Official code for Progressive Voronoi Diagram Subdivision Enables Accurate Data-free Class-Incremental Learning, ICLR'23.
The iVoro method is based on the idea of Voronoi Diagram subdivision from Computational Geometry.
A. Establish Voronoi Diagram based on base model.
B. Insertion of a new class as a new Voronoi cell enables the minimal intervention to the overall structure.
C. Divide-and-conquer (a classical algorithm for Voronoi construction) efficiently introduce a batch of new classes into the system.
The results of MNIST in 2D space below clearly showed different space subdivision results from conventional fine-tuning, PASS, and different variants of iVoro.
Step 1. Training of the base model, please follow PASS (github).
Step 2. Download the feature files. Google Drive
Go to the directory:
cd MNISTThen run analysis/CIFAR_voro.py in following order:
A. iVoro B. iVoro-D C. iVoro-AC/AI D. iVoro-L
See "Release" page for the codebase. Note that the raw code is provided as is, not cleaned and highly messy. If I had time, I would do some cleaning.