Python implementation of the local2global
patch embedding synchronisation
method used in [1]. This code uses embeddings for a set of overlapping
patches as input and aligns them to obtain a global embedding. The alignment step is based on the eigenvector synchronisation method of [2].
The local2global
package can be installed using pip
. Simply run
pip install git+https://github.com/LJeub/Local2Global.git@master
For more information see the documentation.
[1] L. G. S. Jeub, G. Colavizza, X. Dong, M. Bazzi, M. Cucuringu (2021). Local2Global: Scaling global representation learning on graphs via local training. DLG-KDD'21. arXiv:2107.12224 [cs.LG]
[2] M. Cucuringu, Y. Lipman, A. Singer (2012). Sensor network localization by eigenvector synchronization over the euclidean group. ACM Transactions on Sensor Networks 8.3. DOI: 10.1145/2240092.2240093