StopGradientLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
Passes its input through unchanged in the forward direction while blocking gradient flow backwards — the layer form of a stop-gradient / detach barrier.
Inheritance: LayerBase<T> → StopGradientLayer<T>
Implements: IShapeContract
How It Works
Useful wherever an architecture deliberately trains two components as independent functional units inside a single sequential stack. The canonical example is ABINet’s Autonomous principle (Fang et al., CVPR 2021, arXiv:2103.06495), which blocks gradient flow between the vision model and the language model so the language model is forced to learn explicit language modelling rather than becoming an extension of the visual features. The paper measures roughly 0.9% worse accuracy when that gradient is allowed to flow.
Forward is the identity, so inference is completely unaffected; only the backward pass
changes. Implemented with Engine.StopGradient so the tape records the barrier itself
rather than the layer silently detaching outside the graph.
Constructors
| Constructor | Summary |
|---|---|
StopGradientLayer() |
Initializes a new stop-gradient barrier. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
Derived from this type’s [TensorLayout] declarations. |
void ResetState() |