Documentation |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.

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 create a simple 2-layer GCN model using the pre-defined GCNConv:

import tensorlayerx as tlx
from gammagl.layers.conv import GCNConv
from gammagl.datasets import Planetoid

dataset = Planetoid(root='.', name='Cora')

class GCN(tlx.nn.Module):
    def __init__(self, in_channels, hidden_channels, out_channels):
        super().__init__()
        self.conv1 = GCNConv(in_channels, hidden_channels)
        self.conv2 = GCNConv(hidden_channels, out_channels)
        self.relu = tlx.ReLU()

    def forward(self, x, edge_index):
        # x: Node feature matrix of shape [num_nodes, in_channels]
        # edge_index: Graph connectivity matrix of shape [2, num_edges]
        x = self.conv1(x, edge_index)
        x = self.relu(x)
        x = self.conv2(x, edge_index)
        return x

model = GCN(dataset.num_features, 16, dataset.num_classes)

import tensorlayerx as tlx
data = dataset[0]
loss_fn = tlx.losses.softmax_cross_entropy_with_logits
optimizer = tlx.optimizers.Adam(learning_rate=1e-3)
net_with_loss = tlx.model.WithLoss(model, loss_fn)
train_one_step = tlx.model.TrainOneStep(net_with_loss, optimizer, train_weights)

for epoch in range(200):
    loss = train_one_step(data.x, data.y)

import torch.nn.functional as F

data = dataset[0]
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

for epoch in range(200):
    pred = model(data.x, data.edge_index)
    loss = F.cross_entropy(pred[data.train_mask], data.y[data.train_mask])

    # Backpropagation
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

import tensorflow as tf

optimizer = tf.keras.optimizers.Adam()
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
for epoch in range(200):
    with tf.GradientTape() as tape:
        predictions = model(images, training=True)
        loss = loss_fn(labels, predictions)
    gradients = tape.gradient(loss, model.trainable_variables)
    optimizer.apply_gradients(zip(gradients, model.trainable_variables))

import paddle

data = dataset[0]
optim = paddle.optimizer.Adam(parameters=model.parameters())
loss_fn = paddle.nn.CrossEntropyLoss()

model.train()
for epoch in range(200):
    predicts = model(data.x, data.edge_index)
    loss = loss_fn(predicts, y_data)

    # Backpropagation
    loss.backward()
    optim.step()
    optim.clear_grad()

# 1. Generate training dataset
train_dataset = create_dataset(num_data=160, batch_size=16)

# 2.Build a model and define the loss function
net = LinearNet()
loss = nn.MSELoss()

# 3.Connect the network with loss function, and define the optimizer
net_with_loss = nn.WithLossCell(net, loss)
opt = nn.Momentum(net.trainable_params(), learning_rate=0.005, momentum=0.9)

# 4.Define the training network
train_net = nn.TrainOneStepCell(net_with_loss, opt)

# 5.Set the model as training mode
train_net.set_train()

# 6.Training procedure
for epoch in range(200):
    for d in train_dataset.create_dict_iterator():
        result = train_net(d['data'], d['label'])
        print(f"Epoch: [{epoch} / {epochs}], "
              f"step: [{step} / {steps}], "
              f"loss: {result}")
        step = step + 1

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.:

$$x_i^{\prime} ~ = ~ \max_{j \in \mathcal{N}(i)} ~ \textrm{MLP}_{\theta} \left( [ ~ x_i, ~ x_j - x_i ~ ] \right)$$

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)

1. Python environment (Optional): We recommend using Conda package manager

   conda create -n ggl python=3.7
   source activate ggl

2. Install Backend

   # For tensorflow
   pip install tensorflow-gpu # GPU version
   pip install tensorflow # CPU version
   
   # For torch, version 1.10
   # https://pytorch.org/get-started/locally/
   pip3 install torch==1.10.2
   
   # For paddle, any latest stable version
   # https://www.paddlepaddle.org.cn/
   python -m pip install paddlepaddle-gpu
   
   # For mindspore, GammaGL only supports version 1.8.1, GPU-CUDA 11.1
   # https://www.mindspore.cn/install
   pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/1.8.1/MindSpore/gpu/x86_64/cuda-11.1/mindspore_gpu-1.8.1-cp37-cp37m-linux_x86_64.whl --trusted-host ms-release.obs.cn-north-4.myhuaweicloud.com -i https://pypi.tuna.tsinghua.edu.cn/simple

   For other backend with specific version, please check whether TLX supports.

   Install TensorLayerX

   pip install git+https://github.com/tensorlayer/tensorlayerx.git 

   大陆用户如果遇到网络问题，推荐从启智社区安装

   Try to git clone from OpenI

   pip install git+https://git.openi.org.cn/OpenI/TensorLayerX.git

3. Download GammaGL

   git clone --recursive https://github.com/BUPT-GAMMA/GammaGL.git
   python setup.py install

   大陆用户如果遇到网络问题，推荐从启智社区安装

   Try to git clone from OpenI

   git clone --recursive https://git.openi.org.cn/GAMMALab/GammaGL.git

   "--recursive" is required, if you forgot, you can run command below in Gammagl root dir:

   git submodule update --init

Take GCN as an example:

# cd ./examples/gcn
# set parameters if necessary
python gcn_trainer.py --dataset cora --lr 0.01

If you want to use specific backend or GPU, just set environment variable like:

CUDA_VISIBLE_DEVICES="1" TL_BACKEND="paddle" python gcn_trainer.py

Note

The DEFAULT backend is tensorflow and GPU is 0. The backend TensorFlow will take up all GPU left memory by default.

The CANDIDATE backends are tensorflow, paddle, torch and mindspore.

Set CUDA_VISIBLE_DEVICES=" " if you want to run it in CPU.

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{Liu2023gammagl,
  title={GammaGL: A Multi-Backend Library for Graph Neural Networks},
  author={Yaoqi Liu, Cheng Yang, Tianyu Zhao, Hui Han, Siyuan Zhang, Jing Wu, Guangyu Zhou, Hai Huang, Hui Wang, Chuan Shi},
  booktitle={SIGIR},
  year={2023}
}