Jeff Rhoades

PyTorch implementation of Rotary Spatial Embeddings

PyTorch implementation of a Resolution Aware Transformer, designed to take in multiple scales of fixed-resolution images (such as microscopy or medical imaging), with the goal of producing embeddings informed by global context and local details. Utilizes rotary spatial embeddings (RoSE) and spatial grouping attention.

Inspired by the spatial grouping layer in Native Segmentation Vision Transformers (https://arxiv.org/abs/2505.16993), implemented in PyTorch with a modified rotary position embedding generalized to N-dimensions and incorporating real-world pixel spacing.

A repository of experiments inspecting LLM's, particularly focused on detecting signs of human-like self-awareness via psychodynamic-inspired prompting.