chihhuiho / CLAE

Implementation of Contrastive Learning with Adversarial Examples

CLAE

Code for contrastive learning with adversarial examples

Usage

Preprocess

1. Clone the project to directory

git clone https://github.com/chihhuiho/CLAE.git

2. Initiate the conda environment

cd CLAE
conda env create -f environment.yml -n CLAE
conda activate CLAE

3. Download the tinyImagenet dataset.

cd datasets
sh download_tinyImagenet.sh

Run Plain contrastive learning methods

1. Enter to Plain folder

cd Plain

2. Run contrastive learning baseline (use cifar100 [cifar10, tinyImagenet] for example)

python main.py --dataset cifar100 --batch-size 128 --gpu 0 --trial 1 
python eval.py --dataset cifar100 --batch-size 128 --gpu 0 --trial 1 

3. Run contrastive learning with adversarial training (use cifar100 [cifar10, tinyImagenet] for example)

python main.py --dataset cifar100 --batch-size 128 --gpu 0 --trial 1 --adv --bn_adv_momentum 0.01 --eps 0.03 
python eval.py --dataset cifar100 --batch-size 128 --gpu 0 --trial 1 --adv --bn_adv_momentum 0.01 --eps 0.03 

Run UEL contrastive learning methods

1. Enter to UEL folder

cd UEL

2. Run contrastive learning baseline (use cifar100 [cifar10, tinyImagenet] for example)

python main.py --dataset cifar100 --batch-size 128 --gpu 0 --trial 1 
python eval.py --dataset cifar100 --batch-size 128 --gpu 0 --trial 1 

3. Run contrastive learning with adversarial training (use cifar100 [cifar10, tinyImagenet] for example)

python main.py --dataset cifar100 --batch-size 128 --gpu 0 --trial 1 --adv --bn_adv_momentum 0.01 --eps 0.03 
python eval.py --dataset cifar100 --batch-size 128 --gpu 0 --trial 1 --adv --bn_adv_momentum 0.01 --eps 0.03 

Run SimCLR contrastive learning methods

1. Enter to SimCLR folder

cd SimCLR

2. Run contrastive learning baseline (use cifar100 [cifar10, tinyImagenet] for example)

python main.py  --trial 1 --gpu 0  --dataset CIFAR100 
python eval_lr.py  --trial 1  --gpu 0   --dataset CIFAR100 
python eval_knn.py  --trial 1  --gpu 0  --dataset CIFAR100

3. Run contrastive learning with adversarial training (use cifar100 [cifar10, tinyImagenet] for example)

python main.py --alpha 1.0 --trial 1 --gpu 0 --adv  --eps 0.03 --bn_adv_momentum 0.01 --dataset CIFAR100  
python eval_lr.py --alpha 1.0 --trial 1 --adv --gpu 0  --eps 0.03 --bn_adv_momentum 0.01 --dataset CIFAR100 
python eval_knn.py --alpha 1.0 --trial 1 --adv --gpu 0  --eps 0.03 --bn_adv_momentum 0.01 --dataset CIFAR100