public class LayerGradients<T extends AbstractFloat = Float>
Explicit backward result, with parameter derivatives in layer parameter order. Arrays are retained, not copied. No parameter update has been applied.
gradients = layer.differentiate(batch, upstream)
| Modifier and Type | Member | Description |
|---|---|---|
| public readonly property | inputsinputs as NDArray<T> |
Input derivative passed to the preceding layer. |
| public readonly property | parameters | Parameter derivatives in the order defined by Layer.parameter. |
| Modifier and Type | Member | Description |
|---|---|---|
| public | LayerGradientsLayerGradients(inputs as NDArray<T>, parameters as IList<NDArray<T>>) |
Groups derivatives and snapshots their ordering, not their data. |