kaushalshetty/Structured-Self-Attention

Implements multi-hop self-attention with Frobenius norm regularization to generate sentence embeddings for classification tasks, supporting both binary and multiclass problems on IMDB and Reuters datasets. The architecture enables visualization of attention weights across sentence tokens via interactive heatmaps, with configurable parameters for attention hops, gradient clipping, and optional GloVe word embeddings. Achieves 90.2% accuracy on test sets while providing interpretability through attention weight analysis.

494 stars. No commits in the last 6 months.