mwdchang/index.html

## index.html
<!DOCTYPE html>
<html lang="en">
<head>
  <script src="tf.min.js"></script>
</head>
<body>
</body>
<script>

const LEARNING_RATE = 0.07;
const optimizer = tf.train.sgd(LEARNING_RATE);

// Variable and weights definition
const L1w = tf.variable(tf.randomNormal([2, 2]));
const L1b = tf.variable(tf.randomNormal([2]));

const L2w = tf.variable(tf.randomNormal([2, 5]));
const L2b = tf.variable(tf.randomNormal([5]));

const L3w = tf.variable(tf.randomNormal([5, 2]));
const L3b = tf.variable(tf.randomNormal([2]));

const model = (xs) => {
  const l1 = tf.tanh(xs.matMul(L1w).add(L1b));
  const l2 = tf.tanh(l1.matMul(L2w).add(L2b));
  const l3 = tf.tanh(l2.matMul(L3w).add(L3b));
  return l3;
};

// Loss function
const loss = (labels, ys) => {
  // return tf.losses.meanSquaredError(labels, ys).mean();
  return tf.losses.softmaxCrossEntropy(labels.asType('float32'), ys).mean();
};


// Dummy data if the product is even (0) or odd (1)
const createTrainBatch = (num) => {
  const testData = [];
  const testLabel = [];
  for (let i=0; i < num; i++) {
    const a = 1 + Math.floor(Math.random()*10);
    const b = 1 + Math.floor(Math.random()*10);
    testData.push(a);
    testData.push(b);

    if (a > b) {
      testLabel.push(0);
    } else {
      testLabel.push(1);
    }
  }
  return [tf.tensor(testData).as2D(-1, 2), tf.oneHot(tf.tensor1d(testLabel, 'int32'), 2), testLabel];
};


// Traning with SGD/backprop
async function train() {
  L2w.print();
  for (let i=0; i < 50; i++) {
    optimizer.minimize(() => {
      const [input, label] = createTrainBatch(40);
      return loss(label, model(input));
    }, true);
    await tf.nextFrame();
  }
}


// Predict
async function run() {
  await train();
  const axis = 1;
  const [input, label, origLabel] = createTrainBatch(10);
  console.log('A', origLabel);
  console.log('P', Array.from(model(input).argMax(axis).dataSync()));
}

run();


</script>
</html>
	<!DOCTYPE html>
	<html lang="en">
	<head>
	<script src="tf.min.js"></script>
	</head>
	<body>
	</body>
	<script>

	const LEARNING_RATE = 0.07;
	const optimizer = tf.train.sgd(LEARNING_RATE);

	// Variable and weights definition
	const L1w = tf.variable(tf.randomNormal([2, 2]));
	const L1b = tf.variable(tf.randomNormal([2]));

	const L2w = tf.variable(tf.randomNormal([2, 5]));
	const L2b = tf.variable(tf.randomNormal([5]));

	const L3w = tf.variable(tf.randomNormal([5, 2]));
	const L3b = tf.variable(tf.randomNormal([2]));

	const model = (xs) => {
	const l1 = tf.tanh(xs.matMul(L1w).add(L1b));
	const l2 = tf.tanh(l1.matMul(L2w).add(L2b));
	const l3 = tf.tanh(l2.matMul(L3w).add(L3b));
	return l3;
	};

	// Loss function
	const loss = (labels, ys) => {
	// return tf.losses.meanSquaredError(labels, ys).mean();
	return tf.losses.softmaxCrossEntropy(labels.asType('float32'), ys).mean();
	};


	// Dummy data if the product is even (0) or odd (1)
	const createTrainBatch = (num) => {
	const testData = [];
	const testLabel = [];
	for (let i=0; i < num; i++) {
	const a = 1 + Math.floor(Math.random()*10);
	const b = 1 + Math.floor(Math.random()*10);
	testData.push(a);
	testData.push(b);

	if (a > b) {
	testLabel.push(0);
	} else {
	testLabel.push(1);
	}
	}
	return [tf.tensor(testData).as2D(-1, 2), tf.oneHot(tf.tensor1d(testLabel, 'int32'), 2), testLabel];
	};


	// Traning with SGD/backprop
	async function train() {
	L2w.print();
	for (let i=0; i < 50; i++) {
	optimizer.minimize(() => {
	const [input, label] = createTrainBatch(40);
	return loss(label, model(input));
	}, true);
	await tf.nextFrame();
	}
	}


	// Predict
	async function run() {
	await train();
	const axis = 1;
	const [input, label, origLabel] = createTrainBatch(10);
	console.log('A', origLabel);
	console.log('P', Array.from(model(input).argMax(axis).dataSync()));
	}

	run();


	</script>
	</html>