MetaCLIPOptions
Options & Configuration · AiDotNet.VisionLanguage.Encoders
Configuration options for the MetaCLIP model.
Inheritance: ModelOptions → ContrastiveEncoderOptions → MetaCLIPOptions
For Beginners
MetaCLIP focuses on training data quality rather than model architecture. It carefully selects and balances the image-text pairs used for training, making sure the model sees a diverse and representative set of concepts. Better data leads to a better model.
How It Works
MetaCLIP (Xu et al., 2023) improves CLIP training through metadata-driven data curation. Instead of using raw web-scraped data, MetaCLIP balances the training distribution by using metadata from WordNet and Wikipedia to ensure diverse, high-quality image-text pairs. This data-centric approach achieves 70.8% zero-shot on ImageNet with ViT-B/16 on 400M pairs.
Constructors
| Constructor | Summary |
|---|---|
MetaCLIPOptions() |
Initializes default MetaCLIP options. |
MetaCLIPOptions(MetaCLIPOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
ContrastiveLossType LossType { get; set; } |
Gets or sets the contrastive loss type. |
PretrainingDataset Dataset { get; set; } |
Gets or sets the pre-training dataset. |
bool UseSubStringMatching { get; set; } |
Gets or sets whether to use sub-string matching for metadata alignment. |
int MaxEntriesPerConcept { get; set; } |
Gets or sets the maximum number of entries per metadata concept for data balancing. |