CLIPA<T>

Models & Types · AiDotNet.VisionLanguage.Encoders

CLIPA (CLIP with Inverse scaling law and Accelerated training) model using progressive resolution.

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

Implements: IContrastiveVisionLanguageModel<T>

For Beginners

CLIPA discovers that training CLIP at low resolution first, then fine-tuning at full resolution, reduces training cost by 7-8x while maintaining performance. This “inverse scaling law” makes it much cheaper to train strong vision- language models by progressively increasing image resolution during training. Default values follow the original paper settings.

How It Works

CLIPA (Li et al., 2023) discovers an inverse scaling law: training at lower resolution first then fine-tuning at full resolution reduces training cost 7-8x while maintaining performance.

References:

  • Paper: “An Inverse Scaling Law for CLIP Training” (Li et al., 2023)

Architecture layout: Vision encoder (patch embedding + transformer + projection) lives in Layers; text encoder lives in TextEncoderLayers. Default Predict + Train is vision-only; EncodeText walks the text stack independently.

Example

// Create a CLIPA model for accelerated contrastive training
// using progressive resolution for 7-8x training cost reduction
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
            
// ONNX inference mode with pre-trained model
var model = new CLIPA<double>(architecture, "clipa.onnx");
            
// Training mode with native layers
var trainModel = new CLIPA<double>(architecture, new CLIPAOptions());

Constructors

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