ITrainableLayer<T>
Interfaces · AiDotNet.Interfaces
Defines a neural network layer with trainable parameters that can be used with tape-based automatic differentiation (autodiff).
Implements: ILayer<T>
Known implementations: LayerBase<T>
For Beginners
When training a neural network, we need two things:
- Access to the layer’s learnable values (weights and biases) so the optimizer
can update them after computing gradients.
- A way to clear old gradient information before each training step, so gradients
from different batches don’t accumulate incorrectly.
This interface provides both capabilities, enabling the PyTorch-style training loop:
How It Works
This interface bridges layers with GradientTape
by exposing the exact Tensor instances that the layer uses during
its forward pass. The gradient tape tracks operations by tensor reference identity,
so the tensors returned here must be the same objects passed to engine operations
in Forward.
PyTorch equivalent: This combines the role of nn.Module.parameters()
(which yields the same Parameter objects used in forward) and
optimizer.zero_grad() (which clears .grad on each parameter).
Methods
| Method | Summary |
|---|---|
IReadOnlyList<Tensor<T>> GetTrainableParameters() |
Returns the trainable parameter tensors used by this layer during forward execution. |
void SetTrainableParameters(IReadOnlyList<Tensor<T>> parameters) |
Replaces this layer’s trainable parameter tensors with the provided tensors. |
void ZeroGrad() |
Clears all accumulated gradients on this layer’s trainable parameters. |