`DADA: Differentiable Automatic Data Augmentation`

Introduction

The official code for our ECCV 2020 paper DADA: Differentiable Automatic Data Augmentation, which is at least one order of magnitude faster than the state-of-the-art data augmentation (DA) policy search algorithms while achieving very comparable accuracy. The implementation of our training part is based on fast-autoaugment.

License

The project is only free for academic research purposes, but needs authorization for commerce. For commerce permission, please contact wyt@pku.edu.cn.

Citation

If you use our code/model, please consider to cite our ECCV 2020 paper DADA: Differentiable Automatic Data Augmentation [arXiv] [ECCV].

@article{li2020dada,
  author    = {Yonggang Li and
               Guosheng Hu and
               Yongtao Wang and
               Timothy M. Hospedales and
               Neil Martin Robertson and
               Yongxin Yang},
  title     = {{DADA:} Differentiable Automatic Data Augmentation},
  booktitle = {The European Conference on Computer Vision (ECCV)},
  year      = {2020}
}

Model

We provide the checkpoints in BaiduDrive, with fetching code sgap, or GoogleDrive.

CIFAR-10

Search : 0.1 GPU Hours, WResNet-40x2 on Reduced CIFAR-10

Dataset	Model	Baseline	Cutout	AA	PBA	Fast AA	DADA
CIFAR-10	Wide-ResNet-40-2	5.3	4.1	3.7	-	3.6	3.6
CIFAR-10	Wide-ResNet-28-10	3.9	3.1	2.6	2.6	2.7	2.7
CIFAR-10	Shake-Shake(26 2x32d)	3.6	3.0	2.5	2.5	2.7	2.7
CIFAR-10	Shake-Shake(26 2x96d)	2.9	2.6	2.0	2.0	2.0	2.0
CIFAR-10	Shake-Shake(26 2x112d)	2.8	2.6	1.9	2.0	2.0	2.0
CIFAR-10	PyramidNet+ShakeDrop	2.7	2.3	1.5	1.5	1.8	1.7

CIFAR-100

Search : 0.2 GPU Hours, WResNet-40x2 on Reduced CIFAR-100

Dataset	Model	Baseline	Cutout	AA	PBA	Fast AA	DADA
CIFAR-100	Wide-ResNet-40-2	26.0	25.2	20.7	-	20.7	20.9
CIFAR-100	Wide-ResNet-28-10	18.8	18.4	17.1	16.7	17.3	17.5
CIFAR-100	Shake-Shake(26 2x96d)	17.1	16.0	14.3	15.3	14.9	15.3
CIFAR-100	PyramidNet+ShakeDrop	14.0	12.2	10.7	10.9	11.9	11.2

SVHN

Search : 0.1 GPU Hours, WResNet-28x10 on Reduced SVHN

Dataset	Model	Baseline	Cutout	AA	PBA	Fast AA	DADA
SVHN	Wide-ResNet-28-10	1.5	1.3	1.1	1.2	1.1	1.2
SVHN	Shake-Shake(26 2x96d)	1.4	1.2	1.0	1.1	-	1.1

ImageNet

Search : 1.3 GPU Hours, ResNet-50 on Reduced ImageNet

Dataset	Baseline	AA	Fast AA	OHL AA	DADA
ImageNet	23.7 / 6.9	~22.4 / 6.2	22.4 / 6.3	21.1 / 5.7	22.5 / 6.5

Installation

Environment

Ubuntu 16.04 LTS
CUDA 10.0
PyTorch 1.2.0
TorchVision 0.4.0

Install

a. Create a conda virtual environment and activate it.

conda create -n dada-env python=3.6.10
source activate dada-env # or conda activate dada-env

b. Install PyTorch and torchvision following the official instructions, e.g.,

conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit==10.0

c. Install other python package for DADA and fast-autoaugment, e.g.,

# for training and inference
pip install -r fast-autoaugment/requirements.txt

# for searching
pip install -r requirements.txt

Getting Started

Prepare Datasets

The dataset (except ImageNet) will be automatically download if you keep the default setting. You should put the data in ./data as below: (which include the datasets of CIFAR-10, CIFAR-100, SVHN, and ImageNet)

# CIFAR-10
./data/cifar-10-python.tar.gz

# CIFAR-100
./data/cifar-100-python.tar.gz

# SVHN
./data/train_32x32.mat
./data/extra_32x32.mat
./data/test_32x32.mat

# ImageNet
./data/imagenet-pytorch/
./data/imagenet-pytorch/meta.bin
./data/imagenet-pytorch/train
./data/imagenet-pytorch/val

Inference

Download the model-pth provided in , put them in ./fast-autoaugment/weights

cd fast-autoaugment
sh inference.sh

For example, you can test the provided wresnet40x2 model trained on CIFAR-10 as below:

# TITAN
GPUS=0
SEED=0
DATASET=cifar10
CONF=confs/wresnet40x2_cifar10_b512_test.yaml
GENOTYPE=CIFAR10
SAVE=weights/`basename ${CONF} .yaml`_${GENOTYPE}_${DATASET}_${SEED}/test.pth
CUDA_VISIBLE_DEVICES=${GPUS} python FastAutoAugment/train.py -c ${CONF} --dataset ${DATASET} --genotype ${GENOTYPE} --save ${SAVE} --seed ${SEED} --only-eval --batch 32

Train

The training script is provided, including most experiments of our paper.

cd fast-autoaugment
sh train.sh

For example, you can train a wresnet40x2 model on CIFAR-10 as below:

# TITAN
GPUS=0
SEED=0
DATASET=cifar10
CONF=confs/wresnet40x2_cifar10_b512_test.yaml
GENOTYPE=CIFAR10
SAVE=weights/`basename ${CONF} .yaml`_${GENOTYPE}_${DATASET}_${SEED}/test.pth
CUDA_VISIBLE_DEVICES=${GPUS} python FastAutoAugment/train.py -c ${CONF} --dataset ${DATASET} --genotype ${GENOTYPE} --save ${SAVE} --seed ${SEED}

Search

The searching script is provided, including CIFAR10, CIFAR100, SVHN, and ImageNet.

cd search_relax
sh train_paper.sh

For example, you can search a DA policy on the reduced-cifar10 dataset with wresnet40-2 model as below:

# you can change the hyper-parameters as below:
GPU=0
DATASET=reduced_cifar10
MODEL=wresnet40_2
EPOCH=20
BATCH=128
LR=0.1
WD=0.0002
AWD=0.0
ALR=0.005
CUTOUT=16
TEMPERATE=0.5
SAVE=CIFAR10
python train_search_paper.py --unrolled --report_freq 1 --num_workers 0 --epoch ${EPOCH} --batch_size ${BATCH} --learning_rate ${LR} --dataset ${DATASET} --model_name ${MODEL} --save ${SAVE} --gpu ${GPU} --arch_weight_decay ${AWD} --arch_learning_rate ${ALR} --weight_decay ${WD} --cutout --cutout_length ${CUTOUT} --temperature ${TEMPERATE}

The code for DADA with gumbel softmax is also included in this repository.

cd search_gumbel
sh train_paper.sh

Found Policy

We relase the found Data Augmentation policies in CIFAR-10, CIFAR-100, SVHN, and ImageNet by our DADA as below. The origin DA policies have been included in the fast-autoaugment/FastAutoAugment/genotype.py. You can find the genotype used by our paper as below:

vim fast-autoaugment/FastAutoAugment/genotype.py

CIFAR10

Sub-policy	Opeartion 1	Opeartion 2
sub-policy 0	(TranslateX, 0.52, 0.58)	(Rotate, 0.57, 0.53)
sub-policy 1	(ShearX, 0.50, 0.46)	(Sharpness, 0.50, 0.54)
sub-policy 2	(Brightness, 0.56, 0.56)	(Sharpness, 0.52, 0.47)
sub-policy 3	(ShearY, 0.62, 0.48)	(Brightness, 0.47, 0.46)
sub-policy 4	(ShearX, 0.44, 0.58)	(TranslateY, 0.40, 0.51)
sub-policy 5	(Rotate, 0.40, 0.52)	(Equalize, 0.38, 0.36)
sub-policy 6	(AutoContrast, 0.44, 0.48)	(Cutout, 0.49, 0.50)
sub-policy 7	(AutoContrast, 0.56, 0.48)	(Color, 0.45, 0.61)
sub-policy 8	(Rotate, 0.42, 0.64)	(AutoContrast, 0.60, 0.58)
sub-policy 9	(Invert, 0.40, 0.50)	(Color, 0.50, 0.44)
sub-policy 10	(Posterize, 0.56, 0.50)	(Brightness, 0.53, 0.48)
sub-policy 11	(TranslateY, 0.42, 0.51)	(AutoContrast, 0.38, 0.57)
sub-policy 12	(ShearX, 0.38, 0.50)	(Contrast, 0.49, 0.52)
sub-policy 13	(ShearY, 0.54, 0.60)	(Rotate, 0.31, 0.56)
sub-policy 14	(Posterize, 0.42, 0.50)	(Color, 0.45, 0.56)
sub-policy 15	(TranslateX, 0.41, 0.45)	(TranslateY, 0.36, 0.48)
sub-policy 16	(TranslateX, 0.57, 0.50)	(Brightness, 0.54, 0.48)
sub-policy 17	(TranslateX, 0.53, 0.51)	(Cutout, 0.69, 0.49)
sub-policy 18	(ShearX, 0.46, 0.44)	(Invert, 0.42, 0.40)
sub-policy 19	(Rotate, 0.50, 0.42)	(Contrast, 0.49, 0.42)
sub-policy 20	(Rotate, 0.43, 0.47)	(Solarize, 0.50, 0.42)
sub-policy 21	(TranslateY, 0.74, 0.51)	(Color, 0.39, 0.57)
sub-policy 22	(Equalize, 0.42, 0.53)	(Sharpness, 0.40, 0.43)
sub-policy 23	(Solarize, 0.73, 0.42)	(Cutout, 0.51, 0.46)
sub-policy 24	(ShearX, 0.58, 0.56)	(TranslateX, 0.48, 0.49)

CIFAR-100

Sub-policy	Opeartion 1	Opeartion 2
sub-policy 0	(ShearY, 0.56, 0.28)	(Sharpness, 0.49, 0.22)
sub-policy 1	(Rotate, 0.36, 0.19)	(Contrast, 0.56, 0.31)
sub-policy 2	(TranslateY, 0.00, 0.41)	(Brightness, 0.47, 0.52)
sub-policy 3	(AutoContrast, 0.80, 0.44)	(Color, 0.44, 0.37)
sub-policy 4	(Color, 0.94, 0.25)	(Brightness, 0.68, 0.45)
sub-policy 5	(TranslateY, 0.63, 0.40)	(Equalize, 0.82, 0.30)
sub-policy 6	(Equalize, 0.46, 0.71)	(Posterize, 0.50, 0.72)
sub-policy 7	(Color, 0.52, 0.48)	(Sharpness, 0.19, 0.40)
sub-policy 8	(Sharpness, 0.42, 0.38)	(Cutout, 0.55, 0.24)
sub-policy 9	(ShearX, 0.74, 0.56)	(TranslateX, 0.48, 0.67)
sub-policy 10	(Invert, 0.36, 0.59)	(Brightness, 0.50, 0.23)
sub-policy 11	(TranslateX, 0.36, 0.36)	(Posterize, 0.80, 0.32)
sub-policy 12	(TranslateX, 0.48, 0.36)	(Cutout, 0.64, 0.67)
sub-policy 13	(Posterize, 0.31, 0.04)	(Contrast, 1.00, 0.08)
sub-policy 14	(Contrast, 0.42, 0.26)	(Cutout, 0.00, 0.44)
sub-policy 15	(Equalize, 0.16, 0.69)	(Brightness, 0.73, 0.18)
sub-policy 16	(Contrast, 0.45, 0.34)	(Sharpness, 0.59, 0.28)
sub-policy 17	(TranslateX, 0.13, 0.54)	(Invert, 0.33, 0.48)
sub-policy 18	(Rotate, 0.50, 0.58)	(Posterize, 1.00, 0.74)
sub-policy 19	(TranslateX, 0.51, 0.43)	(Rotate, 0.46, 0.48)
sub-policy 20	(ShearX, 0.58, 0.46)	(TranslateY, 0.33, 0.31)
sub-policy 21	(Rotate, 1.00, 0.00)	(Equalize, 0.51, 0.37)
sub-policy 22	(AutoContrast, 0.26, 0.57)	(Cutout, 0.34, 0.35)
sub-policy 23	(ShearX, 0.56, 0.55)	(Color, 0.50, 0.50)
sub-policy 24	(ShearY, 0.46, 0.09)	(Posterize, 0.55, 0.34)

SVHN

Sub-policy	Opeartion 1	Opeartion 2
sub-policy 0	(Solarize, 0.61, 0.53)	(Brightness, 0.64, 0.50)
sub-policy 1	(ShearY, 0.56, 0.54)	(Sharpness, 0.67, 0.50)
sub-policy 2	(AutoContrast, 0.64, 0.50)	(Posterize, 0.49, 0.42)
sub-policy 3	(Invert, 0.43, 0.62)	(Equalize, 0.30, 0.53)
sub-policy 4	(Contrast, 0.49, 0.55)	(Color, 0.51, 0.58)
sub-policy 5	(ShearX, 0.58, 0.50)	(Brightness, 0.56, 0.54)
sub-policy 6	(Rotate, 0.43, 0.50)	(Contrast, 0.47, 0.42)
sub-policy 7	(Brightness, 0.51, 0.57)	(Cutout, 0.48, 0.50)
sub-policy 8	(TranslateY, 0.65, 0.46)	(Rotate, 0.43, 0.46)
sub-policy 9	(ShearY, 0.41, 0.43)	(Contrast, 0.48, 0.49)
sub-policy 10	(ShearY, 0.52, 0.37)	(Brightness, 0.43, 0.37)
sub-policy 11	(ShearY, 0.26, 0.49)	(Posterize, 0.52, 0.56)
sub-policy 12	(TranslateX, 0.67, 0.38)	(TranslateY, 0.45, 0.42)
sub-policy 13	(Posterize, 0.64, 0.43)	(Sharpness, 0.63, 0.54)
sub-policy 14	(Rotate, 0.47, 0.50)	(Sharpness, 0.40, 0.45)
sub-policy 15	(ShearX, 0.47, 0.46)	(Cutout, 0.58, 0.50)
sub-policy 16	(Rotate, 0.58, 0.53)	(Solarize, 0.41, 0.43)
sub-policy 17	(Color, 0.37, 0.44)	(Brightness, 0.52, 0.41)
sub-policy 18	(TranslateX, 0.49, 0.47)	(Posterize, 0.49, 0.52)
sub-policy 19	(TranslateY, 0.50, 0.49)	(Solarize, 0.50, 0.42)
sub-policy 20	(TranslateY, 0.27, 0.50)	(Invert, 0.56, 0.29)
sub-policy 21	(ShearY, 0.64, 0.57)	(Rotate, 0.49, 0.57)
sub-policy 22	(Invert, 0.49, 0.55)	(Contrast, 0.41, 0.50)
sub-policy 23	(ShearX, 0.57, 0.49)	(Color, 0.60, 0.50)
sub-policy 24	(Rotate, 0.54, 0.53)	(Equalize, 0.52, 0.50)

ImageNet

Sub-policy	Opeartion 1	Opeartion 2
sub-policy 0	(TranslateY, 0.85, 0.64)	(Contrast, 0.70, 0.47)
sub-policy 1	(ShearX, 0.69, 0.64)	(Brightness, 0.58, 0.46)
sub-policy 2	(Solarize, 0.33, 0.53)	(Contrast, 0.36, 0.40)
sub-policy 3	(ShearY, 0.54, 0.81)	(Color, 0.65, 0.67)
sub-policy 4	(Rotate, 0.52, 0.28)	(Invert, 0.55, 0.46)
sub-policy 5	(ShearY, 0.76, 0.55)	(AutoContrast, 0.64, 0.46)
sub-policy 6	(TranslateX, 0.32, 0.67)	(Sharpness, 0.45, 0.61)
sub-policy 7	(Brightness, 0.28, 0.54)	(Cutout, 0.29, 0.53)
sub-policy 8	(TranslateY, 0.26, 0.39)	(Brightness, 0.30, 0.57)
sub-policy 9	(ShearX, 0.46, 0.62)	(Rotate, 0.51, 0.59)
sub-policy 10	(TranslateY, 0.63, 0.38)	(Invert, 0.40, 0.33)
sub-policy 11	(TranslateY, 0.49, 0.32)	(Equalize, 0.43, 0.26)
sub-policy 12	(TranslateX, 0.31, 0.46)	(AutoContrast, 0.40, 0.00)
sub-policy 13	(ShearY, 0.57, 0.35)	(Equalize, 0.45, 0.16)
sub-policy 14	(Solarize, 0.78, 0.61)	(Brightness, 0.57, 0.80)
sub-policy 15	(Color, 0.75, 0.40)	(Cutout, 0.54, 0.47)
sub-policy 16	(ShearX, 0.51, 0.67)	(Cutout, 0.37, 0.45)
sub-policy 17	(TranslateX, 0.68, 0.39)	(Rotate, 0.47, 0.16)
sub-policy 18	(Rotate, 0.64, 0.55)	(Sharpness, 0.66, 0.80)
sub-policy 19	(TranslateY, 0.47, 0.75)	(Sharpness, 0.64, 0.52)
sub-policy 20	(AutoContrast, 0.29, 0.54)	(Posterize, 0.35, 0.70)
sub-policy 21	(Invert, 0.55, 0.49)	(Equalize, 0.44, 0.76)
sub-policy 22	(TranslateX, 0.86, 0.29)	(Contrast, 0.41, 0.60)
sub-policy 23	(Invert, 0.28, 0.45)	(Posterize, 0.42, 0.34)
sub-policy 24	(Posterize, 0.15, 0.33)	(Color, 0.50, 0.59)

yuzhang49/DADA

DADA: Differentiable Automatic Data Augmentation

Introduction

License

Citation

Model

CIFAR-10

CIFAR-100

SVHN

ImageNet

Installation

Environment

Install

Getting Started

Prepare Datasets

Inference

Train

Search

Found Policy

CIFAR10

CIFAR-100

SVHN

ImageNet

`DADA: Differentiable Automatic Data Augmentation`