kotobuki/grad-cam-wip.markdown

## grad-cam-wip.markdown

      
    Raw
  

              grad-cam-wip.markdown
            
          
    Grad-CAM [WIP]

A Pen by Shigeru Kobayashi on CodePen.
License.

  
## index.html
<!DOCTYPE html>
<html>

<head>
  <meta charset="utf-8" />
</head>

<body>
  <h1></h1>
  <img id="bird" src="https://ml5js.org/docs/assets/img/bird.jpg" crossorigin="anonymous" />
</body>

</html>

## print-a-summary-for-a-mobilenet-model.markdown

      
    Raw
  

              print-a-summary-for-a-mobilenet-model.markdown
            
          
    Print a Summary for a MobileNet Model

A Pen by Shigeru Kobayashi on CodePen.
License.

  
## script.js
console.clear();

let sourceImg;

function setup() {
  sourceImg = select("#bird");
  sourceImg.class("img").size(224, 224);
  loadMobilenet();
}

async function loadMobilenet() {
  const model = await tf.loadModel(
    "https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_0.25_224/model.json"
  );

  model.summary();

  // Reference:
  // https://github.com/Ankush96/grad-cam.tensorflow/blob/master/main.py#L29-L62
  const nbClasses = 1000;
  model.predict(imgToTensor(sourceImg.elt, [224, 224])).print();
  const predictedClass = model
    .predict(imgToTensor(sourceImg.elt, [224, 224]))
    .as1D()
    .argMax();
  let classId = (await predictedClass.data())[0];
  const convLayer = model.getLayer("conv_pw_13");
  const oneHot = tf.oneHot(tf.tensor1d([classId], "int32"), nbClasses);
  const signal = tf.mul(
    model.getLayer("reshape_2").output,
    oneHot.asType("float32")
  );

  // TODO: Implement following lines...
}

// From: https://github.com/ml5js/ml5-library/blob/5756876f3647cab6ff657dd0db4ec92ed668796e/src/utils/imageUtilities.js#L64-L75
function imgToTensor(input, size = null) {
  return tf.tidy(() => {
    let img = tf.fromPixels(input);
    if (size) {
      img = tf.image.resizeBilinear(img, size);
    }
    // const croppedImage = cropImage(img);
    const batchedImage = img.expandDims(0);
    return batchedImage
      .toFloat()
      .div(tf.scalar(127))
      .sub(tf.scalar(1));
  });
}

## scripts
<script src="https://cdnjs.cloudflare.com/ajax/libs/tensorflow/0.12.5/tf.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/p5.js/0.7.1/p5.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/p5.js/0.7.1/addons/p5.dom.min.js"></script>
	<!DOCTYPE html>
	<html>

	<head>
	<meta charset="utf-8" />
	</head>

	<body>
	<h1></h1>
	<img id="bird" src="https://ml5js.org/docs/assets/img/bird.jpg" crossorigin="anonymous" />
	</body>

	</html>
	console.clear();

	let sourceImg;

	function setup() {
	sourceImg = select("#bird");
	sourceImg.class("img").size(224, 224);
	loadMobilenet();
	}

	async function loadMobilenet() {
	const model = await tf.loadModel(
	"https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_0.25_224/model.json"
	);

	model.summary();

	// Reference:
	// https://github.com/Ankush96/grad-cam.tensorflow/blob/master/main.py#L29-L62
	const nbClasses = 1000;
	model.predict(imgToTensor(sourceImg.elt, [224, 224])).print();
	const predictedClass = model
	.predict(imgToTensor(sourceImg.elt, [224, 224]))
	.as1D()
	.argMax();
	let classId = (await predictedClass.data())[0];
	const convLayer = model.getLayer("conv_pw_13");
	const oneHot = tf.oneHot(tf.tensor1d([classId], "int32"), nbClasses);
	const signal = tf.mul(
	model.getLayer("reshape_2").output,
	oneHot.asType("float32")
	);

	// TODO: Implement following lines...
	}

	// From: https://github.com/ml5js/ml5-library/blob/5756876f3647cab6ff657dd0db4ec92ed668796e/src/utils/imageUtilities.js#L64-L75
	function imgToTensor(input, size = null) {
	return tf.tidy(() => {
	let img = tf.fromPixels(input);
	if (size) {
	img = tf.image.resizeBilinear(img, size);
	}
	// const croppedImage = cropImage(img);
	const batchedImage = img.expandDims(0);
	return batchedImage
	.toFloat()
	.div(tf.scalar(127))
	.sub(tf.scalar(1));
	});
	}
	<script src="https://cdnjs.cloudflare.com/ajax/libs/tensorflow/0.12.5/tf.min.js"></script>
	<script src="https://cdnjs.cloudflare.com/ajax/libs/p5.js/0.7.1/p5.min.js"></script>
	<script src="https://cdnjs.cloudflare.com/ajax/libs/p5.js/0.7.1/addons/p5.dom.min.js"></script>