Efficient and Friendly Graph Neural Network (GNN) Library for TensorFlow 1.x and 2.x.
Inspired by rusty1s/pytorch_geometric, we build a GNN library for TensorFlow.
- Homepage: https://github.com/CrawlScript/tf_geometric
- Documentation: https://tf-geometric.readthedocs.io (中文版)
- Paper: Efficient Graph Deep Learning in TensorFlow with tf_geometric
We use Message Passing mechanism to implement Graph Neural Networks (GNNs), which is way efficient than the dense matrix based implementations and more friendly than the sparse matrix based ones. In addition, we provide easy and elegant APIs for complex GNN operations. The following example constructs a graph and applies a Multi-head Graph Attention Network (GAT) on it:
# coding=utf-8importnumpyasnpimporttf_geometricastfgimporttensorflowastfgraph=tfg.Graph(
x=np.random.randn(5, 20), # 5 nodes, 20 features,edge_index=[[0, 0, 1, 3],
[1, 2, 2, 1]] # 4 undirected edges
)
print("Graph Desc: \n", graph)
graph=graph.to_directed() # pre-process edgesprint("Processed Graph Desc: \n", graph)
print("Processed Edge Index:\n", graph.edge_index)
# Multi-head Graph Attention Network (GAT)gat_layer=tfg.layers.GAT(units=4, num_heads=4, activation=tf.nn.relu)
output=gat_layer([graph.x, graph.edge_index])
print("Output of GAT: \n", output)Output:
Graph Desc:
Graph Shape: x => (5, 20) edge_index => (2, 4) y => None
Processed Graph Desc:
Graph Shape: x => (5, 20) edge_index => (2, 8) y => None
Processed Edge Index:
[[0 0 1 1 1 2 2 3]
[1 2 0 2 3 0 1 1]]
Output of GAT:
tf.Tensor(
[[0.22443159 0. 0.58263206 0.32468423]
[0.29810357 0. 0.19403605 0.35630274]
[0.18071976 0. 0.58263206 0.32468423]
[0.36123228 0. 0.88897204 0.450244 ]
[0. 0. 0.8013462 0. ]], shape=(5, 4), dtype=float32)We recommend you to get started with some demo.
- Graph Convolutional Network (GCN)
- Multi-head Graph Attention Network (GAT)
- Approximate Personalized Propagation of Neural Predictions (APPNP)
- Inductive Representation Learning on Large Graphs (GraphSAGE)
- Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering (ChebyNet)
- Simple Graph Convolution (SGC)
- Topology Adaptive Graph Convolutional Network (TAGCN)
- Deep Graph Infomax (DGI)
- DropEdge: Towards Deep Graph Convolutional Networks on Node Classification (DropEdge)
- Graph Convolutional Networks for Text Classification (TextGCN)
- Simple Spectral Graph Convolution (SSGC/S^2GC)
- MeanPooling
- Graph Isomorphism Network (GIN)
- Self-Attention Graph Pooling (SAGPooling)
- Hierarchical Graph Representation Learning with Differentiable Pooling (DiffPool)
- Order Matters: Sequence to Sequence for Sets (Set2Set)
- ASAP: Adaptive Structure Aware Pooling for Learning Hierarchical Graph Representations (ASAP)
- An End-to-End Deep Learning Architecture for Graph Classification (SortPool)
- Spectral Clustering with Graph Neural Networks for Graph Pooling (MinCutPool)
Requirements:
- Operation System: Windows / Linux / Mac OS
- Python: version >= 3.7
- Python Packages:
- tensorflow/tensorflow-gpu: >= 1.15.0 or >= 2.7.0
- tf_sparse
- numpy >= 1.17.4
- networkx >= 2.1
- scipy >= 1.1.0
Use one of the following commands below:
pip install -U tf_geometric # this will not install the tensorflow/tensorflow-gpu package
pip install -U tf_geometric[tf1-cpu] # this will install TensorFlow 1.x CPU version
pip install -U tf_geometric[tf1-gpu] # this will install TensorFlow 1.x GPU version
pip install -U tf_geometric[tf2-cpu] # this will install TensorFlow 2.x CPU version
pip install -U tf_geometric[tf2-gpu] # this will install TensorFlow 2.x GPU versionWe provide both OOP and Functional API, with which you can make some cool things.
# coding=utf-8importos# Enable GPU 0os.environ["CUDA_VISIBLE_DEVICES"] ="0"importtf_geometricastfgimporttensorflowastfimportnumpyasnp# ==================================== Graph Data Structure ====================================# In tf_geometric, the data of a graph can be represented by either a collections of# tensors (numpy.ndarray or tf.Tensor) or a tfg.Graph object.# A graph usually consists of x(node features), edge_index and edge_weight(optional)# Node Features => (num_nodes, num_features)x=np.random.randn(5, 20).astype(np.float32) # 5 nodes, 20 features# Edge Index => (2, num_edges)# Each column of edge_index (u, v) represents an directed edge from u to v.# Note that it does not cover the edge from v to u. You should provide (v, u) to cover it.# This is not convenient for users.# Thus, we allow users to provide edge_index in undirected form and convert it later.# That is, we can only provide (u, v) and convert it to (u, v) and (v, u) with `convert_edge_to_directed` method.edge_index=np.array([
[0, 0, 1, 3],
[1, 2, 2, 1]
])
# Edge Weight => (num_edges)edge_weight=np.array([0.9, 0.8, 0.1, 0.2]).astype(np.float32)
# Usually, we use a graph object to manager these information# edge_weight is optional, we can set it to None if you don't need it# Using 'to_directed' to obtain a graph with directed edges such that we can use it as the input of GCNgraph=tfg.Graph(x=x, edge_index=edge_index, edge_weight=edge_weight).to_directed()
# Define a Graph Convolutional Layer (GCN)gcn_layer=tfg.layers.GCN(4, activation=tf.nn.relu)
# Perform GCN on the graphh=gcn_layer([graph.x, graph.edge_index, graph.edge_weight])
print("Node Representations (GCN on a Graph): \n", h)
for_inrange(10):
# Using Graph.cache can avoid recomputation of GCN's normalized adjacency matrix,# which can dramatically improve the efficiency of GCN.h=gcn_layer([graph.x, graph.edge_index, graph.edge_weight], cache=graph.cache)
# For algorithms that deal with batches of graphs, we can pack a batch of graph into a BatchGraph object# Batch graph wrap a batch of graphs into a single graph, where each nodes has an unique index and a graph index.# The node_graph_index is the index of the corresponding graph for each node in the batch.# The edge_graph_index is the index of the corresponding edge for each node in the batch.batch_graph=tfg.BatchGraph.from_graphs([graph, graph, graph, graph, graph])
# We can reversely split a BatchGraph object into Graphs objectsgraphs=batch_graph.to_graphs()
# Define a Graph Convolutional Layer (GCN)batch_gcn_layer=tfg.layers.GCN(4, activation=tf.nn.relu)
# Perform GCN on the BatchGraphbatch_h=gcn_layer([batch_graph.x, batch_graph.edge_index, batch_graph.edge_weight])
print("Node Representations (GCN on a BatchGraph): \n", batch_h)
# Graph Pooling algorithms often rely on such batch data structure# Most of them accept a BatchGraph's data as input and output a feature vector for each graph in the batchgraph_h=tfg.nn.mean_pool(batch_h, batch_graph.node_graph_index, num_graphs=batch_graph.num_graphs)
print("Graph Representations (Mean Pooling on a BatchGraph): \n", batch_h)
# Define a Graph Convolutional Layer (GCN) for scoring each nodegcn_score_layer=tfg.layers.GCN(1)
# We provide some advanced graph pooling operations such as topk_poolnode_score=gcn_score_layer([batch_graph.x, batch_graph.edge_index, batch_graph.edge_weight])
node_score=tf.reshape(node_score, [-1])
print("Score of Each Node: \n", node_score)
topk_node_index=tfg.nn.topk_pool(batch_graph.node_graph_index, node_score, ratio=0.6)
print("Top-k Node Index (Top-k Pooling): \n", topk_node_index)
# ==================================== Built-in Datasets ====================================# all graph data are in numpy format# Cora Datasetgraph, (train_index, valid_index, test_index) =tfg.datasets.CoraDataset().load_data()
# PPI Datasettrain_data, valid_data, test_data=tfg.datasets.PPIDataset().load_data()
# TU Datasets# TU Datasets: https://ls11-www.cs.tu-dortmund.de/staff/morris/graphkerneldatasetsgraph_dicts=tfg.datasets.TUDataset("NCI1").load_data()
# ==================================== Basic OOP API ====================================# OOP Style GCN (Graph Convolutional Network)gcn_layer=tfg.layers.GCN(units=20, activation=tf.nn.relu)
forgraphintest_data:
# Cache can speed-up GCN by caching the normed edge informationoutputs=gcn_layer([graph.x, graph.edge_index, graph.edge_weight], cache=graph.cache)
print(outputs)
# OOP Style GAT (Multi-head Graph Attention Network)gat_layer=tfg.layers.GAT(units=20, activation=tf.nn.relu, num_heads=4)
forgraphintest_data:
outputs=gat_layer([graph.x, graph.edge_index])
print(outputs)
# OOP Style Multi-layer GCN ModelclassGCNModel(tf.keras.Model):
def__init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.gcn0=tfg.layers.GCN(16, activation=tf.nn.relu)
self.gcn1=tfg.layers.GCN(7)
self.dropout=tf.keras.layers.Dropout(0.5)
defcall(self, inputs, training=None, mask=None, cache=None):
x, edge_index, edge_weight=inputsh=self.dropout(x, training=training)
h=self.gcn0([h, edge_index, edge_weight], cache=cache)
h=self.dropout(h, training=training)
h=self.gcn1([h, edge_index, edge_weight], cache=cache)
returnhgcn_model=GCNModel()
forgraphintest_data:
outputs=gcn_model([graph.x, graph.edge_index, graph.edge_weight], cache=graph.cache)
print(outputs)
# ==================================== Basic Functional API ====================================# Functional Style GCN# Functional API is more flexible for advanced algorithms# You can pass both data and parameters to functional APIsgcn_w=tf.Variable(tf.random.truncated_normal([test_data[0].num_features, 20]))
forgraphintest_data:
outputs=tfg.nn.gcn(graph.x, graph.adj(), gcn_w, activation=tf.nn.relu)
print(outputs)
# ==================================== Advanced Functional API ====================================# Most APIs are implemented with Map-Reduce Style# This is a gcn without without weight normalization and transformation# Just pass the mapper/reducer/updater functions to the Functional APIforgraphintest_data:
outputs=tfg.nn.aggregate_neighbors(
x=graph.x,
edge_index=graph.edge_index,
edge_weight=graph.edge_weight,
mapper=tfg.nn.identity_mapper,
reducer=tfg.nn.sum_reducer,
updater=tfg.nn.sum_updater
)
print(outputs)If you use tf_geometric in a scientific publication, we would appreciate citations to the following paper:
@inproceedings{DBLP:conf/mm/HuQFWZZX21,
author = {Jun Hu and
Shengsheng Qian and
Quan Fang and
Youze Wang and
Quan Zhao and
Huaiwen Zhang and
Changsheng Xu},
editor = {Heng Tao Shen and
Yueting Zhuang and
John R. Smith and
Yang Yang and
Pablo Cesar and
Florian Metze and
Balakrishnan Prabhakaran},
title = {Efficient Graph Deep Learning in TensorFlow with tf{\_}geometric},
booktitle = {{MM} '21: {ACM} Multimedia Conference, Virtual Event, China, October
20 - 24, 2021},
pages = {3775--3778},
publisher = {{ACM}},
year = {2021},
url = {https://doi.org/10.1145/3474085.3478322},
doi = {10.1145/3474085.3478322},
timestamp = {Wed, 20 Oct 2021 12:40:01 +0200},
biburl = {https://dblp.org/rec/conf/mm/HuQFWZZX21.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}- MIG-GT: "Modality-Independent Graph Neural Networks with Global Transformers for Multimodal Recommendation" (AAAI 2025). URL: https://github.com/CrawlScript/MIG-GT.
- RpHGNN: “Efficient Heterogeneous Graph Learning via Random Projection” (TKDE 2024). URL: https://github.com/CrawlScript/RpHGNN.
- MGDCF: "MGDCF: Distance Learning via Markov Graph Diffusion for Neural Collaborative Filtering" (TKDE 2024). URL: https://github.com/CrawlScript/Torch-MGDCF.
- tf_sparse: We develop TensorFlow Sparse (tf_sparse) to implement efficient and elegant sparse TensorFlow operations for tf_geometric. URL: https://github.com/CrawlScript/tf_sparse.
- GRecX:GRecX is an efficient and unified benchmark for GNN-based recommendation. URL: https://github.com/maenzhier/GRecX.
- 🐈 MMClaw: The Ultra-Lightweight, Pure Python Kernel for Multimodal AI Agents. URL: https://github.com/CrawlScript/MMClaw.