TensorFlow.js with SharedMobility.ai Datasets

dyypgyq.markdown

dyyPgYq

A Pen by Philipp Naderer on CodePen.

License.

index.html

<h1>TensorFlow.js ❤️ SharedMobility.ai</h1>

<button id="train">Train</button>

<ul id="logs">
</ul>

script.js

const DateTime = luxon.DateTime;

const $logs = document.getElementById("logs");
function logLine(line) {
  const $line = document.createElement("li");
  $line.textContent = line;
  $logs.appendChild($line);
}

document.getElementById("train").addEventListener("click", async () => {
  const datasetURL = new URL("https://storage.googleapis.com/smai-public-datasets/citybikewien/service_2-station_42_citybikewien-oper-1055-2019-05-07_2019-05-31.csv");

  train(datasetURL).then(model => {
    logLine(`Finshed training.`)
  });
})

async function train(datasetURL) {
    const INPUT_LENGTH = 48;
  
    logLine(`Loading dataset: ${datasetURL.href}`);
    const trainingSet = tf.data.csv(datasetURL.href).take(150).map(obj => convertCsvRecord(obj));
  
    logLine("Converted the input dataset.");

    const model = tf.sequential();

    // build the layered model
    model.add(tf.layers.dense({units: INPUT_LENGTH * 2, activation: "relu", inputShape: [ INPUT_LENGTH ]}));
    model.add(tf.layers.dense({units: 4, activation: "softmax"}));

    // Compile model to prepare for training.
    model.compile({
        optimizer: tf.train.rmsprop(0.001),
        loss: tf.losses.meanSquaredError,
        metrics: ["accuracy"]
    });
  
    model.summary();

    logLine("Starting the training ...");
  
    const history = await model.fitDataset(trainingSet, {
        epochs: 5,
        batchSize: 64,
        shuffle: true,
        callbacks: {
          onEpochEnd: (epoch, logs) => {
            logLine(`Finished epoch ${epoch}, acc: ${logs.acc}, loss: ${logs.loss}`);
          }
        }
    });

    const loss = history.history.loss.slice(-1)[0];
    const acc = history.history.acc.slice(-1)[0];

    logLine(`Model for ${datasetURL.pathname} => acc ${acc} | loss ${loss}`);
    return model;
}

function convertCsvRecord(record) {
    return {
        xs: tf.tensor2d([recordToInput(record)], [1, 48], "int32"),
        ys: tf.tensor2d([recordToOutput(record)], [1, 4],"int32")
    }
}

function recordToInput(record) {
    const dt = DateTime.fromISO(record.timestamp, { zone: "Europe/Vienna" }).toUTC();

    const dayCategory = new Array(7).fill(0);
    dayCategory[dt.weekday - 1] = 1;

    const hourCategory = new Array(24).fill(0);
    hourCategory[dt.hour] = 1;

    const minuteCategory = new Array(4).fill(0);
    if (dt.minute >= 0 && dt.minute < 15) {
        minuteCategory[0] = 1;
    } else if (dt.minute >= 15 && dt.minute < 30) {
        minuteCategory[1] = 1;
    } else if (dt.minute >= 30 && dt.minute < 45) {
        minuteCategory[2] = 1;
    } else if (dt.minute >= 45 && dt.minute < 60) {
        minuteCategory[3] = 1;
    }

    // weather-based data
    const rainCategory = new Array(4).fill(0);
    if (record.rain >= 1) {
        rainCategory[0] = 1;
    }
    if (record.rain >= 2.5) {
        rainCategory[1] = 1;
    }
    if (record.rain >= 5) {
        rainCategory[2] = 1;
    }
    if (record.rain >= 9) {
        rainCategory[3] = 1;
    }

    const sunshineCategory = new Array(4).fill(0);
    if (record.sunshine >= 25) {
        sunshineCategory[0] = 1;
    }
    if (record.sunshine >= 50) {
        sunshineCategory[1] = 1;
    }
    if (record.sunshine >= 75) {
        sunshineCategory[2] = 1;
    }
    if (record.sunshine === 100) {
        sunshineCategory[3] = 1;
    }

    const temperatureCategory = new Array(4).fill(0);
    if (record.temperature >= 15) {
        temperatureCategory[0] = 1;
    }
    if (record.temperature >= 20) {
        temperatureCategory[1] = 1;
    }
    if (record.temperature >= 25) {
        temperatureCategory[2] = 1;
    }
    if (record.temperature >= 30) {
        temperatureCategory[3] = 1;
    }

    return [
        record.holiday,
        dayCategory,
        hourCategory,
        minuteCategory,
        rainCategory,
        sunshineCategory,
        temperatureCategory
    ].flat(1);
}

function recordToOutput(record) {
    const load = record.vehicles_available / (record.vehicles_available + record.boxes_available);

    return [
        load > 0.8 ? 1 : 0, // full of bikes
        load > 0.5 && load <= 0.8 ? 1 : 0,
        load > 0.2 && load <= 0.5 ? 1 : 0,
        load <= 0.2 ? 1 : 0 // very low load
    ];
}

scripts

<script src="https://cdnjs.cloudflare.com/ajax/libs/tensorflow/1.2.10/tf.min.js"></script>
<script src="https://cdn.jsdelivr.net/npm/luxon@1.19.3/build/global/luxon.min.js"></script>

style.css

ul {
  font-family: monospace;
}

tensorflow-js-with-sharedmobility-ai-datasets.markdown

TensorFlow.js with SharedMobility.ai Datasets

The pen loads a custom dataset by SharedMobility.ai and generates a model.

A Pen by Philipp Naderer on CodePen.

License.