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)
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}
}- train.py : Code to train the model
- train.sh : Script to run train.py
- 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.
- test.sh : Script to run test.py
- makeStyleDistanceMatrix.py : Code to make styleDistanceMatrix10, 100
- makeStyleDistanceMatrix.sh : Script to run makeStyleDistance.sh
- models : The directory containing the pre-trained style transfer encoder and decoder networks.
Modify the contents of train.sh according to each situation and run it.
- 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
- 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
- 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
- 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
- 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
- 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
Modify the contents of test.sh according to each situation and run it.
- 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
- 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
Modify the contents of test.sh according to each situation and run it.
- 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)
- 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)
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.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent