ALIGN<T>

Models & Types · AiDotNet.VisionLanguage.Encoders

ALIGN (A Large-scale ImaGe and Noisy-text embedding) model for zero-shot classification and cross-modal retrieval using EfficientNet as the vision encoder.

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

Implements: IContrastiveVisionLanguageModel<T>

For Beginners

ALIGN is similar to CLIP but uses a different type of image model (EfficientNet, which is a convolutional neural network) instead of a Vision Transformer. It was trained on a very large but noisy dataset, proving that more data beats cleaner data.

How It Works

ALIGN (Jia et al., ICML 2021) demonstrates that simple dual-encoder contrastive learning achieves strong image-text alignment when trained at massive scale. Unlike CLIP which uses a Vision Transformer, ALIGN uses an EfficientNet-B7 as its vision encoder, showing that the contrastive learning recipe is architecture-agnostic.

Key Innovation: ALIGN trains on 1.8 billion noisy alt-text image-text pairs without expensive filtering, showing that scale compensates for data noise. The EfficientNet backbone provides a CNN-based alternative to ViT for vision encoding.

References:

  • Paper: “Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision” (Jia et al., ICML 2021)

Example

// Create an ALIGN model for contrastive image-text alignment
// using EfficientNet vision encoder trained on 1.8B noisy pairs
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 ALIGN<double>(architecture, "align.onnx");
            
// Training mode with native layers
var trainModel = new ALIGN<double>(architecture, new ALIGNOptions());

Constructors

Constructor Summary
ALIGN(NeuralNetworkArchitecture<T> architecture, ALIGNOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates an ALIGN model in native training mode.
ALIGN(NeuralNetworkArchitecture<T> architecture, string imageEncoderModelPath, ALIGNOptions options) Creates an ALIGN model in native training mode.

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)
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
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)