v0.6 Release Highlight

The recent DGL 0.6 release is a major update on many aspects of the project including documentation, APIs, system speed, and scalability. This article highlights some of the new features and enhancements.

A Blitz Introduction to DGL in 120 minutes

The brand new set of tutorials come from our past hands-on tutorials in several major academic conferences (e.g., KDD’19, KDD’20, WWW’20). They start from an end-to-end example of using GNNs for node classification, and gradually unveil the core components in DGL such as DGLGraph, GNN modules, and graph datasets. The tutorials are now available on docs.dgl.ai.

A Gentle Tutorial on Mini-batch Training of GNNs

The scale of real world data can be massive, which demands training GNNs stochastically by mini-batches. However, unlike images or text corpus where data samples are independent, stochastic training of GNNs is more complex because one must handle the dependencies among samples. We observed that stochastic training is one of the most-asked topics on our discuss forum. In 0.6, we summarize the answers to those common questions in a set of tutorials on stochastic training of GNNs, including the insight into neighbor sampling algorithms, training loops and code snippets in DGL to realize them.

More Examples

The release includes 13 new examples, brings a total of 72 models:

• MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing: https://github.com/dmlc/dgl/tree/master/examples/pytorch/mixhop
• Self-Attention Graph Pooling: https://github.com/dmlc/dgl/tree/master/examples/pytorch/sagpool
• GNN-FiLM: Graph Neural Networks with Feature-wise Linear Modulation: https://github.com/dmlc/dgl/tree/master/examples/pytorch/GNN-FiLM
• TensorFlow implementation of Simplifying Graph Convolutional Networks: https://github.com/dmlc/dgl/tree/master/examples/tensorflow/sgc
• Graph Representation Learning via Hard and Channel-Wise Attention Networks: https://github.com/dmlc/dgl/tree/master/examples/pytorch/hardgat
• Graph Random Neural Network for Semi-Supervised Learning on Graphs: https://github.com/dmlc/dgl/tree/master/examples/pytorch/grand
• Hierarchical Graph Pooling with Structure Learning: https://github.com/dmlc/dgl/tree/master/examples/pytorch/hgp_sl
• Towards Deeper Graph Neural Networks: https://github.com/dmlc/dgl/tree/master/examples/pytorch/dagnn
• PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation/PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space (part segmentation): https://github.com/dmlc/dgl/tree/master/examples/pytorch/pointcloud/pointnet
• Graph Cross Networks with Vertex Infomax Pooling: https://github.com/dmlc/dgl/tree/master/examples/pytorch/gxn
• Neural Graph Collaborative Filtering: https://github.com/dmlc/dgl/tree/master/examples/pytorch/NGCF
• Link Prediction Based on Graph Neural Networks: https://github.com/dmlc/dgl/tree/master/examples/pytorch/seal
• Graph Neural Networks with Convolutional ARMA Filters: https://github.com/dmlc/dgl/tree/master/examples/pytorch/arma

The official example folder now indexes the examples by their notable tags such as their targeted tasks and so on.

Usability Enhancements

• Two new APIs DGLGraph.set_batch_num_nodes and DGLGraph.set_batch_num_edges for setting batch information manually, which are useful for transforming a batched graph into another or constructing a new batched graph manually.
• A new API GraphDataLoader, a data loader wrapper for graph classification tasks.
• A new dataset class QM9Dataset.
• A new namespace dgl.nn.functional for hosting NN related utility functions.
• DGL now supports training with half precision and is compatible with PyTorch’s automatic mixed precision package. See the user guide chapter for how to use it.
• (Experimental) Users can now use DistGraph with heterogeneous graph data. This also applies to dgl.sample_neighbors on DistGraph. In addition, DGL supports distributed graph partitioning on a cluster of machines. See the user guide chapter for more details.
• (Experimental) Several new APIs for training sparse embeddings:
  • dgl.nn.NodeEmbedding is a dedicated class for storing trainable node embeddings that can scale to graphs with millions of nodes.
  • dgl.optim.SparseAdagrad and dgl.optim.SparseAdam are two optimizers for the NodeEmbedding class.

System Efficiency Improvements

• With PyTorch backend, DGL will use PyTorch’s native memory management to cache repeated memory allocation and deallocation.
• A new implementation for nn.RelGraphConv when low_mem=True (PyTorch backend). A benchmark on V100 GPU shows it gives a 4.8x boost in training speed on AIFB dataset.
• Faster CPU kernels using AVX512 instructions.
• Faster GPU kernels on CUDA 11.

Further Readings

• Full release note: https://github.com/dmlc/dgl/releases/tag/v0.6.0