CLSS (Contiunal Learning Semantic Segmentation) baselines

• This is CLSS baseline implementation
• Currently, it includes MiB, PLOP, DKD.
• The main implementation is based on dataset VOC2012 and ResNet-101 backbone.
• ADE 20K dataset is not supported currently, but will be supported soon.

Environment

• CUDA 11.1
• python 3.8
• The shell code for setting environment is in scripts/env_create.sh

training

• The shell code for training is in ./z_exp_individual_cmd

    ./train_{dataset}_{scenarios}_{method}.sh
    For examples, ./train_voc_10-1_DKD.sh
    For MiB, PLOP which require 2 GPUs, please pass the argument GPU_NUMBER (ex 0,1) to .sh file

• The configuration for each method is in 'configs/config_{dataset}_{method}.yaml'
• The results are updated in wandb

configuration for each baseline

Pascal VOC 2012

coonfig	MiB	PLOP	DKD	STAR
epoch	30	30	60	60
lr	0.01 / 0.001	0.01 / 0.001	0.001/ 0.0001	0.001/ 0.0001
$\gamma$ (pos weight for BCE Loss)	UnCE	1	2 / 1	4
Optimizer	SGD (momentum 0.9, wd 1e-4, nesterov True)	SGD (momentum 0.9, wd 1e-4)	SGD (momentum 0.9)	Adam (momentum 0.9)
$\alpha,\beta$ (hyperparameter for loss)	10 (lkd)	1 (pod)	5 / 5 (kd / dkd)	5 / 0.05 (pkd/cont)
batch size	24	24	32	24
lr Schedular	PolyLR	PolyLR	PolyLR
GPUs	RTX titian x 2	? x 2	A5000 x 4	RTX 3090 x 2
augmentation	same as [1]

ADE 20K

config MiB PLOP DKD STAR epoch 60 60 100 100 lr 0.01 / 0.001 0.01 / 0.001 0.0025 / 0.00025 0.00025 / 0.000025 $\gamma$ (pos weight for BCE Loss) UnCE 1 35 30 Optimizer SGD (momentum 0.9, wd 1e-4) SGD (momentum 0.9, wd 1e-4) SGD (momentum 0.9) Adam (momentum 0.9) $\alpha,\beta$ (hyperparameter for loss) 10 (lkd) 1 (pod) 5 / 5 (kd / dkd) 5 / 0.05 (pkd/cont) batch size 24 24 24 24 lr Schedular PolyLR PolyLR PolyLR + linear warm up GPUs RTX titian x 2 ? x 2 A5000 x 4 RTX 3090 x 2 augmentation same as [1]

Acknowledgements

• This code is based on DKD (https://github.com/cvlab-yonsei/DKD#decomposed-knowledge-distillation-for-class-incremental-semantic-segmentation) codespaces.
• All implementations have been borrowed from existing code implementations (MiB, DKD, PLOP)