RemOve-And-Retrain is Improper: Data Processing Inequality Perspective

Installation

You have to install MMPretrain with MIM.

pip install openmim==0.3.7
mim install mmpretrain==1.0.0rc7
pip install -e .

CIFAR 10

Data preparation

You have to convert CIFAR10 to CustomDataset for retraining.

python tools/dataset_converters/cifar2custom.py -o data/cifar10

Train

mim train mmpretrain configs/train/resnet18_8xb16_cifar10.py \
    --work-dir cifar10 --gpus 1

Estimate a feature importance

# for train dataset
mim test mmpretrain configs/test/resnet18_8xb16_cifar10.py \
    --checkpoint cifar10/epoch_100.pth \
    --work-dir cifar10/train --gpus 1 --cfg-options test_dataloader.dataset.test_mode=False
# for test dataset
mim test mmpretrain configs/test/resnet18_8xb16_cifar10.py \
    --checkpoint cifar10/epoch_100.pth \
    --work-dir cifar10/test --gpus 1

Retrain

mim gridsearch mmpretrain configs/retrain/resnet18_8xb16_cifar10.py \
    --work-dir cifar10 --gpus 1 \
    --cfg-options load_from=cifar10/epoch_100.pth \
    --search-args '--train_dataloader.dataset.pipeline.1.attr grad gi ig sg vg gc sobl rand
        --train_dataloader.dataset.pipeline.1.ratio 10 30 50 70 90
        --train_dataloader.dataset.pipeline.1.filter none maximum gaussian'

ROAD

mim test mmpretrain configs/debias/resnet18_8xb16_cifar10.py \
    --checkpoint cifar10/epoch_100.pth \
    --work-dir cifar10 --gpus 1 \
    --cfg-options test_dataloader.dataset.pipeline.1.attr=grad \
        test_dataloader.dataset.pipeline.1.ratio=10 \
        test_dataloader.dataset.pipeline.1.filter=none

SVHN

Data preparation

You have to convert SVHN to CustomDataset for retraining.

python tools/dataset_converters/svhn2custom.py -o data/svhn

Train

mim train mmpretrain configs/train/resnet18_8xb16_svhn.py \
    --work-dir svhn --gpus 1

Estimate a feature importance

# for train dataset
mim test mmpretrain configs/test/resnet18_8xb16_svhn.py \
    --checkpoint svhn/epoch_100.pth \
    --work-dir svhn/train --gpus 1 --cfg-options test_dataloader.dataset.test_mode=False
# for test dataset
mim test mmpretrain configs/test/resnet18_8xb16_svhn.py \
    --checkpoint svhn/epoch_100.pth \
    --work-dir svhn/test --gpus 1

Retrain

mim gridsearch mmpretrain configs/retrain/resnet18_8xb16_svhn.py \
    --work-dir svhn --gpus 1 \
    --cfg-options load_from=svhn/epoch_100.pth \
    --search-args '--train_dataloader.dataset.pipeline.1.attr grad gi ig sg vg gc sobl rand
        --train_dataloader.dataset.pipeline.1.ratio 10 30 50 70 90
        --train_dataloader.dataset.pipeline.1.filter none maximum gaussian'

ROAD

mim test mmpretrain configs/debias/resnet18_8xb16_svhn.py \
    --checkpoint svhn/epoch_100.pth \
    --work-dir svhn --gpus 1 \
    --cfg-options test_dataloader.dataset.pipeline.1.attr=grad \
        test_dataloader.dataset.pipeline.1.ratio=10 \
        test_dataloader.dataset.pipeline.1.filter=none

CUB-200

Data preparation

wget https://data.caltech.edu/records/65de6-vp158/files/CUB_200_2011.tgz
tar -xvf CUB_200_2011.tgz
mkdir data
ln -s CUB_200_2011 data

Train (optional)

mim download mmpretrain --config resnet50_8xb8_cub --dest cub
mim train mmpretrain cub/resnet50_8xb8_cub.py \
    --work-dir cub --gpus 1

Estimate a feature importance

mim download mmpretrain --config resnet50_8xb8_cub --dest cub
mim test mmpretrain test/resnet50_8xb8_cub.py \
    --checkpoint cub/resnet50_8xb8_cub_20220307-57840e60.pth \
    --work-dir cub --gpus 1

Retrain

mim gridsearch mmpretrain configs/retrain/resnet50_8xb8_cub.py \
    --work-dir cub --gpus 1 \
    --cfg-options load_from=cub/resnet50_8xb8_cub_20220307-57840e60.pth \
    --search-args '--train_dataloader.dataset.pipeline.1.attr grad gi ig sg vg gc sobl rand
        --train_dataloader.dataset.pipeline.1.ratio 10 30 50 70 90
        --train_dataloader.dataset.pipeline.1.filter none maximum gaussian'

ROAD

mim test mmpretrain configs/debias/resnet50_8xb8_cub.py \
    --checkpoint cub/resnet50_8xb8_cub_20220307-57840e60.pth \
    --work-dir cub --gpus 1 \
    --cfg-options test_dataloader.dataset.pipeline.1.attr=grad \
        test_dataloader.dataset.pipeline.1.ratio=10 \
        test_dataloader.dataset.pipeline.1.filter=none