ViTBackbone class

A Vision Transformer (ViT) backbone for extracting image features.

This model processes an image by dividing it into patches, linearly embedding them, adding positional information, and feeding them through a Transformer Encoder. It outputs the contextualized embeddings of all patches (and optionally a CLS token) for downstream tasks like object detection.

Inheritance

• Object
• Module
• ViTBackbone

Constructors

ViTBackbone({required int imageSize, required int patchSize, int numChannels = 3, required int embedSize, int numLayers = 2, int numHeads = 4})

Properties

clsToken → ValueVector

final

embedSize → int

final

hashCode → int

The hash code for this object.

no setter, inherited

imageSize → int

final

numChannels → int

final

numHeads → int

final

numLayers → int

final

patchProjection → Layer

final

patchSize → int

final

positionEmbeddings → List<ValueVector>

final

runtimeType → Type

A representation of the runtime type of the object.

no setter, inherited

transformerEncoder → TransformerEncoder

final

Methods

forward(List<double> imageData) → List<ValueVector>

The forward pass for the ViT Backbone.

noSuchMethod(Invocation invocation) → dynamic

Invoked when a nonexistent method or property is accessed.

inherited

parameters() → List<Value>

override

toString() → String

A string representation of this object.

inherited

zeroGrad() → void

inherited

Operators

operator ==(Object other) → bool

The equality operator.

inherited