TensorFlow.js Tutorial

Javascript variables

tf: Reference to the TensorFlow.js library
tfvis: Reference to the tfjs-vis library

tfjs-vis

tfjs-vis is a small library for in browser visualization intended for use with TensorFlow.js. Its main features are:

A set of visualizations useful for visualizing model behaviour
A set of high level functions for visualizing objects specific to TensorFlow.js
A way to organize visualizations (the visor) of model behaviour that won't interfere with your web application

The library also aims to be flexible and make it easy for you to incorporate custom visualizations using tools of your choosing, such as d3, Chart.js or plotly.js.

Model Architecture

Model architecture is just a fancy way of saying "which functions will the model run when it is executing", or alternatively "what algorithm will our model use to compute its answers". ML models are algorithms that take an input and produce an output. When using neural networks, the algorithm is a set of layers of neurons with ‘weights' (numbers) governing their output. The training process learns the ideal values for those weights.

Training

Shuffle the data

Shuffling is important because typically during training the dataset is broken up into smaller subsets, called batches, that the model is trained on. Shuffling helps each batch have a variety of data from across the data distribution

You should always shuffle your data before handing it to the training algorithms in TensorFlow.js

Normalize the data

We normalize the data. Here we normalize the data into the numerical range 0-1 using min-max scaling. Normalization is important because the internals of many machine learning models you will build with tensorflow.js are designed to work with numbers that are not too big. Common ranges to normalize data to include 0 to 1 or -1 to 1. You will have more success training your models if you get into the habit of normalizing your data to some reasonable range.

You should always consider normalizing your data before training. Some datasets can be learned without normalization, but normalizing your data will often eliminate a whole class of problems that would prevent effective learning. You can normalize your data before turning it into tensors. We do it afterwards because we can take advantage of vectorization in TensorFlow.js to do the min-max scaling operations without writing any explicit for loops.

Takeaways

Formulate your task:

Is it a regression problem or a classification one?
Can this be done with supervised learning or unsupervised learning?
What is the shape of the input data? What should the output data look like?

Prepare your data:

Clean your data and manually inspect it for patterns when possible
Shuffle your data before using it for training
Normalize your data into a reasonable range for the neural network. Usually 0-1 or -1-1 are good ranges for numerical data.
Convert your data into tensors

Build and run your model:

Define your model using tf.sequential or tf.model then add layers to it using tf.layers.*
Choose an optimizer ( adam is usually a good one), and parameters like batch size and number of epochs.
Choose an appropriate loss function for your problem, and an accuracy metric to help your evaluate progress. meanSquaredError is a common loss function for regression problems.
Monitor training to see whether the loss is going down

Evaluate your model

Choose an evaluation metric for your model that you can monitor while training. Once it's trained, try making some test predictions to get a sense of prediction quality.