Deepbox

The TypeScript Toolkit for AI & Numerical Computing

Deepbox is a zero-runtime-dependency TypeScript framework for tensors, linear algebra, tabular data, machine learning, neural networks, statistics, datasets, and plotting. It is designed for users who want one coherent toolkit instead of stitching together separate numerical and ML libraries.

Docs: deepbox.dev/docs Examples: deepbox.dev/examples Projects: deepbox.dev/projects

Why Deepbox

• Zero runtime dependencies
• ESM and CommonJS builds with bundled type declarations
• Stable subpath exports for each major module
• Broad numerical surface area in a single package
• 315 implementation files under src/**/*.ts excluding *.d.ts, 421 Vitest files matching test/**/*.test.ts, 8,686 tests, 50 example directories, and 9 end-to-end projects in the current v1.0.0 tree (other files under test/ are helpers or benches, not counted here)

Requirements

• Node.js >= 24.13.0 as declared in package.json engines. Deepbox 1.x is built and CI-tested on Node 24.x with a TypeScript ES2024 target; use this line for predictable behavior. Older Node versions are not supported for 1.x.

Installation

npm install deepbox

Import Model

Deepbox is organized around subpath exports. Prefer importing named APIs from the module you actually use:

import { tensor, parameter } from "deepbox/ndarray";
import { LinearRegression } from "deepbox/ml";
import { DataFrame } from "deepbox/dataframe";

The root package exports namespaces, not direct named symbols:

import * as db from "deepbox";

const x = db.ndarray.tensor([1, 2, 3]);
const model = new db.ml.LinearRegression();

Quick Start

Tensors and Autograd

import { parameter, tensor } from "deepbox/ndarray";

const x = parameter([
  [1, 2],
  [3, 4],
]);
const w = parameter([[0.5], [0.25]]);

const y = x.matmul(w).sum();
y.backward();

console.log(x.grad?.toString());
console.log(w.grad?.toString());

const plain = tensor([1, 2, 3]);
console.log(plain.toString());

GPU and WASM Acceleration

Tensors carry a device (cpu, webgpu, wasm). With a registered backend the accelerated op set executes on the device; ops a device cannot run throw a DeviceError with a transfer hint instead of silently computing elsewhere.

The WebGPU device set is training-complete: element-wise arithmetic and activations (incl. gelu, erf, rsqrt, where), matmul and batched matmul (attention), axis reductions (so softmax/logSoftmax/layerNorm compose on-device), 2-D convolution and pooling (incl. MaxPool), and full reductions. The reverse pass (autograd) and every practical optimizer step (SGD, Adam, AdamW, RMSprop, Adagrad, Adamax, Nadam, RAdam, Adadelta, ASGD, Rprop, Lion, LAMB, LARS) also run on the device, so a full forward → backward → update loop for an MLP, transformer, or CNN stays resident on the GPU with no per-step host transfers. Half precision is supported: float16 tensors compute in true on-device half (WGSL shader-f16, halving memory footprint) and bfloat16 carries correct bf16 numerics.

import { registerBackend, WebGpuBackend } from "deepbox/core";
import { dot, relu, tensor } from "deepbox/ndarray";

const gpu = new WebGpuBackend(); // in Node, pass { gpu } from a WebGPU binding
await gpu.init();
if (gpu.info().available) {
  registerBackend("webgpu", gpu);

  const a = tensor([[1, 2], [3, 4]], { device: "webgpu" }); // lives in GPU memory
  const y = relu(dot(a, a));       // WGSL compute kernels
  const host = await y.cpu();      // async readback
  console.log(host.toString());
}

WebGPU kernels are float32/float16 and stride/broadcast-aware (views and transposes execute without copies), verified on GPU hardware against the CPU reference. The WASM backend accelerates contiguous float32 arithmetic with embedded SIMD kernels over zero-copy host storage:

import { registerBackend, WasmBackend } from "deepbox/core";
import { add, tensor } from "deepbox/ndarray";

const wasm = new WasmBackend();
await wasm.init();
if (wasm.info().available) {
  registerBackend("wasm", wasm);
  const a = tensor(new Array(4096).fill(1), { device: "wasm" });
  console.log(add(a, a).at(0)); // SIMD, bit-identical to the CPU result
}

Classical ML

import { tensor } from "deepbox/ndarray";
import { LinearRegression } from "deepbox/ml";

const X = tensor([
  [1],
  [2],
  [3],
  [4],
]);
const y = tensor([2, 4, 6, 8]);

const model = new LinearRegression();
model.fit(X, y);

const predictions = model.predict(tensor([[5], [6]]));
console.log(predictions.toString());

DataFrames

import { DataFrame } from "deepbox/dataframe";

const df = new DataFrame({
  name: ["Alice", "Bob", "Charlie"],
  team: ["A", "A", "B"],
  score: [91, 84, 96],
});

const summary = df.groupBy("team").mean();
console.log(summary.toString());

Modules

v1.0.0 Highlights

Numerical Computing

• Tensor ops spanning arithmetic, broadcasting, reductions, sorting, indexing, signal processing, FFT, and Einstein summation
• Sparse CSR matrices, complex dtypes, half-precision arrays, and NaN-aware reductions
• Linear algebra with SVD, QR, LU, Cholesky, eigensolvers, Schur, polar, Hessenberg, and matrix functions
• Training-complete WebGPU backend: device tensors with PyTorch-style .to(device) transfers, plus axis reductions, batched matmul, convolution/pooling, on-device autograd, and on-device SGD/Adam — MLP, transformer, and CNN training loops run resident on the GPU. WASM SIMD host acceleration for contiguous float32. Strict same-device semantics and loud errors for unaccelerated device ops

Data and Statistics

• DataFrame and Series workflows with grouping, merging, pivoting, rolling, expanding, EWM, string accessors, and datetime tooling
• Statistical distributions, confidence intervals, kernel density estimation, multiple-comparison corrections, and power analysis
• Metrics for classification, regression, clustering, ranking, and pairwise similarity

Machine Learning and Deep Learning

• Expanded estimator surface across ensembles, SVM variants, Naive Bayes, clustering, manifold learning, Gaussian processes, anomaly detection, and model selection
• Pipeline composition with Pipeline, FeatureUnion, ColumnTransformer, GridSearchCV, and RandomizedSearchCV
• Neural network stack with convolutional, recurrent, normalization, attention, transformer, embedding, and utility layers
• Training infrastructure including Trainer, callbacks, clipping, and advanced optimizers and schedulers

Visualization and Data Sources

• Figure-based plotting API with line, scatter, histogram, heatmap, contour, violin, radar, polar, dendrogram, and diagnostic plots
• Real reference datasets (Iris, Wine, Breast Cancer, Diabetes, Digits — values matching scikit-learn), synthetic generators, DataLoader, samplers, and remote dataset helpers
• Streaming / out-of-core datasets: a lazy StreamingDataset with map/shuffle-buffer/batch/prefetch that trains on data larger than RAM via Trainer.fitAsync

For AI Agents

Use SKILL.md as the repo-native agent guide (also included at node_modules/deepbox/SKILL.md when you install from npm). It documents:

• correct import patterns
• module selection guidance
• Deepbox core types
• custom error hierarchy
• common coding patterns and gotchas

Development

npm ci
npm run validate:all

Additional contributor workflow details live in CONTRIBUTING.md. Security reporting instructions live in SECURITY.md.

License

Deepbox is released under the MIT License.