Code repository for "Continual Learning with Guarantees via Weight Interval Constraints" accepted to ICML 2022. Paper link: https://arxiv.org/abs/2206.07996
Build the singularity image if needed:
TMPDIR=~/tmp/singularity_tmpdir singularity build --fakeroot image.sif image.def
(You may need to follow this to be able to do this locally without root)
Create a W&B account and add the following content to your ~/.bashrc:
export WANDB_API_KEY="<YOUR_KEY>"
Populate .env file with settings from .env.example:
cp .env.example .env
edit .env
Ensure the path to the singularity image file is correct.
Scripts for running the experiments from the paper are in the scripts directory.
Main interval training logic is in intervalnet/strategy.py. Model specifics are in intervalnet/models/interval.py.