weakly supervised learning baseline

Methods

semi-supervised

see semi-supervised reimplementations document

• MixMatch
• ReMixMatch
• FixMatch
• CoMatch
• FlexMatch

self-supervised(with linear probe)

• SimCLR
• MoCo
• BYOL

unsupervised

noisy label

• L2R
• IEG
• ULN
• Co-teaching
• DivideMix

supervised(baseline)

• baseline
• baseline with randaugment
• baseline with {mixup, cutout}

Run

python train_{}.py --module={} {args}

such as

python train_sup.py --module=basic --device=0 --model=wrn282

Components

augmentation methods

• cutout
• randaugment
• mixup

datasets

• cifar
  • cifar10
  • cifar100
• svhn
• stl10
• clothing1m
• tiny-imagenet
• imagenet

modules

resnet

• resnet-{18, 34, 50}
• wideresnet-{282}

transformer

• vit

Reference