PyTorch Geometric code for the paper "R-GCN: The R Could Stand for Random"
RR-GCN is an extension of Relational Graph Convolutional Networks (R-GCN) in which the weights are randomly initialised and kept frozen (i.e. no training step is required). As such, our technique is unsupervised and the produced embeddings can be used for any downstream ML task/model. Surprisingly, empirical results indicate that the embeddings produced by our RR-GCN can be competitive to, and even sometimes outperform, end-to-end R-GCNs.
This snippet generates train and test node embeddings for a KG stored as a PyG Data object.
from rrgcn import RRGCNEmbedder
embedder = RRGCNEmbedder(num_nodes=data.num_nodes,
num_relations=dataset.num_relations,
num_layers=2,
emb_size=2048,
device="cuda",
ppv=True)
embeddings = embedder.embeddings(data.edge_index,
data.edge_type,
batch_size=0, # full-batch
idx=torch.hstack((data.train_idx, data.test_idx)))
train_embeddings = embeddings[:len(data.train_idx)]
test_embeddings = embeddings[len(data.train_idx):] Check our documentation for more information
Example notebook for MUTAG (illustration of numeric literal support)
Example notebook for AM (illustration of numeric & textual literal support)
Example notebook for ogbn-mag (illustration of support for initial node features)
After installing the correct PyG version for your PyTorch/CUDA installation as documented in their installation instructions, install the RR-GCN package using:
pip install rrgcn
