Knowledge graphs are large graph-structured databases of facts, which typically suffer from incompleteness. Link prediction is the task of inferring missing relations (links) between entities (nodes) in a knowledge graph. We approach this task using a hypernetwork architecture to generate convolutional layer filters specific to each relation and apply those filters to the subject entity embeddings. This architecture enables a trade-off between non-linear expressiveness and the number of parameters to learn. Our model simplifies the entity and relation embedding interactions introduced by the predecessor convolutional model, while outperforming all previous approaches to link prediction across all standard link prediction datasets.