LiTOptions
Options & Configuration · AiDotNet.VisionLanguage.Encoders
Configuration options for the LiT (Locked-image Tuning) model.
Inheritance: ModelOptions → ContrastiveEncoderOptions → LiTOptions
For Beginners
LiT takes a shortcut: instead of training both image and text encoders from scratch (which is expensive), it takes an already-trained image model and only teaches the text model to align with it. This is much faster and often works just as well.
How It Works
LiT (Zhai et al., CVPR 2022) freezes a pre-trained image encoder and only trains the text encoder and projection layers. This “locked-image tuning” approach achieves strong zero-shot performance while being much cheaper to train than full contrastive learning from scratch.
Constructors
| Constructor | Summary |
|---|---|
LiTOptions() |
Initializes default LiT options. |
LiTOptions(LiTOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
ContrastiveLossType LossType { get; set; } |
Gets or sets the contrastive loss type. |
bool FreezeVisionEncoder { get; set; } |
Gets or sets whether to freeze the vision encoder during training. |
bool InitializeTextFromPretrained { get; set; } |
Gets or sets whether to use a pre-trained text encoder as initialization. |
string PretrainedVisionWeightsPath { get; set; } |
Gets or sets the path to the pre-trained vision encoder weights. |