TensorFlow.js

Preamble

This tutorial is about basic and advanced features of TensorFlow.js. This JavaScript software library a superset of the native WebGL API, which is available in modern browsers. Applications illustrating this tutorial have been designed from TensorFlow.js ver. 144. Accordingly, the code in this tutorial will probably not work for older versions.

For convenience, all applications are written in TypeScript to benefit from type checking since TensorFlow.js provides a TypeScript support.

Headlines

Installation and basic execution of TensorFlow.js

Rule(s)

Availability of TensorFlow.js in applications either relies on the prior installation of Node.js (and thus npm) or by directly downloading the full package from threejs.org. In the former case, TensorFlow.js requires execution of the npm i shell command from the package.json location. This file's content establishes the dependencies from TensorFlow.js as follows:
```
"dependencies": {
    …
    "three": "^0.144.0"
},
```
Common installation in browsers (index.html file) requires the inclusion of library files plus application files that reuse libraries. Recently, TensorFlow.js may be (re)used by means of the export and import JavaScript clauses, but this approach is appropriate for very new applications built from scratch.
Common execution leads to a template JavaScript as shown below.

Example (HTML)

<script src="js/lib/three.js"></script>

Example (JavaScript)

Data pre-processing

Fundamentals of TensorFlow.js

Train versus test versus validation data

Rule(s)

Train data are key inputs of the tf.LayersModel.fit instance method. By definition, neuronal network (i.e., model) fitting changes weights (randomly defined at neuronal network construction time). After (enough) epochs (epoch number ⧴ hyper-parameter), fitting aims at providing a neuronal network as a reliable calculator for upcoming predictions.

Test data are key inputs of the tf.LayersModel.evaluate instance method. When fitted, a neuronal network allows predictions, which should lead to test data. tf.LayersModel.evaluate executes the loss function without, compared to tf.LayersModel.fit, changing weights. activation function, e.g., sigmoïde function

Validation data are key inputs of hyper-parameter tuning. tf.LayersModel.fit can be parameterized such that a percentage of train data is not used ⧴ these are validation data.

Model compilation: loss function: error measurement optimization function like Stochastic Gradient Descent (SGD) changes weights.

Application(s)

Convulational Neural Network, CNN, ConvNet p.343 (Aurélien VANNIEUWENHUYZE

Example

Given front as a blue half sphere, compute mouth_geometry (in pink: raw location of a hypothetical mouth) from scratch.

Listing 1 (copy of a subset of faces from source geometry, i.e., front.geometry)
Listing 2 (copy of a subset of vertices from source geometry and consistently rearrange dependency between faces and vertices in target geometry , i.e., mouth_geometry)

TensorFlow.js

Rule(s)

Example (HTML)

Example (JavaScript)

Rule(s)

Train versus test versus validation data

Rule(s)

Application(s)

Example

Listing 1 (copy of a subset of faces from source geometry, i.e., `front.geometry`)

Listing 2 (copy of a subset of vertices from source geometry and consistently rearrange dependency between faces and vertices in target geometry , i.e., `mouth_geometry`)

TensorFlow.js

Rule(s)

Example (HTML)

Example (JavaScript)

Rule(s)

Train versus test versus validation data

Rule(s)

Application(s)

Example

Listing 1 (copy of a subset of faces from source geometry, i.e., front.geometry)

Listing 2 (copy of a subset of vertices from source geometry and consistently rearrange dependency between faces and vertices in target geometry , i.e., mouth_geometry)

Listing 1 (copy of a subset of faces from source geometry, i.e., `front.geometry`)

Listing 2 (copy of a subset of vertices from source geometry and consistently rearrange dependency between faces and vertices in target geometry , i.e., `mouth_geometry`)