CLIPAOptions

Options & Configuration · AiDotNet.VisionLanguage.Encoders

Configuration options for the CLIPA (CLIP with Inverse scaling law and Accelerated training) model.

Inheritance: ModelOptions → ContrastiveEncoderOptions → CLIPAOptions

For Beginners

CLIPA finds that you can train CLIP much faster by using lower resolution images and shorter text during most of the training, then switching to full resolution at the end. This is like studying cliff notes first to get the big picture, then reading the full text for details.

How It Works

CLIPA (Li et al., 2023) discovers an “inverse scaling law” for CLIP training: using shorter image sequences and text lengths during the bulk of training, then fine-tuning at full resolution. This reduces training cost by 7-8x while maintaining performance, enabling efficient scaling.

Constructors

Constructor Summary
CLIPAOptions() Initializes default CLIPA options.
CLIPAOptions(CLIPAOptions 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 IsFineTuningPhase { get; set; } Gets or sets whether the model is in the fine-tuning phase (full resolution).
double ReducedResolutionFraction { get; set; } Gets or sets the fraction of training done at reduced resolution before switching to full.
int InitialImageSize { get; set; } Gets or sets the initial (reduced) image size for the bulk of training.
int InitialSequenceLength { get; set; } Gets or sets the initial (reduced) text sequence length for the bulk of training.