A minor release that brings PyTorch 1.9.0 and Python 3.9 support to PyTorch Geometric. In case you are in the process of updating to PyTorch 1.9.0, please re-install the external dependencies for PyTorch 1.9.0 as well (torch-scatter and torch-sparse).

Features

• EGConv (thanks to @shyam196)
• GATv2Conv (thanks to @shakedbr)
• GraphNorm normalization layer
• GNNExplainer now supports explaining graph-level predictions (thanks to @wsad1)
• bro and gini regularization (thanks to @rhsimplex)
• train_test_split_edges() and to_undirected() can now edge features (thanks to @saiden89 and @SherylHYX)
• Datasets can now be accessed with np.ndarray as well (thanks to @josephenguehard)
• dense_to_sparse can now handle batched adjacency matrices
• numba is now an optional dependency

Datasets

• The tree-structured fake news propagation UPFD dataset (thanks to @YingtongDou)
• The large-scale AmazonProducts graph from the GraphSAINT paper
• Added support for two more datasets in the SNAPDataset benchmark suite (thanks to @SherylHYX)

Issues

• Fixed an issue in which SuperGATConv used all positive edges for computing the auxiliary loss (thanks to @anniekmyatt)
• Fixed a bug in which MemPooling produced NaN gradients (thanks to @wsad1)
• Fixed an issue in which the schnetpack package was required for training SchNet (thanks to @mshuaibii)
• Modfied XConv to sample without replacement in case dilation > 1 (thanks to @mayur-ag)
• GraphSAINTSampler can now be used in combination with PyTorch Lightning
• Fixed a bug in HypergraphConv in case num_nodes > num_edges (thanks to @THinnerichs)