kimhc6028/relational-networks

Implements relational reasoning through a specialized neural module that compares object pairs using concatenated feature representations, evaluated on the Sort-of-CLEVR visual question-answering task with both binary and ternary relation types. The architecture combines a CNN feature extractor with a relation module that processes pairwise object interactions, achieving 89% accuracy on relational questions compared to 66% for standard CNN+MLP baselines. Supports PyTorch training with configurable relation types and includes dataset generation utilities for the synthetic CLEVR benchmark.

818 stars. No commits in the last 6 months.