DualEncoderArchitecture<T>
Base Classes · AiDotNet.NeuralNetworks
Modality-agnostic base for two-encoder neural network architectures.
Inheritance: NeuralNetworkArchitecture<T> → DualEncoderArchitecture<T>
Known derived types: AudioTextDualStreamArchitecture<T>, DualStreamArchitecture<T>
How It Works
Dual-encoder networks (CLIP, ALIGN, CLAP, ImageBind, …) train two parallel
encoders that produce embeddings into a shared space for contrastive
learning. The flat Layers list
describes only one graph, so when callers want to customise both encoders
they must reach for a richer architecture descriptor — this abstract base
holds the two layer stacks under modality-neutral names so consumers can
program against the generic dual-encoder shape, while concrete subclasses
expose semantically-named aliases (VisionLayers/TextLayers,
AudioLayers/TextLayers, etc.) for the call-site clarity each
modality pair deserves.
SOLID rationale:
- SRP: each concrete subclass describes exactly one
modality pairing.
- OCP: new modality pairs add new subclasses without
touching existing CLIP-family or audio-text consumers.
- LSP: tape-training and parameter-iteration code
that walks EncoderALayers + EncoderBLayers works against any
subclass.
- ISP: modality-specific aliases let consumers
depend only on the names that apply to them — vision-language models never
see AudioLayers.
Properties
| Property | Summary |
|---|---|
List<ILayer<T>> EncoderALayers { get; } |
Gets the layer stack for the first encoder stream. |
List<ILayer<T>> EncoderBLayers { get; } |
Gets the layer stack for the second encoder stream. |