RUA

Automating Augmentation Through Random Unidimensional Search.

Pre-requisites

• TensorFlow == 2.4.1
• PyTorch == 1.7.1
• FastEstimator == 1.2.0

Run PyramidNet on Cifar10:

• RUA search:

cd pyramidnet_cifar10/rua
fastestimator run pyramidnet_cifar10_rua.py

• After finding optimal augmentation level:

cd pyramidnet_cifar10/final
fastestimator train pyramidnet_cifar10_final.py

Run WRN-28-10 on Cifar10:

• RUA search:

cd wrn2810_cifar10/rua
fastestimator run wrn2810_cifar10_rua.py

• After finding optimal augmentation level:

cd wrn2810_cifar10/final
fastestimator train wrn2810_cifar10_final.py

Run WRN-28-10 on Cifar100:

• RUA search:

cd wrn2810_cifar100/rua
fastestimator run wrn2810_cifar100_rua.py

• After finding optimal augmentation level:

cd wrn2810_cifar100/final
fastestimator train wrn2810_cifar100_final.py

Run WRN-28-2 on SVHN_Cropped:

• RUA search:

cd wrn282_svhn/rua
fastestimator run wrn282_svhn_rua.py

• After finding optimal augmentation level:

cd  wrn282_svhn/final
fastestimator train wrn282_svhn_final.py

Run Resnet50 on ImageNet:

First please download the ImageNet dataset here. Then organize your folder like this:

- /data/imagenet/train
    |- class1
        |- image1.png
        |- image2.png
        |- ...
    |- ...
    |- class1000

- /data/imagenet/val
    |- class1
        |- image1.png
        |- image2.png
        |- ...
    |- ...
    |- class1000

• RUA search:

cd resnet50_imagenet/rua
fastestimator run resnet50_imagenet_rua.py --data_dir /data/imagenet

• After finding optimal augmentation level:

cd  wrn282_svhn/final
fastestimator train wrn282_svhn_final.py --data_dir /data/imagenet