UACL-Unsupervised Adversarial Contrastive Learning

This repository is the official PyTorch implementation of "Unsupervised Adversarial Contrastive Learning" by Yanjie Xu, Hao Sun, Jin Chen, Lin Lei, Kefeng Ji and Gangyao Kuang.

Requirements

Currently, requires following packages

• python 3.6+
• torch 1.6+
• torchvision 0.7+
• CUDA 10.1+
• tqdm 4.41+
• PIL 6.2+
• kornia 0.4+

Training

To train the model(s) in the paper, run this command:

python -m  UACL_train --data='train_list.txt' --batch_size=8 --image_size=128 --class_num=10 --epochs=200 --checkpoint='checkpoints' --save_frep=50 --attack_strength=8 --attack_step_size=2 --attack_iter=10  --lr=5.0 --weight_decay=1e-6 --momentum=0.9 --mean=0.184 --std=0.119 --seed=1

To finetune the pretrained model(s) in the paper, run this command:

python -m  finetune --data_dir='MSTAR' --pretrained_model_dir='checkpoints/UACL_encoder.pkl' --checkpoint='checkpoints' --image_size=128 --batch_size=8 --epochs=100 --mean=0.184 --std=0.119