The project is based on the course I took at Udemy Machine Learning in Javascript with Tensorflow.js
I gave it a twist, the author of the course used the Tensorflowjs library that can be used in the browser,
I think that library is pretty cool because it uses WebGL(gpu) to run the Tensor processes, which is much faster than simple JavaScript.
I thought it could be cool if I tried to use the Nodejs library which has CUDA support and serve the prediction as endpoint
My laptop has a CUDA GPU(not a good one), but I notice that runs a bit faster than the Nodejs lib that has only C++ bindings,
So I need to set the CUDA-enabled features first
Once you have the settings, the rest is easy😉
Nodejs lib with Cuda support:
npm i @tensorflow/tfjs-node-gpuWe need a Dataset for that go to Kaggle and download kc_house_data.csv (798 kB)
It is interesting to note that we have options, we can host the dataset(.csv) prepare the features/labels,
train the model then save that model to a local file and finally load that same model to make predictions.
The other option may be that the model has been worked on, trained, tested elsewhere(python)
and with our Nodejs server we can simply load it from the url. For me the idea I wanted to test was to
do the round trip here with Nodejs and using Typescript (the Tensorflowjs lib is written in Typescript).
So doing everything in node will look like:
- Save dataset(.csv) file locally and from there prepare data see: create-features interesting to note the format for saving csv file.
// import dataset from CSV file
const houseSalesDataset = data.csv("file://data/kc_house_data.csv");- Our problem will be: given the living area of a house in square feet, we will predict the price of that house,
so we will get only two columns of the dataset:
const pointsDataset = await houseSalesDataset.map((record) => ({
//@ts-ignore
x: record.sqft_living,
//@ts-ignore
y: record.price,
}));- Will split the data in 2 parts 1 for training and the other for testing/validating
To do that we need to have an even number of records, so if out number of records is odd we need to remove one:
const points = await pointsDataset.toArray();
// if odd number will eliminate last one so can split in two exactly
if (points.length % 2 !== 0) {
points.pop();
}- We'll need to shuffle data, to avoid bias on records order
// shuffle the points
shuffle(points);- This simple linear regression problem will have a 2D Tensor for both Features(input) and Labels(output)
// Features (input)
const featureValues = points.map(({ x }) => x as number);
const featureTensor = tensor2d(featureValues, [featureValues.length, 1]);
// Labels (output)
const labelValues = points.map(({ y }) => y as number);
const labelTensor = tensor2d(labelValues, [labelValues.length, 1]);
// normalize Tensors
let {
tensor: normalizedFeatureTensor,
min: minFeature,
max: maxFeature,
} = normalize(featureTensor);
let {
tensor: normalizedLabelTensor,
min: minLabel,
max: maxLabel,
} = normalize(labelTensor);
return {
normalizedFeatureTensor,
normalizedLabelTensor,
minFeature,
maxFeature,
minLabel,
maxLabel,
};For our simple model will have only one layer(Layers API) and that layer will have only one node as Sequential model
our activition function will be a linear input = output, our loss function is **Mean Squared Error"
src/utils/create-model.ts:
function createModel(): Sequential {
const model = sequential();
model.add(
layers.dense({ units: 1, useBias: true, activation: "linear", inputDim: 1 })
);
const optimizer = train.adam();
model.compile({ loss: "meanSquaredError", optimizer });
return model;
}Having our features/labels
and running only 20 epochs
src/utils/train-model.ts:
async function trainModel(
model: Sequential,
trainingFeatureTensor: Tensor<Rank>,
trainingLabelTensor: Tensor<Rank>
) {
return model.fit(trainingFeatureTensor, trainingLabelTensor, {
epochs: 20,
validationSplit: 0.2,
callbacks: {
onEpochEnd: (epoch, log) =>
console.log(`Epoch ${epoch}: loss = ${log?.["loss"]}`),
},
});
}Those 20 epochs took no more than a minute to execute.
src/housesalesdataset.ts (line#35)
const lossTensor = model.evaluate(testingFeature, testingLabel);
const loss = (lossTensor as Scalar).dataSync();
console.log(`Testing set loss: ${loss}`);Having the model trained and tested, we only have to load it from the saved location and run a prediction.
src/utils/predict-house-price.ts:
async function predictHousePrice(houseArea: number): Promise<number> {
if (isNaN(houseArea)) {
throw new Error("House are must be a number value");
}
const inputTensor = tensor1d([houseArea]);
let { minFeature, maxFeature, minLabel, maxLabel } = await createFeatures();
const { tensor: inputTensorNormalized } = normalize(
inputTensor,
minFeature,
maxFeature
);
const loadedModel = await loadLayersModel(
"file://models/house-price-regression-v1/model.json"
);
const normalizedOutputTensor = loadedModel.predict(inputTensorNormalized);
if (!minLabel || !maxLabel) {
throw new Error("Missing parameters to run de-noramalization");
}
const outputTensor = deNormalize(
normalizedOutputTensor as Tensor<Rank>,
minLabel,
maxLabel
);
const outputValue = outputTensor.dataSync()[0];
if (!outputValue) {
throw new Error("Error getting predicted price");
}
const roundedOutput = +(outputValue / 1000).toFixed(0) * 1000;
return roundedOutput;
}That's easy! just expose the prediction as an ExpressJS endpoint
src/server.ts:
app.use("/prediction/house-price-by-squareft", housePriceRouter);src/controllers/house-price.ts:
const housePriceRouter = Router();
interface PredictionBody {
area: number;
}
interface PredictionResponse {
data: { price: number };
}
housePriceRouter.post(
"/",
async (
req: Request<core.ParamsDictionary, PredictionResponse, PredictionBody>,
res: Response
) => {
const { area } = req.body;
const price = await predictHousePrice(area);
res.status(200).json({ data: { price } });
}
);
export { housePriceRouter };