Module mlp 

Multi-layer perceptron (MLP) core.

The low-level API is intentionally allocation-free:

• Mlp::forward writes activations into a reusable Scratch and returns a slice.
• Mlp::backward writes gradients into a reusable Gradients.

Shape mismatches are treated as programmer error and will panic via assert!.

Structs

BatchBackpropScratch

Reusable buffers for Mlp::backward_batch.

BatchScratch

Reusable buffers for Mlp::forward_batch.

Gradients

Parameter gradients for an Mlp (overwrite semantics).

Mlp

A feed-forward multi-layer perceptron composed of dense layers.

Scratch

Reusable buffers for Mlp::forward.

Trainer

Reusable buffers for training a specific Mlp.