@numrs/node

NumRs Node is the native Node.js binding for the NumRs numerical engine. It provides high-performance, hardware-accelerated (Metal/Accelerate on macOS, MKL/OpenBLAS elsewhere) tensor operations and deep learning capabilities.

🚀 Features

• Hardware Acceleration: Uses Metal on Apple Silicon for GPU-like performance.
• Multithreading: Native parallelization via Rayon.
• Zero-Copy: Efficient memory sharing between Node.js and Rust.
• Deep Learning: Full Autograd engine and Neural Network modules.
• Type Safe: Full TypeScript support.

📦 Installation

npm install @numrs/node

Note: Requires Rust to be installed on your system to compile the native extension.

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

🎯 Usage

Quick Start: Tensors & Autograd

import { Tensor, nn } from '@numrs/node';

// Create tensors with autograd enabled
const x = Tensor.randn([2, 5], true); 
const w = Tensor.randn([5, 10], true);

// Fast matrix multiplication
const y = x.matmul(w);

// Backward pass
y.backward();
console.log("Gradient:", x.grad());

Deep Learning Model

import { Sequential, nn } from '@numrs/node';

const model = new Sequential();
model.add_linear(new nn.Linear(10, 32));
model.add_relu(new nn.ReLU());
model.add_linear(new nn.Linear(32, 2));

// Forward pass
const input = Tensor.randn([100, 10]);
const output = model.forward(input);

Training Loop

Use the TrainerBuilder for a highly optimized, native training loop.

import { TrainerBuilder, Dataset } from '@numrs/node';

// 1. Prepare Data
const input = Tensor.randn([100, 10]);
const target = Tensor.randn([100, 2]);
const dataset = new Dataset(input, target);

// 2. Train
const trainer = new TrainerBuilder(model, dataset)
    .batch_size(16)
    .learning_rate(0.01)
    .optimizer("adam") // Supports: adam, sgd, rmsprop, etc.
    .loss("mse")       // Supports: mse, cross_entropy
    .max_epochs(50)
    .build();

const history = trainer.fit();
console.log("Final Loss:", history.final_loss);

🧠 Optimizers & Loss Functions

| Category | Supported Algorithms | | :----------------- | :--------------------------------------------------------------------------------------------------------------------------------------- | | Optimizers | "sgd", "adam", "adamw", "nadam", "radam", "rmsprop", "adagrad", "adadelta", "lamb", "adabound", "lbfgs", "rprop" | | Loss Functions | "mse" (Regression), "cross_entropy" (Classification) |

🔍 Advanced Features

Zero-Copy Float32Array Ops

For raw numerical computing, you can operate directly on Float32Array buffers without Tensor overhead.

const numrs = require('@numrs/node');
const a = new Float32Array([1, 2, 3, 4]);
const b = new Float32Array([5, 6, 7, 8]);
// Element-wise addition
const c = numrs.add(a, [2, 2], b, [2, 2]); 

ONNX Export

Export your trained models to ONNX for verifying graphs or interoperability.

model.save_onnx(dummy_input, "model.onnx");

📚 Documentation

For full API documentation, please refer to the main NumRs Repository.

License

AGPL-3.0-only