Class Incremental Learning Zoo - pytorch version

Unofficial pytorch version of reproducing the method of class incremental learning (Working) in classification.
Original code is from this repo.

Currently only CIFAR100 dataset is available.

• How to Implement (iCaRL)

    python main.py train --train-mode icarl --gpu-ids 0 --model resnet32

• How to Implement (EEIL, with BiC setting)

    python main.py train --train-mode eeil --gpu-ids 2 --task-size 10 --model resnet32 --batch-size 128 --lr 0.1 --gamma 0.1 --epochs 250 --lr-steps 100,150,200 --weight-decay 0.0002

• How to Implement (EEIL, with EEIL setting)

    python main.py train --train-mode eeil --gpu-ids 2 --task-size 10 --model resnet32 --batch-size 128 --lr 0.1 --gamma 0.1 --epochs 40 --lr-steps 10,20,30 --weight-decay 0.0001

• How to Implement (BiC)

    python main.py train --train-mode bic --gpu-ids 2 --task-size 10 --model resnet32 --batch-size 128 --lr 0.1 --gamma 0.1 --epochs 250 --lr-steps 100,150,200 --weight-decay 0.0002

• 5 task (Top-1 Accuracy)

5 task	1	2	3	4	5
iCaRL	78.35	65.25	54.15	47.94	38.7
EEIL	82.5	72.12	65.75	59.08	54.23
BiC	81.15	71.43	64.28	58.56	54.91

EEIL use BiC settings for reproducing, we provide EEIL settings also.

• 10 task

10 task	1	2	3	4	5	6	7	8	9	10
iCaRL	80.6	66.7	59.67	52.88	48.36	44.68	41.76	38.19	36.03	33.19
EEIL	87	77.7	73.7	66.5	61.66	58.37	55.27	51.21	48.59	45.47
BiC	87.7	80.4	72.17	66.45	61.92	57.78	54.61	51.4	50.28	47.78

• 20 task

20 task	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20
iCaRL	79.6	73.2	67.8	63.35	63.92	58.93	57.6	53.9	50.51	49	47.22	42.97	41.46	38.04	32.43	30.28	27.13	25.53	23.06	20
EEIL	84.4	83.3	75.4	69.65	70.36	65.6	63.86	56.8	54.29	52.3	50.42	48.32	47.95	45.39	43.12	42.81	40.58	40.46	39.65	36.12
BiC	87.4	85.4	78.47	76.4	73.84	68.87	65.54	61.68	57.67	54.8	53	50.68	51.21	49.62	48.77	45.76	45.29	43.67	41.27	39.64

Reference

• iCaRL construct exemplar set is borrowed from iCaRL-pytorch.
• repo

Contributor

• Minsoo Kang