kyegomez/VisionMamba

Bidirectional state space models replace traditional attention mechanisms to achieve linear complexity, enabling efficient processing of high-resolution images with significantly reduced memory footprint. The architecture uses patch embedding with configurable depth and state dimensions, integrated with PyTorch for straightforward model instantiation and forward pass inference. Supports flexible configuration across vision tasks with parameters for model capacity (dim, depth), state representation (d_state, dt_rank), and image processing (patch_size, image_size).

482 stars.