Awesome

StyleMix: Separating Content and Style for Enhanced Data Augmentation (CVPR 2021)

This repository contains the official PyTorch implementation for our CVPR 2021 paper.

• Minui Hong*, Jinwoo Choi* and Gunhee Kim. StyleMix: Separating Content and Style for Enhanced Data Augmentation. In CVPR, 2021. (* equal contribution)

[paper][supp]

Reference

If you cite this paper, please refer to the following:

@InProceedings{hong2021stylemix,
    author    = {Minui Hong and Jinwoo Choi and Gunhee Kim},
    title     = {StyleMix: Separating Content and Style for Enhanced Data Augmentation},
    booktitle = {CVPR},
    year      = {2021}
}

Usage

1. train.py : Code to train the model
2. train.sh : Script to run train.py
3. test.py : Code to check CIFAR-100 and CIFAR-10 classification performance (if fgsm option is set to False) or experiment with FGSM Attack (if fgsm option is set to True) in Pyramid200 model.
4. test.sh : Script to run test.py
5. makeStyleDistanceMatrix.py : Code to make styleDistanceMatrix10, 100
6. makeStyleDistanceMatrix.sh : Script to run makeStyleDistance.sh
7. models : The directory containing the pre-trained style transfer encoder and decoder networks.

Train Model

Modify the contents of train.sh according to each situation and run it.

1. StyleCutMix_Auto_Gamma + CIFAR-100

python train.py \
    --net_type pyramidnet \
    --dataset cifar100 \
    --depth 200 \
    --alpha 240 \
    --batch_size 64 \
    --lr 0.25 \
    --expname PyraNet200 \
    --epochs 300 \
    --prob 0.5 \
    --r 0.7 \
    --delta 3.0 \
    --method StyleCutMix_Auto_Gamma \
    --save_dir /set/your/save/dir \
    --data_dir /set/your/data/dir \
    --no-verbose

2. StyleCutMix_Auto_Gamma + CIFAR-10

python train.py \
    --net_type pyramidnet \
    --dataset cifar10 \
    --depth 200 \
    --alpha 240 \
    --batch_size 64 \
    --lr 0.25 \
    --expname PyraNet200 \
    --epochs 300 \
    --prob 0.5 \
    --r 0.7 \
    --delta 1.0 \
    --method StyleCutMix_Auto_Gamma \
    --save_dir /set/your/save/dir \
    --data_dir /set/your/data/dir \
    --no-verbose

3. StyleCutMix + CIFAR-100

python train.py \
    --net_type pyramidnet \
    --dataset cifar100 \
    --depth 200 \
    --alpha 240 \
    --batch_size 64 \
    --lr 0.25 \
    --expname PyraNet200 \
    --epochs 300 \
    --prob 0.5 \
    --r 0.7 \
    --alpha2 0.8 \
    --method StyleCutMix \
    --save_dir /set/your/save/dir \
    --data_dir /set/your/data/dir \
    --no-verbose

4. StyleCutMix + CIFAR-10

python train.py \
    --net_type pyramidnet \
    --dataset cifar10 \
    --depth 200 \
    --alpha 240 \
    --batch_size 64 \
    --lr 0.25 \
    --expname PyraNet200 \
    --epochs 300 \
    --prob 0.5 \
    --r 0.7 \
    --alpha2 0.8 \
    --method StyleCutMix \
    --save_dir /set/your/save/dir \
    --data_dir /set/your/data/dir \
    --no-verbose

5. StyleMix + CIFAR-100

python train.py \
    --net_type pyramidnet \
    --dataset cifar100 \
    --depth 200 \
    --alpha 240 \
    --batch_size 64 \
    --lr 0.25 \
    --expname PyraNet200 \
    --epochs 300 \
    --alpha1 0.5 \
    --prob 0.2 \
    --r 0.7 \
    --method StyleMix \
    --save_dir /set/your/save/dir \
    --data_dir /set/your/data/dir \
    --no-verbose

6. StyleMix + CIFAR-10

python train.py \
    --net_type pyramidnet \
    --dataset cifar10 \
    --depth 200 \
    --alpha 240 \
    --batch_size 64 \
    --lr 0.25 \
    --expname PyraNet200 \
    --epochs 300 \
    --alpha1 0.5 \
    --prob 0.2 \
    --r 0.7 \
    --method StyleMix \
    --save_dir /set/your/save/dir \
    --data_dir /set/your/data/dir \
    --no-verbose

Test classification performance

Modify the contents of test.sh according to each situation and run it.

1. CIFAR-10

test.sh :
  python test.py \
    --net_type pyramidnet \
    --dataset cifar10 \
    --batch_size 128 \
    --depth 200 \
    --alpha 240 \
    --fgsm False \
    --data_dir /set/your/data/dir \
    --pretrained /set/pretrained/model/dir

2. CIFAR-100

test.sh :
  python test.py \
    --net_type pyramidnet \
    --dataset cifar100 \
    --batch_size 128 \
    --depth 200 \
    --alpha 240 \
    --fgsm False \
    --data_dir /set/your/data/dir \
    --pretrained /set/pretrained/model/dir

Test FGSM Attack

Modify the contents of test.sh according to each situation and run it.

1. FGSM Attack on CIFAR-10

test.sh :
  python test.py \
    --net_type pyramidnet \
    --dataset cifar10 \
    --batch_size 128 \
    --depth 200 \
    --alpha 240 \
    --fgsm True \
    --eps 1 \
    --data_dir /set/your/data/dir \
    --pretrained /set/pretrained/model/dir

(You can change eps to 1, 2, 4)

2. FGSM Attack on CIFAR-100

test.sh :
  python test.py \
    --net_type pyramidnet \
    --dataset cifar100 \
    --batch_size 128 \
    --depth 200 \
    --alpha 240 \
    --fgsm True \
    --eps 1 \
    --data_dir /set/your/data/dir \
    --pretrained /set/pretrained/model/dir

(You can change eps to 1, 2, 4)

Acknowledgments

This code is based on the implementations for CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features, Puzzle Mix: Exploiting Saliency and Local Statistics for Optimal Mixup, and Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization.