// Cloned by Rajasi Chavan on 2 Dec 2022 from World "CharRecognition_UsingCNN (clone by Dheera(21261395))" by Dheera
// Please leave this clone trail here.
//RC : Understanding of the CovNetJS code from "https://cs.stanford.edu/people/karpathy/convnetjs/docs.html"
const PIXELS = 28,
PIXELSSQUARED = PIXELS * PIXELS,
NOTRAIN = 124800,
NOTEST = 20800,
noinput = PIXELSSQUARED,
nohidden = 64,
nooutput = 10,
learningrate = .1;
let do_training = !0;
const TRAINPERSTEP = 15,
TESTPERSTEP = 5,
ZOOMFACTOR = 7,
ZOOMPIXELS = 7 * PIXELS,
canvaswidth = PIXELS + ZOOMPIXELS + 50,
canvasheight = 3 * ZOOMPIXELS + 100,
DOODLE_THICK = 18,
DOODLE_BLUR = 3;
let mnist, mycnn, mycnnTrain, mycnnModel, doodle, demo,
alphabets = ["A", "B", "C", "D", "E", "F", "G", "H", "I", "J", "K", "L", "M", "N", "O", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z"],
trainrun = 1,
train_index = 0,
testrun = 1,
test_index = 0,
total_tests = 0,
total_correct = 0,
doodle_exists = !1,
demo_exists = !1,
mousedrag = !1;
var train_inputs, test_inputs, demo_inputs, doodle_inputs, thehtml;
function randomWeight() {
return AB.randomFloatAtoB(-.5, .5)
}
AB.headerCSS({
"max-height": "95vh"
}), thehtml = "<hr> <h1> 1. Doodle </h1> Top row: Doodle (left) and shrunk (right). <br> Draw your doodle in top LHS. <button onclick='wipeDoodle();' class='normbutton' >Clear and Re-draw</button> <br> ", AB.msg(thehtml, 1), thehtml = "<hr> <h1> 2. Training </h1> Middle row: Training image magnified (left) and original (right). <br> <button onclick='do_training = false;' class='normbutton' >Pause training</button> <button onclick='do_training = true;' class='normbutton' >Resume training</button> <br> ", AB.msg(thehtml, 3), thehtml = "<h3> Hidden tests </h3> ", AB.msg(thehtml, 5), thehtml = "<hr> <h1> 3. Demo </h1> Bottom row: Test image magnified (left) and original (right). <br> The network is <i>not</i> trained on any of these images. <br> <button onclick='makeDemo();' class='normbutton' >Demo test image</button> <br> ", AB.msg(thehtml, 7);
const bluespan = "<span style='font-weight:bold; font-size:large; color:darkblue'> ";
function setup() {
createCanvas(canvaswidth, canvasheight),
(doodle = createGraphics(ZOOMPIXELS, ZOOMPIXELS)).pixelDensity(1), wipeDoodle(), AB.loadingScreen(),
$.getScript("/uploads/codingtrain/matrix.js", function() {
$.getScript("/uploads/dheera0704/convnet3.js", function() {
$.getScript("/uploads/dheera0704/mnistDhe.js", function() {
console.log("All JS Files loaded");
let t = [];
//RC: Instantiate a Network and Trainer
//RC: This will create mycnnModel and mycnnTrain variables. mycnnModel is given input as 't' to make layers and
//mycnnTrain has the model trained using SGDTrainer of convnetjs.
t.push({
type: "input",
out_sx: 28,
out_sy: 28,
out_depth: 1
}), t.push({
type: "conv",
sx: 5,
filters: 8,
stride: 1,
pad: 2,
activation: "relu"
}), t.push({
type: "pool",
sx: 2,
stride: 2
}), t.push({
type: "conv",
sx: 5,
filters: 16,
stride: 1,
pad: 2,
activation: "relu"
}), t.push({
type: "pool",
sx: 3,
stride: 3
}), t.push({
type: "softmax",
num_classes: 26
}), (mycnnModel = new convnetjs.Net).makeLayers(t), mycnnTrain = new convnetjs.SGDTrainer(mycnnModel, {
method: "adadelta", //RC: Adadelta is used which is one of per-parameter adaptive step size methods, so that the change of learning rates or momentum over time is managed automatically.
momentum: .9,
batch_size: 10,
l2_decay: .001
}), loadData()
})
})
})
}
function loadData() {
loadMNIST(function(t) {
mnist = t;
let e = 0;
//RC: Mnist alphabet dataset is loaded does not have images aligned correctly hence rotating images for correct learning purpose.
for (; e < NOTRAIN; e++) rotateImage(mnist.train_images[e]);
for (e = 0; e < NOTEST; e++) rotateImage(mnist.test_images[e]);
console.log("All data loaded into Emnist object."), console.log(mnist), AB.removeLoading()
})
}
function getImage(t) {
let e = createImage(PIXELS, PIXELS); // make a P5 image object from a raw data array
e.loadPixels(); // make blank image, then populate it
for (let n = 0; n < PIXELSSQUARED; n++) {
let o = t[n],
s = 4 * n;
e.pixels[s + 0] = o, e.pixels[s + 1] = o, e.pixels[s + 2] = o, e.pixels[s + 3] = 255
}
return e.updatePixels(), e
}
function getInputs(t) {
let e = [];
for (let n = 0; n < PIXELSSQUARED; n++) {
let o = t[n];
e[n] = o / 255 // RC: To check the brightness/color value of each pixel in the image array
}
return e
}
//RC: Mnist alphabet dataset is loaded does not have images aligned correctly hence rotating images for correct learning purpose.
function rotateImage(t) {
for (let e = 0; e < PIXELS; e++)
for (let n = e; n < PIXELS; n++) {
let o = e * PIXELS + n,
s = n * PIXELS + e,
i = t[o];
t[o] = t[s], t[s] = i
}
}
//RC : Function to train minst data
function trainit(t) {
let e = mnist.train_images[train_index],
n = mnist.train_labels[train_index];
if (t) {
var o = getImage(e);
image(o, 0, ZOOMPIXELS + 50, ZOOMPIXELS, ZOOMPIXELS), //magnified
image(o, ZOOMPIXELS + 50, ZOOMPIXELS + 50, PIXELS, PIXELS) //original
}
// set up the inputs
let s = getInputs(e);
train_inputs = s; {
let t = getmycnnInputs(s);
mycnnTrain.train(t, n);
}
thehtml = " trainrun: " + trainrun + "<br> no: " + train_index, AB.msg(thehtml, 4),
++train_index == NOTRAIN && (train_index = 0, console.log("finished trainrun: " + trainrun), trainrun++);
}
function getmycnnInputs(t) {
//RC : create a 28x28x2 volume of input activations, intialised to zero's
//RC : Iterates to each pixel in inputimage and assigns its value to e.w[n] and n is pixel position
for (var e = new convnetjs.Vol(28, 28, 1, 0), n = 0; n < PIXELSSQUARED; n++) e.w[n] = t[n];
return e;
}
function testit() {
let t = mnist.test_images[test_index],
e = mnist.test_labels[test_index],
n = getInputs(t),
o = getmycnnInputs(n);
test_inputs = n;
let s = findMax(mycnnModel.forward(o).w); //RC : To pass input data in forward direction only
var i = getImage(t);
image(i, 0, ZOOMPIXELS + 50, ZOOMPIXELS, ZOOMPIXELS), //maginified image
image(i, ZOOMPIXELS + 50, ZOOMPIXELS + 50, PIXELS, PIXELS), //original image
total_tests++, s == e && total_correct++;
//RC : total_correct has correct number of outputs predicted
//RC : a is ratio of total correct predicted outcomes to total outcomes
let a = total_correct / total_tests * 100;
thehtml = " testrun: " + testrun + "<br> no: " + total_tests + " <br> correct: " + total_correct + "<br> score: " + bluespan + a.toFixed(2) + "</span>", AB.msg(thehtml, 6), ++test_index == NOTEST && (console.log("finished testrun: " + testrun + " score: " + a.toFixed(2)), testrun++, test_index = 0, total_tests = 0, total_correct = 0)
}
function find12(t) {
let e = 0,
n = 0,
o = 0,
s = 0;
for (let i = 0; i < t.length; i++) t[i] > o ? (n = e, s = o, e = i, o = t[i]) : t[i] > s && (n = i, s = t[i]);
return [e, n]
}
function findMax(t) {
let e = 0,
n = 0;
for (let o = 0; o < t.length; o++) t[o] > n && (e = o, n = t[o]);
return e
}
function draw() {
if (void 0 !== mnist) {
if (background("black"), strokeWeight(1), stroke("green"), rect(0, 0, ZOOMPIXELS, ZOOMPIXELS), textSize(10), textAlign(CENTER), text("DOODLE HERE", ZOOMPIXELS / 2, ZOOMPIXELS / 2), do_training) {
for (let t = 0; t < TRAINPERSTEP; t++) trainit(0 === t);
for (let t = 0; t < TESTPERSTEP; t++) testit()
}
if (demo_exists && (drawDemo(), guessDemo()), doodle_exists && (drawDoodle(), guessDoodle()), mouseIsPressed) {
var t = ZOOMPIXELS + 20;
mouseX < t && mouseY < t && pmouseX < t && pmouseY < t && (mousedrag = !0, doodle_exists = !0, doodle.stroke("red"), strokeJoin(ROUND), doodle.strokeWeight(DOODLE_THICK), doodle.line(mouseX, mouseY, pmouseX, pmouseY))
} else mousedrag && (mousedrag = !1, doodle.filter(BLUR, DOODLE_BLUR))
}
}
function makeDemo() {
demo_exists = !0;
var t = AB.randomIntAtoB(0, NOTEST - 1);
demo = mnist.test_images[t];
var e = mnist.test_labels[t];
thehtml = "Test image no: " + t + "<br>Classification: " + alphabets[e - 1] + "<br>", AB.msg(thehtml, 8)
}
function drawDemo() {
var t = getImage(demo);
image(t, 0, canvasheight - ZOOMPIXELS, ZOOMPIXELS, ZOOMPIXELS), image(t, ZOOMPIXELS + 50, canvasheight - ZOOMPIXELS, PIXELS, PIXELS)
}
function guessDemo() {
let t = getInputs(demo);
demo_inputs = t;
let e = getmycnnInputs(t),
n = findMax(mycnnModel.forward(e).w);
thehtml = " We classify it as: " + bluespan + alphabets[n - 1] + "</span>", AB.msg(thehtml, 9)
}
function drawDoodle() {
let t = doodle.get();
image(t, 0, 0, ZOOMPIXELS, ZOOMPIXELS), image(t, ZOOMPIXELS + 20, 0, PIXELS, PIXELS)
}
function guessDoodle() {
let t = doodle.get();
t.resize(PIXELS, PIXELS), t.loadPixels();
let e = [];
for (let n = 0; n < PIXELSSQUARED; n++) e[n] = t.pixels[4 * n] / 255;
doodle_inputs = e;
let n = getmycnnInputs(e),
o = find12(mycnnModel.forward(n).w);
thehtml = " Our 1st Guess is: " + bluespan + alphabets[o[0] - 1] + "</span> <br> Our 2nd Guess is: " + bluespan + alphabets[o[1] - 1] + "</span>", AB.msg(thehtml, 2)
}
function wipeDoodle() {
doodle_exists = !1, doodle.background("black")
}
function showInputs(t) {
var e = "";
for (let n = 0; n < t.length; n++) {
n % PIXELS == 0 && (e += "\n"), e = e + " " + t[n].toFixed(2)
}
console.log(e)
}