Learning a Unified Classifier Incrementally via Rebalancing

Unofficial Implementation of paper Learning a Unified Classifier Incrementally via Rebalancing paper

Requirements

python3 -m venv lucir-env

source lucir-env/bin/activate

pip install requirements.txt

Running

python main.py --args

Hyper-parameters

--dataset Dataset [CIFAR100, IMAGENET]

--start Number of classes of first task

--increment Number of classes at each next task

--rehearsal Number of example stored per each class

--selection Selection of exemplar [Herding, Random, Closest to Mean]

--exR if True, Exemplar are stored

--class_balance_finetuning if True, a class balance fine-tuning is performed at the end of each task

--less_forg if True, less-forget constraint is used

--lambda_base weight factor of less-forget loss

--ranking if True, margin ranking loss constraint is used

Comparison with original results

CIFAR-100

Starting Classes	Increment	Average Incremental Accuracy
50	50	69.81	Here
50	25	66.76	Here
50	10	63.42	Here
50	5	60.18	Here

IMAGENET-100