ViLTOptions
Options & Configuration · AiDotNet.VisionLanguage.Foundational
Configuration options for ViLT (Vision-and-Language Transformer).
Inheritance: ModelOptions → FoundationalVLMOptions → ViLTOptions
How It Works
ViLT (Kim et al., ICML 2021) is a minimal architecture that removes the CNN/object detector entirely. Raw image patches are linearly embedded and concatenated with text tokens in a single transformer, making it 60x faster than region-feature-based models at comparable accuracy.
Constructors
| Constructor | Summary |
|---|---|
ViLTOptions() |
Initializes a new instance with default values. |
ViLTOptions(ViLTOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool UseRandAugment { get; set; } |
Gets or sets whether to use image augmentation during training. |
bool UseWholeWordMasking { get; set; } |
Gets or sets whether to use whole-word masking for MLM pre-training. |
double WarmupInitialLearningRate { get; set; } |
Gets or sets the non-zero initial rate used at the start of warmup. |
int PatchSize { get; set; } |
Gets or sets the patch size for image tokenization. |
int TotalTrainingSteps { get; set; } |
Gets or sets the training horizon used by the linear decay schedule. |
int WarmupSteps { get; set; } |
Gets or sets the number of linear learning-rate warmup steps. |