Florence2ProjectorLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
Florence-2’s visual projector (HF Florence2MultiModalProjector). It turns DaViT’s last feature map
into the image tokens that are spliced into the language encoder.
Inheritance: LayerBase<T> → Florence2ProjectorLayer<T>
Implements: IShapeContract
How It Works
The steps are:
- Add a learned 2-D position embedding. Its first half comes from a column table and its second half from
a row table.
- Flatten the map and add the 1-D cosine temporal embedding of frame 0.
- Prepend the mean over all positions as one extra token.
- Project with a bias-free linear layer, then apply LayerNorm. The result is
1 + h*wtokens.
Constructors
| Constructor | Summary |
|---|---|
Florence2ProjectorLayer(int visionDim, int projectionDim, int grid, int maxPositions) |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
|
int TokenCount { get; } |
Number of image tokens produced: the mean token plus one per position. |
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). |