UNITEROptions
Options & Configuration · AiDotNet.VisionLanguage.Foundational
Configuration options for UNITER (Universal Image-TExt Representation).
Inheritance: ModelOptions → FoundationalVLMOptions → UNITEROptions
How It Works
UNITER (Chen et al., ECCV 2020) uses conditional masking pre-training where either image regions or text tokens are masked, forcing the model to learn cross-modal alignment. It uses a single-stream transformer for joint image-text encoding.
Constructors
| Constructor | Summary |
|---|---|
UNITEROptions() |
Initializes a new instance with default values. |
UNITEROptions(UNITEROptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
double EndLearningRate { get; set; } |
Gets or sets the final learning rate after linear decay. |
double ImageMaskProbability { get; set; } |
Gets or sets the image region masking probability during training. |
double MaxGradientNorm { get; set; } |
Gets or sets the global gradient-norm clipping threshold. |
double TextMaskProbability { get; set; } |
Gets or sets the text token masking probability during training. |
double? WarmupInitialLearningRate { get; set; } |
Gets or sets the learning rate at the first warmup update, or null to derive the first positive increment from LearningRate and WarmupSteps. |
int MaxImageRegions { get; set; } |
Gets or sets the maximum number of image regions. |
int TotalTrainingSteps { get; set; } |
Gets or sets the training horizon used by the post-warmup linear decay. |
int WarmupSteps { get; set; } |
Gets or sets the number of linear learning-rate warmup steps. |
Methods
| Method | Summary |
|---|---|
void Validate() |
Validates the native optimizer and learning-rate schedule contract. |