Awesome

TriRE: A Multi-Mechanism Learning Paradigm for Continual Knowledge Retention and Promotion

The official repository for NeurIPS'23 paper. We extended the original repo DER++ with our method.

How to run TriRE?

• Use python main.py to run experiments.

• For example, for dataset Seq-CIFAR100, run
  python main.py --dataset seq-cifar100 --model trire --buffer_size 200 --load_best_args --img_size 32 --tensorboard --reservoir_buffer --kwinner_sparsity 0.3 --pruning_technique CWI --sparsity 0.2 --lr_fl 0.002 --lr_sl 0.0001 --reset_act_counters --train_budget_1 0.6 --train_budget_2 0.2 --reparameterize --reinit_technique rewind --use_cl_mask --reg_weight 0.05 --stable_model_update_freq 0.1 --rewind_tuning_incl --use_het_drop

  where, kwinner_sparsity and sparsity represents the percentage of most activated neurons and corresponding most important weights to be retained at the end of Retain stage respectively.
  pruning_technique : {'CWI', 'Magnitude Pruning', 'Fisher Information'}
  lr_fl is the learning rate for Retain and Revise stages and lr_sl is the slowed learning rate for Revise stage.
  train_budget_1 and train_budget_2 are the percentages of epochs dedicated to Retain and Revise stages which also implicates that the rest of the epochs are used for Rewind stage.
  use_cl_mask indicates that the model is using a single head classifier.
  reinit_technique : {xavier, rewind}

Setup

• Use argument --load_best_args to 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.

Datasets

Class-Il / Task-IL settings

• Seq-CIFAR10
• Seq-CIFAR100
• SeqTinyImageNet

Cite Our Work

If you find the code useful in your research, please consider citing our paper:

@article{vijayan2024trire,
  title={TriRE: A Multi-Mechanism Learning Paradigm for Continual Knowledge Retention and Promotion},
  author={Vijayan, Preetha and Bhat, Prashant and Zonooz, Bahram and Arani, Elahe},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}