public override backwardTraining(context as LayerPass<T>, upstream as NDArray<T>) as LayerGradients<T>
Reuses exactly the keep flags of the corresponding forward pass.
| Parameter | Description |
|---|---|
context | Context generated by this layer. |
upstream | Same-shaped output gradient. |
Independent input derivative, or upstream unchanged when rate=0.
dx = dropout.backwardTraining(context, upstream).inputs