DADA: Differentiable Automatic Data Augmentation

Contact us with [email protected], [email protected].

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 [email protected].

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

CIFAR-100

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

SVHN

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

ImageNet

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

Installation

Environment

1. Ubuntu 16.04 LTS
2. CUDA 10.0
3. PyTorch 1.2.0
4. 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

CIFAR-100

SVHN

ImageNet