Enum ImageAugmentationOp 

pub enum ImageAugmentationOp {
    HorizontalFlip {
        probability: f32,
    },
    VerticalFlip {
        probability: f32,
    },
    RandomRotate90 {
        probability: f32,
    },
    BrightnessJitter {
        max_delta: f32,
    },
    ContrastJitter {
        max_scale_delta: f32,
    },
    GammaJitter {
        max_gamma_delta: f32,
    },
    GaussianNoise {
        probability: f32,
        std_dev: f32,
    },
    BoxBlur3x3 {
        probability: f32,
    },
    RandomResizedCrop {
        probability: f32,
        min_scale: f32,
        max_scale: f32,
    },
    Cutout {
        probability: f32,
        max_height_fraction: f32,
        max_width_fraction: f32,
        fill_value: f32,
    },
    ChannelNormalize {
        mean: Vec<f32>,
        std: Vec<f32>,
    },
    Custom(Arc<dyn Fn(&Tensor) -> Result<Tensor, ModelError> + Send + Sync>),
    RandomCrop {
        height: usize,
        width: usize,
    },
    GaussianBlur {
        kernel_size: usize,
    },
}

Per-sample image augmentations for rank-4 NHWC training tensors.

Variants

HorizontalFlip

Flip image across width axis with configured probability.

Fields

probability: f32

VerticalFlip

Flip image across height axis with configured probability.

Fields

probability: f32

RandomRotate90

Rotate sample by random multiples of 90 degrees with configured probability.

For square samples, rotation is sampled from {0, 90, 180, 270} degrees. For non-square samples, rotation is sampled from {0, 180} degrees to preserve shape.

Fields

probability: f32

BrightnessJitter

Add random uniform brightness delta in [-max_delta, +max_delta] and clamp to [0, 1].

Fields

max_delta: f32

ContrastJitter

Scale contrast around per-sample mean by factor in [1-max_scale_delta, 1+max_scale_delta].

Fields

max_scale_delta: f32

GammaJitter

Apply gamma correction with gamma sampled in [1-max_gamma_delta, 1+max_gamma_delta].

Fields

max_gamma_delta: f32

GaussianNoise

Add per-value Gaussian noise with configured standard deviation and clamp to [0, 1].

Fields

probability: f32

std_dev: f32

BoxBlur3x3

Apply 3x3 box blur with configured probability.

Fields

probability: f32

RandomResizedCrop

Crop a random window and resize it back to original sample size.

Fields

probability: f32

min_scale: f32

max_scale: f32

Cutout

Apply random rectangular erasing with configured max size fractions and fill value.

Fields

probability: f32

max_height_fraction: f32

max_width_fraction: f32

fill_value: f32

ChannelNormalize

Per-channel normalization in HWC layout: (x - mean[c]) / std[c].

Fields

mean: Vec<f32>

std: Vec<f32>

Custom(Arc<dyn Fn(&Tensor) -> Result<Tensor, ModelError> + Send + Sync>)

User-provided closure for custom augmentation logic.

RandomCrop

Random crop from larger image (does not resize back; changes spatial dims).

Fields

height: usize

width: usize

GaussianBlur

Apply gaussian blur with the given square kernel size (must be odd and >= 1).

Fields

kernel_size: usize

Implementations

impl ImageAugmentationOp

pub fn from_transform<T: Transform + Send + Sync + 'static>(t: T) -> Self

Create an augmentation op from any Transform implementation.

Trait Implementations

impl Clone for ImageAugmentationOp

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ImageAugmentationOp

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl PartialEq for ImageAugmentationOp

fn eq(&self, other: &Self) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.