UniformAugment

Unofficial PyTorch Reimplementation of UniformAugment. Most of codes are from Fast AutoAugment and PyTorch RandAugment.

Introduction

UniformAugment is an automated data augmentation approach that completely avoids a search phase. UniformAugment’s effectiveness is comparable to the known methods, while still being highly efficient by virtue of not requiring any search.

Install

pip install git+https://github.com/tgilewicz/uniformaugment/

Usage

from torchvision.transforms import transforms
from UniformAugment import UniformAugment

transform_train = transforms.Compose([
    transforms.RandomCrop(32, padding=4),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize(_CIFAR_MEAN, _CIFAR_STD),
])
# Add UniformAugment with num_ops hyperparameter (num_ops=2 is optimal)
transform_train.transforms.insert(0, UniformAugment())

Experiment

The details of the experiment were consulted with the authors of the UniformAugment paper.

You can run an example experiment with,

$ python UniformAugment/train.py -c confs/wresnet28x10_cifar.yaml --dataset cifar10 \
    --save cifar10_wres28x10.pth --dataroot ~/data --tag v1

CIFAR-10 Classification, TOP1 Accuracy

CIFAR-100 Classification, TOP1 Accuracy

ImageNet Classification

Core class

class UniformAugment:
    def __init__(self, ops_num=2):
        self._augment_list = augment_list(for_autoaug=False)
        self._ops_num = ops_num

    def __call__(self, img):
        # Selecting unique num_ops transforms for each image would help the
        #   training procedure.
        ops = random.choices(self._augment_list, k=self._ops_num)

        for op in ops:
            augment_fn, low, high = op
            probability = random.random()
            if random.random() < probability:
                img = augment_fn(img.copy(), random.uniform(low, high))

        return img

References

• UniformAugment : Paper
• Fast AutoAugment : Code Paper
• Pytorch RandAugment: Code