Batch-efficient EigenDecomposition for Small and Medium Matrices
A Pytorch-based batch-efficient ED solver for small and medium matrices (dim<32), which is dedicated to the application scenarios of computer vision.
Main Characteristic
The core part of the algorithm is based on the QR iteration with Double Wilkinson shifts and some other acceleration techniques carefully designed for the best batch efficiency. Our Pytorch-implemented solver performs the ED entirely via batched matrix-matrix multiplication, which processes all the matrices simultaneously and thus fully exploits the parallel computational power of GPUs.
Research areas
Integrative AI
Last updated
26.01.2024 - 10:22