salemohamedo/tasksim

Continual Learning Task Similarity Experiments

1. Create virtual env
2. Install requirements.txt
3. Add tiny dataset to data/ folder. You can download the dataset here. The other datasets will download automatically when invoked.
4. Run!

Example: python main.py --model resnet50 --n-classes-per-task 2 --n-tasks 4 --dataset cifar-100 --metrics

The --metrics command will evaluate all metrics implemented in metrics.py. However, task2vec is implemented elsewhere and needs to be invoked with --task2vec.

See util/tasksim_args.py for available CLI args. Runs can also be invoked with wandb sweeps, see sweeps/pretrained/class_inc_all_pretrained.yml for an example.