DFNCLIPOptions

Options & Configuration · AiDotNet.VisionLanguage.Encoders

Configuration options for the DFN-CLIP (Data Filtering Networks for CLIP) model.

Inheritance: ModelOptions → ContrastiveEncoderOptions → DFNCLIPOptions

For Beginners

DFN-CLIP uses a clever bootstrapping trick: first train a small model, then use that model to find the best training data from a huge noisy pool, and finally train a bigger model on just the good data. It’s like having a teacher pre-screen study materials.

How It Works

DFN-CLIP (Fang et al., 2023) from Apple uses a small CLIP model as a “data filtering network” to score and select high-quality image-text pairs from a large noisy pool. The filtered data is then used to train a larger CLIP model, achieving 83.0% zero-shot on ImageNet with ViT-H/14.

Constructors

Constructor Summary
DFNCLIPOptions() Initializes default DFN-CLIP options.
DFNCLIPOptions(DFNCLIPOptions 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.
double FilteringThreshold { get; set; } Gets or sets the filtering threshold score for data selection.
int FilteredDatasetSizeMillions { get; set; } Gets or sets the target dataset size after filtering (in millions).