Update README.md
Browse files# GMT: Graph Matching Transformer
**GMT** (Graph Matching Transformer) is a PyTorch-based framework for matching and aligning 2D curves (graphs) using rich geometric embeddings and a cross‑attention Transformer architecture. It supports multiple model variants—`tiny`, `small`, `medium`, `large`, `xl`, and `2xl`—to scale computational complexity and capacity.
## Key Features
* **Curvature & Ray Embeddings**: Computes curvature, ray distances, incidence angles, and hit flags for each point.
* **Index & Initial Shift Embedding**: Includes normalized index, curvature, and initial displacement as features.
* **Cross‑Attention Transformer**: Two-stream self‑attention on target & baseline, followed by cross‑attention for fine-grained alignment.
* **Variants**: Six predefined configurations (`tiny` … `2xl`) with adjustable `d_model`, depth, and feed‑forward dimensions.
* **Metal/CUDA/CPU**: Auto-selects MPS (Apple Silicon), CUDA, or CPU device.
* **Visualizations**: Built-in training loss curves, inference progression plots, and error distribution histograms.