ViLBERTOptions
Options & Configuration · AiDotNet.VisionLanguage.Foundational
Configuration options for ViLBERT (Vision-and-Language BERT) with co-attention.
Inheritance: ModelOptions → FoundationalVLMOptions → ViLBERTOptions
How It Works
ViLBERT (Lu et al., NeurIPS 2019) processes images and text in two parallel BERT-like streams connected by co-attention transformer layers, enabling rich cross-modal interaction.
Constructors
| Constructor | Summary |
|---|---|
ViLBERTOptions() |
Initializes a new instance with default values. |
ViLBERTOptions(ViLBERTOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
int FusionIntermediateDim { get; set; } |
Gets or sets the co-attention feed-forward width. |
int MaxVisualRegions { get; set; } |
Gets or sets the maximum number of visual regions (object proposals). |
int NumCoAttentionLayers { get; set; } |
Gets or sets the number of co-attention layers between streams. |
int NumFusionHeads { get; set; } |
Gets or sets the number of heads in the bidirectional co-attention blocks. |
int NumTextHeads { get; set; } |
Gets or sets the number of self-attention heads in the BERT text stream. |
int NumVisionHeads { get; set; } |
Gets or sets the number of self-attention heads in the visual stream. |
int TextIntermediateDim { get; set; } |
Gets or sets the BERT text-stream feed-forward width. |
int VisionIntermediateDim { get; set; } |
Gets or sets the visual-stream feed-forward width. |
int VisualFeatureDim { get; set; } |
Gets or sets the Faster R-CNN detector feature width before visual embedding. |