|
1 |
| -class SVMWorker extends BaseWorker { |
2 |
| - constructor() { |
3 |
| - super('model/worker/svm_worker.js'); |
| 1 | +const Kernel = { |
| 2 | + gaussian: |
| 3 | + (d = 1) => |
| 4 | + (a, b) => { |
| 5 | + let r = a.reduce((acc, v, i) => acc + (v - b[i]) ** 2, 0) |
| 6 | + return Math.exp(-r / (2 * d * d)) |
| 7 | + }, |
| 8 | + linear: () => (a, b) => a.reduce((acc, v, i) => acc + v * b[i], 0), |
| 9 | +} |
| 10 | + |
| 11 | +class SVM { |
| 12 | + constructor(kernel) { |
| 13 | + this._n = 0 |
| 14 | + this._a = [] |
| 15 | + this._x = [] |
| 16 | + this._t = [] |
| 17 | + this._b = 0 |
| 18 | + |
| 19 | + this._C = 1000 |
| 20 | + this._eps = 0.001 |
| 21 | + this._tolerance = 0.001 |
| 22 | + this._err = [] |
| 23 | + |
| 24 | + this._kernel = kernel |
4 | 25 | }
|
5 | 26 |
|
6 |
| - initialize(kernel, train_x, train_y, method = "oneone") { |
7 |
| - this._postMessage({ |
8 |
| - "mode": "init", |
9 |
| - "method": method, |
10 |
| - "kernel": kernel, |
11 |
| - "x": train_x, |
12 |
| - "y": train_y |
13 |
| - }); |
| 27 | + init(train_x, train_y) { |
| 28 | + this._n = train_x.length |
| 29 | + this._a = Array(this._n).fill(0) |
| 30 | + this._x = train_x.map((x) => x) |
| 31 | + this._t = train_y |
| 32 | + this._err = Array(this._n).fill(0) |
| 33 | + this._alldata = true |
14 | 34 | }
|
15 | 35 |
|
16 |
| - fit(iteration, cb) { |
17 |
| - this._postMessage({ |
18 |
| - "mode": "fit", |
19 |
| - "iteration": iteration |
20 |
| - }, cb); |
| 36 | + fit() { |
| 37 | + let changed = this._fitOnce(this._alldata) |
| 38 | + if (this._alldata) { |
| 39 | + this._alldata = false |
| 40 | + if (changed == 0) { |
| 41 | + return |
| 42 | + } |
| 43 | + } else if (changed == 0) { |
| 44 | + this._alldata = true |
| 45 | + } |
21 | 46 | }
|
22 | 47 |
|
23 |
| - predict(x, cb) { |
24 |
| - this._postMessage({ |
25 |
| - "mode": "predict", |
26 |
| - "x": x |
27 |
| - }, cb); |
| 48 | + _fitOnce(all = false) { |
| 49 | + let change = 0 |
| 50 | + |
| 51 | + const between_eps = (v) => this._eps < v && v < this._C - this._eps |
| 52 | + for (let i = 0; i < this._n; i++) { |
| 53 | + let ei = 0 |
| 54 | + if (between_eps(this._a[i])) { |
| 55 | + ei = this._err[i] |
| 56 | + } else if (all) { |
| 57 | + ei = this.predict(this._x[i]) - this._t[i] |
| 58 | + } else { |
| 59 | + continue |
| 60 | + } |
| 61 | + const yfi = ei * this._t[i] |
| 62 | + |
| 63 | + if ( |
| 64 | + (this._a[i] >= this._C - this._eps || yfi >= -this._tolerance) && |
| 65 | + (this._a[i] <= this._eps || yfi <= this._tolerance) |
| 66 | + ) { |
| 67 | + continue |
| 68 | + } |
| 69 | + |
| 70 | + let max_e = 0 |
| 71 | + let max_j = -1 |
| 72 | + |
| 73 | + const offset = Math.floor(Math.random() * (this._n + 1)) |
| 74 | + let in_eps = [] |
| 75 | + let out_eps = [] |
| 76 | + for (let j = 0; j < this._n; j++) { |
| 77 | + const p = (j + offset) % this._n |
| 78 | + if (p === i) { |
| 79 | + continue |
| 80 | + } |
| 81 | + if (between_eps(this._a[p])) { |
| 82 | + const ej = this._err[p] |
| 83 | + if (Math.abs(ei - ej) > max_e) { |
| 84 | + max_e = Math.abs(ei - ej) |
| 85 | + if (max_j >= 0) in_eps.push(max_j) |
| 86 | + max_j = p |
| 87 | + } else { |
| 88 | + in_eps.push(p) |
| 89 | + } |
| 90 | + } else { |
| 91 | + out_eps.push(p) |
| 92 | + } |
| 93 | + } |
| 94 | + const checks = max_j >= 0 ? [].concat(max_j, in_eps, out_eps) : [].concat(in_eps, out_eps) |
| 95 | + for (let ck = 0; ck < checks.length; ck++) { |
| 96 | + const j = checks[ck] |
| 97 | + |
| 98 | + const ai_old = this._a[i] |
| 99 | + const aj_old = this._a[j] |
| 100 | + let u, v |
| 101 | + if (this._t[i] != this._t[j]) { |
| 102 | + u = Math.max(0, ai_old - aj_old) |
| 103 | + v = Math.min(this._C, this._C + ai_old - aj_old) |
| 104 | + } else { |
| 105 | + u = Math.max(0, ai_old + aj_old - this._C) |
| 106 | + v = Math.min(this._C, ai_old + aj_old) |
| 107 | + } |
| 108 | + if (u == v) { |
| 109 | + continue |
| 110 | + } |
| 111 | + |
| 112 | + const kii = this._kernel(this._x[i], this._x[i]) |
| 113 | + const kjj = this._kernel(this._x[j], this._x[j]) |
| 114 | + const kij = this._kernel(this._x[i], this._x[j]) |
| 115 | + const k = kii + kjj - 2 * kij |
| 116 | + const ej = between_eps(this._a[j]) ? this._err[j] : this.predict(this._x[j]) - this._t[j] |
| 117 | + |
| 118 | + let bClip = false |
| 119 | + let ai_new = 0, |
| 120 | + aj_new = 0 |
| 121 | + if (k <= 0) { |
| 122 | + let lh = [u, v].map((t) => { |
| 123 | + let ai_n = t |
| 124 | + let aj_n = aj_old + this._t[i] * this._t[j] * (ai_old - ai_n) |
| 125 | + this._a[i] = ai_n |
| 126 | + this._a[j] = aj_n |
| 127 | + const v1 = |
| 128 | + this.predict(this._x[j]) + this._b - this._t[j] * aj_old * kjj - this._t[i] * ai_old * kij |
| 129 | + const v2 = |
| 130 | + this.predict(this._x[i]) + this._b - this._t[j] * aj_old * kij - this._t[i] * ai_old * kii |
| 131 | + const lobj = |
| 132 | + aj_n + |
| 133 | + ai_n - |
| 134 | + (kjj * aj_n ** 2) / 2 - |
| 135 | + (kii * ai_n ** 2) / 2 - |
| 136 | + this._t[j] * this._t[i] * kij * aj_n * ai_n - |
| 137 | + this._t[j] * aj_n * v1 - |
| 138 | + this._t[i] * ai_n * v2 |
| 139 | + }) |
| 140 | + this._a[i] = ai_old |
| 141 | + this._a[j] = aj_old |
| 142 | + |
| 143 | + ai_new = lh[0] > lh[1] + this._eps ? u : lh[0] < lh[1] - this._eps ? v : ai_old |
| 144 | + bClip = true |
| 145 | + } else { |
| 146 | + ai_new = ai_old + (this._t[i] * (ej - ei)) / k |
| 147 | + if (ai_new > v) { |
| 148 | + bClip = true |
| 149 | + ai_new = v |
| 150 | + } else if (ai_new < u) { |
| 151 | + bClip = true |
| 152 | + ai_new = u |
| 153 | + } |
| 154 | + } |
| 155 | + |
| 156 | + if (Math.abs(ai_new - ai_old) < this._eps * (ai_new + ai_old + this._eps)) { |
| 157 | + continue |
| 158 | + } |
| 159 | + aj_new = aj_old + this._t[i] * this._t[j] * (ai_old - ai_new) |
| 160 | + const b_old = this._b |
| 161 | + if (between_eps(this._a[i])) { |
| 162 | + this._b += ei + (ai_new - ai_old) * this._t[i] * kii + (aj_new - aj_old) * this._t[j] * kij |
| 163 | + } else if (between_eps(this._a[j])) { |
| 164 | + this._b += ej + (ai_new - ai_old) * this._t[i] * kij + (aj_new - aj_old) * this._t[j] * kjj |
| 165 | + } else { |
| 166 | + this._b += |
| 167 | + (ei + |
| 168 | + ej + |
| 169 | + (ai_new - ai_old) * this._t[i] * (kii + kij) + |
| 170 | + (aj_new - aj_old) * this._t[j] * (kij + kjj)) / |
| 171 | + 2 |
| 172 | + } |
| 173 | + |
| 174 | + for (let m = 0; m < this._n; m++) { |
| 175 | + if (m == i || m == j) { |
| 176 | + continue |
| 177 | + } |
| 178 | + this._err[m] += |
| 179 | + this._t[j] * (aj_new - aj_old) * this._kernel(this._x[j], this._x[m]) + |
| 180 | + this._t[i] * (ai_new - ai_old) * this._kernel(this._x[i], this._x[m]) + |
| 181 | + b_old - |
| 182 | + this._b |
| 183 | + } |
| 184 | + |
| 185 | + this._a[i] = ai_new |
| 186 | + this._a[j] = aj_new |
| 187 | + |
| 188 | + if (!bClip) { |
| 189 | + this._err[i] = 0 |
| 190 | + } else if (between_eps(ai_new)) { |
| 191 | + this._err[i] = this.predict(this._x[i]) - this._t[i] |
| 192 | + } |
| 193 | + this._err[j] = this.predict(this._x[j]) - this._t[j] |
| 194 | + |
| 195 | + change++ |
| 196 | + break |
| 197 | + } |
| 198 | + } |
| 199 | + return change |
| 200 | + } |
| 201 | + |
| 202 | + predict(data) { |
| 203 | + const f = (v) => { |
| 204 | + let y = 0 |
| 205 | + for (let n = 0; n < this._n; n++) { |
| 206 | + if (this._a[n]) y += this._a[n] * this._t[n] * this._kernel(v, this._x[n]) |
| 207 | + } |
| 208 | + return y - this._b |
| 209 | + } |
| 210 | + return !Array.isArray(data[0]) ? f(data) : data.map(f) |
28 | 211 | }
|
29 | 212 | }
|
30 | 213 |
|
31 |
| -var dispSVM = function(elm, platform) { |
32 |
| - const step = 4; |
33 |
| - let model = new SVMWorker(); |
34 |
| - let learn_epoch = 0; |
| 214 | +var dispSVM = function (elm, platform) { |
| 215 | + const step = 4 |
| 216 | + let model = null |
| 217 | + let learn_epoch = 0 |
35 | 218 |
|
36 |
| - const calcSVM = function(cb) { |
| 219 | + const calcSVM = function (cb) { |
37 | 220 | if (platform.datas.length == 0) {
|
38 |
| - return; |
| 221 | + return |
39 | 222 | }
|
40 |
| - const iteration = +elm.select("[name=iteration]").property("value"); |
| 223 | + const iteration = +elm.select('[name=iteration]').property('value') |
41 | 224 | platform.fit((tx, ty, fit_cb) => {
|
42 |
| - model.fit(iteration, e => { |
43 |
| - platform.predict((px, pred_cb) => { |
44 |
| - model.predict(px, e => { |
45 |
| - let data = e.data; |
46 |
| - pred_cb(data); |
47 |
| - learn_epoch += iteration |
48 |
| - cb && cb(); |
49 |
| - }); |
50 |
| - }, step) |
51 |
| - }); |
52 |
| - }); |
53 |
| - }; |
54 |
| - |
55 |
| - elm.append("select") |
56 |
| - .attr("name", "method") |
57 |
| - .selectAll("option") |
58 |
| - .data(["oneone", "onerest"]) |
| 225 | + for (let i = 0; i < iteration; i++) { |
| 226 | + model.fit() |
| 227 | + } |
| 228 | + platform.predict((px, pred_cb) => { |
| 229 | + const data = model.predict(px) |
| 230 | + pred_cb(data) |
| 231 | + learn_epoch += iteration |
| 232 | + cb && cb() |
| 233 | + }, step) |
| 234 | + }) |
| 235 | + } |
| 236 | + |
| 237 | + elm.append('select') |
| 238 | + .attr('name', 'method') |
| 239 | + .selectAll('option') |
| 240 | + .data(['oneone', 'onerest']) |
59 | 241 | .enter()
|
60 |
| - .append("option") |
61 |
| - .property("value", d => d) |
62 |
| - .text(d => d); |
63 |
| - elm.append("select") |
64 |
| - .attr("name", "kernel") |
65 |
| - .on("change", function() { |
66 |
| - const k = d3.select(this).property("value"); |
67 |
| - if (k == "gaussian") { |
68 |
| - elm.select("input[name=gamma]").style("display", "inline"); |
| 242 | + .append('option') |
| 243 | + .property('value', (d) => d) |
| 244 | + .text((d) => d) |
| 245 | + elm.append('select') |
| 246 | + .attr('name', 'kernel') |
| 247 | + .on('change', function () { |
| 248 | + const k = d3.select(this).property('value') |
| 249 | + if (k == 'gaussian') { |
| 250 | + elm.select('input[name=gamma]').style('display', 'inline') |
69 | 251 | } else {
|
70 |
| - elm.select("input[name=gamma]").style("display", "none"); |
| 252 | + elm.select('input[name=gamma]').style('display', 'none') |
71 | 253 | }
|
72 | 254 | })
|
73 |
| - .selectAll("option") |
74 |
| - .data(["gaussian", "linear"]) |
| 255 | + .selectAll('option') |
| 256 | + .data(['gaussian', 'linear']) |
75 | 257 | .enter()
|
76 |
| - .append("option") |
77 |
| - .property("value", d => d) |
78 |
| - .text(d => d); |
79 |
| - elm.append("input") |
80 |
| - .attr("type", "number") |
81 |
| - .attr("name", "gamma") |
82 |
| - .attr("value", 1) |
83 |
| - .attr("min", 0.01) |
84 |
| - .attr("max", 10.0) |
85 |
| - .attr("step", 0.01); |
| 258 | + .append('option') |
| 259 | + .property('value', (d) => d) |
| 260 | + .text((d) => d) |
| 261 | + elm.append('input') |
| 262 | + .attr('type', 'number') |
| 263 | + .attr('name', 'gamma') |
| 264 | + .attr('value', 1) |
| 265 | + .attr('min', 0.01) |
| 266 | + .attr('max', 10.0) |
| 267 | + .attr('step', 0.01) |
86 | 268 | const slbConf = platform.setting.ml.controller.stepLoopButtons().init(() => {
|
87 |
| - let kernel = elm.select("[name=kernel]").property("value"); |
88 |
| - if (kernel == "gaussian") { |
89 |
| - kernel = [kernel, +elm.select("input[name=gamma]").property("value")]; |
| 269 | + const kernel = elm.select('[name=kernel]').property('value') |
| 270 | + const kernel_args = [] |
| 271 | + if (kernel == 'gaussian') { |
| 272 | + kernel_args.push(+elm.select('input[name=gamma]').property('value')) |
| 273 | + } |
| 274 | + const method = elm.select('[name=method]').property('value') |
| 275 | + if (method == 'onerest') { |
| 276 | + model = new OneVsRestModel(SVM, null, [Kernel[kernel](...kernel_args)]) |
| 277 | + } else { |
| 278 | + model = new OneVsOneModel(SVM, null, [Kernel[kernel](...kernel_args)]) |
90 | 279 | }
|
91 | 280 | platform.fit((tx, ty) => {
|
92 |
| - model.initialize(kernel, tx, ty.map(v => v[0]), elm.select("[name=method]").property("value")); |
| 281 | + model.init( |
| 282 | + tx, |
| 283 | + ty.map((v) => v[0]) |
| 284 | + ) |
93 | 285 | })
|
94 | 286 | learn_epoch = 0
|
95 | 287 | platform.init()
|
96 | 288 | })
|
97 |
| - elm.append("span") |
98 |
| - .text(" Iteration "); |
99 |
| - elm.append("select") |
100 |
| - .attr("name", "iteration") |
101 |
| - .selectAll("option") |
| 289 | + elm.append('span').text(' Iteration ') |
| 290 | + elm.append('select') |
| 291 | + .attr('name', 'iteration') |
| 292 | + .selectAll('option') |
102 | 293 | .data([1, 10, 100, 1000])
|
103 | 294 | .enter()
|
104 |
| - .append("option") |
105 |
| - .property("value", d => d) |
106 |
| - .text(d => d); |
| 295 | + .append('option') |
| 296 | + .property('value', (d) => d) |
| 297 | + .text((d) => d) |
107 | 298 | slbConf.step(calcSVM).epoch(() => learn_epoch)
|
108 |
| - |
109 |
| - return () => { |
110 |
| - model.terminate(); |
111 |
| - }; |
112 | 299 | }
|
113 | 300 |
|
114 |
| -export default function(platform) { |
| 301 | +export default function (platform) { |
115 | 302 | platform.setting.ml.usage = 'Click and add data point. Then, click "Calculate".'
|
116 |
| - platform.setting.terminate = dispSVM(platform.setting.ml.configElement, platform); |
| 303 | + dispSVM(platform.setting.ml.configElement, platform) |
117 | 304 | }
|
0 commit comments