Awesome
AGILE - Mitigating Interference in Incremental Learning through Attention-Guided Rehearsal
The official repository for CoLLAs'24 paper. We extended the original repo DER++ with our method.
<img width="878" alt="Screenshot 2024-05-27 at 13 30 25" src="https://github.com/NeurAI-Lab/AGILE/assets/27284368/46109324-6ed1-48ef-a87b-a7d37d0bec2c">How to run?
- python main.py --seed 10 --dataset seq-cifar10 --model agile --buffer_size 200 --load_best_args
--tensorboard --notes 'AGILE baseline'
Setup
- Use
./utils/main.pyto run experiments. - Use argument
--load_best_argsto use the best hyperparameters from the paper. - New models can be added to the
models/folder. - New datasets can be added to the
datasets/folder.
Models
- Attention-Guided Incremental Learning (AGILE)
Datasets
Class-Il / Task-IL settings
+ Sequential CIFAR-10
+ Sequential CIFAR-100
+ Sequential Tiny ImageNet