imgaug-tf

An image augmentation library for tensorflow. This library is implemented in TF native and has been tested to work with TPU.

Installation

!pip install -U imgaugtf

or for nightly version

!pip install -U git+https://github.com/hirune924/imgaug-tf

Required packages:

• tensorflow (2.6.3 or higher recommended)
• tensorflow_addons (0.14.0 or higher recommended)
• tensorflow_probability (0.14.1 or higher recommended)

Quick Start

imgaugtf is implemented to work simply with tf.data. Example of use single transform.

import imgaugtf

image = imgaugtf.random_solarize_add(image, addition=30, prob=0.5)

You can also apply transform for a mask as same as a image.

import imgaugtf

image = imgaugtf.seg.random_solarize_add(image, mask, addition=30, prob=0.5)

You can also randomly select n of multiple transformations to apply, as shown below. You can use mixup or cutmix on batched images.

import imgaugtf

def augmentation(example):
    example['image'] = imgaugtf.random_resized_crop(example['image'], size=[256, 256], prob=1.0)
    example['image'] = imgaugtf.apply_n(example['image'], functions=imgaugtf.OPERATORS, num_ops=2, prob=1.0)
    return example

def batch_augmentation(example, num_class=120):
    image = example['image']
    label = tf.one_hot(example['label'], num_class)
    
    image, label = imgaugtf.cutmix(image, label)
    return image, label

result = next(iter(dataset.map(augmentation).batch(15).map(batch_augmentation)))

for i in range(10):
    plt.imshow(result[0][i])
    plt.show()

functions is list of dict like this example. dict has keys of 'func' and 'option'. you can customize it you like.

[
    {"func": imgaugtf.random_cutout, "option": {"num_holes": 8, "hole_size": 20, "replace": 0}},
    {"func": imgaugtf.random_solarize, "option": {"threshold": 128}},
    {"func": imgaugtf.random_solarize_add, "option": {"addition": 30, "threshold": 128}},
]

Examples

• calssification
• segmentation

Augmentations

pixel

	image	mask
original
random_solarize
random_solarize_add
random_color
random_contrast
random_brightness
random_posterize
random_invert
random_equalize
random_sharpness
random_autocontrast
random_hsv_in_yiq
random_gaussian_filter2d
random_mean_filter2d
random_median_filter2d
random_cutout
random_gray
random_hue
random_saturation
random_gamma
random_jpeg_quality
random_gaussian_noise
random_speckle_noise

geometory

	image	mask
original
random_flip_left_right
random_flip_up_down
random_resized_crop
random_rotate
random_translate_x
random_translate_y
random_shear_x
random_shear_y
random_zoom
random_grid_shuffle
random_affine
random_elastic_deform
random_sparse_warp
random_crop

blend

• mixup
• cutmix

compose

• apply_one
• apply_n