DGL PyTorch backend >>> graphgallery.set_backend("dgl") # Switch to PyTorch backend >>> graphgallery.set_backend("th") # "torch", "pytorch"
But your codes don't even need to change.
This is motivated by gnn-benchmark
from graphgallery.data import Graph # Load the adjacency matrix A, attribute (feature) matrix X and labels vector y # A - scipy.sparse.csr_matrix of shape [num_nodes, num_nodes] # X - scipy.sparse.csr_matrix or numpy.ndarray of shape [num_nodes, num_feats] # y - numpy.ndarray of shape [num_nodes] mydataset = Graph(adj_matrix=A, attr_matrix=X, label=y) # save dataset mydataset.to_npz('path/to/mydataset.npz') # load dataset mydataset = Graph.from_npz('path/to/mydataset.npz')
- Add PyTorch trainers support
- Add other frameworks (PyG and DGL) support
- set tensorflow as optional dependency when using graphgallery
- Add more GNN trainers
- Support for more tasks, e.g.,
graph Classificationandlink prediction - Support for more types of graphs, e.g., Heterogeneous graph
- Add Docstrings and Documentation (Building)
- Comprehensive tutorials
Please fell free to contact me if you have any troubles.
This project is motivated by Pytorch Geometric, Stellargraph and DGL, etc., and the original implementations of the authors, thanks for their excellent works!
Please cite our paper (and the respective papers of the methods used) if you use this code in your own work:
@inproceedings{li2021graphgallery, author = {Jintang Li and Kun Xu and Liang Chen and Zibin Zheng and Xiao Liu}, booktitle = {2021 IEEE/ACM 43rd International Conference on Software Engineering: Companion Proceedings (ICSE-Companion)}, title = {GraphGallery: A Platform for Fast Benchmarking and Easy Development of Graph Neural Networks Based Intelligent Software}, year = {2021}, pages = {13-16}, publisher = {IEEE Computer Society}, address = {Los Alamitos, CA, USA}, }