v0.4.3

We are thrilled to announce DGL v0.4.3, which provides many new features that enhance usability and efficiency.

Major Features and Improvements

• TensorFlow backend is now an official feature. Add an environment variable USE_OFFICIAL_TFDLPACK for switching to the official TensorFlow DLPack support from tensorflow/tensorflow#36862.
• New graph sampling APIs compatible with DGLHeteroGraph:
  • Redesigned dgl.random_walk with the support of metapaths.
  • A new API dgl.random_walk_with_restart for random walk with restart probability.
  • A new API dgl.sample_neighbors for sampling among neighbors with or without replacement.
  • A new API dgl.sample_neighbors_topk for picking K neighbors with the largest weight.
  • A new API dgl.in_subgraph for extracting all the neighbors connected by the in-edges.
  • A new API dgl.out_subgraph for extracting all the neighbors connected by the out-edges.
• Accompany utilities for graph sampling:
  • A new API dgl.create_from_paths for creating a graph from sampled random walk traces.
  • A new API dgl.compact_graphs for converting a sampled subgraph to a smaller graph with no isolated nodes.
  • A new API dgl.to_block for converting a sampled subgraph to a bipartite graph suitable for computation.
  • Reworked dgl.to_simple_graph to support heterogeneous graph.
  • Reworked dgl.remove_edges to support heterogeneous graph.
• When constructing a DGLHeteroGraph to be unidirectional bipartite — there are two node types and one edge type, where all edges are from one node type to another, the following APIs are enabled:
  • A new API DGLHeteroGraph.is_unibipartite
  • New APIs DGLHeteroGraph.num_src_nodes and DGLHeteroGraph.num_dst_nodes
  • New APIs DGLHeteroGraph.srcnodes and DGLHeteroGraph.dstnodes for getting a view of source and destination nodes, respectively.
  • New APIs DGLHeteroGraph.srcdata and DGLHeteroGraph.dstdata for getting the data of source and destination nodes, respectively.
• NN module changes:
  • Users can now directly use dgl.nn.SomeModule instead of dgl.nn.<backend>.SomeModule.
  • Extend dgl.nn.GraphConv to support asymmetric normalizer. It now also accepts an external weight matrix instead of creating its own.
  • Extend all the NN modules to support bipartite graph inputs which enable them for sampling-based GNN training. The input node feature argument now can be a pair of tensors.
  • A new wrapper module dgl.nn.HeteroGraphConv for leveraging DGL NN modules on heterogeneous graphs.
• Model examples using the new sampling APIs
  • Train the GraphSAGE model by neighbor sampling and scale it to multiple GPUs (link).
  • Train the Relational GCN model on heterogeneous graphs by sampling for both node classification and link prediction (link).
  • Train the PinSAGE model by random walk sampling for item recommendation (link).
  • Train the GCMC model by sampling for MovieLens rating prediction (link).
  • Implement the variance reduction technique for neighbor sampling (link) proposed by Chen et al.
• DGL-KE
  • Move to a standalone repository under awslab: https://github.com/awslabs/dgl-ke
  • See the release note for more details.
• DGL-LifeSci
  • Spun off into a standalone package.
  • See the project page for more details.
• A new example for scene graph extraction: https://github.com/dmlc/dgl/tree/master/examples/mxnet/scenegraph
• A new API dgl.metis_partition for partitioning a DGLGraph by the Metis algorithm.
• New APIs dgl.as_immutable_graph and dgl.as_heterograph for casting between DGLGraph and DGLHeteroGraph efficiently.
• A new API dgl.rand_graph for constructing a random graph with specified number of nodes and edges.
• A new API dgl.random.choice for more efficient non-uniform random choice.
• Replaced DGLHeteroGraph.__setitem__ and DGLHeteroGraph.__getitem__ with a more efficient implementation.
• dgl.data.save_graphs and dgl.data.load_graphs now support heterogeneous graphs.
• UDFs now have the access to node types and edge types.

API Breaking Changes

• The type of the norm argument in dgl.nn.GraphConv is changed from bool to string, with "none" indicating no normalization, "both" indicating the original symmetric normalizer proposed by GCN and "right" indicating normalizing by degrees.
• DGLSubGraph and BatchedDGLGraph classes are removed and merged to DGLGraph. All their methods are ported to DGLGraph too, so typical usages will not be affected by this change.
• The multigraph flag in dgl.DGLGraph is deprecated and will be removed in the future.
• Rename the card argument in dgl.graph and dgl.bipartite to num_nodes.

Bug Fixes and Others

• Fix a bug in remove_edges when the graph has no edge.
• Fix a bug in creating DGLGraph from scipy coo matrix that has duplicate entries.
• Improve the speed of sorting a COO format graph.
• Improve the speed of dgl.to_bidirected .
• Fix a bug in building DGL on MacOS using clang.
• Fix a bug in NodeFlow when apply_edges is called.
• Fix a bug in the stack cross-type reducer in DGLHeteroGraph.multi_update_all, DGLHeteroGraph.multi_pull and DGLHeteroGraph.multi_recv to make the stacking order consistent and to remove a redundant dimension.
• Fix a bug in the loss function of the RGCN example.
• Fix a bug in the MXNet backend when using the new Deep NumPy feature.
• Fix a memory leak bug in the PyTorch backend when retain_graph is True.