ScratchAttribute
Attributes · AiDotNet.Attributes
Marks a tensor field as transient working state: not trained, not counted, not saved.
Inheritance: Attribute → ScratchAttribute
For Beginners
if deleting this tensor and rebuilding it on the next forward pass would lose nothing, it is scratch.
How It Works
Scratch is everything a layer keeps for its own convenience — a cached activation held for the backward pass, an intermediate buffer reused across calls, a workspace for a fused kernel. None of it is part of the model; recreating it from scratch after a load changes nothing.
The framework deliberately does not infer a role from Tensor<T>. A cache and a weight have
the same storage type, so inference would silently hand activations to an optimizer. This
declaration tells the shared analyzer and generators that the storage is transient.
Nullability is only a storage fact and does not imply scratch. Apply [Scratch] to cached
activations and workspaces whether nullable or non-nullable. A matching generated gradient field
is the one narrow convention recognized because its owning trainable parameter supplies the
semantic relationship.
Constructors
| Constructor | Summary |
|---|---|
ScratchAttribute() |