@numrs/wasm
v0.1.23
Published
NumRs WebAssembly bindings for browser and Node.js
Maintainers
rjaguiluzReadme
@numrs/wasm
NumRs WASM is the WebAssembly binding for the NumRs numerical engine, bringing high-performance tensor operations and deep learning to the browser and Node.js.
🚀 Features
- Zero FFI overhead: Direct WASM calls.
- Deep Learning: Full support for Autograd, Neural Networks, and Optimizers.
- Universal: Works in Browser (ESM), Node.js, Deno, and Bundlers.
- TypeScript: Full type definitions included.
📦 Installation
npm install @numrs/wasm🎯 Usage
🌐 Browser (No Bundler)
To use @numrs/wasm directly in the browser without a bundler (like Vite/Webpack), you must use an Import Map to point to the pkg-web (ES Modules) build.
Add this to your HTML <head>:
<script type="importmap">
{
"imports": {
"@numrs/wasm": "./node_modules/@numrs/wasm/pkg-web/numrs_wasm.js"
}
}
</script>
<script type="module">
import init, { Tensor, nn } from '@numrs/wasm';
async function run() {
await init(); // Initialize the WASM module
console.log("NumRs WASM loaded!");
// Create a tensor
let x = Tensor.randn([10, 5]);
// Define a model
let model = new nn.Sequential();
model.add_linear(new nn.Linear(5, 10));
model.add_relu(new nn.ReLU());
model.add_linear(new nn.Linear(10, 2));
// Forward pass
let output = model.forward(x);
console.log("Output shape:", output.shape());
}
run();
</script>📦 Bundlers (Vite, Webpack) & Node.js
import init, { Tensor, Sequential, Linear, ReLU, Trainer } from '@numrs/wasm';
// Initialize WASM
await init();
// 1. Define Model
const model = new Sequential();
model.add_linear(new Linear(10, 32));
model.add_relu(new ReLU());
model.add_linear(new Linear(32, 1));
// 2. Training Loop
const trainer = new Trainer(model, "adam", "mse", 0.01);
// Assuming xTrain and yTrain are Tensors
trainer.fit(xTrain, yTrain, 10);🧠 Optimizers & Loss Functions
NumRs supports a wide range of optimizers and loss functions for training:
| Context | Supported Values |
| :------------- | :--------------------------------------------------------------------------------------------------------------------------------------- |
| Optimizers | "sgd", "adam", "adamw", "nadam", "radam", "rmsprop", "adagrad", "adadelta", "lamb", "adabound", "lbfgs", "rprop" |
| Losses | "mse" (Regression), "cross_entropy" (Classification) |
🔍 Advanced Topics
Time Series (1D CNN)
Use Conv1d for sequence processing. input shape should be [Batch, Channels, Length].
model.add_conv1d(new nn.Conv1d(in_channels, out_channels, kernel_size));
model.add_relu(new nn.ReLU());
model.add_flatten(new nn.Flatten(1, -1));
model.add_linear(new nn.Linear(hidden_size, output_size));ONNX Support
NumRs WASM can export models to JSON representation compatible with web inference.
- Export:
OnnxModelWrapper.export_model_to_json(...) - Inference:
OnnxModelWrapper.load_from_json(...)
📚 Documentation
For full API documentation, please refer to the main NumRs Repository or the detailed docs in DOCS.md.
License
AGPL-3.0-only