A TypeScript neural network framework

Write the graph.
Compile the work.

TypeNN gives Tensor and Module code a deferred graph. TypeShade compiles its kernels and executes them on CPU and supported GPU backends.

Try the playground Read the guide

Pre-release · source and examples are the current usage reference.

Tensor APIexample.ts

import * as nn from "typenn";

const device = await nn.device();
const input = device.tensor([1, 2, 3, 4], { shape: [2, 2] });
const weight = device.tensor([5, 6, 7, 8], { shape: [2, 2] });
const output = input.matmul(weight).relu();

console.log(await output.data());
// Float32Array [19, 22, 43, 50]
await output.disposeGraph();
await device.dispose();

Float32Array[19, 22, 43, 50]Example output

From TypeScript to execution

Three layers, one visible path.

How execution works

Define

Tensor calls describe shapes and operations. TypeNN records them as graph nodes.

tensor.matmul(weight).relu()

Lower

Each node holds a TypeShade kernel function that TypeNN calls on resident arrays.

Tensor graph → kernel functions

Run

One fuse scope runs the reachable kernels on the selected execution tier.

CPU · WebGPU · WebGL2*

* WebGL2 and the CPU tier run the same kernel functions apart. Only WebGPU fuses an elementwise chain into one dispatch.

Explore TypeNN

Start with a runnable operation. Continue into the system.

Try matrix operations in the browser, then follow the guide into graph execution, automatic differentiation, and model training.

PlaygroundRun real TypeNN operations in this browser.Train a modelBuild the documented MNIST MLP path.API referenceBrowse the current Tensor and Module surface.ExamplesOpen complete source examples and current limits.

Current implementation

Every op is a kernel function. Training is not tuned yet.

Each data() runs as one TypeShade fuse scope, and device.lastReport shows its plan. Optimizers still update host arrays, so Adam is not a device kernel and tiled MatMul remains future work.

Read implementation status