// Port of Character recognition neural network from here:
// https://github.com/CodingTrain/Toy-Neural-Network-JS/tree/master/examples/mnist
// with many modifications
// --- defined by MNIST - do not change these ---------------------------------------
const CROP_PIXELS = 24; // AGT: Random 24x24 crops are used to train WebCNN
const PIXELS = 28; // images in data set are tiny
const PIXELSSQUARED = PIXELS * PIXELS;
// number of training and test exemplars in the data set:
const NOTRAIN = 60000;
const NOTEST = 10000;
//--- can modify all these --------------------------------------------------
// no of nodes in network
const noinput = PIXELSSQUARED;
const nohidden = 64;
const nooutput = 10;
const learningrate = 0.1; // default 0.1
// should we train every timestep or not
let do_training = true;
// how many to train and test per timestep
const BATCH_SIZE = 50;
const TRAINPERSTEP = BATCH_SIZE;
const TESTPERSTEP = 5;
// multiply it by this to magnify for display
const ZOOMFACTOR = 7;
const ZOOMPIXELS = ZOOMFACTOR * PIXELS;
// 3 rows of
// large image + 50 gap + small image
// 50 gap between rows
const canvaswidth = ( PIXELS + ZOOMPIXELS ) + 50;
const canvasheight = ( ZOOMPIXELS * 3 ) + 100;
const DOODLE_THICK = 18; // thickness of doodle lines
// AGT:
const DOODLE_BLUR = 0; // blur factor applied to doodles
let mnist;
// all data is loaded into this
// mnist.train_images
// mnist.train_labels
// mnist.test_images
// mnist.test_labels
// AGT: pick the network to use in The World
// 0: 1-hidden-layer network with ReLU activation (for the hidden layer)
// 1: 1-hidden-layer network with Tanh activation (for the hidden layer)
// 2: Original WebCNN
// 3: Original WebCNN (pretrained)
const theNN = 3;
let nn;
let trainrun = 1;
let train_index = 0;
let testrun = 1;
let test_index = 0;
let total_tests = 0;
let total_correct = 0;
// images in LHS:
let doodle, demo;
let doodle_exists = false;
let demo_exists = false;
let mousedrag = false; // are we in the middle of a mouse drag drawing?
// save inputs to global var to inspect
// type these names in console
var train_inputs, test_inputs, demo_inputs, doodle_inputs;
// Matrix.randomize() is changed to point to this. Must be defined by user of Matrix.
function randomWeight()
{
return ( AB.randomFloatAtoB ( -0.5, 0.5 ) );
// Coding Train default is -1 to 1
}
// CSS trick
// make run header bigger
$("#runheaderbox").css ( { "max-height": "95vh" } );
//--- start of AB.msgs structure: ---------------------------------------------------------
// We output a serious of AB.msgs to put data at various places in the run header
var thehtml;
// 1 Doodle header
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 doodle</button> <br> ";
AB.msg ( thehtml, 1 );
// 2 Doodle variable data (guess)
// 3 Training header
thehtml = "<hr> <h1> 2. Training </h1> Middle row: Training image magnified (left) and original (right). <br> " +
" <button onclick='do_training = false;' class='normbutton' >Stop training</button> <br> ";
AB.msg ( thehtml, 3 );
// 4 variable training data
// 5 Testing header
thehtml = "<h3> Hidden tests </h3> " ;
AB.msg ( thehtml, 5 );
// 6 variable testing data
// 7 Demo header
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 );
// 8 Demo variable data (random demo ID)
// 9 Demo variable data (changing guess)
const greenspan = "<span style='font-weight:bold; font-size:x-large; color:darkgreen'> " ;
//--- end of AB.msgs structure: ---------------------------------------------------------
function setup()
{
createCanvas ( canvaswidth, canvasheight );
doodle = createGraphics ( ZOOMPIXELS, ZOOMPIXELS ); // doodle on larger canvas
doodle.pixelDensity(1);
// JS load other JS
// maybe have a loading screen while loading the JS and the data set
AB.loadingScreen();
$.getScript ( "/uploads/codingtrain/mnist.js", function()
{
// AGT: Load WebCNN Math utils, network definition, pre-defined weights.
$.getScript ( "/uploads/atotev/mathutils.js", function()
{
$.getScript ( "/uploads/atotev/webcnn.js", function()
{
$.getJSON ( "/uploads/atotev/cnn_mnist_10_20_98accuracy.json", function(networkJSON)
{
console.log ("All JS loaded");
if (theNN===0) {
nn = createShallowNetwork(ACTIVATION_RELU);
} else if (theNN===1) {
nn = createShallowNetwork(ACTIVATION_TANH);
} else if (theNN===2) {
nn = createDefaultNetwork();
} else if (theNN===3) {
nn = loadNetworkFromJSON(networkJSON);
} else {
console.log("Unknown NN type: " + theNN);
}
loadData();
});
});
});
});
}
// AGT: Pre-trained WebCNN
function loadNetworkFromJSON( networkJSON )
{
let result = new WebCNN();
if (networkJSON.momentum !== undefined) result.setMomentum( networkJSON.momentum );
if (networkJSON.lambda !== undefined) result.setLambda( networkJSON.lambda );
if (networkJSON.learningRate !== undefined) result.setLearningRate( networkJSON.learningRate );
for ( let layerIndex = 0; layerIndex < networkJSON.layers.length; ++layerIndex )
{
let layerDesc = networkJSON.layers[ layerIndex ];
console.log( layerDesc );
result.newLayer( layerDesc );
}
for ( let layerIndex = 0; layerIndex < networkJSON.layers.length; ++layerIndex )
{
let layerDesc = networkJSON.layers[ layerIndex ];
switch ( networkJSON.layers[ layerIndex ].type )
{
case LAYER_TYPE_CONV:
case LAYER_TYPE_FULLY_CONNECTED:
{
if (layerDesc.weights !== undefined && layerDesc.biases !== undefined )
{
result.layers[ layerIndex ].setWeightsAndBiases( layerDesc.weights, layerDesc.biases );
}
break;
}
}
}
result.initialize();
return result;
}
// AGT: Default WebCNN with randmonly initialized weights
function createDefaultNetwork()
{
let result = new WebCNN();
result.newLayer( { name: "image", type: LAYER_TYPE_INPUT_IMAGE, width: CROP_PIXELS, height: CROP_PIXELS, depth: 1 } );
result.newLayer( { name: "conv1", type: LAYER_TYPE_CONV, units: 10, kernelWidth: 5, kernelHeight: 5, strideX: 1, strideY: 1, padding: false } );
result.newLayer( { name: "pool1", type: LAYER_TYPE_MAX_POOL, poolWidth: 2, poolHeight: 2, strideX: 2, strideY: 2 } );
result.newLayer( { name: "conv2", type: LAYER_TYPE_CONV, units: 20, kernelWidth: 5, kernelHeight: 5, strideX: 1, strideY: 1, padding: false } );
result.newLayer( { name: "pool2", type: LAYER_TYPE_MAX_POOL, poolWidth: 2, poolHeight: 2, strideX: 2, strideY: 2 } );
result.newLayer( { name: "out", type: LAYER_TYPE_FULLY_CONNECTED, units: 10, activation: ACTIVATION_SOFTMAX } );
result.initialize();
result.setLearningRate( 0.01 );
result.setMomentum( 0.9 );
result.setLambda( 0.0 );
return result;
}
// AGT: A single-hidden-layer network (similar to the original NN) with tanh activation for the hidden layers
function createShallowNetwork(activation)
{
let result = new WebCNN();
result.newLayer( { name: "image", type: LAYER_TYPE_INPUT_IMAGE, width: CROP_PIXELS, height: CROP_PIXELS, depth: 1 } );
result.newLayer( { name: "hidden", type: LAYER_TYPE_FULLY_CONNECTED, units: nohidden, activation: activation } );
result.newLayer( { name: "out", type: LAYER_TYPE_FULLY_CONNECTED, units: nooutput, activation: ACTIVATION_SOFTMAX } );
result.initialize();
result.setLearningRate( 0.01 );
result.setMomentum( 0.9 );
result.setLambda( 0.0 );
return result;
}
// load data set from local file (on this server)
function loadData()
{
loadMNIST ( function(data)
{
mnist = data;
console.log ("All data loaded into mnist object:")
console.log(mnist);
AB.removeLoading(); // if no loading screen exists, this does nothing
});
}
// AGT: Calculates bounding box and translates to centre
function centerImage(pixels, size) {
// convert to matrix
let m = [];
for (let i = 0; i < size; i++)
{
m[i] = [];
for (let j = 0; j < size; j++)
{
m[i][j] = pixels[(i*size + j) * 4];
}
}
// find bounding box
var topmost = Number.MAX_VALUE;
var leftmost = Number.MAX_VALUE;
var bottommost = -1;
var rightmost = -1;
for(var y = 0; y < m.length; y++)
{
var l = m[y].indexOf(255);
var r = m[y].lastIndexOf(255);
if (l >= 0 && l < leftmost) leftmost = l;
if (r >= 0 && r > rightmost) rightmost = r;
// only check if some 1 found
if (l >= 0 && y < topmost) topmost = y;
if (l >= 0 && y > bottommost) bottommost = y;
}
//console.log("AGT: " + leftmost + " " + topmost + " " + rightmost + " " + bottommost)
// translate to centre
let transY = Math.floor((size - bottommost - topmost) / 2);
let transX = Math.floor((size - rightmost - leftmost) / 2);
let result = Array(size).fill().map(() => Array(size).fill(0));
for (i=topmost; i<=bottommost; i++) {
for(j=leftmost; j<=rightmost; j++) {
result[i + transY][j + transX] = m[i][j];
}
}
// Convert back to 1D array
let m1D = [];
for (let i = 0; i < size; i++)
{
for (let j = 0; j < size; j++)
{
m1D[i*size + j] = result[i][j];
}
}
return m1D;
}
// AGT: Modified to work with both original and cropped input size
function getImage ( img, size=PIXELS ) // make a P5 image object from a raw data array
{
let theimage = createImage (size, size); // make blank image, then populate it
theimage.loadPixels();
for (let i = 0; i < size*size ; i++)
{
let bright = img[i];
let index = i * 4;
theimage.pixels[index + 0] = bright;
theimage.pixels[index + 1] = bright;
theimage.pixels[index + 2] = bright;
theimage.pixels[index + 3] = 255;
}
theimage.updatePixels();
return theimage;
}
// AGT: Crop a random size*size part from img
function randomCrop ( img, size ) // convert img array into normalised input array
{
const maxStartIndex = PIXELS - size;
let xrand = Math.floor( Math.random() * maxStartIndex );
let yrand = Math.floor( Math.random() * maxStartIndex );
return crop(img, size, xrand, yrand);
}
// AGT: Crop a size*size part from image starting at (x, y)
function crop ( img, newSize, x=2, y=2 ) // convert img array into normalised input array
{
const originalSize = PIXELS;
let xEndIndex = x + newSize;
let yEndIndex = y + newSize;
let inputs = [];
for (let i = x; i < xEndIndex ; i++)
{
for (let j = y; j < yEndIndex ; j++)
{
inputs.push(img[ i*originalSize + j ])
}
}
// console.log(inputs);
return ( inputs );
}
function getInputs ( img ) // convert img array into normalised input array
{
let inputs = [];
for (let i = 0; i < PIXELSSQUARED ; i++)
{
let bright = img[i];
inputs[i] = bright / 255; // normalise to 0 to 1
}
return ( inputs );
}
// AGT: Utility function to convert image to object expected by the active NN
function toModelFormat(image, size) {
return {
"width": size,
"height": size,
"data": getImage (randomCrop(image, size), size).pixels
};
}
// AGT: Modified to train on mini batches
function trainit (show) // train the network with a single exemplar, from global var "train_index", show visual on or off
{
if (train_index%TRAINPERSTEP!==0) {
train_index++;
return;
}
// Train on new batch
let img = mnist.train_images[train_index];
let label = mnist.train_labels[train_index];
let imgs = [];
let labels = [];
for (i=0; i<TRAINPERSTEP; i++) {
imgs.push(toModelFormat(mnist.train_images[train_index+i], CROP_PIXELS));
labels.push(mnist.train_labels[train_index+i]);
}
// set up the inputs
let inputs = img; // get inputs from data array
// AGT: Always show
var theimage = getImage ( img ); // get image from data array
image ( theimage, 0, ZOOMPIXELS+50, ZOOMPIXELS, ZOOMPIXELS ); // magnified
image ( theimage, ZOOMPIXELS+50, ZOOMPIXELS+50, PIXELS, PIXELS ); // original
// console.log(train_index);
// console.log(inputs);
// console.log(targets);
train_inputs = inputs; // can inspect in console
nn.trainCNNClassifier ( imgs, labels );
thehtml = " trainrun: " + trainrun + "<br> no: " + train_index ;
AB.msg ( thehtml, 4 );
train_index++;
if ( train_index+TRAINPERSTEP >= NOTRAIN )
{
train_index = 0;
console.log( "finished trainrun: " + trainrun );
trainrun++;
}
}
// AGT: Modified to work with WebCNN
function testit() // test the network with a single exemplar, from global var "test_index"
{
let img = mnist.test_images[test_index];
let label = mnist.test_labels[test_index];
// set up the inputs
let inputs = getInputs ( img );
test_inputs = inputs; // can inspect in console
let result = nn.classifyImages( [ toModelFormat(mnist.test_images[test_index], CROP_PIXELS) ] );
let guess = findMax(result)
total_tests++;
if (guess == label) total_correct++;
let percent = (total_correct / total_tests) * 100 ;
thehtml = " testrun: " + testrun + "<br> no: " + total_tests + " <br> " +
" correct: " + total_correct + "<br>" +
" score: " + greenspan + percent.toFixed(2) + "</span>";
AB.msg ( thehtml, 6 );
test_index++;
if ( test_index == NOTEST )
{
console.log( "finished testrun: " + testrun + " score: " + percent.toFixed(2) );
testrun++;
test_index = 0;
total_tests = 0;
total_correct = 0;
}
}
//--- find no.1 (and maybe no.2) output nodes ---------------------------------------
// (restriction) assumes array values start at 0 (which is true for output nodes)
// AGT: Modified to work with WebCNN
function find12 (result) // return array showing indexes of no.1 and no.2 values in array
{
let guess1 = 0;
let max1 = 0;
let guess2 = 0;
let max2 = 0;
for ( var i = 0; i < 10; ++i )
{
if ( result[ 0 ].getValue( 0, 0, i ) > max1 )
{
max1 = result[ 0 ].getValue( 0, 0, i );
guess1 = i;
} else if ( result[ 0 ].getValue( 0, 0, i ) > max2 )
{
max2 = result[ 0 ].getValue( 0, 0, i );
guess2 = i;
}
}
return [guess1, guess2];
}
// AGT: Modified to work with WebCNN
// just get the maximum - separate function for speed - done many times
// find our guess - the max of the output nodes array
function findMax (result)
{
let guess = 0;
let max = 0;
for ( var i = 0; i < 10; ++i )
{
if ( result[ 0 ].getValue( 0, 0, i ) > max )
{
max = result[ 0 ].getValue( 0, 0, i );
guess = i;
}
}
return guess;
}
// --- the draw function -------------------------------------------------------------
// every step:
// AGT: Modify function in order to pause training while user is drawing a doodle
function draw()
{
// check if libraries and data loaded yet:
if ( typeof mnist == 'undefined' ) return;
// how can we get white doodle on black background on yellow canvas?
// background('#ffffcc'); doodle.background('black');
background ('black');
// keep drawing demo and doodle images
// and keep guessing - we will update our guess as time goes on
if ( demo_exists )
{
drawDemo();
guessDemo();
}
if ( doodle_exists )
{
drawDoodle();
guessDoodle();
}
// detect doodle drawing
// (restriction) the following assumes doodle starts at 0,0
if ( mouseIsPressed ) // gets called when we click buttons, as well as if in doodle corner
{
// console.log ( mouseX + " " + mouseY + " " + pmouseX + " " + pmouseY );
var MAX = ZOOMPIXELS + 20; // can draw up to this pixels in corner
if ( (mouseX < MAX) && (mouseY < MAX) && (pmouseX < MAX) && (pmouseY < MAX) )
{
mousedrag = true; // start a mouse drag
doodle_exists = true;
doodle.stroke('white');
doodle.strokeWeight( DOODLE_THICK );
doodle.line(mouseX, mouseY, pmouseX, pmouseY);
}
}
else if ( mousedrag )
{
mousedrag = false;
console.log ("Exiting draw. Now blurring.");
doodle.filter (BLUR, DOODLE_BLUR); // just blur once
// console.log (doodle);
} else if ( do_training ) {
// do some training per step
for (let i = 0; i < TRAINPERSTEP; i++)
{
if (i === 0) trainit(true); // show only one per step - still flashes by
else trainit(false);
}
// do some testing per step
for (let i = 0; i < TESTPERSTEP; i++)
testit();
}
}
//--- demo -------------------------------------------------------------
// demo some test image and predict it
// get it from test set so have not used it in training
function makeDemo()
{
demo_exists = true;
var i = AB.randomIntAtoB ( 0, NOTEST - 1 );
demo = mnist.test_images[i];
var label = mnist.test_labels[i];
thehtml = "Test image no: " + i + "<br>" +
"Classification: " + label + "<br>" ;
AB.msg ( thehtml, 8 );
// type "demo" in console to see raw data
}
function drawDemo()
{
var theimage = getImage ( demo );
// console.log (theimage);
image ( theimage, 0, canvasheight - ZOOMPIXELS, ZOOMPIXELS, ZOOMPIXELS ); // magnified
image ( theimage, ZOOMPIXELS+50, canvasheight - ZOOMPIXELS, PIXELS, PIXELS ); // original
}
function guessDemo()
{
let inputs = getInputs ( demo );
demo_inputs = inputs; // can inspect in console
let result = nn.classifyImages( [ toModelFormat(demo, CROP_PIXELS) ] );
let guess = findMax(result);
thehtml = " We classify it as: " + greenspan + guess + "</span>" ;
AB.msg ( thehtml, 9 );
}
//--- doodle -------------------------------------------------------------
function drawDoodle()
{
// doodle is createGraphics not createImage
let theimage = doodle.get();
// console.log (theimage);
image ( theimage, 0, 0, ZOOMPIXELS, ZOOMPIXELS ); // original
image ( theimage, ZOOMPIXELS+50, 0, PIXELS, PIXELS ); // shrunk
}
function guessDoodle()
{
// doodle is createGraphics not createImage
let img = doodle.get();
img.resize ( PIXELS, PIXELS );
img.loadPixels();
// feed forward to make prediction
const result = nn.classifyImages( [ toModelFormat(centerImage(img.pixels, PIXELS), CROP_PIXELS)] );
let guess = find12(result);
thehtml = " We classify it as: " + greenspan + guess[0] + "</span> <br>" +
" No.2 guess is: " + greenspan + guess[1] + "</span>";
AB.msg ( thehtml, 2 );
}
function wipeDoodle()
{
doodle_exists = false;
doodle.background('black');
}
// --- debugging --------------------------------------------------
// in console
// showInputs(demo_inputs);
// showInputs(doodle_inputs);
function showInputs ( inputs )
// display inputs row by row, corresponding to square of pixels
{
var str = "";
for (let i = 0; i < inputs.length; i++)
{
if ( i % PIXELS === 0 ) str = str + "\n"; // new line for each row of pixels
var value = inputs[i];
str = str + " " + value.toFixed(2) ;
}
console.log (str);
}