Official Pytorch Implementation of ReBac

Preparation

Requirements

• CUDA>=11.8
• torch>=2.0.0
• torchvision>=0.15.0
• numpy
• pillow
• scikit-learn
• tqdm
• matplotlib

Datasets

We use the Pascal VOC 2012 and ADE20K datasets for evaluation following the previous methods. You can download the datasets from the following links:

Download Pascal VOC 2012 dataset:

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar

Download Additional Segmentation Class Annotations:

wget https://www.dropbox.com/s/oeu149j8qtbs1x0/SegmentationClassAug.zip

Download ADE20K dataset:

wget http://data.csail.mit.edu/places/ADEchallenge/ADEChallengeData2016.zip

data_root/
   ├── VOC2012/
   │   ├── Annotations/
   │   ├── ImageSet/
   │   ├── JPEGImages/
   │   ├── SegmentationClassAug/
   │   └── saliency_map/
   └── ADEChallengeData2016
       ├── annotations
       │   ├── training
       │   └── validation
       └── images
           ├── training
           └── validation

Getting Started

Class-Incremental Segmentation Segmentation on VOC 2012

Run our scripts run_init.sh and run.sh for class-incremental segmentation on VOC 2012 dataset, or follow the instructions below.

Initial step:

MODEL=deeplabv3bga_resnet101
DATA_ROOT= # Your dataset root path
DATASET=voc
TASK=15-1
EPOCH=50
BATCH=16
LOSS=bce_loss
LR=0.001
THRESH=0.7
SUBPATH=BARM
CURR=0

CUDA_VISIBLE_DEVICES=0 \
python train.py --data_root ${DATA_ROOT} --model ${MODEL} --crop_val --lr ${LR} \
    --batch_size ${BATCH} --train_epoch ${EPOCH}  --loss_type ${LOSS} \
    --dataset ${DATASET} --task ${TASK} --lr_policy poly \
    --pseudo --pseudo_thresh ${THRESH}  --bn_freeze  --amp \
    --curr_step ${CURR} --subpath ${SUBPATH} --initial --overlap 

Incremental steps:

MODEL=deeplabv3bga_resnet101
DATA_ROOT= # Your dataset root path
DATASET=voc
TASK=15-1
EPOCH=20
BATCH=8
LOSS=bce_loss
LR=0.001
THRESH=0.7
SUBPATH=BARM
CURR=1

CUDA_VISIBLE_DEVICES=0,1 \
torchrun --nproc_per_node=2 --master_port=11451 \
train.py --data_root ${DATA_ROOT} --model ${MODEL} --crop_val --lr ${LR} \
    --batch_size ${BATCH} --train_epoch ${EPOCH}  --loss_type ${LOSS} \
    --dataset ${DATASET} --task ${TASK} --lr_policy poly \
    --pseudo --pseudo_thresh ${THRESH}  --bn_freeze  --amp\
    --curr_step ${CURR} --subpath ${SUBPATH} --overlap

Class-Incremental Segmentation Segmentation on ADE20K

Run our scripts run_init.sh and run.sh for class-incremental segmentation on ADE20K dataset, or follow the instructions below.

Initial step:

MODEL=deeplabv3bga_resnet101
DATA_ROOT= # Your dataset root path
DATASET=ade
TASK=100-5
EPOCH=60
BATCH=8
LOSS=bce_loss
LR=0.001
THRESH=0.7
SUBPATH=BARM
CURR=0

CUDA_VISIBLE_DEVICES=0 \
python train.py --data_root ${DATA_ROOT} --model ${MODEL} --crop_val --lr ${LR} \
    --batch_size ${BATCH} --train_epoch ${EPOCH}  --loss_type ${LOSS} \
    --dataset ${DATASET} --task ${TASK} --lr_policy poly \
    --pseudo --pseudo_thresh ${THRESH}  --bn_freeze  --amp \
    --curr_step ${CURR} --subpath ${SUBPATH} --initial --overlap 

Incremental steps:

MODEL=deeplabv3bga_resnet101
DATA_ROOT= # Your dataset root path
DATASET=ade
TASK=100-10
EPOCH=100
BATCH=4
LOSS=bce_loss
LR=0.01
THRESH=0.7
SUBPATH=BARM
CURR=1

CUDA_VISIBLE_DEVICES=0,1 \
torchrun --nproc_per_node=2 --master_port=19198 \
train.py --data_root ${DATA_ROOT} --model ${MODEL} --crop_val --lr ${LR} \
    --batch_size ${BATCH} --train_epoch ${EPOCH}  --loss_type ${LOSS} \
    --dataset ${DATASET} --task ${TASK} --lr_policy poly \
    --pseudo --pseudo_thresh ${THRESH}  --bn_freeze  --amp \
    --curr_step ${CURR} --subpath ${SUBPATH} --overlap

Acknowledgement

Our implementation is based on these repositories: DeepLabV3Plus-Pytorch, SSUL. Thanks for their great work!