This is the repo associated to the paper Multi-Objective Evolutionary Optimization of Virtualized Fast Feedforward Networks, to be published at EvoStar 2025.
To reproduce the experiments, you can run:
Virtualization via NSGA-II:
python src/main.py --multirun loader=mnist,har,sc exp=movirt Virtualizatin via random search:
python src/main.py --multirun loader=mnist,har,sc exp=randomvirt Landscape/Virtualization via manual tuning:
python src/main.py --multirun loader=mnist,har,sc exp=weightvirt exp.page_size=0,1,2,3,4 exp.amount=0.5,0.6,0.7,0.8,0.9 Note that these scripts run multiple experiments. If you want to run a single experiment, you can run:
python src/main.py loader=loader-name exp=experiment-name - Options for loader-name: mnist, har, sc
- Options for expriment-name: train, movirt, randomvirt, weightvirt, generate_fim
Please see YAML files inside conf directory to see more details regarding configurations. Code, data (including pretrained models and their corresponding fisher information matrices), and results are available at: https://dagshub.com/berab/MOE-VFFF.