🧱 @dxo/nn

Developer preview — API unstable (0.0.x).

Pure TypeScript modules over @dxo/core.

Contract (G3 preview)

| Type | Behavior | |---------------------------------|--------------------------------------------------------------------| | Linear(in, out) | y = x @ W + b only — no implicit ReLU | | Relu / relu(x) | Explicit nonlinearity | | Sequential(layers) | Chains forward | | parameters() | Collects Tensor fields (and nested modules) in enumeration order | | zeroGrad() | Calls zeroGrad on each parameter leaf | | Linear.loadParameters([W, b]) | Reassign leaves after optimizer.step | | Linear.state / loadState | Plain { shape, data } snapshots |

import {Linear, Relu, Sequential} from '@dxo/nn';
import {SGD} from '@dxo/optimizer';

const model = new Sequential([new Linear(2, 4), new Relu(), new Linear(4, 1)]);
const opt = new SGD(0.05);
// ... loss.backward()
model.layers[0] // update Linear leaves via loadParameters after step

Training step pattern:

model.zeroGrad();
const loss = /* scalar Tensor */;
loss.backward();
const linear = model.layers[0] as Linear;
linear.loadParameters(opt.step(linear.parameters()));