tf_geometric使用消息传递机制来实现图神经网络：相比于基于稠密矩阵的实现，它具有更高的效率；相比于基于稀疏矩阵的实现，它具有更友好的API。除此之外，tf_geometric还为复杂的图神经网络操作提供了简易优雅的API。下面的示例展现了使用tf_geometric构建一个图结构的数据，并使用多头图注意力网络（Multi-headGAT）对图数据进行处理的流程：
#coding=utf-8importnumpyasnpimporttf_geometricastfgimporttensorflowastfgraph=tfg.Graph(x=np.random.randn(5,20),#5个节点,20维特征edge_index=[[0,0,1,3],[1,2,2,1]]#4个无向边)print("GraphDesc:\n",graph)graph.convert_edge_to_directed()#预处理边数据，将无向边表示转换为有向边表示print("ProcessedGraphDesc:\n",graph)print("ProcessedEdgeIndex:\n",graph.edge_index)#多头图注意力网络（Multi-headGAT）gat_layer=tfg.layers.GAT(units=4,num_heads=4,activation=tf.nn.relu)output=gat_layer([graph.x,graph.edge_index])print("OutputofGAT:\n",output)输出:
GraphDesc:GraphShape:x=>(5,20)edge_index=>(2,4)y=>NoneProcessedGraphDesc:GraphShape:x=>(5,20)edge_index=>(2,8)y=>NoneProcessedEdgeIndex:[[00111223][12023011]]OutputofGAT:tf.Tensor([[0.224431590.0.582632060.32468423][0.298103570.0.194036050.35630274][0.180719760.0.582632060.32468423][0.361232280.0.888972040.450244][0.0.0.80134620.]],shape=(5,4),dtype=float32)入门教程教程列表安装环境要求与依赖库使用pip一键安装tf_geometric及依赖快速入门使用简单示例快速入门面向对象接口（OOPAPI）和函数式接口（FunctionalAPI）使用示例进行快速入门强烈建议您通过下面的示例代码来快速入门tf_geometric：
节点分类图卷积网络GraphConvolutionalNetwork(GCN)多头图注意力网络Multi-headGraphAttentionNetwork(GAT)ApproximatePersonalizedPropagationofNeuralPredictions(APPNP)InductiveRepresentationLearningonLargeGraphs(GraphSAGE)切比雪夫网络ConvolutionalNeuralNetworksonGraphswithFastLocalizedSpectralFiltering(ChebyNet)SimpleGraphConvolution(SGC)TopologyAdaptiveGraphConvolutionalNetwork(TAGCN)DeepGraphInfomax(DGI)DropEdge:TowardsDeepGraphConvolutionalNetworksonNodeClassification(DropEdge)基于图卷积网络的文本分类GraphConvolutionalNetworksforTextClassification(TextGCN)图分类平均池化MeanPoolingGraphIsomorphismNetwork(GIN)自注意力图池化Self-AttentionGraphPooling(SAGPooling)可微池化HierarchicalGraphRepresentationLearningwithDifferentiablePooling(DiffPool)OrderMatters:SequencetoSequenceforSets(Set2Set)ASAP:AdaptiveStructureAwarePoolingforLearningHierarchicalGraphRepresentations(ASAP)AnEnd-to-EndDeepLearningArchitectureforGraphClassification(SortPool)最小割池化SpectralClusteringwithGraphNeuralNetworksforGraphPooling(MinCutPool)链接预测图自编码器GraphAuto-Encoder(GAE)保存和载入模型模型保存和载入使用tf.train.Checkpoint进行模型保存和载入分布式训练分布式图卷积网络（节点分类）分布式平均池化（图分类）引用如果您在科研出版物中使用了tf_geometric，欢迎引用下方的论文：
@misc{hu2021efficient,title={EfficientGraphDeepLearninginTensorFlowwithtf_geometric},author={JunHuandShengshengQianandQuanFangandYouzeWangandQuanZhaoandHuaiwenZhangandChangshengXu},year={2021},eprint={2101.11552},archivePrefix={arXiv},primaryClass={cs.LG}}
评论