How to run a pre-trained model in TensorFlow.js

The tfjs-models repository has several JavaScript libraries that expose a pre-trained model as a friendly JavaScript API. Here I show what’s going on internally in those libraries.

BlazeFace will be my example, as it is perhaps the smallest of those libraries. It’s “a lightweight model that detects faces in images”. It has a single method, estimateFaces, that takes an image and returns a promise of an array of “face landmarks”.

The BlazeFace library is a thin wrapper around TensorFlow.js. Here I show the essence of the BlazeFace wrapper using just the TensorFlow.js API:

// 1: Load the model
const blazefaceGraphModel = await tf.loadGraphModel(
  'https://tfhub.dev/tensorflow/tfjs-model/blazeface/1/default/1', { fromTFHub: true });

async function estimateFaces(inputImage) {
  // 2: Convert input to a tensor with `fromPixels`
  const inputImageTensor = tf.expandDims(tf.cast(tf.browser.fromPixels(inputImage), 'float32'), 0);

  // 3: Convert image tensor to size and format expected by pre-trained model
  const resizedImage = tf.image.resizeBilinear(inputImageTensor, [WIDTH, HEIGHT]);
  const normalizedImage = tf.mul(tf.sub(tf.div(resizedImage, 255), 0.5), 2);

  // 4: Run inference with .predict()
  const prediction = tf.squeeze(blazefaceGraphModel.predict(normalizedImage));

  // 5: Decode the prediction tensor into friendly JavaScript objects
  // ... omitted for brevity ...
  return faces;
}

// Now to use it:
const imgEl = document.getElementById("someImage");
const faces = await estimateFaces(imgEl);
console.log(faces);

This shows the basic pattern of nearly all model wrapper libraries:

1. Load the pre-trained model. In this case, the Blazeface model is in the undocumented TensorFlow.js “web-friendly format”, so we use tf.loadGraphModel.
2. Convert the input to a tensor. Most models work on images, so we use tf.browser.fromPixels. It returns a 3D tensor with shape [height, width, numChannels], with RGB values in the range [0, 255]. (Note the surprising shape: height is the first dimension!)
3. Convert the input tensor to the format expected by the model. The Blazeface model only works on 128×128 square images, but input images (e.g. from a webcam) can be any size and shape. The Blazeface library chooses to squish the whole input image into a 128×128 square. This destroys the aspect ratio, but appears to not affect accuracy (presumably, the model has been trained on various squished images). The Blazeface model also expects RGB channels in the range [-1, 1], so we have to convert from the range [0, 255].
4. Run the model on the input tensor with .predict. We get back a prediction tensor.
5. Decode the prediction into friendly JavaScript objects. This at least requires running .array() on tensors to get back JavaScript values. For Blazeface, the decoding is a bit involved, because the model architecture is a “Single Shot Multibox Detector”. To see how this works, read the next post.

This page copyright James Fisher 2021. Content is not associated with my employer. Found an error? Edit this page.

Jim Fisher

CV

Speaking

Blogroll

RSS

TigYog

Kickabout