Dataset-Generation

This repository provides simple PyTorch implementations for generating representation based on adversarial training and generative adversarial networks on CIFAR-10.

1. Dataset Generation Base on Adversarial Training

step 1: Training a robust Model

--attack-steps ATTACK_STEPS
                   number of steps for adversarial attack (default: 7)
--constraint {inf,2,unconstrained}
                      adv constraint (choices: {arg_type}, default:
                    required)
--eps EPS             adversarial perturbation budget (default: required)
--attack-lr ATTACK_LR
                      step size for PGD (default: required)
--use-best {0,1}      if 1 (0) use best (final) PGD step as example
                      (choices: {arg_type}, default: 1)
--random-restarts RANDOM_RESTARTS
                      number of random PGD restarts for eval (default: 0)
--custom-eps-multiplier EPS_SCHEDULE
                      epsilon multiplier sched (same format as LR schedule)

example:Training a robust ResNet-50 for the Cifar10 dataset

CUDA_VISIBLE_DEVICES=1,2,3,4 python  main_adv.py --dataset cifar --data $Cifar10_PATH --adv-train 1 --arch resnet50 \
   --out-dir ./output/adv/ --eps 0.5 --attack-lr 0.1 --attack-steps 7 --constraint 2 --exp-name train_time:1

step 2: Evaluating trained models

To evaluate a trained model, we use the --eval-only flag when calling main_adv.py. To evaluate the model for just standard (not adversarial) accuracy, only the following arguments are required:

python -m main_adv.py --dataset DATASET --data /path/to/dataset \
   --eval-only 1 --out-dir OUT_DIR --arch arch --adv-eval 0 \
   --resume PATH_TO_TRAINED_MODEL_CHECKPOINT

example：Testing the standard and adversarial accuracy of a trained CIFAR-10 model with L2 norm constraint of 0.5 and 100 L2-PGD steps:

python -m main_adv.py --dataset cifar --eval-only 1 --out-dir ./output/adv/eval_time:1/ --arch resnet50 --adv-eval 1 --constraint 2 --eps 0.5 --attack-lr 0.1 --attack-steps 100 --resume PATH_TO_TRAINED_MODEL_CHECKPOINT

step 3: Generate new dataset base on trained model

python gen_cifar10_adv.py --out-dir=./output/adv/ --cuda --dataroot=$Cifar10_PATH --resume-path PATH_TO_TRAINED_MODEL_CHECKPOINT --ngpu 4 --gen-time 1

2. Dataset Generation Base on Standard Training

step 1: Training a non-robust Model

set --adv-train option 0

example：Training a non-robust ResNet-18 for the CIFAR dataset

 python main_adv.py --dataset cifar --data  $Cifar10_PATH --adv-train 0 --arch resnet50 --out-dir ./output/standard/  --exp-name train_time:1

step 2: Evaluating trained models

similar to above

step 2: Generate new dataset base on trained model

python gen_cifar10_adv.py --out-dir=./output/standard/ --cuda --dataroot=$Cifar10_PATH --resume-path PATH_TO_TRAINED_MODEL_CHECKPOINT --ngpu 4 --gen-time 1

3. Dataset Generation Base on ACGAN

step1: Train ACGAN

(bug: 数据集大小应该是batchsize的倍数)

python main_acgan.py --out-dir=./output/acgan/ --niter=200 --batchsize=100 --cuda --dataset=cifar10 --imageSize=32 --dataroot=$Cifar10_PATH --gpu=0,1,2,3 --train_time 1 

step2: Generate dataset

python gen_cifar10_acgan.py --out-dir=./output/acgan/ --cuda --dataset=cifar10 --ngpu 4 --gen-time 1

4. Temp Result

new datase	test【resnet18】mixup
adv 【resnet50】	train_time:3
standard 【resnet50】	train_time:3
acgan	train_time:2

Reference

robustness package

clvrai/ACGAN-PyTorch

MadryLab/robust_representations