code after part 2

Author: shiffman

Diff for: .gitignore

@@ -1,5 +1,6 @@
 data/
 *.png
+*.jpg
 node_modules/
 .env
 .DS_Store

Diff for: AutoEncoder_TrainingData/AutoEncoder_TrainingData.pde

@@ -0,0 +1,20 @@
+int counter = 0;
+void setup() {
+  size(280, 280);
+}
+
+void draw() {
+  background(255);
+  float r = random(100, 200);
+  strokeWeight(16);
+  rectMode(CENTER);
+  square(width/2, height/2, r);
+  PImage img = get();
+  img.resize(28, 28);
+  img.save("data/square" + nf(counter,3) + ".png");
+  counter++;
+  if (counter == 550) {
+    exit();
+  }
+  // noLoop();
+}

Diff for: index.js

@@ -1,58 +1,119 @@
 console.log("Hello Autoencoder 🚂");
 
-import * as tf from '@tensorflow/tfjs-node'
-
-const autoencoder = tf.sequential();
-
-const encoder = tf.layers.dense({
-  units: 32,
-  inputShape: [784],
-  activation: 'relu'
-});
-const decoder = tf.layers.dense({
-  units: 784,
-  activation: 'sigmoid'
-  // inputShape: [32]
-});
-
-autoencoder.add(encoder);
-autoencoder.add(decoder);
-
-autoencoder.compile({
-  optimizer: 'adam',
-  loss: 'binaryCrossentropy',
-  metrics: ['accuracy'],
-});
-
-
-function generateImage() {
-  const img = [];
-  for (let i = 0; i < 784; i++) {
-    img[i] = Math.random();
-  }
-  return img;
+import * as tf from "@tensorflow/tfjs-node";
+// import canvas from "canvas";
+// const { loadImage } = canvas;
+import Jimp from "jimp";
+import numeral from "numeral";
+
+main();
+
+async function main() {
+  // Build the model
+  const autoencoder = buildModel();
+  // load all image data
+  const images = await loadImages(550);
+
+  // train the model
+  const x_train = tf.tensor2d(images.slice(500));
+  await trainModel(autoencoder, x_train, 250);
+
+  // test the model
+  const x_test = tf.tensor2d(images.slice(500, 550));
+  await generateTests(autoencoder, x_test);
 }
 
-const x_inputs = [];
-for (let i = 0; i < 1000; i++) {
-  x_inputs[i] = generateImage();
+async function generateTests(autoencoder, x_test) {
+  const output = autoencoder.predict(x_test);
+  // output.print();
+
+  const newImages = await output.array();
+  for (let i = 0; i < newImages.length; i++) {
+    const img = newImages[i];
+    const buffer = [];
+    for (let n = 0; n < img.length; n++) {
+      const val = Math.floor(img[n] * 255);
+      buffer[n * 4 + 0] = val;
+      buffer[n * 4 + 1] = val;
+      buffer[n * 4 + 2] = val;
+      buffer[n * 4 + 3] = 255;
+    }
+    const image = new Jimp(
+      {
+        data: Buffer.from(buffer),
+        width: 28,
+        height: 28,
+      },
+      (err, image) => {
+        const num = numeral(i).format("000");
+        image.write(`output/square${num}.png`);
+      }
+    );
+  }
 }
 
-const x_train =  tf.tensor2d(x_inputs);
-x_train.print();
-
-trainModel();
+function buildModel() {
+  const autoencoder = tf.sequential();
+  // Build the model
+  autoencoder.add(
+    tf.layers.dense({
+      units: 256,
+      inputShape: [784],
+      activation: "relu",
+    })
+  );
+  autoencoder.add(
+    tf.layers.dense({
+      units: 128,
+      activation: "relu",
+    })
+  );
+
+  autoencoder.add(
+    tf.layers.dense({
+      units: 256,
+      activation: "sigmoid",
+    })
+  );
+
+  autoencoder.add(
+    tf.layers.dense({
+      units: 784,
+      activation: "sigmoid",
+    })
+  );
+  autoencoder.compile({
+    optimizer: "adam",
+    loss: "binaryCrossentropy",
+    metrics: ["accuracy"],
+  });
+  return autoencoder;
+}
 
-async function trainModel() {
+async function trainModel(autoencoder, x_train, epochs) {
   await autoencoder.fit(x_train, x_train, {
-    epochs: 50,
-    batch_size: 256,
+    epochs: epochs,
+    batch_size: 32,
     shuffle: true,
-    verbose: true
+    verbose: true,
   });
 }
 
-
-
-
-
+async function loadImages(total) {
+  const allImages = [];
+  for (let i = 0; i < total; i++) {
+    const num = numeral(i).format("000");
+    const img = await Jimp.read(`data/square${num}.png`);
+
+    let rawData = [];
+    for (let n = 0; n < 28 * 28; n++) {
+      let index = n * 4;
+      let r = img.bitmap.data[index + 0];
+      // let g = img.bitmap.data[n + 1];
+      // let b = img.bitmap.data[n + 2];
+      rawData[n] = r / 255.0;
+    }
+    allImages[i] = rawData;
+  }
+  return allImages;
+}