Class SSLAugmentationPolicies<T>

Namespace

AiDotNet.SelfSupervisedLearning

Assembly

AiDotNet.dll

Provides standard augmentation policies for self-supervised learning methods.

public class SSLAugmentationPolicies<T>

Type Parameters

T

The numeric type used for computations (typically float or double).

Inheritance

object

SSLAugmentationPolicies<T>

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

For Beginners: Self-supervised learning methods rely heavily on data augmentation to create different "views" of the same image. These views should be different enough to be challenging, but similar enough that they can still be recognized as the same image.

Common augmentations for SSL:

• Random crop and resize: Most important - crops different regions
• Color jitter: Changes brightness, contrast, saturation, hue
• Grayscale: Removes color information
• Gaussian blur: Adds blur (important for SimCLR, BYOL)
• Horizontal flip: Random left-right flip
• Solarization: Inverts pixels above threshold (BYOL, DINO)

Method-specific policies:

• SimCLR: Strong augmentation with blur and color jitter
• MoCo v2: Similar to SimCLR with slightly different parameters
• BYOL: Asymmetric augmentation (view 1 and view 2 different)
• DINO: Multi-crop strategy (global + local crops)

Constructors

SSLAugmentationPolicies(int?)

Initializes a new instance of the SSLAugmentationPolicies class.

public SSLAugmentationPolicies(int? seed = null)

Parameters

seed int?

Optional random seed for reproducibility.

Methods

ApplyBYOL(Tensor<T>)

Applies BYOL-style asymmetric augmentation.

public (Tensor<T> view1, Tensor<T> view2) ApplyBYOL(Tensor<T> image)

Parameters

image Tensor<T>

Input image tensor.

Returns

(Tensor<T> grad1, Tensor<T> grad2)

Two asymmetrically augmented views.

Remarks

BYOL uses asymmetric augmentation:

View 1: Full augmentation pipeline

View 2: Lighter augmentation (may skip some transforms)

ApplyDINO(Tensor<T>, int, int)

Applies DINO-style multi-crop augmentation.

public (Tensor<T>[] globalViews, Tensor<T>[] localViews) ApplyDINO(Tensor<T> image, int numGlobalCrops = 2, int numLocalCrops = 8)

Parameters

image Tensor<T>

Input image tensor.

numGlobalCrops int

Number of global (large) crops (default: 2).

numLocalCrops int

Number of local (small) crops (default: 8).

Returns

(Tensor<T>[] globalViews, Tensor<T>[] localViews)

Global and local crop views.

Remarks

DINO multi-crop strategy:

• Global crops: 224x224, covering 50-100% of image
• Local crops: 96x96, covering 5-50% of image
• Teacher sees only global crops, student sees all crops

ApplyMinimal(Tensor<T>)

Applies minimal augmentation (used by MAE which relies on masking rather than augmentation).

public Tensor<T> ApplyMinimal(Tensor<T> image)

Parameters

image Tensor<T>

Input image tensor.

Returns

Tensor<T>

Minimally augmented image.

Remarks

For Beginners: MAE uses minimal augmentation because the masking itself provides the learning signal. Only basic normalization and optional horizontal flip.

ApplyMoCoV2(Tensor<T>)

Applies MoCo v2 style augmentation.

public (Tensor<T> query, Tensor<T> key) ApplyMoCoV2(Tensor<T> image)

Parameters

image Tensor<T>

Input image tensor.

Returns

(Tensor<T> grad1, Tensor<T> grad2)

Two augmented views.

ApplySimCLR(Tensor<T>)

Applies SimCLR-style augmentation to create two views.

public (Tensor<T> view1, Tensor<T> view2) ApplySimCLR(Tensor<T> image)

Parameters

image Tensor<T>

Input image tensor [C, H, W] or [B, C, H, W].

Returns

(Tensor<T> grad1, Tensor<T> grad2)

Two augmented views of the image.

Remarks

SimCLR augmentation pipeline:

RandomResizedCrop(224, scale=(0.08, 1.0))
RandomHorizontalFlip(p=0.5)
ColorJitter(brightness=0.8, contrast=0.8, saturation=0.8, hue=0.2, p=0.8)
RandomGrayscale(p=0.2)
GaussianBlur(kernel_size=23, sigma=(0.1, 2.0), p=0.5)

ForBYOL(int?)

Creates a standard BYOL augmentation policy.

public static SSLAugmentationPolicies<T> ForBYOL(int? seed = null)

Parameters

seed int?

Returns

SSLAugmentationPolicies<T>

ForDINO(int?)

Creates a standard DINO augmentation policy.

public static SSLAugmentationPolicies<T> ForDINO(int? seed = null)

Parameters

seed int?

Returns

SSLAugmentationPolicies<T>

ForSimCLR(int?)

Creates a standard SimCLR augmentation policy.

public static SSLAugmentationPolicies<T> ForSimCLR(int? seed = null)

Parameters

seed int?

Returns

SSLAugmentationPolicies<T>