Description
The lightaime/deep_gcns_torch repository is a comprehensive PyTorch implementation for training exceptionally deep Graph Convolutional Networks (GCNs). Inspired by the success of deep Convolutional Neural Networks (CNNs), this project introduces novel methods to overcome the challenges of training GCNs with a large number of layers. Key architectural adaptations include the integration of residual and dense connections, along with dilated convolutions, which are fundamental to achieving depth in CNNs. These techniques are carefully adapted to the unique structure of graph data, allowing for more powerful and scalable GNN models.
Read More
The repository hosts code for several significant GNN models and research contributions: DeepGCNs (published at ICCV'2019 and TPAMI'2021), DeeperGCN (presented in arXiv'2020), and GNN1000 (introduced at ICML'2021). These models have been evaluated on various benchmark datasets, including S3DIS, PartNet, ModelNet40, PPI, and OGB datasets, demonstrating their effectiveness in tasks such as point cloud classification, semantic segmentation, and node/graph property prediction.
Users can find detailed instructions for setting up the environment, running experiments, and evaluating models within the README files located in the respective example folders. The code architecture is organized to facilitate modularity and extensibility, with dedicated directories for GCN libraries (sparse and dense data), memory-efficient modules, and various example applications. The project emphasizes extensive experimentation and ablation studies to understand the impact of different architectural components on GCN performance.
This resource is invaluable for researchers and practitioners in the fields of computer vision, bioinformatics, cheminformatics, and social network analysis who are working with graph-structured data. By providing well-implemented GNN models and clear experimental setups, the repository aims to advance the state-of-the-art in deep learning on graphs and encourage further exploration into deeper and more effective graph neural network architectures.
- CategoryMachine Learning Platforms
- Pricing ModelOpen source
- DeploymentCloud, Self-hosted
- PlatformsWeb
DeepGCNs's Core Features
PyTorch implementation for DeepGCNs
Code for DeeperGCN (arXiv'2020)
Implementation of GNN1000 (ICML'2021)
Adaptation of CNN concepts (residual/dense connections, dilation) for GCNs
Extensive experimental results and ablation studies
Support for various graph datasets (S3DIS, PartNet, ModelNet40, PPI, OGB)
Code for point cloud classification and semantic segmentation
Code for node and graph property prediction
Memory-efficient GNN modules
Modular code architecture for GCN libraries (sparse and dense)
Clear instructions for environment setup and model execution
Open-source under MIT License
Getting Started with DeepGCNs
Clone repository: Download the code from the GitHub repository.
Install environment: Run the `deepgcn_env_install.sh` script to set up required libraries.
Configure experiments: Navigate to the `examples` directory and select a task (e.g., `modelnet_cls`, `sem_seg_dense`).
Run models: Follow instructions in task-specific READMEs to train, test, and evaluate models.
Integrate modules: Utilize GCN libraries and efficient modules in custom graph neural network projects.
Adapt for research: Modify and extend existing architectures for novel graph learning tasks.
DeepGCNs's Use Cases
- Point Cloud Classification
- Semantic Segmentation
- Node Classification
- Graph Property Prediction
- Part Segmentation
- GNN Research
