forked from RyanLinXiang/neural-network-with-js
-
Notifications
You must be signed in to change notification settings - Fork 0
/
nn-with-stop.js
436 lines (351 loc) · 12.6 KB
/
nn-with-stop.js
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
/** these short cuts of mathJS functions make our life easier */
const mmap = math.map; // to be used to pass each element of a matrix to a function
const rand = math.random;
const transp = math.transpose;
const mat = math.matrix;
const e = math.evaluate;
const sub = math.subtract;
const sqr = math.square;
const sum = math.sum;
class NeuralNetwork {
constructor(inputnodes, hiddennodes, outputnodes, learningrate, wih, who) {
this.inputnodes = inputnodes;
this.hiddennodes = hiddennodes;
this.outputnodes = outputnodes;
this.learningrate = learningrate;
/* initialise the weights either randomly or, if passed in as arguments, with pretrained values */
/* wih = weights of input-to-hidden layer */
/* who = weights of hidden-to-output layer */
this.wih = wih || sub(mat(rand([hiddennodes, inputnodes])), 0.5);
this.who = who || sub(mat(rand([outputnodes, hiddennodes])), 0.5);
/* the sigmoid activation function */
this.act = (matrix) => mmap(matrix, (x) => 1 / (1 + Math.exp(-x)));
}
static normalizeData = (data) => {
return data.map((e) => (e / 255) * 0.99 + 0.01);
};
cache = { loss: [] };
forward = (input) => {
const wih = this.wih;
const who = this.who;
const act = this.act;
input = transp(mat([input]));
/* hidden layer */
const h_in = e("wih * input", { wih, input });
const h_out = act(h_in);
/* output layer */
const o_in = e("who * h_out", { who, h_out });
const actual = act(o_in);
/* these values are needed later in "backward" */
this.cache.input = input;
this.cache.h_out = h_out;
this.cache.actual = actual;
return actual;
};
backward = (target) => {
const who = this.who;
const input = this.cache.input;
const h_out = this.cache.h_out;
const actual = this.cache.actual;
target = transp(mat([target]));
// calculate the gradient of the error function (E) w.r.t the activation function (A)
const dEdA = sub(target, actual);
// calculate the gradient of the activation function (A) w.r.t the weighted sums (Z) of the output layer
const o_dAdZ = e("actual .* (1 - actual)", {
actual,
});
// calculate the error gradient of the loss function w.r.t the weights of the hidden-to-output layer
const dwho = e("(dEdA .* o_dAdZ) * h_out'", {
dEdA,
o_dAdZ,
h_out,
});
// calculate the weighted error for the hidden layer
const h_err = e("who' * (dEdA .* o_dAdZ)", {
who,
dEdA,
o_dAdZ,
});
// calculate the gradient of the activation function (A) w.r.t the weighted sums (Z) of the hidden layer
const h_dAdZ = e("h_out .* (1 - h_out)", {
h_out,
});
// calculate the error gradient of the loss function w.r.t the weights of the input-to-hidden layer
const dwih = e("(h_err .* h_dAdZ) * input'", {
h_err,
h_dAdZ,
input,
});
this.cache.dwih = dwih;
this.cache.dwho = dwho;
this.cache.loss.push(sum(sqr(dEdA)));
};
update = () => {
const wih = this.wih;
const who = this.who;
const dwih = this.cache.dwih;
const dwho = this.cache.dwho;
const r = this.learningrate;
/* update the current weights of each layer with their corresponding error gradients */
/* error gradients are negated by using the positve sign */
this.wih = e("wih + (r .* dwih)", { wih, r, dwih });
this.who = e("who + (r .* dwho)", { who, r, dwho });
};
predict = (input) => {
return this.forward(input);
};
train = (input, target) => {
this.forward(input);
this.backward(target);
this.update();
};
}
/* neural network's hyper parameters */
const inputnodes = 784;
const hiddennodes = 100;
const outputnodes = 10;
const learningrate = 0.2;
const threshold = 0.5;
let iter = 0;
const iterations = 5;
/* path to the data sets */
const trainingDataPath = "./mnist/mnist_train_100.csv";
const testDataPath = "./mnist/mnist_test_10.csv";
const weightsFilename = "weights.json";
const savedWeightsPath = `./dist/${weightsFilename}`;
/* these constants will be filled during data loading and preparation */
const trainingData = [];
const trainingLabels = [];
const testData = [];
const testLabels = [];
const savedWeights = {};
/* states after how many trained samples a log message should appear */
const printSteps = 1000;
/* save all setTimeouts to clear them when training or testing should be aborted */
const timeouts = [];
/* save references to the stop function to clear up event listeners when training or testing is aborted and restarted */
const stopTraining = () => {
stop("training", "Aborted training.", true);
};
const stopTesting = () => {
stop("esting", "Aborted testing.", false);
};
/* button labels for different states: either in training/testing or aborted */
const trainButtonLabel = "Train";
const testButtonLabel = "Test";
const stopButtonLabel = "Stop!";
let myNN;
window.onload = async () => {
/* Instantiate an entity from the NeuralNetwork class */
myNN = new NeuralNetwork(inputnodes, hiddennodes, outputnodes, learningrate);
trainButton.disabled = true;
testButton.disabled = true;
loadWeightsButton.disabled = true;
status.innerHTML = "Loading the data sets. Please wait ...<br>";
/* get all the data set files and do the preparations */
const trainCSV = await loadData(trainingDataPath, "CSV");
if (trainCSV) {
prepareData(trainCSV, trainingData, trainingLabels);
status.innerHTML += "Training data successfully loaded...<br>";
}
const testCSV = await loadData(testDataPath, "CSV");
if (testCSV) {
prepareData(testCSV, testData, testLabels);
status.innerHTML += "Test data successfully loaded...<br>";
}
if (!trainCSV || !testCSV) {
status.innerHTML +=
"Error loading train/test data set. Please check your file path!";
return;
}
trainButton.disabled = false;
testButton.disabled = false;
const weightsJSON = await loadData(savedWeightsPath, "JSON");
/* if there is a saved JSON file with pretrained weights existing, save the content in the weightsJSON constant */
if (weightsJSON) {
savedWeights.wih = weightsJSON.wih;
savedWeights.who = weightsJSON.who;
loadWeightsButton.disabled = false;
}
status.innerHTML += "Ready.<br><br>";
};
async function loadData(path, type) {
try {
const result = await fetch(path, {
mode: "no-cors",
});
switch (type) {
case "CSV":
return await result.text();
break;
case "JSON":
return await result.json();
break;
default:
return false;
}
} catch {
return false;
}
}
function prepareData(rawData, target, labels) {
rawData = rawData.split("\n"); // create an array where each element correspondents to one line in the CSV file
rawData.pop(); // remove the last element which is empty because it refers to a last blank line in the CSV file
rawData.forEach((current) => {
let sample = current.split(",").map((x) => +x); // create an array where each element has a grey color value
labels.push(sample[0]); // extract the first element of the sample which is (mis)used as the label
sample.shift(); // remove the first element
sample = NeuralNetwork.normalizeData(sample);
target.push(sample);
});
}
function train(_, inTraining = false) {
trainButton.innerHTML = stopButtonLabel;
if (!inTraining) {
trainButton.removeEventListener("click", train);
trainButton.addEventListener("click", stopTraining);
}
testButton.disabled = true;
loadWeightsButton.disabled = true;
download.innerHTML = "";
if (iter < iterations) {
iter++;
status.innerHTML += "Starting training ...<br>";
status.innerHTML += "Iteration " + iter + " of " + iterations + "<br>";
trainingData.forEach((current, index) => {
timeouts.push(
setTimeout(() => {
/* create one-hot encoding of the label */
const label = trainingLabels[index];
const oneHotLabel = Array(10).fill(0);
oneHotLabel[label] = 0.99;
myNN.train(current, oneHotLabel);
/* check if the defined interval for showing a message on the training progress is reached */
if (index > 0 && !((index + 1) % printSteps)) {
status.innerHTML += `finished ${index + 1} samples ... <br>`;
}
/* check if the end of the training iteration is reached */
if (index === trainingData.length - 1) {
status.innerHTML += `Loss: ${
sum(myNN.cache.loss) / trainingData.length
}<br><br>`;
myNN.cache.loss = [];
test("", true); // true to signal "test" that it is called from within training
}
}, 0)
);
});
}
}
function test(_, inTraining = false) {
// skip the first parameter as it includes data from the event listener which we don't need here
if (!inTraining) {
testButton.innerHTML = stopButtonLabel;
testButton.removeEventListener("click", test);
testButton.addEventListener("click", stopTesting);
trainButton.disabled = true;
}
loadWeightsButton.disabled = true;
status.innerHTML += "Starting testing ...<br>";
let correctPredicts = 0;
testData.forEach((current, index) => {
timeouts.push(
setTimeout(() => {
const actual = testLabels[index];
const predict = formatPrediction(myNN.predict(current));
predict === actual ? correctPredicts++ : null;
/* check if the defined interval for showing a message on the testing progress is reached */
if (index > 0 && !((index + 1) % printSteps)) {
status.innerHTML += " finished " + (index + 1) + " samples ...<br>";
}
/* check if testing is complete */
if (index >= testData.length - 1) {
status.innerHTML +=
"Accuracy: " +
Math.round((correctPredicts / testData.length) * 100) +
" %<br><br>";
/* check if training is complete */
if (iter + 1 > iterations) {
createDownloadLink();
enableAllButtons();
resetButtons("training");
status.innerHTML += "Finished training.<br><br>";
iter = 0;
} else if (inTraining) {
// if test is called from within training and the training is not complete yet, continue training
train("", true);
} else {
enableAllButtons();
resetButtons("testing");
}
}
}, 0)
);
});
}
function predict() {
/* resize the canvas to the training data image size */
const tempCanvas = document.createElement("canvas");
const tempCtx = tempCanvas.getContext("2d");
tempCtx.drawImage(canvas, 0, 0, 150, 150, 0, 0, 28, 28);
/* convert the canvas image */
const img = tempCtx.getImageData(0, 0, 28, 28);
/* remove the alpha channel and convert to grayscale */
let sample = [];
for (let i = 0, j = 0; i < img.data.length; i += 4, j++) {
sample[j] = (img.data[i + 0] + img.data[i + 1] + img.data[i + 2]) / 3;
}
img.data = NeuralNetwork.normalizeData(img.data);
const predict = formatPrediction(myNN.predict(sample));
prediction.innerHTML = predict;
}
function formatPrediction(prediction) {
const flattened = prediction.toArray().map((x) => x[0]);
/* get the index of the highest number in the flattened array */
return flattened.indexOf(Math.max(...flattened));
}
function loadWeights() {
myNN.wih = savedWeights.wih;
myNN.who = savedWeights.who;
status.innerHTML += "Weights successfully loaded.";
}
function createDownloadLink() {
const wih = myNN.wih.toArray();
const who = myNN.who.toArray();
const weights = { wih, who };
download.innerHTML = `<a download="${weightsFilename}" id="downloadLink" href="data:text/json;charset=utf-8,${encodeURIComponent(
JSON.stringify(weights)
)}">Download model weights</a>`;
}
/* UI helper functions */
function enableAllButtons() {
trainButton.disabled = false;
testButton.disabled = false;
loadWeightsButton.disabled = false;
}
function stop(mode, message = "", downloadLink = false) {
timeouts.forEach((e) => clearTimeout(e));
timeouts.splice(0, timeouts.length);
status.innerHTML += `${message}<br><br>`;
iter = 0;
if (mode === "training") {
resetButtons("training");
} else {
resetButtons("testing");
}
enableAllButtons();
if (downloadLink) {
createDownloadLink();
}
}
function resetButtons(mode) {
if (mode === "training") {
trainButton.removeEventListener("click", stopTraining);
trainButton.addEventListener("click", train);
trainButton.innerHTML = trainButtonLabel;
} else {
testButton.removeEventListener("click", stopTesting);
testButton.addEventListener("click", test);
testButton.innerHTML = testButtonLabel;
}
}