RADIOv25<T>
Models & Types · AiDotNet.VisionLanguage.Encoders
RADIOv2.5 agglomerative vision foundation model distilling DINOv2, SAM, SigLIP, and CLIP.
Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → RADIOv25<T>
Implements: IVisualEncoder<T>
For Beginners
RADIO v2.5 from NVIDIA is a “best of all worlds” vision encoder that distills knowledge from multiple teacher models (DINOv2, SAM, SigLIP, and CLIP) into a single student model. The result is a universal vision backbone that can be used as a drop-in replacement for any of those individual models, producing compatible features for all of them. Default values follow the original paper settings.
How It Works
RADIOv2.5 (Ranzinger et al., NVIDIA 2025) distills multiple teacher vision models into a single student via multi-teacher distillation. The student produces features compatible with all teachers, serving as a universal drop-in replacement for DINOv2, SAM, SigLIP, and CLIP vision encoders.
References:
- Paper: “AM-RADIO: Agglomerative Vision Foundation Model” (Ranzinger et al., 2025)
Example
// Create a RADIOv2.5 model distilling DINOv2, SAM, SigLIP, and CLIP
// into a single universal vision encoder backbone
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 RADIOv25<double>(architecture, "radiov25.onnx");
// Training mode with native layers
var trainModel = new RADIOv25<double>(architecture, new RADIOv25Options());
Constructors
| Constructor | Summary |
|---|---|
RADIOv25(NeuralNetworkArchitecture<T> architecture, RADIOv25Options options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) |
|
RADIOv25(NeuralNetworkArchitecture<T> architecture, string modelPath, RADIOv25Options options) |
Properties
| Property | Summary |
|---|---|
int EmbeddingDimension { get; } |
Methods
| Method | Summary |
|---|---|
ModelMetadata<T> GetModelMetadata() |
|
Tensor<T> EncodeImage(Tensor<T> image) |