Oxyzen_node NPM

node-tflite

node-tflite is an unofficial TensorFlow Lite 2.2 bindings for Node.js.

It can run TensorFlow Lite models (.tflite) in Node.js environment (including Electron).

Differences to TensorFlow.js

node-tflite can run TensorFlow Lite models (.tflite) while TensorFlow.js cannot (you'll need a different way of model conversion in TensorFlow.js)
node-tflite may or may not be faster than TensorFlow.js
- I didn't run any benchmarks yet, but the example below runs faster than TF.js
node-tflite only supports model inference, not training
node-tflite doesn't support Web environments
node-tflite doesn't support GPU execution (now) while TensorFlow.js supports through WebGL or tfjs-node-gpu

Supported Platforms

macOS
Windows
Linux

Install

npm install node-tflite

Use

import { Interpreter } from "node-tflite";

const modelData = fs.readFileSync("/path/to/model.tflite");
const interpreter = new Interpreter(modelData);

interpreter.allocateTensors();

interpreter.inputs[0].copyFrom(inputData);

interpreter.invoke();

interpreter.outputs[0].copyTo(outputData);

Examples

BlazeFace face detection in Electron
- Uses the BlazeFace model from MediaPipe
- It runs in 60 FPS in MacBook Pro 16'' 2019, which is faster than BlazeFace TF.js demo (around 40 FPS in both wasm and WebGL)