Tensors, Models & Inference

Tensors & Backends

import * as tf from '@tensorflow/tfjs';

const t = tf.tensor([1, 2, 3, 4], [2, 2]);
console.log(await t.array());  // async — the WebGL backend may not have
                                 // finished the GPU computation yet

// Backends: cpu (plain JS), webgl (GPU, browser), wasm, or tfjs-node
// (native TensorFlow C++ binding — far faster than WebGL for server use)
await tf.setBackend('webgl');
console.log(tf.getBackend());

Loading & Running a Pretrained Model

const model = await tf.loadLayersModel('model.json');
// Python-trained models need tensorflowjs_converter first:
// tensorflowjs_converter --input_format=keras model.h5 web_model/

const input = tf.tensor2d([[0.5, 0.2, 0.1, 0.9]]);
const prediction = model.predict(input);
console.log(await prediction.data());

input.dispose();
prediction.dispose();  // WebGL memory is NOT garbage-collected by JS —
                       // forgetting this leaks GPU memory over time

Memory Management with tf.tidy

// In a per-frame webcam inference loop, unmanaged intermediate tensors
// accumulate fast — tf.tidy auto-disposes everything except the return value
function classifyFrame(model, videoElement) {
  return tf.tidy(() => {
    const img = tf.browser.fromPixels(videoElement);
    const resized = tf.image.resizeBilinear(img, [224, 224]);
    const normalized = resized.div(255.0).expandDims(0);
    return model.predict(normalized);  // only this survives tf.tidy's cleanup
  });
}

// Without tf.tidy, a webcam demo running at 30fps can crash the tab
// within minutes from GPU memory exhaustion.