TextConditioningBase<T>

Base Classes · AiDotNet.Diffusion.Conditioning

Base class for text conditioning modules used in diffusion models.

Inheritance: NeuralNetworkBase<T> → TextConditioningBase<T>

Implements: IConditioningModule<T>, IShapeContract

Known derived types: CLIPTextConditioner<T>, ChatGLM3TextConditioner<T>, DistilledT5TextConditioner<T>, GemmaTextConditioner<T>, Qwen2TextConditioner<T>, SigLIP2TextConditioner<T>, SigLIPTextConditioner<T>, T5TextConditioner<T>

How It Works

Inherits NeuralNetworkBase so concrete subclasses use the codebase’s golden “Architecture.Layers ? : LayerHelper.CreateDefaultXxxTextLayers” pattern — the same shape that the recent CLAP / AST / PANNs rewrites established. The forward path walks the inherited Layers list via Predict, so engine fast-paths, gradient tape, and weight-streaming auto-detect (#1222, merged) all apply automatically.

Tokenisation is delegated to an injected ITokenizer from src/Tokenization. Each concrete subclass constructs its paper- canonical tokenizer (CLIP byte-level BPE, T5 SentencePiece-Unigram, Gemma SentencePiece, etc.) in its constructor; this base class never fabricates token IDs from characters.

Properties

Property Summary
ConditioningType ConditioningType { get; }
bool ProducesPooledOutput { get; }
int EmbeddingDimension { get; }
int MaxSequenceLength { get; }
int VocabSize { get; } Vocabulary size of the configured tokenizer.

Methods

Method Summary
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
ModelMetadata<T> GetModelMetadata()
Tensor<T> Encode(Tensor<T> input)
Tensor<T> EncodeText(Tensor<T> tokenIds, Tensor<T> attentionMask)
Tensor<T> GetPooledEmbedding(Tensor<T> sequenceEmbeddings) Default pooling: mean over the sequence axis.
Tensor<T> GetUnconditionalEmbedding(int batchSize)
Tensor<T> Tokenize(string text)
Tensor<T> TokenizeBatch(string[] texts)