EVACLIPOptions

Options & Configuration · AiDotNet.VisionLanguage.Encoders

Configuration options for the EVA-CLIP model.

Inheritance: ModelOptions → ContrastiveEncoderOptions → EVACLIPOptions

For Beginners

EVA-CLIP is like CLIP but with a better-trained image encoder. The EVA vision model first learns to understand images on its own (by filling in masked patches), and then learns to connect images with text. This two-step process produces stronger results.

How It Works

EVA-CLIP (Sun et al., 2023) combines the EVA-02 Vision Transformer backbone with CLIP-style contrastive learning. EVA-02 uses masked image modeling (MIM) pre-training to produce a stronger vision encoder, which is then used for contrastive image-text alignment. The largest EVA-CLIP variant (ViT-E/14, 4.4B params) achieves 82.0% zero-shot on ImageNet.

Constructors

Constructor Summary
EVACLIPOptions() Initializes default EVA-CLIP options.
EVACLIPOptions(EVACLIPOptions 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 UseEVA02 { get; set; } Gets or sets whether to use EVA-02 (improved) backbone instead of EVA-01.
bool UseRoPE { get; set; } Gets or sets whether to use rotary position embeddings (RoPE) in the vision encoder.
bool UseSwiGLU { get; set; } Gets or sets whether to use SwiGLU activation in the FFN layers.
string MIMPretraining { get; set; } Gets or sets the MIM pre-training method used for the vision encoder.