BartEncoderDecoderLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

The BART encoder-decoder Florence-2 uses as its language model (Lewis et al. 2020; HF BartModel).

Inheritance: LayerBase<T> → BartEncoderDecoderLayer<T>

Implements: IShapeContract

How It Works

The architecture follows the reference:

  • Learned absolute positions with BART’s offset of 2, then layernorm_embedding.
  • Post-norm blocks with biased projections and GELU feed-forwards.
  • A causal decoder with cross-attention.
  • A head tied to the shared embedding.

The encoder takes precomputed embeddings, so Florence-2 can splice image tokens ahead of its text. BART’s final_logits_bias is a zero buffer and is omitted. Dropout is applied to the residual branches in training mode.

Constructors

Constructor Summary
BartEncoderDecoderLayer(int vocabSize, int dim, int numHeads, int ffnDim, int numEncoderLayers, int numDecoderLayers, int maxPositions, double dropoutRate)

Properties

Property Summary
bool SupportsTraining { get; }
int Dim { get; }
int MaxPositions { get; } Longest encoder or decoder sequence the learned positions cover.
int VocabSize { get; }

Methods

Method Summary
Dictionary<string, string> GetParameterRoles() Returns parameter roles for per-role optimizer configuration (e.g., weight decay exemption for biases).
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
IReadOnlyList<Tensor<T>> GetTrainableParameters() Returns all trainable parameter tensors marked with [TrainableParameter].
void ResetState()
void SetTrainableParameters(IReadOnlyList<Tensor<T>> parameters) Replaces trainable parameter tensors (e.g., with ParameterBuffer views).
void ZeroGrad() Clears all gradient fields discovered by convention ({paramName}Gradient).

Fields

Field Summary
int DecoderStartTokenId BART’s decoder start token (also EOS).