Nithanaroy

async function loadModel() {
    google.script.run.withSuccessHandler(async modelFiles => {
        const modelJson = new File([modelFiles[0]], "model.json", { type: "application/json" })
        const modelWeights = new File([Uint8Array.from(modelFiles[1])], "group1-shard1of1.bin")
        model = await tf.loadLayersModel(tf.io.browserFiles([modelJson, modelWeights]))
        model.summary();
    }).withFailureHandler(onFailure).loadModelFromDrive();
}

Nithanaroy

function loadModelFromDrive() {
  const f = DriveApp.getFileById("1iO0Lz3022AHgTOR4Xd_WLuR1FVXNUIdU")
  const modeljson = f.getAs("application/json").getDataAsString();
  
  const f2 = DriveApp.getFileById("1bDZaBUjweWy-KwGpQP68eoJLMulUfXYZ")
  const weights = f2.getBlob().getBytes()
  
  return [modeljson, weights]
}

Nithanaroy

const c = document.getElementById("myCanvas");
c.addEventListener("mousedown", setLastCoords); // fires before mouse left btn is released
c.addEventListener("mousemove", freeForm);

const ctx = c.getContext("2d");

function setLastCoords(e) {
  const {x, y} = c.getBoundingClientRect();
    lastX = e.clientX - x;
    lastY = e.clientY - y;

Nithanaroy

class VisionModel {
    // ... other methods
    async run({ batchSize = 1024, epochs = 1, trainExisting = true } = {}) {
        // ... same as before
        
        await this.visionModelWorker.create(true);
        const modelInstance = await tf.loadLayersModel(Constants.MODEL_DISK_PATH());
        tfvis.show.modelSummary(this.tensorboardDiv, modelInstance);
        
        // ... training callbacks like before

Nithanaroy

// In the main thread
class VisionModel {
    /**
    * Creates an instance of a wrapper of Vision Model that resides in the worker thread
    * This has access to the DOM and can show visualizations, unlike the worker
    * @param {HTMLElement} tensorboardDiv Instance of a Div tag where to show live model training
    */
    constructor(tensorboardDiv) {
        this.tensorboardDiv = tensorboardDiv;
    }

Nithanaroy

class VisionModelWorker {

   ... // create and getData functions
   
    async train(epochs, batchSize) {
        this.model.compile({ optimizer: tf.train.adam(), loss: 'categoricalCrossentropy', metrics: ['accuracy'] });
        const historyObj = await this.model.fit(this.dataBunch.trainX, this.dataBunch.trainY, {
            batchSize: batchSize,
            validationData: [this.dataBunch.testX, this.dataBunch.testY],
            epochs: epochs,

Nithanaroy

importScripts("https://cdnjs.cloudflare.com/ajax/libs/tensorflow/1.3.2/tf.min.js", "data.js");

class VisionModelWorker {

  constructor() {
      this.model = null; // holds tfjs model
      this.dataBunch = null; // holds the X and y datasets
  }
  
  create() {

Nithanaroy

static parseBuffer(buffer, offset = 16, isBigEndianProcessor = false) {
    let allData = null;
    if (isBigEndianProcessor) {
        // Let native JavaScript decode the bytes as MNIST dataset is encoded in Big Endian format 
        allData = new Uint8Array(buffer, offset)
    }
    else {
        const dataView = new DataView(buffer);
        const numBytes = dataView.byteLength - offset;
        allData = new Uint8Array(numBytes);

Nithanaroy

async function main() {
  console.log(`Start: ${performance.now()}`);
  await sleep(2000);
  console.log(`End: ${performance.now()}`);
}

Nithanaroy