TensorFlow.js Guide

TensorFlow.js is unique because it allows training in the browser, not just inference.

Key Feature: Transfer Learning

You can take a pre-trained model (like MobileNet) and retrain just the last layer to recognize your specific images (e.g., "Hot Dog vs Not Hot Dog") in seconds.

import * as tf from '@tensorflow/tfjs';
import * as mobilenet from '@tensorflow-models/mobilenet';

// 1. Load the base model
const baseModel = await mobilenet.load();

// 2. Create a classifier (KNN)
const classifier = knnClassifier.create();

// 3. Add examples
const img0 = tf.browser.fromPixels(document.getElementById('hotdog'));
classifier.addExample(baseModel.infer(img0, true), 'hot_dog');

const img1 = tf.browser.fromPixels(document.getElementById('not_hotdog'));
classifier.addExample(baseModel.infer(img1, true), 'not_hot_dog');

// 4. Predict
const result = await classifier.predictClass(baseModel.infer(webcamElement, true));

Model Conversion

If you trained a model in Python (Keras/PyTorch), convert it to JSON format for the web.

pip install tensorflowjs
tensorflowjs_converter --input_format=keras model.h5 /web_model

When to use?

• When you need to fine-tune on user data without uploading it.
• When using older, standard architectures (CNNs, LSTMs) rather than Transformers.