What is Albumentations?

https://github.com/albumentations-team/albumentations
Python library for data extension for machine learning
A wealth of features commonly used in Data augmentation
pip install albumentations

This article

Confirmed the operation while changing the parameters of about 50 Augmentations.
version 0.4.6
Nodebook written for testing (Jupyter, Colab) ――If you prepare an image, you can check everything in the same way.

Blur

Blur with a randomly sized kernel
blur_limit (int) – Maximum blurr kernel size. Default: (3, 7)

MotionBlur

Apply motion blur (box filter) with random kernel size
blur_limit (int) – Maximum blurr kernel size. Default: (3, 7).

GaussianBlur

Apply Gaussian filter with random kernel size
blur_limit (int) – Maximum kernel size to blur, must be odd. Default: (3, 7).

GlassBlur

Adds glass noise (effect like frosted glass)
sigma (float) – Standard deviation of the Gaussian kernel. Defalt: 0.7.
max_delta (int) – Maximum distance between swapped pixels. . Defalt: 4.
iterations (int) – Number of repeats. Default: (2).
mode (str) – Computation mode. Fast or exact. Default: “fast”.

Noise, Compression

GaussNoise

Add Gaussian noise
var_limit ((float, float) or float) – Noise distribution. Default: (10.0, 50.0).
mean (float) – Noise average. Default: 0

JpegCompression

Apply Jpeg compression noise
quality_lower (float) – The lower limit of quality.
quality_upper (float) – Quality limit.

ImageCompression

Apply Jpeg / WebP compression noise
quality_lower (float) – The lower limit of quality.
quality_upper (float) – Quality limit.
compression_type (ImageCompressionType) – Compression type (JPEG / WEBP). should be ImageCompressionType.JPEG or ImageCompressionType.WEBP. Default: ImageCompressionType.JPEG

ISONoise

Add camera sensor noise
color_shift (float, float) – amount of hue change.
intensity ((float, float) – Intensity of color / luminance noise.

MultiplicativeNoise

Multiply by random number array
multiplier (float or tuple of floats) – A range of numbers to multiply. Default: (0.9, 1.1).
per_channel (bool) – False: Same for all channels, True: Use sample values for each channel. Default False.
elementwise (bool) – False: Multiply all pixels in the image, True: Randomly sample and multiply pixel by pixel. Defaule: False.

Downscale

Downscale and then upscale to reduce image quality
scale_min (float) – The lower limit of the scale. Should be <1. Default: 0.25.
scale_max (float) – The upper limit of the scale. Default: 0.25.

Simple geometric change system (Flip, Crop, Rotate, Scale, Transpose)

Flip

Randomly flips horizontally, vertically, or both horizontally and vertically

VerticalFlip

Flip vertically around the x-axis

HorizontalFlip

Flip horizontally around the y-axis

Crop

Cut out the area
x_min (int) – The minimum value of the upper left x coordinate.
y_min (int) – The minimum value of the upper left y coordinate.
x_max (int) – Maximum value of the lower right x coordinate.
y_max (int) – Maximum value of the lower right y coordinate.

RandomCrop

Randomly cut
height (int) – The height to cut.
width (int) – The width to cut.

CenterCrop

Crop the center part
height (int) – The height to crop.
width (int) – The width to crop.

RandomSizedCrop

Randomly cut and rescale to a specific size
min_max_height ((int, int)) – Crop size range.
height (int) – Height after resizing.
width (int) – Width after resizing.
w2h_ratio (float) – Crop aspect ratio. Default: 1.0.

RandomResizedCrop

Randomly cut and rescale to a specific size (Torchvision variant)
height (int) – Height after resizing.
width (int) – Width after resizing.
scale ((float, float)) – The size range of the area to crop. Default: (0.08, 1.0).
ratio ((float, float)) – The range of aspect ratios for the area to be cropped. Default: (0.75, 1.3333333333333333).

Rotate

Rotate at random angles
limit: Range of angles, (-limit, limit) for single numbers. Default: (-90, 90)

RandomScale

Randomly change the image size
scale_limit ((float, float) or float) – Scaling range (note that 0 is unchanged). Default: (0.9, 1.1).

Since it is previewed in the same size, the original image size has changed although it is not visible.

RandomRotate90

Randomly rotate in 90 ° increments

Transpose

Transpose rows and columns

Advanced geometric transformation system (Affine, Distortion)

ShiftScaleRotate

Randomly apply affine transformations (translation, scaling, rotation)
shift_limit: The range of translation. Default: (-0.0625, 0.0625).
scale_limit: Scale range (note that 0 is unchanged). Default: (-0.1, 0.1).
rotate_limit: rotation range. The range of rotation. Default: (-45, 45).

OpticalDistortion

Reproduce optical distortion
distort_limit (float, (float, float)) – Range of distortion. Default: (-0.05, 0.05).
shift_limit (float, (float, float))) – The range to shift. Default: (-0.05, 0.05).
Not applicable to Bounding Box and Keypoints

GridDistortion

Reproduce Grid distortion
num_steps (int) – Specify the number of grid cells on each side. Default: 5.
distort_limit (float, (float, float)) – Range of distortion. Default: (-0.03, 0.03).
Not applicable to Bounding Box and Keypoints

ElasticTransform

Elastic deformation
alpha (float) – transformation parameters. Default: 1.
sigma (float) – Gaussian filter parameters. Default: 50.
alpha_affine (float) – Range of alpha_affine. Default: 50.
Not applicable to Bounding Box and Keypoints

RandomGridShuffle

Randomly shuffle cells in the grid
grid ((int, int)) – The size of the grid that divides the image. Default: (3,3).
Not applicable to Bounding Box and Keypoints

Color tinkering system

HueSaturationValue

Randomly change hue, saturation, and brightness
hue_shift_limit ((int, int) or int) – Hue range. Default: (-20, 20).
sat_shift_limit ((int, int) or int) – Saturation range. Default: (-30, 30).
val_shift_limit ((int, int) or int) – Brightness range. Default: (-20, 20).

RGBShift

Randomly change the value of each RGB channel
r_shift_limit ((int, int) or int) – Red channel range of change. Default: (-20, 20).
g_shift_limit ((int, int) or int) – Green channel range of change. Default: (-20, 20).
b_shift_limit ((int, int) or int) – Blue channel range of change. Default: (-20, 20).

ChannelShuffle

Randomly sort RGB channels

ChannelDropout

Randomly drop channels
channel_drop_range (int, int) – The range of channels to drop.
fill_value (int, float) – Pixel value to fill the dropped channel instead.

Posterize

Reduce the number of bits in each color channel
num_bits ((int, int) – range of bits. Default: 4.

ToGray

Convert RGB image to grayscale

ToSepia

Apply sepia filter

Brightness and contrast tinkering system

InvertImg

Invert the pixel value

Normalize

Divide the pixel value by 255 → subtract the average value for each channel → divide by the standard deviation for each channel
** Mean and standard deviation are just parameters and are not implicitly calculated internally **
mean (float, list of float) – Mean. Dafault: (0.485, 0.456, 0.406).
std (float, list of float) – Standard deviation. Dafault: (0.229, 0.224, 0.225).
max_pixel_value (float) – Maximum pixel value. Dafault: 255.0

RandomGamma

Randomly apply gamma conversion
gamma_limit (float or (float, float)) – The upper limit of gamma. Default: (80, 120).

RandomBrightness

Randomly change the brightness
limit ((float, float) or float) – Range of change in brightness. Default: (-0.2, 0.2).

RandomContrast

Randomly change the contrast
limit ((float, float) or float) – Range of change in contrast. Default: (-0.2, 0.2).

RandomBrightnessContrast

Randomly change brightness and contrast
brightness_limit ((float, float) or float) – The range of change in brightness. Default: (-0.2, 0.2).
contrast_limit ((float, float) or float) – Contrast range of change. Default: (-0.2, 0.2).
brightness_by_max (Boolean) – How to adjust the contrast. Default: True.

CLAHE

Contrast limit adaptive histogram equalization
clip_limit (float or (float, float)) – The upper threshold of the contrast limit. Default: (1, 4).
tile_grid_size ((int, int)) – Grid size for histogram equalization. Default: (8, 8).

Solarize

Invert pixel values above the threshold (solar)
threshold ((int, int) or int, or (float, float) or float) – Inversion threshold. Default: 128.

Dropout system

Cutout

Rough Dropout in rectangular area
num_holes (int) – Number of regions to drop to zero. Defalt: 8.
max_h_size (int) – Maximum height of the area. Defalt: 8.
max_w_size (int) – Maximum width of the area. Defalt: 8.
fill_value (int, float, lisf of int, list of float). Pixel value of the dropped area. Defalt: 0.

CoarseDropout

Rough dropout of rectangular area (minimum value can be specified)
max_holes (int) – Maximum number of regions to drop to zero.
max_height (int) – Maximum height of the area. Defalt: 8.
max_width (int) – Maximum width of the area. Defalt: 8.
min_holes (int) – The minimum number of regions to drop to zero. Default: None.
min_height (int) – Minimum height of the area. Default: None.
min_width – The minimum width of the area. Default: None.
fill_value (int, float, lisf of int, list of float). Pixel value of the dropped area. Defalt: 0.

Weather change / environment / optical reproduction system

RandomSnow

Simulate snow
snow_point_lower (float) – The lower limit for snow. Default: 0.1.
snow_point_upper (float) – Upper limit for snow. Default: 0.3.
brightness_coeff (float) – Higher values result in more snow. Should be> = 0. Default: 2.5.

RandomRain

Rain effect
slant_lower – Diagonal down condition. Default: -10.
slant_upper – Diagonal up. Default: 10.
drop_length – The length of the rain. Default: 20.
drop_width – The width of the rain. Default: 1.
drop_color (list of (r, g, b)) – The color of the rain line. Default: (200, 200, 200).
blur_value (int) – A raindrop blur. Default: 7.
brightness_coefficient (float) – Brightness. Default: 0.7.
rain_type – The type of rain. [None, “drizzle”, “heavy”, “torrestial”]. Default: None.

RandomFog

Simulate fog
fog_coef_lower (float) – Lower limit of fog intensity.
fog_coef_upper (float) – Upper limit of fog intensity.

RandomSunFlare

Simulate solar flares
flare_roi (float, float, float, float) – The area where flares appear (x_min, y_min, x_max, y_max).
angle_lower (float) – Lower limit of angle.
angle_upper (float) – Upper limit of angle.
num_flare_circles_lower (int) – Lower limit for the number of flares.
num_flare_circles_upper (int) – Maximum number of flares.
src_radius (int) – Flare radius.
src_color ((int, int, int)) – Flare color.

RandomShadow

Simulate shadows
shadow_roi (float, float, float, float) – The area where the shadow appears (x_min, y_min, x_max, y_max).
num_shadows_lower (int) – Lower limit for the number of shadows.
num_shadows_upper (int) – Maximum number of shadows.
shadow_dimension (int) – The number of sides of the shadow polygon.

other

FancyPCA

Extension with Fancy PCA
alpha (float) – How much to scale the eigenvectors and eigenvalues.

PadIfNeeded

Pad the edges of the image for the desired resolution
min_height (int) – Minimum image height. Default: 1024.
min_width (int) – Minimum image width. Default: 1024.

[PYTHON] Move the Augmentaiton of Albumentations earnestly