METEROptions
Options & Configuration · AiDotNet.VisionLanguage.Foundational
Configuration options for METER (Multimodal End-to-end TransformER).
Inheritance: ModelOptions → FoundationalVLMOptions → METEROptions
How It Works
METER (Dou et al., CVPR 2022) is a systematic study of vision-language pre-training components. It uses separate CLIP ViT vision encoder and RoBERTa text encoder, connected by a co-attention transformer fusion module, providing an optimized combination of architecture choices for VLP.
Constructors
| Constructor | Summary |
|---|---|
METEROptions() |
Initializes a new instance with default values. |
METEROptions(METEROptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
TextEncoderVariant TextEncoder { get; set; } |
Gets or sets the text encoder type. |
ViTVariant VisionEncoder { get; set; } |
Gets or sets the vision encoder type. |
int NumCrossAttentionLayers { get; set; } |
Gets or sets the number of cross-attention fusion layers. |