EVACLIP<T>

Models & Types · AiDotNet.VisionLanguage.Encoders

EVA-CLIP model combining EVA-02 masked image modeling backbone with CLIP contrastive training.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → EVACLIP<T>

Implements: IContrastiveVisionLanguageModel<T>

For Beginners

EVA-CLIP improves CLIP by first pre-training the vision encoder with masked image modeling (learning to reconstruct hidden image patches), then applying contrastive image-text training. The largest variant has 4.4 billion parameters and achieves 82.0% zero-shot ImageNet accuracy. It also adds rotary position embeddings (RoPE) and SwiGLU activations for stronger feature learning. Default values follow the original paper settings.

How It Works

EVA-CLIP (Sun et al., 2023) uses the EVA-02 ViT backbone pre-trained with masked image modeling (MIM) for stronger visual representations. The largest variant (ViT-E/14, 4.4B params) achieves 82.0% zero-shot on ImageNet. EVA-02 adds RoPE positional embeddings and SwiGLU FFN activations.

References:

  • Paper: “EVA-CLIP: Improved Training Techniques for CLIP at Scale” (Sun et al., 2023)

Example

var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
var trainModel = new EVACLIP<double>(architecture, new EVACLIPOptions());

Constructors

Constructor Summary
EVACLIP(NeuralNetworkArchitecture<T> architecture, EVACLIPOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
EVACLIP(NeuralNetworkArchitecture<T> architecture, string imageEncoderModelPath, EVACLIPOptions options)

Properties

Property Summary
T Temperature { get; }
int EmbeddingDimension { get; }
int MaxSequenceLength { get; }
int ProjectionDimension { get; }
int TextEmbeddingDimension { get; }

Methods

Method Summary
Dictionary<string, T> ZeroShotClassify(Tensor<T> image, string[] labels)
ModelMetadata<T> GetModelMetadata()
T ComputeSimilarity(Tensor<T> image, string text)
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> EncodeText(string text)
Tensor<T>[] EncodeTexts(string[] texts)