Documentation | Get Started | Quick Tour | Supported Models | Examples | Contributing | 启智社区
GammaGL is a multi-backend graph learning library based on TensorLayerX, which supports TensorFlow, PyTorch, PaddlePaddle, MindSpore as the backends.
We give a development tutorial in Chinese on wiki.
GammaGL supports multiple deep learning backends, such as TensorFlow, PyTorch, Paddle and MindSpore. Different from DGL, the GammaGL's examples are implemented with the same code on different backend. It allows users to run the same code on different hardwares like Nvidia-GPU and Huawei-Ascend. Besides, users could use a particular framework API based on preferences for different frameworks.
Following PyTorch Geometric(PyG), GammaGL utilizes a tensor-centric API. If you are familiar with PyG, it will be friendly and maybe a TensorFlow Geometric, Paddle Geometric, or MindSpore Geometric to you.
2026-07-05 release v0.6.0
We release GammaGL v0.6.0.
- Use one
gammaglpackage for CPU and GPU environments, with source builds selected byGAMMAGL_WITH_CUDA=0/1/auto. - Use the GAMMA Lab maintained TensorLayerX nightly branch for source builds.
- Keep LLM and graph foundation model dependencies optional through
llm,gfm, andllm-gfmextras. - Improve public API exports for common layers, datasets, transforms, loaders, models and utilities.
- Update installation guidance and release package metadata for Python 3.9+ Linux environments.
2024-07-29 release v0.5
We release version v0.5.
- 70 GNN models
- More fused operators
- Support GPU sample
- Support GraphStore and FeatureStore
2024-01-24 release v0.4
We release version v0.4.
- 60 GNN models
- More fused operators and users can truly use these operators
- Support the latest version of PyTorch and MindSpore
- Support for graph database like neo4j
2023-07-12 release v0.3
We release version v0.3.
- 50 GNN models
- Efficient message passing operators and fused operator
- Rebuild sampling architecture.
2023-04-01 paper accepted
Our paper GammaGL: A Multi-Backend Library for Graph Neural Networks is accpeted at SIGIR 2023 resource paper track.
2023-02-21 中国电子学会科技进步一等奖
算法库支撑了北邮牵头，蚂蚁、中移动、海致科技等参与的“大规模复杂异质图数据智能分析技术与规模化应用”项目。该项目获得了2022年电子学会科技进步一等奖。
2023-01-17 release v0.2
We release version v0.2.
- 40 GNN models
- 20 datasets
- Efficient message passing operators and fused operator
- GPU sampling and heterogeneous graphs samplers.
2022-06-20 release v0.1
We release version v0.1.
- Framework-agnostic design
- PyG-like
- Graph data structures, message passing module and sampling module
- 20+ GNN models
GammaGL 0.6.0 requires Python >= 3.9 and is supported on Linux. Use the
same gammagl package for CPU and GPU; choose the backend wheel and extension
build mode during installation.
For the released package, install a backend first, then install GammaGL:
pip install torch torchvision torchaudio
pip install gammaglFor CPU-only PyTorch:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
pip install gammaglFor the latest source build, use the GAMMA Lab maintained TensorLayerX branch and the source installation commands below.
conda create -n gammagl-cpu python=3.10
conda activate gammagl-cpu
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
pip install git+https://github.com/dddg617/tensorlayerx.git@nightly
git clone --recursive https://github.com/BUPT-GAMMA/GammaGL.git
cd GammaGL
pip install pybind11 ninja
GAMMAGL_WITH_CUDA=0 TL_BACKEND=torch pip install -e ".[build]" --no-build-isolation
TL_BACKEND=torch python examples/gcn/gcn_trainer.py --dataset cora --n_epoch 1 --gpu -1Choose the PyTorch CUDA wheel that matches your driver and CUDA runtime. For example, with CUDA 12.1 wheels:
conda create -n gammagl-cu python=3.10
conda activate gammagl-cu
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install git+https://github.com/dddg617/tensorlayerx.git@nightly
git clone --recursive https://github.com/BUPT-GAMMA/GammaGL.git
cd GammaGL
pip install pybind11 ninja
GAMMAGL_WITH_CUDA=auto TL_BACKEND=torch pip install -e ".[build]" --no-build-isolation
TL_BACKEND=torch python examples/gcn/gcn_trainer.py --dataset cora --n_epoch 1 --gpu 0GAMMAGL_WITH_CUDA accepts 0, 1, or auto. CPU-only installs should use
0; CUDA builds can use auto or 1 when CUDA headers and nvcc are
available.
For other backends, install the backend first, then install the GAMMA Lab TensorLayerX branch:
pip install git+https://github.com/dddg617/tensorlayerx.git@nightlyPyTorch must be installed before installing this TensorLayerX branch. This TensorLayerX is maintained by the BUPT GAMMA Lab Team.
GraphGPT, LLaGA, LLMRec, WalkLM, NLGraph and related LLM/GFM utilities require additional heavy dependencies. Install them only when using those features:
For the released package:
pip install "gammagl[llm-gfm]"For a source checkout:
pip install pybind11 ninja
GAMMAGL_WITH_CUDA=auto TL_BACKEND=torch pip install -e ".[build,llm-gfm]" --no-build-isolationCore GammaGL installation does not require transformers, torch_geometric,
openai, or sentence_transformers.
大陆用户如果遇到网络问题，推荐从启智社区安装：
git clone --recursive https://git.openi.org.cn/GAMMALab/GammaGL.gitIf
--recursivewas omitted, rungit submodule update --initin the GammaGL root directory.
In this quick tour, we highlight the ease of creating and training a GNN model with only a few lines of code.
In the first glimpse of GammaGL, we implement the training of a GNN for classifying papers in a citation graph.
For this, we load the Cora dataset and train a 2-layer GCN with TensorLayerX's backend-neutral training API. The full version is available in examples/gcn/gcn_trainer.py.
import tensorlayerx as tlx
from gammagl.datasets import Planetoid
from tensorlayerx.model import TrainOneStep, WithLoss
from gammagl.models import GCNModel
from gammagl.utils import add_self_loops, mask_to_index
class SemiSpvzLoss(WithLoss):
def forward(self, data, y):
logits = self.backbone_network(
data["x"], data["edge_index"], None, data["num_nodes"]
)
train_logits = tlx.gather(logits, data["train_idx"])
train_y = tlx.gather(data["y"], data["train_idx"])
return self._loss_fn(train_logits, train_y)
dataset = Planetoid(root="./data", name="cora")
graph = dataset[0]
edge_index, _ = add_self_loops(graph.edge_index, num_nodes=graph.num_nodes)
model = GCNModel(
feature_dim=dataset.num_node_features,
hidden_dim=16,
num_class=dataset.num_classes,
drop_rate=0.5,
num_layers=2,
)
optimizer = tlx.optimizers.Adam(lr=0.01, weight_decay=5e-4)
train_one_step = TrainOneStep(
SemiSpvzLoss(model, tlx.losses.softmax_cross_entropy_with_logits),
optimizer,
model.trainable_weights,
)
data = {
"x": graph.x,
"y": graph.y,
"edge_index": edge_index,
"train_idx": mask_to_index(graph.train_mask),
"num_nodes": graph.num_nodes,
}
for epoch in range(200):
model.set_train()
loss = train_one_step(data, graph.y)More information about evaluating final model performance can be found in the corresponding example.
In addition to the easy application of existing GNNs, GammaGL makes it simple to implement custom Graph Neural Networks (see here for the accompanying tutorial). For example, this is all it takes to implement the edge convolutional layer from Wang et al.:
import tensorlayerx as tlx
from tensorlayerx.nn import Sequential as Seq, Linear, ReLU
from gammagl.layers import MessagePassing
class EdgeConv(MessagePassing):
def __init__(self, in_channels, out_channels):
super().__init__()
self.mlp = Seq(Linear(2 * in_channels, out_channels),
ReLU(),
Linear(out_channels, out_channels))
def forward(self, x, edge_index):
# x has shape [N, in_channels]
# edge_index has shape [2, E]
return self.propagate(x=x, edge_index,aggr_type='max')
def message(self, x_i, x_j):
# x_i has shape [E, in_channels]
# x_j has shape [E, in_channels]
tmp = tlx.concat([x_i, x_j - x_i], axis=1) # tmp has shape [E, 2 * in_channels]
return self.mlp(tmp)Take GCN as an example:
cd examples/gcn
TL_BACKEND=torch python gcn_trainer.py --dataset cora --lr 0.01 --n_epoch 200 --gpu 0For CPU:
TL_BACKEND=torch python gcn_trainer.py --dataset cora --n_epoch 200 --gpu -1For a specific GPU:
CUDA_VISIBLE_DEVICES=1 TL_BACKEND=torch python gcn_trainer.py --dataset cora --gpu 0For another backend, install that backend first and set TL_BACKEND explicitly:
TL_BACKEND=paddle python gcn_trainer.py --dataset cora --gpu 0Note
When
TL_BACKENDis not set, GammaGL usestorchby default.Use
--gpu -1for CPU execution.The CANDIDATE backends are
tensorflow,paddle,torchandmindspore.
Now, GammaGL supports about 70 models, we welcome everyone to use or contribute models.
| TensorFlow | PyTorch | Paddle | MindSpore | |
|---|---|---|---|---|
| GCN [ICLR 2017] | ✔️ | ✔️ | ✔️ | ✔️ |
| GAT [ICLR 2018] | ✔️ | ✔️ | ✔️ | ✔️ |
| GraphSAGE [NeurIPS 2017] | ✔️ | ✔️ | ✔️ | ✔️ |
| ChebNet [NeurIPS 2016] | ✔️ | ✔️ | ✔️ | ✔️ |
| GCNII [ICLR 2017] | ✔️ | ✔️ | ✔️ | ✔️ |
You may see the other models here.
| Contrastive Learning | TensorFlow | PyTorch | Paddle | MindSpore |
|---|---|---|---|---|
| DGI [ICLR 2019] | ✔️ | ✔️ | ✔️ | ✔️ |
| GRACE [ICML 2020 Workshop] | ✔️ | ✔️ | ✔️ | ✔️ |
| GRADE [NeurIPS 2022] | ✔️ | ✔️ | ✔️ | ✔️ |
| MVGRL [ICML 2020] | ✔️ | ✔️ | ✔️ | ✔️ |
| InfoGraph [ICLR 2020] | ✔️ | ✔️ | ✔️ | ✔️ |
| MERIT [IJCAI 2021] | ✔️ | ✔️ | ✔️ | |
| GNN-POT [NeurIPS 2023] | ✔️ | |||
| MAGCL [AAAI 2023] | ✔️ | ✔️ | ✔️ | ✔️ |
| Sp2GCL [NeurIPS 2023] | ✔️ |
| Heterogeneous Graph Learning | TensorFlow | PyTorch | Paddle | MindSpore |
|---|---|---|---|---|
| RGCN [ESWC 2018] | ✔️ | ✔️ | ✔️ | ✔️ |
| HAN [WWW 2019] | ✔️ | ✔️ | ✔️ | ✔️ |
| HGT [WWW 2020] | ✔️ | ✔️ | ✔️ | ✔️ |
| SimpleHGN [KDD 2021] | ✔️ | ✔️ | ||
| CompGCN [ICLR 2020] | ✔️ | ✔️ | ✔️ | |
| HPN [TKDE 2021] | ✔️ | ✔️ | ✔️ | ✔️ |
| ieHGCN [TKDE 2021] | ✔️ | ✔️ | ✔️ | ✔️ |
| MetaPath2Vec [KDD 2017] | ✔️ | ✔️ | ✔️ | ✔️ |
| HERec [TKDE 2018] | ✔️ | ✔️ | ✔️ | ✔️ |
| HeCo [KDD 2021] | ✔️ | ✔️ | ✔️ | |
| DHN [TKDE 2023] | ✔️ | |||
| HEAT [T-ITS 2023] | ✔️ |
Note
The models can be run in mindspore backend. Howerver, the results of experiments are not satisfying due to training component issue, which will be fixed in future.
GammaGL Team[GAMMA LAB] and Peng Cheng Laboratory.
See more in CONTRIBUTING.
Contribution is always welcomed. Please feel free to open an issue or email to [email protected].
If you use GammaGL in a scientific publication, we would appreciate citations to the following paper:
@inproceedings{10.1145/3539618.3591891,
author = {Liu, Yaoqi and Yang, Cheng and Zhao, Tianyu and Han, Hui and Zhang, Siyuan and Wu, Jing and Zhou, Guangyu and Huang, Hai and Wang, Hui and Shi, Chuan},
title = {GammaGL: A Multi-Backend Library for Graph Neural Networks},
year = {2023},
isbn = {9781450394086},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3539618.3591891},
doi = {10.1145/3539618.3591891},
booktitle = {Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval},
pages = {2861–2870},
numpages = {10},
keywords = {graph neural networks, frameworks, deep learning},
location = {, Taipei, Taiwan, },
series = {SIGIR '23}
}