I am following this TensorFlow JS tutorial where you load car data. The data looks like this:

[{x:100, y:20}, {x:80, y:33}]

X is the horsepower of a car, Y is the expected miles per gallon usage. After creating the model I save it locally using:

async function saveModel(){
    await model.save('downloads://cars-model');

Next, I load the model in a separate project, to make predictions without needing the original data.


async function app(){
    let model = await tf.loadLayersModel('./cars-model.json');
    console.log("car model is loaded!");

I expect to be able to run predict here, on a single number (say, 120)

model.predict(tf.tensor2d([120], [1, 1]))


I think the number 120 needs to be normalised to a number between 0-1, just like the training data was. But how do I know the inputMin, inputMax, labelMin, labelMax values from the loaded model?

To un-normalise the prediction (in this case 0.6) I also need those original values.

How do I normalise/un-normalise data when loading a model?

original prediction code uses label and input values from the original data

function testModel(model, inputData, normalizationData) {
    const { inputMax, inputMin, labelMin, labelMax } = normalizationData;

    // Generate predictions for a uniform range of numbers between 0 and 1;
    // We un-normalize the data by doing the inverse of the min-max scaling 
    // that we did earlier.
    const [xs, preds] = tf.tidy(() => {

        const xs = tf.linspace(0, 1, 100);
        const preds = model.predict(xs.reshape([100, 1]));

        const unNormXs = xs

        const unNormPreds = preds

        // Un-normalize the data
        return [unNormXs.dataSync(), unNormPreds.dataSync()];

    const predictedPoints = Array.from(xs).map((val, i) => {
        return { x: val, y: preds[i] }


How do I convert this to a number between 0-1 without having access to the original car data?

You save the normalisation parameters (typically an offset and a multiplier for each column), and consider that part of the model. Typically you do this when you originally scale training data.

When you want to re-use the model, as well as loading the neural network architecture and weights, you need to load the normalisation parameters in order to re-use those too and scale your inputs.

When tutorials present a self-contained neural network that loads training data, builds a model, then tests that model, all in the same process, then often this step is not shown. However, saving the normalisation data is important, and basically should be considered part of the model, even though it is not directly part of the neural network parameters or hyper-parameters.

  • $\begingroup$ OK, thanks for clearing that up. So do I save that data manually, in a separate file, or should it be part of the saved model? $\endgroup$ – Kokodoko Apr 1 '19 at 13:08
  • $\begingroup$ @Kokodoko: It's up to you as the developer which you prefer, and may depend on your use case. For instance if you always use the same training data, scaling it the same way and want to move different NNs around for testing, then separate files will be slightly easier and quicker to develop. But a self-contained file with supporting data for the model might make more sense in other situations $\endgroup$ – Neil Slater Apr 1 '19 at 13:18
  • $\begingroup$ But if I use model.save, is there an option to include that data? Can I add it manually before saving? For example, model.min = 100; model.max = 200. $\endgroup$ – Kokodoko Apr 1 '19 at 14:55
  • $\begingroup$ @Kokodoko: I don't know, that depends on how function save is written $\endgroup$ – Neil Slater Apr 1 '19 at 14:56
  • 1
    $\begingroup$ From the description it does not look like it will accept added data, so you will need to arrange to store the normalisation parameters separately $\endgroup$ – Neil Slater Apr 1 '19 at 17:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.