Unofficial RandAugment[1] implementation for image and keypoint augmentation.

Why RandAugment? RandAugment achieves state-of-the-art performance with a greatly reduced augmentation parameter search space. It's received additional attention due to its use in self-supervised learning models such as Unsupervised Data Augmentation(Xie et al., 2020)[3] and FixMatch(Sohn et al., 2020)[2].

Why this package? This package makes three contributions:

1. Provides an easily extendable framework in order to explore different magnitudes, policies, and augmentation operations.
2. Abstracts an "Augmentation Plan" that consists of magnitude, operations, and directions which can then be applied to both images and keypoints.
3. Enables inverse augmentations for keypoints, which moves them back to their original position.

Support for Consistency Training: Consistency Training has been used to attain state-of-the-art results on image classification problems [2, 3]. One challenge in adapting the classification technique proposed in [1] to handle keypoints is that the augmentations cause the keypoints to become misaligned. In order to deal with this, we implement a RandAugment.apply_keypoints_inv, which takes keypoint predictions from augmented images and normalizes them so that they can be compared between augmentations.

A 'visibility' column can be added to keypoints (eg. keypoints become [[x, y, visible]]). If an operation causes a keypoint to be moved outside the image then its visibility will be changed to 0.

python setup.py install

Using RandAugment Keypoints is simple. First use plan_augment in order to create (magnitude, operation, directions) and then apply them using apply_image or apply_keypoints.

Example:

import augkey

image = ... # your PIL.Image
randaug = augkey.RandAugment()
plan = randaug.plan_augment() 
augmented_image = randaug.apply_image(image, *plan)

See the demo notebook for more examples.

By default, RandAugment uses the operations discussed in the original paper[1].

In response to discussion about the large impact of augmentations and their magnitude ranges in the original paper[1], this package abstracts operations to make them easier to create and modify.

import augkey
from augkey import operations as ops

# Create your new operation.
class Crop(ops.Operation):
    """Crop an image"""
    def transform_image(self, image, magnitude=0, direction=1):
        value = self.magnitude_range[magnitude]
        width, height = im.size
        return image.crop((value, height - value, width - value, value))

    # if using keypoints, similarly implement `transform_keypoints`.

operations = {
    'crop': Crop(np.arange(30)), # Pass the magnitude_range in as an argument.
    'rotate': ops.Rotate(np.linspace(0, math.radians(30), 30)),
}

randaug = ra.RandAug(operations=operations)

Tests require opencv-python, install with pip install opencv-python

To run tests: python -m unittest -v

1. Cubuk, Ekin D., Barret Zoph, Jonathon Shlens, and Quoc V. Le. “RandAugment: Practical Automated Data Augmentation with a Reduced Search Space.” ArXiv:1909.13719 [Cs], November 13, 2019. [http://arxiv.org/abs/1909.13719].

2. Sohn, Kihyuk, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. “FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence.” ArXiv:2001.07685 [Cs, Stat], November 25, 2020. [http://arxiv.org/abs/2001.07685].

3. Xie, Qizhe, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. “Unsupervised Data Augmentation for Consistency Training.” ArXiv:1904.12848 [Cs, Stat], November 5, 2020. [http://arxiv.org/abs/1904.12848].