[8607] | 1 | /*
|
---|
| 2 | * Copyright (c) 2000-2012 Chih-Chung Chang and Chih-Jen Lin
|
---|
| 3 | * All rights reserved.
|
---|
| 4 | *
|
---|
| 5 | * Redistribution and use in source and binary forms, with or without
|
---|
| 6 | * modification, are permitted provided that the following conditions
|
---|
| 7 | * are met:
|
---|
| 8 | *
|
---|
| 9 | * 1. Redistributions of source code must retain the above copyright
|
---|
| 10 | * notice, this list of conditions and the following disclaimer.
|
---|
| 11 | *
|
---|
| 12 | * 2. Redistributions in binary form must reproduce the above copyright
|
---|
| 13 | * notice, this list of conditions and the following disclaimer in the
|
---|
| 14 | * documentation and/or other materials provided with the distribution.
|
---|
| 15 | *
|
---|
| 16 | * 3. Neither name of copyright holders nor the names of its contributors
|
---|
| 17 | * may be used to endorse or promote products derived from this software
|
---|
| 18 | * without specific prior written permission.
|
---|
| 19 | *
|
---|
| 20 | *
|
---|
| 21 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
---|
| 22 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
---|
| 23 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
---|
| 24 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR
|
---|
| 25 | * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
---|
| 26 | * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
---|
| 27 | * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
---|
| 28 | * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
|
---|
| 29 | * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
|
---|
| 30 | * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
|
---|
| 31 | * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
---|
| 32 | *
|
---|
| 33 | * C# port from the original java sources by Gabriel Kronberger (Sept. 2012)
|
---|
| 34 | */
|
---|
| 35 |
|
---|
| 36 | using System.Globalization;
|
---|
| 37 | using System.IO;
|
---|
| 38 | using System.Linq;
|
---|
| 39 | using System.Threading;
|
---|
| 40 |
|
---|
| 41 | namespace LibSVM {
|
---|
| 42 | //
|
---|
| 43 | // Kernel Cache
|
---|
| 44 | //
|
---|
| 45 | // l is the number of total data items
|
---|
| 46 | // size is the cache size limit in bytes
|
---|
| 47 | //
|
---|
| 48 |
|
---|
| 49 | using System;
|
---|
| 50 |
|
---|
| 51 | class Cache {
|
---|
| 52 | private readonly int l;
|
---|
| 53 | private long size;
|
---|
| 54 | private sealed class head_t {
|
---|
| 55 | public head_t prev, next; // a cicular list
|
---|
| 56 | public float[] data;
|
---|
| 57 | public int len; // data[0,len) is cached in this entry
|
---|
| 58 | }
|
---|
| 59 | private readonly head_t[] head;
|
---|
| 60 | private head_t lru_head;
|
---|
| 61 |
|
---|
| 62 | public Cache(int l_, long size_) {
|
---|
| 63 | l = l_;
|
---|
| 64 | size = size_;
|
---|
| 65 | head = new head_t[l];
|
---|
| 66 | for (int i = 0; i < l; i++) head[i] = new head_t();
|
---|
| 67 | size /= 4;
|
---|
| 68 | size -= l * (16 / 4); // sizeof(head_t) == 16
|
---|
| 69 | size = Math.Max(size, 2 * (long)l); // cache must be large enough for two columns
|
---|
| 70 | lru_head = new head_t();
|
---|
| 71 | lru_head.next = lru_head.prev = lru_head;
|
---|
| 72 | }
|
---|
| 73 |
|
---|
| 74 | private void lru_delete(head_t h) {
|
---|
| 75 | // delete from current location
|
---|
| 76 | h.prev.next = h.next;
|
---|
| 77 | h.next.prev = h.prev;
|
---|
| 78 | }
|
---|
| 79 |
|
---|
| 80 | private void lru_insert(head_t h) {
|
---|
| 81 | // insert to last position
|
---|
| 82 | h.next = lru_head;
|
---|
| 83 | h.prev = lru_head.prev;
|
---|
| 84 | h.prev.next = h;
|
---|
| 85 | h.next.prev = h;
|
---|
| 86 | }
|
---|
| 87 |
|
---|
| 88 | // request data [0,len)
|
---|
| 89 | // return some position p where [p,len) need to be filled
|
---|
| 90 | // (p >= len if nothing needs to be filled)
|
---|
| 91 | // java: simulate pointer using single-element array
|
---|
| 92 | public int get_data(int index, float[][] data, int len) {
|
---|
| 93 | head_t h = head[index];
|
---|
| 94 | if (h.len > 0) lru_delete(h);
|
---|
| 95 | int more = len - h.len;
|
---|
| 96 |
|
---|
| 97 | if (more > 0) {
|
---|
| 98 | // free old space
|
---|
| 99 | while (size < more) {
|
---|
| 100 | head_t old = lru_head.next;
|
---|
| 101 | lru_delete(old);
|
---|
| 102 | size += old.len;
|
---|
| 103 | old.data = null;
|
---|
| 104 | old.len = 0;
|
---|
| 105 | }
|
---|
| 106 |
|
---|
| 107 | // allocate new space
|
---|
| 108 | float[] new_data = new float[len];
|
---|
| 109 | if (h.data != null) {
|
---|
| 110 | Array.Copy(h.data, 0, new_data, 0, h.len);
|
---|
| 111 | }
|
---|
| 112 | h.data = new_data;
|
---|
| 113 | size -= more; { int _ = h.len; h.len = len; len = _; }
|
---|
| 114 | }
|
---|
| 115 |
|
---|
| 116 | lru_insert(h);
|
---|
| 117 | data[0] = h.data;
|
---|
| 118 | return len;
|
---|
| 119 | }
|
---|
| 120 |
|
---|
| 121 | public void swap_index(int i, int j) {
|
---|
| 122 | if (i == j) return;
|
---|
| 123 |
|
---|
| 124 | if (head[i].len > 0) lru_delete(head[i]);
|
---|
| 125 | if (head[j].len > 0) lru_delete(head[j]); { float[] _ = head[i].data; head[i].data = head[j].data; head[j].data = _; }
|
---|
| 126 | { int _ = head[i].len; head[i].len = head[j].len; head[j].len = _; }
|
---|
| 127 | if (head[i].len > 0) lru_insert(head[i]);
|
---|
| 128 | if (head[j].len > 0) lru_insert(head[j]);
|
---|
| 129 |
|
---|
| 130 | if (i > j) { int _ = i; i = j; j = _; }
|
---|
| 131 | for (head_t h = lru_head.next; h != lru_head; h = h.next) {
|
---|
| 132 | if (h.len > i) {
|
---|
| 133 | if (h.len > j) { float _ = h.data[i]; h.data[i] = h.data[j]; h.data[j] = _; } else {
|
---|
| 134 | // give up
|
---|
| 135 | lru_delete(h);
|
---|
| 136 | size += h.len;
|
---|
| 137 | h.data = null;
|
---|
| 138 | h.len = 0;
|
---|
| 139 | }
|
---|
| 140 | }
|
---|
| 141 | }
|
---|
| 142 | }
|
---|
| 143 | }
|
---|
| 144 |
|
---|
| 145 | //
|
---|
| 146 | // Kernel evaluation
|
---|
| 147 | //
|
---|
| 148 | // the static method k_function is for doing single kernel evaluation
|
---|
| 149 | // the constructor of Kernel prepares to calculate the l*l kernel matrix
|
---|
| 150 | // the member function get_Q is for getting one column from the Q Matrix
|
---|
| 151 | //
|
---|
| 152 | abstract class QMatrix {
|
---|
| 153 | public abstract float[] get_Q(int column, int len);
|
---|
| 154 | public abstract double[] get_QD();
|
---|
| 155 | public abstract void swap_index(int i, int j);
|
---|
| 156 | };
|
---|
| 157 |
|
---|
| 158 | abstract class Kernel : QMatrix {
|
---|
| 159 | private svm_node[][] x;
|
---|
| 160 | private readonly double[] x_square;
|
---|
| 161 |
|
---|
| 162 | // svm_parameter
|
---|
| 163 | private readonly int kernel_type;
|
---|
| 164 | private readonly int degree;
|
---|
| 165 | private readonly double gamma;
|
---|
| 166 | private readonly double coef0;
|
---|
| 167 |
|
---|
| 168 | public override abstract float[] get_Q(int column, int len);
|
---|
| 169 | public override abstract double[] get_QD();
|
---|
| 170 |
|
---|
| 171 | public override void swap_index(int i, int j) {
|
---|
| 172 | { svm_node[] _ = x[i]; x[i] = x[j]; x[j] = _; }
|
---|
| 173 | if (x_square != null) { double _ = x_square[i]; x_square[i] = x_square[j]; x_square[j] = _; }
|
---|
| 174 | }
|
---|
| 175 |
|
---|
| 176 | private static double powi(double @base, int times) {
|
---|
| 177 | double tmp = @base, ret = 1.0;
|
---|
| 178 |
|
---|
| 179 | for (int t = times; t > 0; t /= 2) {
|
---|
| 180 | if (t % 2 == 1) ret *= tmp;
|
---|
| 181 | tmp = tmp * tmp;
|
---|
| 182 | }
|
---|
| 183 | return ret;
|
---|
| 184 | }
|
---|
| 185 |
|
---|
| 186 | protected virtual double kernel_function(int i, int j) {
|
---|
| 187 | switch (kernel_type) {
|
---|
| 188 | case svm_parameter.LINEAR:
|
---|
| 189 | return dot(x[i], x[j]);
|
---|
| 190 | case svm_parameter.POLY:
|
---|
| 191 | return powi(gamma * dot(x[i], x[j]) + coef0, degree);
|
---|
| 192 | case svm_parameter.RBF:
|
---|
| 193 | return Math.Exp(-gamma * (x_square[i] + x_square[j] - 2 * dot(x[i], x[j])));
|
---|
| 194 | case svm_parameter.SIGMOID:
|
---|
| 195 | return Math.Tanh(gamma * dot(x[i], x[j]) + coef0);
|
---|
| 196 | case svm_parameter.PRECOMPUTED:
|
---|
| 197 | return x[i][(int)(x[j][0].value)].value;
|
---|
| 198 | default:
|
---|
| 199 | return 0; // java
|
---|
| 200 | }
|
---|
| 201 | }
|
---|
| 202 |
|
---|
| 203 | public Kernel(int l, svm_node[][] x_, svm_parameter param) {
|
---|
| 204 | this.kernel_type = param.kernel_type;
|
---|
| 205 | this.degree = param.degree;
|
---|
| 206 | this.gamma = param.gamma;
|
---|
| 207 | this.coef0 = param.coef0;
|
---|
| 208 |
|
---|
| 209 | x = (svm_node[][])x_.Clone();
|
---|
| 210 |
|
---|
| 211 | if (kernel_type == svm_parameter.RBF) {
|
---|
| 212 | x_square = new double[l];
|
---|
| 213 | for (int i = 0; i < l; i++)
|
---|
| 214 | x_square[i] = dot(x[i], x[i]);
|
---|
| 215 | } else x_square = null;
|
---|
| 216 | }
|
---|
| 217 |
|
---|
| 218 | static double dot(svm_node[] x, svm_node[] y) {
|
---|
| 219 | double sum = 0;
|
---|
| 220 | int xlen = x.Length;
|
---|
| 221 | int ylen = y.Length;
|
---|
| 222 | int i = 0;
|
---|
| 223 | int j = 0;
|
---|
| 224 | while (i < xlen && j < ylen) {
|
---|
| 225 | if (x[i].index == y[j].index)
|
---|
| 226 | sum += x[i++].value * y[j++].value;
|
---|
| 227 | else {
|
---|
| 228 | if (x[i].index > y[j].index)
|
---|
| 229 | ++j;
|
---|
| 230 | else
|
---|
| 231 | ++i;
|
---|
| 232 | }
|
---|
| 233 | }
|
---|
| 234 | return sum;
|
---|
| 235 | }
|
---|
| 236 |
|
---|
| 237 | public static double k_function(svm_node[] x, svm_node[] y,
|
---|
| 238 | svm_parameter param) {
|
---|
| 239 | switch (param.kernel_type) {
|
---|
| 240 | case svm_parameter.LINEAR:
|
---|
| 241 | return dot(x, y);
|
---|
| 242 | case svm_parameter.POLY:
|
---|
| 243 | return powi(param.gamma * dot(x, y) + param.coef0, param.degree);
|
---|
| 244 | case svm_parameter.RBF: {
|
---|
| 245 | double sum = 0;
|
---|
| 246 | int xlen = x.Length;
|
---|
| 247 | int ylen = y.Length;
|
---|
| 248 | int i = 0;
|
---|
| 249 | int j = 0;
|
---|
| 250 | while (i < xlen && j < ylen) {
|
---|
| 251 | if (x[i].index == y[j].index) {
|
---|
| 252 | double d = x[i++].value - y[j++].value;
|
---|
| 253 | sum += d * d;
|
---|
| 254 | } else if (x[i].index > y[j].index) {
|
---|
| 255 | sum += y[j].value * y[j].value;
|
---|
| 256 | ++j;
|
---|
| 257 | } else {
|
---|
| 258 | sum += x[i].value * x[i].value;
|
---|
| 259 | ++i;
|
---|
| 260 | }
|
---|
| 261 | }
|
---|
| 262 |
|
---|
| 263 | while (i < xlen) {
|
---|
| 264 | sum += x[i].value * x[i].value;
|
---|
| 265 | ++i;
|
---|
| 266 | }
|
---|
| 267 |
|
---|
| 268 | while (j < ylen) {
|
---|
| 269 | sum += y[j].value * y[j].value;
|
---|
| 270 | ++j;
|
---|
| 271 | }
|
---|
| 272 |
|
---|
| 273 | return Math.Exp(-param.gamma * sum);
|
---|
| 274 | }
|
---|
| 275 | case svm_parameter.SIGMOID:
|
---|
| 276 | return Math.Tanh(param.gamma * dot(x, y) + param.coef0);
|
---|
| 277 | case svm_parameter.PRECOMPUTED:
|
---|
| 278 | return x[(int)(y[0].value)].value;
|
---|
| 279 | default:
|
---|
| 280 | return 0; // java
|
---|
| 281 | }
|
---|
| 282 | }
|
---|
| 283 | }
|
---|
| 284 |
|
---|
| 285 | // An SMO algorithm in Fan et al., JMLR 6(2005), p. 1889--1918
|
---|
| 286 | // Solves:
|
---|
| 287 | //
|
---|
| 288 | // min 0.5(\alpha^T Q \alpha) + p^T \alpha
|
---|
| 289 | //
|
---|
| 290 | // y^T \alpha = \delta
|
---|
| 291 | // y_i = +1 or -1
|
---|
| 292 | // 0 <= alpha_i <= Cp for y_i = 1
|
---|
| 293 | // 0 <= alpha_i <= Cn for y_i = -1
|
---|
| 294 | //
|
---|
| 295 | // Given:
|
---|
| 296 | //
|
---|
| 297 | // Q, p, y, Cp, Cn, and an initial feasible point \alpha
|
---|
| 298 | // l is the size of vectors and matrices
|
---|
| 299 | // eps is the stopping tolerance
|
---|
| 300 | //
|
---|
| 301 | // solution will be put in \alpha, objective value will be put in obj
|
---|
| 302 | //
|
---|
| 303 | class Solver {
|
---|
| 304 | protected int active_size;
|
---|
| 305 | protected short[] y;
|
---|
| 306 | protected double[] G; // gradient of objective function
|
---|
| 307 | protected const byte LOWER_BOUND = 0;
|
---|
| 308 | protected const byte UPPER_BOUND = 1;
|
---|
| 309 | protected const byte FREE = 2;
|
---|
| 310 | protected byte[] alpha_status; // LOWER_BOUND, UPPER_BOUND, FREE
|
---|
| 311 | protected double[] alpha;
|
---|
| 312 | protected QMatrix Q;
|
---|
| 313 | protected double[] QD;
|
---|
| 314 | protected double eps;
|
---|
| 315 | protected double Cp, Cn;
|
---|
| 316 | protected double[] p;
|
---|
| 317 | protected int[] active_set;
|
---|
| 318 | protected double[] G_bar; // gradient, if we treat free variables as 0
|
---|
| 319 | protected int l;
|
---|
| 320 | protected bool unshrink; // XXX
|
---|
| 321 |
|
---|
| 322 | protected const double INF = double.PositiveInfinity;
|
---|
| 323 |
|
---|
| 324 | protected virtual double get_C(int i) {
|
---|
| 325 | return (y[i] > 0) ? Cp : Cn;
|
---|
| 326 | }
|
---|
| 327 | protected virtual void update_alpha_status(int i) {
|
---|
| 328 | if (alpha[i] >= get_C(i))
|
---|
| 329 | alpha_status[i] = UPPER_BOUND;
|
---|
| 330 | else if (alpha[i] <= 0)
|
---|
| 331 | alpha_status[i] = LOWER_BOUND;
|
---|
| 332 | else alpha_status[i] = FREE;
|
---|
| 333 | }
|
---|
| 334 | protected virtual bool is_upper_bound(int i) { return alpha_status[i] == UPPER_BOUND; }
|
---|
| 335 | protected virtual bool is_lower_bound(int i) { return alpha_status[i] == LOWER_BOUND; }
|
---|
| 336 | protected virtual bool is_free(int i) { return alpha_status[i] == FREE; }
|
---|
| 337 |
|
---|
| 338 | // java: information about solution except alpha,
|
---|
| 339 | // because we cannot return multiple values otherwise...
|
---|
| 340 | public class SolutionInfo {
|
---|
| 341 | public double obj;
|
---|
| 342 | public double rho;
|
---|
| 343 | public double upper_bound_p;
|
---|
| 344 | public double upper_bound_n;
|
---|
| 345 | public double r; // for Solver_NU
|
---|
| 346 | }
|
---|
| 347 |
|
---|
| 348 | protected virtual void swap_index(int i, int j) {
|
---|
| 349 | Q.swap_index(i, j); { short _ = y[i]; y[i] = y[j]; y[j] = _; }
|
---|
| 350 | { double _ = G[i]; G[i] = G[j]; G[j] = _; }
|
---|
| 351 | { byte _ = alpha_status[i]; alpha_status[i] = alpha_status[j]; alpha_status[j] = _; }
|
---|
| 352 | { double _ = alpha[i]; alpha[i] = alpha[j]; alpha[j] = _; }
|
---|
| 353 | { double _ = p[i]; p[i] = p[j]; p[j] = _; }
|
---|
| 354 | { int _ = active_set[i]; active_set[i] = active_set[j]; active_set[j] = _; }
|
---|
| 355 | { double _ = G_bar[i]; G_bar[i] = G_bar[j]; G_bar[j] = _; }
|
---|
| 356 | }
|
---|
| 357 |
|
---|
| 358 | protected virtual void reconstruct_gradient() {
|
---|
| 359 | // reconstruct inactive elements of G from G_bar and free variables
|
---|
| 360 |
|
---|
| 361 | if (active_size == l) return;
|
---|
| 362 |
|
---|
| 363 | int i, j;
|
---|
| 364 | int nr_free = 0;
|
---|
| 365 |
|
---|
| 366 | for (j = active_size; j < l; j++)
|
---|
| 367 | G[j] = G_bar[j] + p[j];
|
---|
| 368 |
|
---|
| 369 | for (j = 0; j < active_size; j++)
|
---|
| 370 | if (is_free(j))
|
---|
| 371 | nr_free++;
|
---|
| 372 |
|
---|
| 373 | if (2 * nr_free < active_size)
|
---|
| 374 | svm.info("WARNING: using -h 0 may be faster" + Environment.NewLine + Environment.NewLine);
|
---|
| 375 |
|
---|
| 376 | if (nr_free * l > 2 * active_size * (l - active_size)) {
|
---|
| 377 | for (i = active_size; i < l; i++) {
|
---|
| 378 | float[] Q_i = Q.get_Q(i, active_size);
|
---|
| 379 | for (j = 0; j < active_size; j++)
|
---|
| 380 | if (is_free(j))
|
---|
| 381 | G[i] += alpha[j] * Q_i[j];
|
---|
| 382 | }
|
---|
| 383 | } else {
|
---|
| 384 | for (i = 0; i < active_size; i++)
|
---|
| 385 | if (is_free(i)) {
|
---|
| 386 | float[] Q_i = Q.get_Q(i, l);
|
---|
| 387 | double alpha_i = alpha[i];
|
---|
| 388 | for (j = active_size; j < l; j++)
|
---|
| 389 | G[j] += alpha_i * Q_i[j];
|
---|
| 390 | }
|
---|
| 391 | }
|
---|
| 392 | }
|
---|
| 393 |
|
---|
| 394 | public virtual void Solve(int l, QMatrix Q, double[] p_, short[] y_,
|
---|
| 395 | double[] alpha_, double Cp, double Cn, double eps, SolutionInfo si, int shrinking) {
|
---|
| 396 | this.l = l;
|
---|
| 397 | this.Q = Q;
|
---|
| 398 | QD = Q.get_QD();
|
---|
| 399 | p = (double[])p_.Clone();
|
---|
| 400 | y = (short[])y_.Clone();
|
---|
| 401 | alpha = (double[])alpha_.Clone();
|
---|
| 402 | this.Cp = Cp;
|
---|
| 403 | this.Cn = Cn;
|
---|
| 404 | this.eps = eps;
|
---|
| 405 | this.unshrink = false;
|
---|
| 406 |
|
---|
| 407 | // initialize alpha_status
|
---|
| 408 | {
|
---|
| 409 | alpha_status = new byte[l];
|
---|
| 410 | for (int i = 0; i < l; i++)
|
---|
| 411 | update_alpha_status(i);
|
---|
| 412 | }
|
---|
| 413 |
|
---|
| 414 | // initialize active set (for shrinking)
|
---|
| 415 | {
|
---|
| 416 | active_set = new int[l];
|
---|
| 417 | for (int i = 0; i < l; i++)
|
---|
| 418 | active_set[i] = i;
|
---|
| 419 | active_size = l;
|
---|
| 420 | }
|
---|
| 421 |
|
---|
| 422 | // initialize gradient
|
---|
| 423 | {
|
---|
| 424 | G = new double[l];
|
---|
| 425 | G_bar = new double[l];
|
---|
| 426 | int i;
|
---|
| 427 | for (i = 0; i < l; i++) {
|
---|
| 428 | G[i] = p[i];
|
---|
| 429 | G_bar[i] = 0;
|
---|
| 430 | }
|
---|
| 431 | for (i = 0; i < l; i++)
|
---|
| 432 | if (!is_lower_bound(i)) {
|
---|
| 433 | float[] Q_i = Q.get_Q(i, l);
|
---|
| 434 | double alpha_i = alpha[i];
|
---|
| 435 | int j;
|
---|
| 436 | for (j = 0; j < l; j++)
|
---|
| 437 | G[j] += alpha_i * Q_i[j];
|
---|
| 438 | if (is_upper_bound(i))
|
---|
| 439 | for (j = 0; j < l; j++)
|
---|
| 440 | G_bar[j] += get_C(i) * Q_i[j];
|
---|
| 441 | }
|
---|
| 442 | }
|
---|
| 443 |
|
---|
| 444 | // optimization step
|
---|
| 445 |
|
---|
| 446 | int iter = 0;
|
---|
| 447 | int max_iter = Math.Max(10000000, l > int.MaxValue / 100 ? int.MaxValue : 100 * l);
|
---|
| 448 | int counter = Math.Min(l, 1000) + 1;
|
---|
| 449 | int[] working_set = new int[2];
|
---|
| 450 |
|
---|
| 451 | while (iter < max_iter) {
|
---|
| 452 | // show progress and do shrinking
|
---|
| 453 |
|
---|
| 454 | if (--counter == 0) {
|
---|
| 455 | counter = Math.Min(l, 1000);
|
---|
| 456 | if (shrinking != 0) do_shrinking();
|
---|
| 457 | svm.info(".");
|
---|
| 458 | }
|
---|
| 459 |
|
---|
| 460 | if (select_working_set(working_set) != 0) {
|
---|
| 461 | // reconstruct the whole gradient
|
---|
| 462 | reconstruct_gradient();
|
---|
| 463 | // reset active set size and check
|
---|
| 464 | active_size = l;
|
---|
| 465 | svm.info("*");
|
---|
| 466 | if (select_working_set(working_set) != 0)
|
---|
| 467 | break;
|
---|
| 468 | else
|
---|
| 469 | counter = 1; // do shrinking next iteration
|
---|
| 470 | }
|
---|
| 471 |
|
---|
| 472 | int i = working_set[0];
|
---|
| 473 | int j = working_set[1];
|
---|
| 474 |
|
---|
| 475 | ++iter;
|
---|
| 476 |
|
---|
| 477 | // update alpha[i] and alpha[j], handle bounds carefully
|
---|
| 478 |
|
---|
| 479 | float[] Q_i = Q.get_Q(i, active_size);
|
---|
| 480 | float[] Q_j = Q.get_Q(j, active_size);
|
---|
| 481 |
|
---|
| 482 | double C_i = get_C(i);
|
---|
| 483 | double C_j = get_C(j);
|
---|
| 484 |
|
---|
| 485 | double old_alpha_i = alpha[i];
|
---|
| 486 | double old_alpha_j = alpha[j];
|
---|
| 487 |
|
---|
| 488 | if (y[i] != y[j]) {
|
---|
| 489 | double quad_coef = QD[i] + QD[j] + 2 * Q_i[j];
|
---|
| 490 | if (quad_coef <= 0)
|
---|
| 491 | quad_coef = 1e-12;
|
---|
| 492 | double delta = (-G[i] - G[j]) / quad_coef;
|
---|
| 493 | double diff = alpha[i] - alpha[j];
|
---|
| 494 | alpha[i] += delta;
|
---|
| 495 | alpha[j] += delta;
|
---|
| 496 |
|
---|
| 497 | if (diff > 0) {
|
---|
| 498 | if (alpha[j] < 0) {
|
---|
| 499 | alpha[j] = 0;
|
---|
| 500 | alpha[i] = diff;
|
---|
| 501 | }
|
---|
| 502 | } else {
|
---|
| 503 | if (alpha[i] < 0) {
|
---|
| 504 | alpha[i] = 0;
|
---|
| 505 | alpha[j] = -diff;
|
---|
| 506 | }
|
---|
| 507 | }
|
---|
| 508 | if (diff > C_i - C_j) {
|
---|
| 509 | if (alpha[i] > C_i) {
|
---|
| 510 | alpha[i] = C_i;
|
---|
| 511 | alpha[j] = C_i - diff;
|
---|
| 512 | }
|
---|
| 513 | } else {
|
---|
| 514 | if (alpha[j] > C_j) {
|
---|
| 515 | alpha[j] = C_j;
|
---|
| 516 | alpha[i] = C_j + diff;
|
---|
| 517 | }
|
---|
| 518 | }
|
---|
| 519 | } else {
|
---|
| 520 | double quad_coef = QD[i] + QD[j] - 2 * Q_i[j];
|
---|
| 521 | if (quad_coef <= 0)
|
---|
| 522 | quad_coef = 1e-12;
|
---|
| 523 | double delta = (G[i] - G[j]) / quad_coef;
|
---|
| 524 | double sum = alpha[i] + alpha[j];
|
---|
| 525 | alpha[i] -= delta;
|
---|
| 526 | alpha[j] += delta;
|
---|
| 527 |
|
---|
| 528 | if (sum > C_i) {
|
---|
| 529 | if (alpha[i] > C_i) {
|
---|
| 530 | alpha[i] = C_i;
|
---|
| 531 | alpha[j] = sum - C_i;
|
---|
| 532 | }
|
---|
| 533 | } else {
|
---|
| 534 | if (alpha[j] < 0) {
|
---|
| 535 | alpha[j] = 0;
|
---|
| 536 | alpha[i] = sum;
|
---|
| 537 | }
|
---|
| 538 | }
|
---|
| 539 | if (sum > C_j) {
|
---|
| 540 | if (alpha[j] > C_j) {
|
---|
| 541 | alpha[j] = C_j;
|
---|
| 542 | alpha[i] = sum - C_j;
|
---|
| 543 | }
|
---|
| 544 | } else {
|
---|
| 545 | if (alpha[i] < 0) {
|
---|
| 546 | alpha[i] = 0;
|
---|
| 547 | alpha[j] = sum;
|
---|
| 548 | }
|
---|
| 549 | }
|
---|
| 550 | }
|
---|
| 551 |
|
---|
| 552 | // update G
|
---|
| 553 |
|
---|
| 554 | double delta_alpha_i = alpha[i] - old_alpha_i;
|
---|
| 555 | double delta_alpha_j = alpha[j] - old_alpha_j;
|
---|
| 556 |
|
---|
| 557 | for (int k = 0; k < active_size; k++) {
|
---|
| 558 | G[k] += Q_i[k] * delta_alpha_i + Q_j[k] * delta_alpha_j;
|
---|
| 559 | }
|
---|
| 560 |
|
---|
| 561 | // update alpha_status and G_bar
|
---|
| 562 |
|
---|
| 563 | {
|
---|
| 564 | bool ui = is_upper_bound(i);
|
---|
| 565 | bool uj = is_upper_bound(j);
|
---|
| 566 | update_alpha_status(i);
|
---|
| 567 | update_alpha_status(j);
|
---|
| 568 | int k;
|
---|
| 569 | if (ui != is_upper_bound(i)) {
|
---|
| 570 | Q_i = Q.get_Q(i, l);
|
---|
| 571 | if (ui)
|
---|
| 572 | for (k = 0; k < l; k++)
|
---|
| 573 | G_bar[k] -= C_i * Q_i[k];
|
---|
| 574 | else
|
---|
| 575 | for (k = 0; k < l; k++)
|
---|
| 576 | G_bar[k] += C_i * Q_i[k];
|
---|
| 577 | }
|
---|
| 578 |
|
---|
| 579 | if (uj != is_upper_bound(j)) {
|
---|
| 580 | Q_j = Q.get_Q(j, l);
|
---|
| 581 | if (uj)
|
---|
| 582 | for (k = 0; k < l; k++)
|
---|
| 583 | G_bar[k] -= C_j * Q_j[k];
|
---|
| 584 | else
|
---|
| 585 | for (k = 0; k < l; k++)
|
---|
| 586 | G_bar[k] += C_j * Q_j[k];
|
---|
| 587 | }
|
---|
| 588 | }
|
---|
| 589 |
|
---|
| 590 | }
|
---|
| 591 |
|
---|
| 592 | if (iter >= max_iter) {
|
---|
| 593 | if (active_size < l) {
|
---|
| 594 | // reconstruct the whole gradient to calculate objective value
|
---|
| 595 | reconstruct_gradient();
|
---|
| 596 | active_size = l;
|
---|
| 597 | svm.info("*");
|
---|
| 598 | }
|
---|
| 599 | svm.info("WARNING: reaching max number of iterations" + Environment.NewLine);
|
---|
| 600 | }
|
---|
| 601 |
|
---|
| 602 | // calculate rho
|
---|
| 603 |
|
---|
| 604 | si.rho = calculate_rho();
|
---|
| 605 |
|
---|
| 606 | // calculate objective value
|
---|
| 607 | {
|
---|
| 608 | double v = 0;
|
---|
| 609 | int i;
|
---|
| 610 | for (i = 0; i < l; i++)
|
---|
| 611 | v += alpha[i] * (G[i] + p[i]);
|
---|
| 612 |
|
---|
| 613 | si.obj = v / 2;
|
---|
| 614 | }
|
---|
| 615 |
|
---|
| 616 | // put back the solution
|
---|
| 617 | {
|
---|
| 618 | for (int i = 0; i < l; i++)
|
---|
| 619 | alpha_[active_set[i]] = alpha[i];
|
---|
| 620 | }
|
---|
| 621 |
|
---|
| 622 | si.upper_bound_p = Cp;
|
---|
| 623 | si.upper_bound_n = Cn;
|
---|
| 624 |
|
---|
| 625 | svm.info("optimization finished, #iter = " + iter + Environment.NewLine);
|
---|
| 626 | }
|
---|
| 627 |
|
---|
| 628 | // return 1 if already optimal, return 0 otherwise
|
---|
| 629 | protected virtual int select_working_set(int[] working_set) {
|
---|
| 630 | // return i,j such that
|
---|
| 631 | // i: maximizes -y_i * grad(f)_i, i in I_up(\alpha)
|
---|
| 632 | // j: mimimizes the decrease of obj value
|
---|
| 633 | // (if quadratic coefficeint <= 0, replace it with tau)
|
---|
| 634 | // -y_j*grad(f)_j < -y_i*grad(f)_i, j in I_low(\alpha)
|
---|
| 635 |
|
---|
| 636 | double Gmax = -INF;
|
---|
| 637 | double Gmax2 = -INF;
|
---|
| 638 | int Gmax_idx = -1;
|
---|
| 639 | int Gmin_idx = -1;
|
---|
| 640 | double obj_diff_min = INF;
|
---|
| 641 |
|
---|
| 642 | for (int t = 0; t < active_size; t++)
|
---|
| 643 | if (y[t] == +1) {
|
---|
| 644 | if (!is_upper_bound(t))
|
---|
| 645 | if (-G[t] >= Gmax) {
|
---|
| 646 | Gmax = -G[t];
|
---|
| 647 | Gmax_idx = t;
|
---|
| 648 | }
|
---|
| 649 | } else {
|
---|
| 650 | if (!is_lower_bound(t))
|
---|
| 651 | if (G[t] >= Gmax) {
|
---|
| 652 | Gmax = G[t];
|
---|
| 653 | Gmax_idx = t;
|
---|
| 654 | }
|
---|
| 655 | }
|
---|
| 656 |
|
---|
| 657 | int i = Gmax_idx;
|
---|
| 658 | float[] Q_i = null;
|
---|
| 659 | if (i != -1) // null Q_i not accessed: Gmax=-INF if i=-1
|
---|
| 660 | Q_i = Q.get_Q(i, active_size);
|
---|
| 661 |
|
---|
| 662 | for (int j = 0; j < active_size; j++) {
|
---|
| 663 | if (y[j] == +1) {
|
---|
| 664 | if (!is_lower_bound(j)) {
|
---|
| 665 | double grad_diff = Gmax + G[j];
|
---|
| 666 | if (G[j] >= Gmax2)
|
---|
| 667 | Gmax2 = G[j];
|
---|
| 668 | if (grad_diff > 0) {
|
---|
| 669 | double obj_diff;
|
---|
| 670 | double quad_coef = QD[i] + QD[j] - 2.0 * y[i] * Q_i[j];
|
---|
| 671 | if (quad_coef > 0)
|
---|
| 672 | obj_diff = -(grad_diff * grad_diff) / quad_coef;
|
---|
| 673 | else
|
---|
| 674 | obj_diff = -(grad_diff * grad_diff) / 1e-12;
|
---|
| 675 |
|
---|
| 676 | if (obj_diff <= obj_diff_min) {
|
---|
| 677 | Gmin_idx = j;
|
---|
| 678 | obj_diff_min = obj_diff;
|
---|
| 679 | }
|
---|
| 680 | }
|
---|
| 681 | }
|
---|
| 682 | } else {
|
---|
| 683 | if (!is_upper_bound(j)) {
|
---|
| 684 | double grad_diff = Gmax - G[j];
|
---|
| 685 | if (-G[j] >= Gmax2)
|
---|
| 686 | Gmax2 = -G[j];
|
---|
| 687 | if (grad_diff > 0) {
|
---|
| 688 | double obj_diff;
|
---|
| 689 | double quad_coef = QD[i] + QD[j] + 2.0 * y[i] * Q_i[j];
|
---|
| 690 | if (quad_coef > 0)
|
---|
| 691 | obj_diff = -(grad_diff * grad_diff) / quad_coef;
|
---|
| 692 | else
|
---|
| 693 | obj_diff = -(grad_diff * grad_diff) / 1e-12;
|
---|
| 694 |
|
---|
| 695 | if (obj_diff <= obj_diff_min) {
|
---|
| 696 | Gmin_idx = j;
|
---|
| 697 | obj_diff_min = obj_diff;
|
---|
| 698 | }
|
---|
| 699 | }
|
---|
| 700 | }
|
---|
| 701 | }
|
---|
| 702 | }
|
---|
| 703 |
|
---|
| 704 | if (Gmax + Gmax2 < eps)
|
---|
| 705 | return 1;
|
---|
| 706 |
|
---|
| 707 | working_set[0] = Gmax_idx;
|
---|
| 708 | working_set[1] = Gmin_idx;
|
---|
| 709 | return 0;
|
---|
| 710 | }
|
---|
| 711 |
|
---|
| 712 | private bool be_shrunk(int i, double Gmax1, double Gmax2) {
|
---|
| 713 | if (is_upper_bound(i)) {
|
---|
| 714 | if (y[i] == +1)
|
---|
| 715 | return (-G[i] > Gmax1);
|
---|
| 716 | else
|
---|
| 717 | return (-G[i] > Gmax2);
|
---|
| 718 | } else if (is_lower_bound(i)) {
|
---|
| 719 | if (y[i] == +1)
|
---|
| 720 | return (G[i] > Gmax2);
|
---|
| 721 | else
|
---|
| 722 | return (G[i] > Gmax1);
|
---|
| 723 | } else
|
---|
| 724 | return (false);
|
---|
| 725 | }
|
---|
| 726 |
|
---|
| 727 | protected virtual void do_shrinking() {
|
---|
| 728 | int i;
|
---|
| 729 | double Gmax1 = -INF; // max { -y_i * grad(f)_i | i in I_up(\alpha) }
|
---|
| 730 | double Gmax2 = -INF; // max { y_i * grad(f)_i | i in I_low(\alpha) }
|
---|
| 731 |
|
---|
| 732 | // find maximal violating pair first
|
---|
| 733 | for (i = 0; i < active_size; i++) {
|
---|
| 734 | if (y[i] == +1) {
|
---|
| 735 | if (!is_upper_bound(i)) {
|
---|
| 736 | if (-G[i] >= Gmax1)
|
---|
| 737 | Gmax1 = -G[i];
|
---|
| 738 | }
|
---|
| 739 | if (!is_lower_bound(i)) {
|
---|
| 740 | if (G[i] >= Gmax2)
|
---|
| 741 | Gmax2 = G[i];
|
---|
| 742 | }
|
---|
| 743 | } else {
|
---|
| 744 | if (!is_upper_bound(i)) {
|
---|
| 745 | if (-G[i] >= Gmax2)
|
---|
| 746 | Gmax2 = -G[i];
|
---|
| 747 | }
|
---|
| 748 | if (!is_lower_bound(i)) {
|
---|
| 749 | if (G[i] >= Gmax1)
|
---|
| 750 | Gmax1 = G[i];
|
---|
| 751 | }
|
---|
| 752 | }
|
---|
| 753 | }
|
---|
| 754 |
|
---|
| 755 | if (unshrink == false && Gmax1 + Gmax2 <= eps * 10) {
|
---|
| 756 | unshrink = true;
|
---|
| 757 | reconstruct_gradient();
|
---|
| 758 | active_size = l;
|
---|
| 759 | }
|
---|
| 760 |
|
---|
| 761 | for (i = 0; i < active_size; i++)
|
---|
| 762 | if (be_shrunk(i, Gmax1, Gmax2)) {
|
---|
| 763 | active_size--;
|
---|
| 764 | while (active_size > i) {
|
---|
| 765 | if (!be_shrunk(active_size, Gmax1, Gmax2)) {
|
---|
| 766 | swap_index(i, active_size);
|
---|
| 767 | break;
|
---|
| 768 | }
|
---|
| 769 | active_size--;
|
---|
| 770 | }
|
---|
| 771 | }
|
---|
| 772 | }
|
---|
| 773 |
|
---|
| 774 | protected virtual double calculate_rho() {
|
---|
| 775 | double r;
|
---|
| 776 | int nr_free = 0;
|
---|
| 777 | double ub = INF, lb = -INF, sum_free = 0;
|
---|
| 778 | for (int i = 0; i < active_size; i++) {
|
---|
| 779 | double yG = y[i] * G[i];
|
---|
| 780 |
|
---|
| 781 | if (is_lower_bound(i)) {
|
---|
| 782 | if (y[i] > 0)
|
---|
| 783 | ub = Math.Min(ub, yG);
|
---|
| 784 | else
|
---|
| 785 | lb = Math.Max(lb, yG);
|
---|
| 786 | } else if (is_upper_bound(i)) {
|
---|
| 787 | if (y[i] < 0)
|
---|
| 788 | ub = Math.Min(ub, yG);
|
---|
| 789 | else
|
---|
| 790 | lb = Math.Max(lb, yG);
|
---|
| 791 | } else {
|
---|
| 792 | ++nr_free;
|
---|
| 793 | sum_free += yG;
|
---|
| 794 | }
|
---|
| 795 | }
|
---|
| 796 |
|
---|
| 797 | if (nr_free > 0)
|
---|
| 798 | r = sum_free / nr_free;
|
---|
| 799 | else
|
---|
| 800 | r = (ub + lb) / 2;
|
---|
| 801 |
|
---|
| 802 | return r;
|
---|
| 803 | }
|
---|
| 804 |
|
---|
| 805 | }
|
---|
| 806 |
|
---|
| 807 | //
|
---|
| 808 | // Solver for nu-svm classification and regression
|
---|
| 809 | //
|
---|
| 810 | // additional constraint: e^T \alpha = constant
|
---|
| 811 | //
|
---|
| 812 | internal sealed class Solver_NU : Solver {
|
---|
| 813 | private SolutionInfo si;
|
---|
| 814 |
|
---|
| 815 | public override void Solve(int l, QMatrix Q, double[] p, short[] y,
|
---|
| 816 | double[] alpha, double Cp, double Cn, double eps,
|
---|
| 817 | SolutionInfo si, int shrinking) {
|
---|
| 818 | this.si = si;
|
---|
| 819 | base.Solve(l, Q, p, y, alpha, Cp, Cn, eps, si, shrinking);
|
---|
| 820 | }
|
---|
| 821 |
|
---|
| 822 | // return 1 if already optimal, return 0 otherwise
|
---|
| 823 | protected override int select_working_set(int[] working_set) {
|
---|
| 824 | // return i,j such that y_i = y_j and
|
---|
| 825 | // i: maximizes -y_i * grad(f)_i, i in I_up(\alpha)
|
---|
| 826 | // j: minimizes the decrease of obj value
|
---|
| 827 | // (if quadratic coefficeint <= 0, replace it with tau)
|
---|
| 828 | // -y_j*grad(f)_j < -y_i*grad(f)_i, j in I_low(\alpha)
|
---|
| 829 |
|
---|
| 830 | double Gmaxp = -INF;
|
---|
| 831 | double Gmaxp2 = -INF;
|
---|
| 832 | int Gmaxp_idx = -1;
|
---|
| 833 |
|
---|
| 834 | double Gmaxn = -INF;
|
---|
| 835 | double Gmaxn2 = -INF;
|
---|
| 836 | int Gmaxn_idx = -1;
|
---|
| 837 |
|
---|
| 838 | int Gmin_idx = -1;
|
---|
| 839 | double obj_diff_min = INF;
|
---|
| 840 |
|
---|
| 841 | for (int t = 0; t < active_size; t++)
|
---|
| 842 | if (y[t] == +1) {
|
---|
| 843 | if (!is_upper_bound(t))
|
---|
| 844 | if (-G[t] >= Gmaxp) {
|
---|
| 845 | Gmaxp = -G[t];
|
---|
| 846 | Gmaxp_idx = t;
|
---|
| 847 | }
|
---|
| 848 | } else {
|
---|
| 849 | if (!is_lower_bound(t))
|
---|
| 850 | if (G[t] >= Gmaxn) {
|
---|
| 851 | Gmaxn = G[t];
|
---|
| 852 | Gmaxn_idx = t;
|
---|
| 853 | }
|
---|
| 854 | }
|
---|
| 855 |
|
---|
| 856 | int ip = Gmaxp_idx;
|
---|
| 857 | int @in = Gmaxn_idx;
|
---|
| 858 | float[] Q_ip = null;
|
---|
| 859 | float[] Q_in = null;
|
---|
| 860 | if (ip != -1) // null Q_ip not accessed: Gmaxp=-INF if ip=-1
|
---|
| 861 | Q_ip = Q.get_Q(ip, active_size);
|
---|
| 862 | if (@in != -1)
|
---|
| 863 | Q_in = Q.get_Q(@in, active_size);
|
---|
| 864 |
|
---|
| 865 | for (int j = 0; j < active_size; j++) {
|
---|
| 866 | if (y[j] == +1) {
|
---|
| 867 | if (!is_lower_bound(j)) {
|
---|
| 868 | double grad_diff = Gmaxp + G[j];
|
---|
| 869 | if (G[j] >= Gmaxp2)
|
---|
| 870 | Gmaxp2 = G[j];
|
---|
| 871 | if (grad_diff > 0) {
|
---|
| 872 | double obj_diff;
|
---|
| 873 | double quad_coef = QD[ip] + QD[j] - 2 * Q_ip[j];
|
---|
| 874 | if (quad_coef > 0)
|
---|
| 875 | obj_diff = -(grad_diff * grad_diff) / quad_coef;
|
---|
| 876 | else
|
---|
| 877 | obj_diff = -(grad_diff * grad_diff) / 1e-12;
|
---|
| 878 |
|
---|
| 879 | if (obj_diff <= obj_diff_min) {
|
---|
| 880 | Gmin_idx = j;
|
---|
| 881 | obj_diff_min = obj_diff;
|
---|
| 882 | }
|
---|
| 883 | }
|
---|
| 884 | }
|
---|
| 885 | } else {
|
---|
| 886 | if (!is_upper_bound(j)) {
|
---|
| 887 | double grad_diff = Gmaxn - G[j];
|
---|
| 888 | if (-G[j] >= Gmaxn2)
|
---|
| 889 | Gmaxn2 = -G[j];
|
---|
| 890 | if (grad_diff > 0) {
|
---|
| 891 | double obj_diff;
|
---|
| 892 | double quad_coef = QD[@in] + QD[j] - 2 * Q_in[j];
|
---|
| 893 | if (quad_coef > 0)
|
---|
| 894 | obj_diff = -(grad_diff * grad_diff) / quad_coef;
|
---|
| 895 | else
|
---|
| 896 | obj_diff = -(grad_diff * grad_diff) / 1e-12;
|
---|
| 897 |
|
---|
| 898 | if (obj_diff <= obj_diff_min) {
|
---|
| 899 | Gmin_idx = j;
|
---|
| 900 | obj_diff_min = obj_diff;
|
---|
| 901 | }
|
---|
| 902 | }
|
---|
| 903 | }
|
---|
| 904 | }
|
---|
| 905 | }
|
---|
| 906 |
|
---|
| 907 | if (Math.Max(Gmaxp + Gmaxp2, Gmaxn + Gmaxn2) < eps)
|
---|
| 908 | return 1;
|
---|
| 909 |
|
---|
| 910 | if (y[Gmin_idx] == +1)
|
---|
| 911 | working_set[0] = Gmaxp_idx;
|
---|
| 912 | else
|
---|
| 913 | working_set[0] = Gmaxn_idx;
|
---|
| 914 | working_set[1] = Gmin_idx;
|
---|
| 915 |
|
---|
| 916 | return 0;
|
---|
| 917 | }
|
---|
| 918 |
|
---|
| 919 | private bool be_shrunk(int i, double Gmax1, double Gmax2, double Gmax3, double Gmax4) {
|
---|
| 920 | if (is_upper_bound(i)) {
|
---|
| 921 | if (y[i] == +1)
|
---|
| 922 | return (-G[i] > Gmax1);
|
---|
| 923 | else
|
---|
| 924 | return (-G[i] > Gmax4);
|
---|
| 925 | } else if (is_lower_bound(i)) {
|
---|
| 926 | if (y[i] == +1)
|
---|
| 927 | return (G[i] > Gmax2);
|
---|
| 928 | else
|
---|
| 929 | return (G[i] > Gmax3);
|
---|
| 930 | } else
|
---|
| 931 | return (false);
|
---|
| 932 | }
|
---|
| 933 |
|
---|
| 934 | protected override void do_shrinking() {
|
---|
| 935 | double Gmax1 = -INF; // max { -y_i * grad(f)_i | y_i = +1, i in I_up(\alpha) }
|
---|
| 936 | double Gmax2 = -INF; // max { y_i * grad(f)_i | y_i = +1, i in I_low(\alpha) }
|
---|
| 937 | double Gmax3 = -INF; // max { -y_i * grad(f)_i | y_i = -1, i in I_up(\alpha) }
|
---|
| 938 | double Gmax4 = -INF; // max { y_i * grad(f)_i | y_i = -1, i in I_low(\alpha) }
|
---|
| 939 |
|
---|
| 940 | // find maximal violating pair first
|
---|
| 941 | int i;
|
---|
| 942 | for (i = 0; i < active_size; i++) {
|
---|
| 943 | if (!is_upper_bound(i)) {
|
---|
| 944 | if (y[i] == +1) {
|
---|
| 945 | if (-G[i] > Gmax1) Gmax1 = -G[i];
|
---|
| 946 | } else if (-G[i] > Gmax4) Gmax4 = -G[i];
|
---|
| 947 | }
|
---|
| 948 | if (!is_lower_bound(i)) {
|
---|
| 949 | if (y[i] == +1) {
|
---|
| 950 | if (G[i] > Gmax2) Gmax2 = G[i];
|
---|
| 951 | } else if (G[i] > Gmax3) Gmax3 = G[i];
|
---|
| 952 | }
|
---|
| 953 | }
|
---|
| 954 |
|
---|
| 955 | if (unshrink == false && Math.Max(Gmax1 + Gmax2, Gmax3 + Gmax4) <= eps * 10) {
|
---|
| 956 | unshrink = true;
|
---|
| 957 | reconstruct_gradient();
|
---|
| 958 | active_size = l;
|
---|
| 959 | }
|
---|
| 960 |
|
---|
| 961 | for (i = 0; i < active_size; i++)
|
---|
| 962 | if (be_shrunk(i, Gmax1, Gmax2, Gmax3, Gmax4)) {
|
---|
| 963 | active_size--;
|
---|
| 964 | while (active_size > i) {
|
---|
| 965 | if (!be_shrunk(active_size, Gmax1, Gmax2, Gmax3, Gmax4)) {
|
---|
| 966 | swap_index(i, active_size);
|
---|
| 967 | break;
|
---|
| 968 | }
|
---|
| 969 | active_size--;
|
---|
| 970 | }
|
---|
| 971 | }
|
---|
| 972 | }
|
---|
| 973 |
|
---|
| 974 | protected override double calculate_rho() {
|
---|
| 975 | int nr_free1 = 0, nr_free2 = 0;
|
---|
| 976 | double ub1 = INF, ub2 = INF;
|
---|
| 977 | double lb1 = -INF, lb2 = -INF;
|
---|
| 978 | double sum_free1 = 0, sum_free2 = 0;
|
---|
| 979 |
|
---|
| 980 | for (int i = 0; i < active_size; i++) {
|
---|
| 981 | if (y[i] == +1) {
|
---|
| 982 | if (is_lower_bound(i))
|
---|
| 983 | ub1 = Math.Min(ub1, G[i]);
|
---|
| 984 | else if (is_upper_bound(i))
|
---|
| 985 | lb1 = Math.Max(lb1, G[i]);
|
---|
| 986 | else {
|
---|
| 987 | ++nr_free1;
|
---|
| 988 | sum_free1 += G[i];
|
---|
| 989 | }
|
---|
| 990 | } else {
|
---|
| 991 | if (is_lower_bound(i))
|
---|
| 992 | ub2 = Math.Min(ub2, G[i]);
|
---|
| 993 | else if (is_upper_bound(i))
|
---|
| 994 | lb2 = Math.Max(lb2, G[i]);
|
---|
| 995 | else {
|
---|
| 996 | ++nr_free2;
|
---|
| 997 | sum_free2 += G[i];
|
---|
| 998 | }
|
---|
| 999 | }
|
---|
| 1000 | }
|
---|
| 1001 |
|
---|
| 1002 | double r1, r2;
|
---|
| 1003 | if (nr_free1 > 0)
|
---|
| 1004 | r1 = sum_free1 / nr_free1;
|
---|
| 1005 | else
|
---|
| 1006 | r1 = (ub1 + lb1) / 2;
|
---|
| 1007 |
|
---|
| 1008 | if (nr_free2 > 0)
|
---|
| 1009 | r2 = sum_free2 / nr_free2;
|
---|
| 1010 | else
|
---|
| 1011 | r2 = (ub2 + lb2) / 2;
|
---|
| 1012 |
|
---|
| 1013 | si.r = (r1 + r2) / 2;
|
---|
| 1014 | return (r1 - r2) / 2;
|
---|
| 1015 | }
|
---|
| 1016 | }
|
---|
| 1017 |
|
---|
| 1018 | //
|
---|
| 1019 | // Q matrices for various formulations
|
---|
| 1020 | //
|
---|
| 1021 | class SVC_Q : Kernel {
|
---|
| 1022 | private readonly short[] y;
|
---|
| 1023 | private readonly Cache cache;
|
---|
| 1024 | private readonly double[] QD;
|
---|
| 1025 |
|
---|
| 1026 | public SVC_Q(svm_problem prob, svm_parameter param, short[] y_)
|
---|
| 1027 | : base(prob.l, prob.x, param) {
|
---|
| 1028 | y = (short[])y_.Clone();
|
---|
| 1029 | cache = new Cache(prob.l, (long)(param.cache_size * (1 << 20)));
|
---|
| 1030 | QD = new double[prob.l];
|
---|
| 1031 | for (int i = 0; i < prob.l; i++)
|
---|
| 1032 | QD[i] = kernel_function(i, i);
|
---|
| 1033 | }
|
---|
| 1034 |
|
---|
| 1035 | public override float[] get_Q(int i, int len) {
|
---|
| 1036 | float[][] data = new float[1][];
|
---|
| 1037 | int start, j;
|
---|
| 1038 | if ((start = cache.get_data(i, data, len)) < len) {
|
---|
| 1039 | for (j = start; j < len; j++)
|
---|
| 1040 | data[0][j] = (float)(y[i] * y[j] * kernel_function(i, j));
|
---|
| 1041 | }
|
---|
| 1042 | return data[0];
|
---|
| 1043 | }
|
---|
| 1044 |
|
---|
| 1045 | public override double[] get_QD() {
|
---|
| 1046 | return QD;
|
---|
| 1047 | }
|
---|
| 1048 |
|
---|
| 1049 | public override void swap_index(int i, int j) {
|
---|
| 1050 | cache.swap_index(i, j);
|
---|
| 1051 | base.swap_index(i, j); { short _ = y[i]; y[i] = y[j]; y[j] = _; }
|
---|
| 1052 | { double _ = QD[i]; QD[i] = QD[j]; QD[j] = _; }
|
---|
| 1053 | }
|
---|
| 1054 | }
|
---|
| 1055 |
|
---|
| 1056 | class ONE_CLASS_Q : Kernel {
|
---|
| 1057 | private readonly Cache cache;
|
---|
| 1058 | private readonly double[] QD;
|
---|
| 1059 |
|
---|
| 1060 | public ONE_CLASS_Q(svm_problem prob, svm_parameter param)
|
---|
| 1061 | : base(prob.l, prob.x, param) {
|
---|
| 1062 | cache = new Cache(prob.l, (long)(param.cache_size * (1 << 20)));
|
---|
| 1063 | QD = new double[prob.l];
|
---|
| 1064 | for (int i = 0; i < prob.l; i++)
|
---|
| 1065 | QD[i] = kernel_function(i, i);
|
---|
| 1066 | }
|
---|
| 1067 |
|
---|
| 1068 | public override float[] get_Q(int i, int len) {
|
---|
| 1069 | float[][] data = new float[1][];
|
---|
| 1070 | int start, j;
|
---|
| 1071 | if ((start = cache.get_data(i, data, len)) < len) {
|
---|
| 1072 | for (j = start; j < len; j++)
|
---|
| 1073 | data[0][j] = (float)kernel_function(i, j);
|
---|
| 1074 | }
|
---|
| 1075 | return data[0];
|
---|
| 1076 | }
|
---|
| 1077 |
|
---|
| 1078 | public override double[] get_QD() {
|
---|
| 1079 | return QD;
|
---|
| 1080 | }
|
---|
| 1081 |
|
---|
| 1082 | public override void swap_index(int i, int j) {
|
---|
| 1083 | cache.swap_index(i, j);
|
---|
| 1084 | base.swap_index(i, j); { double _ = QD[i]; QD[i] = QD[j]; QD[j] = _; }
|
---|
| 1085 | }
|
---|
| 1086 | }
|
---|
| 1087 |
|
---|
| 1088 | class SVR_Q : Kernel {
|
---|
| 1089 | private int l;
|
---|
| 1090 | private Cache cache;
|
---|
| 1091 | private short[] sign;
|
---|
| 1092 | private int[] index;
|
---|
| 1093 | private int next_buffer;
|
---|
| 1094 | private float[][] buffer;
|
---|
| 1095 | private readonly double[] QD;
|
---|
| 1096 |
|
---|
| 1097 | public SVR_Q(svm_problem prob, svm_parameter param)
|
---|
| 1098 | : base(prob.l, prob.x, param) {
|
---|
| 1099 | l = prob.l;
|
---|
| 1100 | cache = new Cache(l, (long)(param.cache_size * (1 << 20)));
|
---|
| 1101 | QD = new double[2 * l];
|
---|
| 1102 | sign = new short[2 * l];
|
---|
| 1103 | index = new int[2 * l];
|
---|
| 1104 | for (int k = 0; k < l; k++) {
|
---|
| 1105 | sign[k] = 1;
|
---|
| 1106 | sign[k + l] = -1;
|
---|
| 1107 | index[k] = k;
|
---|
| 1108 | index[k + l] = k;
|
---|
| 1109 | QD[k] = kernel_function(k, k);
|
---|
| 1110 | QD[k + l] = QD[k];
|
---|
| 1111 | }
|
---|
| 1112 | buffer = new float[2][];
|
---|
| 1113 | buffer[0] = new float[2 * l];
|
---|
| 1114 | buffer[1] = new float[2 * l];
|
---|
| 1115 | next_buffer = 0;
|
---|
| 1116 | }
|
---|
| 1117 |
|
---|
| 1118 | public override void swap_index(int i, int j) {
|
---|
| 1119 | { short _ = sign[i]; sign[i] = sign[j]; sign[j] = _; }
|
---|
| 1120 | { int _ = index[i]; index[i] = index[j]; index[j] = _; }
|
---|
| 1121 | { double _ = QD[i]; QD[i] = QD[j]; QD[j] = _; }
|
---|
| 1122 | }
|
---|
| 1123 |
|
---|
| 1124 | public override float[] get_Q(int i, int len) {
|
---|
| 1125 | float[][] data = new float[1][];
|
---|
| 1126 | int j, real_i = index[i];
|
---|
| 1127 | if (cache.get_data(real_i, data, l) < l) {
|
---|
| 1128 | for (j = 0; j < l; j++)
|
---|
| 1129 | data[0][j] = (float)kernel_function(real_i, j);
|
---|
| 1130 | }
|
---|
| 1131 |
|
---|
| 1132 | // reorder and copy
|
---|
| 1133 | float[] buf = buffer[next_buffer];
|
---|
| 1134 | next_buffer = 1 - next_buffer;
|
---|
| 1135 | short si = sign[i];
|
---|
| 1136 | for (j = 0; j < len; j++)
|
---|
| 1137 | buf[j] = (float)si * sign[j] * data[0][index[j]];
|
---|
| 1138 | return buf;
|
---|
| 1139 | }
|
---|
| 1140 |
|
---|
| 1141 | public override double[] get_QD() {
|
---|
| 1142 | return QD;
|
---|
| 1143 | }
|
---|
| 1144 | }
|
---|
| 1145 |
|
---|
| 1146 | public class svm {
|
---|
| 1147 | //
|
---|
| 1148 | // construct and solve various formulations
|
---|
| 1149 | //
|
---|
| 1150 | public static readonly int LIBSVM_VERSION = 312;
|
---|
| 1151 | public static readonly Random rand = new Random();
|
---|
| 1152 |
|
---|
| 1153 | private static Action<string> svm_print_string = (s) => {
|
---|
| 1154 | Console.Out.Write(s);
|
---|
| 1155 | Console.Out.Flush();
|
---|
| 1156 | };
|
---|
| 1157 |
|
---|
| 1158 | public static void info(String s) {
|
---|
| 1159 | svm_print_string(s);
|
---|
| 1160 | }
|
---|
| 1161 |
|
---|
| 1162 | private static void solve_c_svc(svm_problem prob, svm_parameter param,
|
---|
| 1163 | double[] alpha, Solver.SolutionInfo si,
|
---|
| 1164 | double Cp, double Cn) {
|
---|
| 1165 | int l = prob.l;
|
---|
| 1166 | double[] minus_ones = new double[l];
|
---|
| 1167 | short[] y = new short[l];
|
---|
| 1168 |
|
---|
| 1169 | int i;
|
---|
| 1170 |
|
---|
| 1171 | for (i = 0; i < l; i++) {
|
---|
| 1172 | alpha[i] = 0;
|
---|
| 1173 | minus_ones[i] = -1;
|
---|
| 1174 | if (prob.y[i] > 0) y[i] = +1;
|
---|
| 1175 | else y[i] = -1;
|
---|
| 1176 | }
|
---|
| 1177 |
|
---|
| 1178 | Solver s = new Solver();
|
---|
| 1179 | s.Solve(l, new SVC_Q(prob, param, y), minus_ones, y,
|
---|
| 1180 | alpha, Cp, Cn, param.eps, si, param.shrinking);
|
---|
| 1181 |
|
---|
| 1182 | double sum_alpha = 0;
|
---|
| 1183 | for (i = 0; i < l; i++)
|
---|
| 1184 | sum_alpha += alpha[i];
|
---|
| 1185 |
|
---|
| 1186 | if (Cp == Cn)
|
---|
| 1187 | svm.info("nu = " + sum_alpha / (Cp * prob.l) + Environment.NewLine);
|
---|
| 1188 |
|
---|
| 1189 | for (i = 0; i < l; i++)
|
---|
| 1190 | alpha[i] *= y[i];
|
---|
| 1191 | }
|
---|
| 1192 |
|
---|
| 1193 | private static void solve_nu_svc(svm_problem prob, svm_parameter param,
|
---|
| 1194 | double[] alpha, Solver.SolutionInfo si) {
|
---|
| 1195 | int i;
|
---|
| 1196 | int l = prob.l;
|
---|
| 1197 | double nu = param.nu;
|
---|
| 1198 |
|
---|
| 1199 | short[] y = new short[l];
|
---|
| 1200 |
|
---|
| 1201 | for (i = 0; i < l; i++)
|
---|
| 1202 | if (prob.y[i] > 0)
|
---|
| 1203 | y[i] = +1;
|
---|
| 1204 | else
|
---|
| 1205 | y[i] = -1;
|
---|
| 1206 |
|
---|
| 1207 | double sum_pos = nu * l / 2;
|
---|
| 1208 | double sum_neg = nu * l / 2;
|
---|
| 1209 |
|
---|
| 1210 | for (i = 0; i < l; i++)
|
---|
| 1211 | if (y[i] == +1) {
|
---|
| 1212 | alpha[i] = Math.Min(1.0, sum_pos);
|
---|
| 1213 | sum_pos -= alpha[i];
|
---|
| 1214 | } else {
|
---|
| 1215 | alpha[i] = Math.Min(1.0, sum_neg);
|
---|
| 1216 | sum_neg -= alpha[i];
|
---|
| 1217 | }
|
---|
| 1218 |
|
---|
| 1219 | double[] zeros = new double[l];
|
---|
| 1220 |
|
---|
| 1221 | for (i = 0; i < l; i++)
|
---|
| 1222 | zeros[i] = 0;
|
---|
| 1223 |
|
---|
| 1224 | Solver_NU s = new Solver_NU();
|
---|
| 1225 | s.Solve(l, new SVC_Q(prob, param, y), zeros, y,
|
---|
| 1226 | alpha, 1.0, 1.0, param.eps, si, param.shrinking);
|
---|
| 1227 | double r = si.r;
|
---|
| 1228 |
|
---|
| 1229 | svm.info("C = " + 1 / r + Environment.NewLine);
|
---|
| 1230 |
|
---|
| 1231 | for (i = 0; i < l; i++)
|
---|
| 1232 | alpha[i] *= y[i] / r;
|
---|
| 1233 |
|
---|
| 1234 | si.rho /= r;
|
---|
| 1235 | si.obj /= (r * r);
|
---|
| 1236 | si.upper_bound_p = 1 / r;
|
---|
| 1237 | si.upper_bound_n = 1 / r;
|
---|
| 1238 | }
|
---|
| 1239 |
|
---|
| 1240 | private static void solve_one_class(svm_problem prob, svm_parameter param,
|
---|
| 1241 | double[] alpha, Solver.SolutionInfo si) {
|
---|
| 1242 | int l = prob.l;
|
---|
| 1243 | double[] zeros = new double[l];
|
---|
| 1244 | short[] ones = new short[l];
|
---|
| 1245 | int i;
|
---|
| 1246 |
|
---|
| 1247 | int n = (int)(param.nu * prob.l); // # of alpha's at upper bound
|
---|
| 1248 |
|
---|
| 1249 | for (i = 0; i < n; i++)
|
---|
| 1250 | alpha[i] = 1;
|
---|
| 1251 | if (n < prob.l)
|
---|
| 1252 | alpha[n] = param.nu * prob.l - n;
|
---|
| 1253 | for (i = n + 1; i < l; i++)
|
---|
| 1254 | alpha[i] = 0;
|
---|
| 1255 |
|
---|
| 1256 | for (i = 0; i < l; i++) {
|
---|
| 1257 | zeros[i] = 0;
|
---|
| 1258 | ones[i] = 1;
|
---|
| 1259 | }
|
---|
| 1260 |
|
---|
| 1261 | Solver s = new Solver();
|
---|
| 1262 | s.Solve(l, new ONE_CLASS_Q(prob, param), zeros, ones,
|
---|
| 1263 | alpha, 1.0, 1.0, param.eps, si, param.shrinking);
|
---|
| 1264 | }
|
---|
| 1265 |
|
---|
| 1266 | private static void solve_epsilon_svr(svm_problem prob, svm_parameter param,
|
---|
| 1267 | double[] alpha, Solver.SolutionInfo si) {
|
---|
| 1268 | int l = prob.l;
|
---|
| 1269 | double[] alpha2 = new double[2 * l];
|
---|
| 1270 | double[] linear_term = new double[2 * l];
|
---|
| 1271 | short[] y = new short[2 * l];
|
---|
| 1272 | int i;
|
---|
| 1273 |
|
---|
| 1274 | for (i = 0; i < l; i++) {
|
---|
| 1275 | alpha2[i] = 0;
|
---|
| 1276 | linear_term[i] = param.p - prob.y[i];
|
---|
| 1277 | y[i] = 1;
|
---|
| 1278 |
|
---|
| 1279 | alpha2[i + l] = 0;
|
---|
| 1280 | linear_term[i + l] = param.p + prob.y[i];
|
---|
| 1281 | y[i + l] = -1;
|
---|
| 1282 | }
|
---|
| 1283 |
|
---|
| 1284 | Solver s = new Solver();
|
---|
| 1285 | s.Solve(2 * l, new SVR_Q(prob, param), linear_term, y,
|
---|
| 1286 | alpha2, param.C, param.C, param.eps, si, param.shrinking);
|
---|
| 1287 |
|
---|
| 1288 | double sum_alpha = 0;
|
---|
| 1289 | for (i = 0; i < l; i++) {
|
---|
| 1290 | alpha[i] = alpha2[i] - alpha2[i + l];
|
---|
| 1291 | sum_alpha += Math.Abs(alpha[i]);
|
---|
| 1292 | }
|
---|
| 1293 | svm.info("nu = " + sum_alpha / (param.C * l) + Environment.NewLine);
|
---|
| 1294 | }
|
---|
| 1295 |
|
---|
| 1296 | private static void solve_nu_svr(svm_problem prob, svm_parameter param,
|
---|
| 1297 | double[] alpha, Solver.SolutionInfo si) {
|
---|
| 1298 | int l = prob.l;
|
---|
| 1299 | double C = param.C;
|
---|
| 1300 | double[] alpha2 = new double[2 * l];
|
---|
| 1301 | double[] linear_term = new double[2 * l];
|
---|
| 1302 | short[] y = new short[2 * l];
|
---|
| 1303 | int i;
|
---|
| 1304 |
|
---|
| 1305 | double sum = C * param.nu * l / 2;
|
---|
| 1306 | for (i = 0; i < l; i++) {
|
---|
| 1307 | alpha2[i] = alpha2[i + l] = Math.Min(sum, C);
|
---|
| 1308 | sum -= alpha2[i];
|
---|
| 1309 |
|
---|
| 1310 | linear_term[i] = -prob.y[i];
|
---|
| 1311 | y[i] = 1;
|
---|
| 1312 |
|
---|
| 1313 | linear_term[i + l] = prob.y[i];
|
---|
| 1314 | y[i + l] = -1;
|
---|
| 1315 | }
|
---|
| 1316 |
|
---|
| 1317 | Solver_NU s = new Solver_NU();
|
---|
| 1318 | s.Solve(2 * l, new SVR_Q(prob, param), linear_term, y,
|
---|
| 1319 | alpha2, C, C, param.eps, si, param.shrinking);
|
---|
| 1320 |
|
---|
| 1321 | svm.info("epsilon = " + (-si.r) + Environment.NewLine);
|
---|
| 1322 |
|
---|
| 1323 | for (i = 0; i < l; i++)
|
---|
| 1324 | alpha[i] = alpha2[i] - alpha2[i + l];
|
---|
| 1325 | }
|
---|
| 1326 |
|
---|
| 1327 | //
|
---|
| 1328 | // decision_function
|
---|
| 1329 | //
|
---|
| 1330 | private sealed class decision_function {
|
---|
| 1331 | public double[] alpha;
|
---|
| 1332 | public double rho;
|
---|
| 1333 | };
|
---|
| 1334 |
|
---|
| 1335 | private static decision_function svm_train_one(
|
---|
| 1336 | svm_problem prob, svm_parameter param,
|
---|
| 1337 | double Cp, double Cn) {
|
---|
| 1338 | double[] alpha = new double[prob.l];
|
---|
| 1339 | Solver.SolutionInfo si = new Solver.SolutionInfo();
|
---|
| 1340 | switch (param.svm_type) {
|
---|
| 1341 | case svm_parameter.C_SVC:
|
---|
| 1342 | solve_c_svc(prob, param, alpha, si, Cp, Cn);
|
---|
| 1343 | break;
|
---|
| 1344 | case svm_parameter.NU_SVC:
|
---|
| 1345 | solve_nu_svc(prob, param, alpha, si);
|
---|
| 1346 | break;
|
---|
| 1347 | case svm_parameter.ONE_CLASS:
|
---|
| 1348 | solve_one_class(prob, param, alpha, si);
|
---|
| 1349 | break;
|
---|
| 1350 | case svm_parameter.EPSILON_SVR:
|
---|
| 1351 | solve_epsilon_svr(prob, param, alpha, si);
|
---|
| 1352 | break;
|
---|
| 1353 | case svm_parameter.NU_SVR:
|
---|
| 1354 | solve_nu_svr(prob, param, alpha, si);
|
---|
| 1355 | break;
|
---|
| 1356 | }
|
---|
| 1357 |
|
---|
| 1358 | svm.info("obj = " + si.obj + ", rho = " + si.rho + Environment.NewLine);
|
---|
| 1359 |
|
---|
| 1360 | // output SVs
|
---|
| 1361 |
|
---|
| 1362 | int nSV = 0;
|
---|
| 1363 | int nBSV = 0;
|
---|
| 1364 | for (int i = 0; i < prob.l; i++) {
|
---|
| 1365 | if (Math.Abs(alpha[i]) > 0) {
|
---|
| 1366 | ++nSV;
|
---|
| 1367 | if (prob.y[i] > 0) {
|
---|
| 1368 | if (Math.Abs(alpha[i]) >= si.upper_bound_p)
|
---|
| 1369 | ++nBSV;
|
---|
| 1370 | } else {
|
---|
| 1371 | if (Math.Abs(alpha[i]) >= si.upper_bound_n)
|
---|
| 1372 | ++nBSV;
|
---|
| 1373 | }
|
---|
| 1374 | }
|
---|
| 1375 | }
|
---|
| 1376 |
|
---|
| 1377 | svm.info("nSV = " + nSV + ", nBSV = " + nBSV + Environment.NewLine);
|
---|
| 1378 |
|
---|
| 1379 | decision_function f = new decision_function();
|
---|
| 1380 | f.alpha = alpha;
|
---|
| 1381 | f.rho = si.rho;
|
---|
| 1382 | return f;
|
---|
| 1383 | }
|
---|
| 1384 |
|
---|
| 1385 | // Platt's binary SVM Probablistic Output: an improvement from Lin et al.
|
---|
| 1386 | private static void sigmoid_train(int l, double[] dec_values, double[] labels,
|
---|
| 1387 | double[] probAB) {
|
---|
| 1388 | double A, B;
|
---|
| 1389 | double prior1 = 0, prior0 = 0;
|
---|
| 1390 | int i;
|
---|
| 1391 |
|
---|
| 1392 | for (i = 0; i < l; i++)
|
---|
| 1393 | if (labels[i] > 0) prior1 += 1;
|
---|
| 1394 | else prior0 += 1;
|
---|
| 1395 |
|
---|
| 1396 | int max_iter = 100; // Maximal number of iterations
|
---|
| 1397 | double min_step = 1e-10; // Minimal step taken in line search
|
---|
| 1398 | double sigma = 1e-12; // For numerically strict PD of Hessian
|
---|
| 1399 | double eps = 1e-5;
|
---|
| 1400 | double hiTarget = (prior1 + 1.0) / (prior1 + 2.0);
|
---|
| 1401 | double loTarget = 1 / (prior0 + 2.0);
|
---|
| 1402 | double[] t = new double[l];
|
---|
| 1403 | double fApB, p, q, h11, h22, h21, g1, g2, det, dA, dB, gd, stepsize;
|
---|
| 1404 | double newA, newB, newf, d1, d2;
|
---|
| 1405 | int iter;
|
---|
| 1406 |
|
---|
| 1407 | // Initial Point and Initial Fun Value
|
---|
| 1408 | A = 0.0;
|
---|
| 1409 | B = Math.Log((prior0 + 1.0) / (prior1 + 1.0));
|
---|
| 1410 | double fval = 0.0;
|
---|
| 1411 |
|
---|
| 1412 | for (i = 0; i < l; i++) {
|
---|
| 1413 | if (labels[i] > 0) t[i] = hiTarget;
|
---|
| 1414 | else t[i] = loTarget;
|
---|
| 1415 | fApB = dec_values[i] * A + B;
|
---|
| 1416 | if (fApB >= 0)
|
---|
| 1417 | fval += t[i] * fApB + Math.Log(1 + Math.Exp(-fApB));
|
---|
| 1418 | else
|
---|
| 1419 | fval += (t[i] - 1) * fApB + Math.Log(1 + Math.Exp(fApB));
|
---|
| 1420 | }
|
---|
| 1421 | for (iter = 0; iter < max_iter; iter++) {
|
---|
| 1422 | // Update Gradient and Hessian (use H' = H + sigma I)
|
---|
| 1423 | h11 = sigma; // numerically ensures strict PD
|
---|
| 1424 | h22 = sigma;
|
---|
| 1425 | h21 = 0.0;
|
---|
| 1426 | g1 = 0.0;
|
---|
| 1427 | g2 = 0.0;
|
---|
| 1428 | for (i = 0; i < l; i++) {
|
---|
| 1429 | fApB = dec_values[i] * A + B;
|
---|
| 1430 | if (fApB >= 0) {
|
---|
| 1431 | p = Math.Exp(-fApB) / (1.0 + Math.Exp(-fApB));
|
---|
| 1432 | q = 1.0 / (1.0 + Math.Exp(-fApB));
|
---|
| 1433 | } else {
|
---|
| 1434 | p = 1.0 / (1.0 + Math.Exp(fApB));
|
---|
| 1435 | q = Math.Exp(fApB) / (1.0 + Math.Exp(fApB));
|
---|
| 1436 | }
|
---|
| 1437 | d2 = p * q;
|
---|
| 1438 | h11 += dec_values[i] * dec_values[i] * d2;
|
---|
| 1439 | h22 += d2;
|
---|
| 1440 | h21 += dec_values[i] * d2;
|
---|
| 1441 | d1 = t[i] - p;
|
---|
| 1442 | g1 += dec_values[i] * d1;
|
---|
| 1443 | g2 += d1;
|
---|
| 1444 | }
|
---|
| 1445 |
|
---|
| 1446 | // Stopping Criteria
|
---|
| 1447 | if (Math.Abs(g1) < eps && Math.Abs(g2) < eps)
|
---|
| 1448 | break;
|
---|
| 1449 |
|
---|
| 1450 | // Finding Newton direction: -inv(H') * g
|
---|
| 1451 | det = h11 * h22 - h21 * h21;
|
---|
| 1452 | dA = -(h22 * g1 - h21 * g2) / det;
|
---|
| 1453 | dB = -(-h21 * g1 + h11 * g2) / det;
|
---|
| 1454 | gd = g1 * dA + g2 * dB;
|
---|
| 1455 |
|
---|
| 1456 |
|
---|
| 1457 | stepsize = 1; // Line Search
|
---|
| 1458 | while (stepsize >= min_step) {
|
---|
| 1459 | newA = A + stepsize * dA;
|
---|
| 1460 | newB = B + stepsize * dB;
|
---|
| 1461 |
|
---|
| 1462 | // New function value
|
---|
| 1463 | newf = 0.0;
|
---|
| 1464 | for (i = 0; i < l; i++) {
|
---|
| 1465 | fApB = dec_values[i] * newA + newB;
|
---|
| 1466 | if (fApB >= 0)
|
---|
| 1467 | newf += t[i] * fApB + Math.Log(1 + Math.Exp(-fApB));
|
---|
| 1468 | else
|
---|
| 1469 | newf += (t[i] - 1) * fApB + Math.Log(1 + Math.Exp(fApB));
|
---|
| 1470 | }
|
---|
| 1471 | // Check sufficient decrease
|
---|
| 1472 | if (newf < fval + 0.0001 * stepsize * gd) {
|
---|
| 1473 | A = newA;
|
---|
| 1474 | B = newB;
|
---|
| 1475 | fval = newf;
|
---|
| 1476 | break;
|
---|
| 1477 | } else
|
---|
| 1478 | stepsize = stepsize / 2.0;
|
---|
| 1479 | }
|
---|
| 1480 |
|
---|
| 1481 | if (stepsize < min_step) {
|
---|
| 1482 | svm.info("Line search fails in two-class probability estimates" + Environment.NewLine);
|
---|
| 1483 | break;
|
---|
| 1484 | }
|
---|
| 1485 | }
|
---|
| 1486 |
|
---|
| 1487 | if (iter >= max_iter)
|
---|
| 1488 | svm.info("Reaching maximal iterations in two-class probability estimates" + Environment.NewLine);
|
---|
| 1489 | probAB[0] = A;
|
---|
| 1490 | probAB[1] = B;
|
---|
| 1491 | }
|
---|
| 1492 |
|
---|
| 1493 | private static double sigmoid_predict(double decision_value, double A, double B) {
|
---|
| 1494 | double fApB = decision_value * A + B;
|
---|
| 1495 | if (fApB >= 0)
|
---|
| 1496 | return Math.Exp(-fApB) / (1.0 + Math.Exp(-fApB));
|
---|
| 1497 | else
|
---|
| 1498 | return 1.0 / (1 + Math.Exp(fApB));
|
---|
| 1499 | }
|
---|
| 1500 |
|
---|
| 1501 | // Method 2 from the multiclass_prob paper by Wu, Lin, and Weng
|
---|
| 1502 | private static void multiclass_probability(int k, double[][] r, double[] p) {
|
---|
| 1503 | int t, j;
|
---|
| 1504 | int iter = 0, max_iter = Math.Max(100, k);
|
---|
| 1505 | double[][] Q = new double[k][];
|
---|
| 1506 | double[] Qp = new double[k];
|
---|
| 1507 | double pQp, eps = 0.005 / k;
|
---|
| 1508 |
|
---|
| 1509 | for (t = 0; t < k; t++) {
|
---|
| 1510 | Q[t] = new double[k];
|
---|
| 1511 | p[t] = 1.0 / k; // Valid if k = 1
|
---|
| 1512 | Q[t][t] = 0;
|
---|
| 1513 | for (j = 0; j < t; j++) {
|
---|
| 1514 | Q[t][t] += r[j][t] * r[j][t];
|
---|
| 1515 | Q[t][j] = Q[j][t];
|
---|
| 1516 | }
|
---|
| 1517 | for (j = t + 1; j < k; j++) {
|
---|
| 1518 | Q[t][t] += r[j][t] * r[j][t];
|
---|
| 1519 | Q[t][j] = -r[j][t] * r[t][j];
|
---|
| 1520 | }
|
---|
| 1521 | }
|
---|
| 1522 | for (iter = 0; iter < max_iter; iter++) {
|
---|
| 1523 | // stopping condition, recalculate QP,pQP for numerical accuracy
|
---|
| 1524 | pQp = 0;
|
---|
| 1525 | for (t = 0; t < k; t++) {
|
---|
| 1526 | Qp[t] = 0;
|
---|
| 1527 | for (j = 0; j < k; j++)
|
---|
| 1528 | Qp[t] += Q[t][j] * p[j];
|
---|
| 1529 | pQp += p[t] * Qp[t];
|
---|
| 1530 | }
|
---|
| 1531 | double max_error = 0;
|
---|
| 1532 | for (t = 0; t < k; t++) {
|
---|
| 1533 | double error = Math.Abs(Qp[t] - pQp);
|
---|
| 1534 | if (error > max_error)
|
---|
| 1535 | max_error = error;
|
---|
| 1536 | }
|
---|
| 1537 | if (max_error < eps) break;
|
---|
| 1538 |
|
---|
| 1539 | for (t = 0; t < k; t++) {
|
---|
| 1540 | double diff = (-Qp[t] + pQp) / Q[t][t];
|
---|
| 1541 | p[t] += diff;
|
---|
| 1542 | pQp = (pQp + diff * (diff * Q[t][t] + 2 * Qp[t])) / (1 + diff) / (1 + diff);
|
---|
| 1543 | for (j = 0; j < k; j++) {
|
---|
| 1544 | Qp[j] = (Qp[j] + diff * Q[t][j]) / (1 + diff);
|
---|
| 1545 | p[j] /= (1 + diff);
|
---|
| 1546 | }
|
---|
| 1547 | }
|
---|
| 1548 | }
|
---|
| 1549 | if (iter >= max_iter)
|
---|
| 1550 | svm.info("Exceeds max_iter in multiclass_prob" + Environment.NewLine);
|
---|
| 1551 | }
|
---|
| 1552 |
|
---|
| 1553 | // Cross-validation decision values for probability estimates
|
---|
| 1554 | private static void svm_binary_svc_probability(svm_problem prob, svm_parameter param, double Cp, double Cn,
|
---|
| 1555 | double[] probAB) {
|
---|
| 1556 | int i;
|
---|
| 1557 | int nr_fold = 5;
|
---|
| 1558 | int[] perm = new int[prob.l];
|
---|
| 1559 | double[] dec_values = new double[prob.l];
|
---|
| 1560 |
|
---|
| 1561 | // random shuffle
|
---|
| 1562 | for (i = 0; i < prob.l; i++) perm[i] = i;
|
---|
| 1563 | for (i = 0; i < prob.l; i++) {
|
---|
| 1564 | int j = i + rand.Next(prob.l - i);
|
---|
| 1565 | {
|
---|
| 1566 | int _ = perm[i];
|
---|
| 1567 | perm[i] = perm[j];
|
---|
| 1568 | perm[j] = _;
|
---|
| 1569 | }
|
---|
| 1570 | }
|
---|
| 1571 | for (i = 0; i < nr_fold; i++) {
|
---|
| 1572 | int begin = i * prob.l / nr_fold;
|
---|
| 1573 | int end = (i + 1) * prob.l / nr_fold;
|
---|
| 1574 | int j, k;
|
---|
| 1575 | svm_problem subprob = new svm_problem();
|
---|
| 1576 |
|
---|
| 1577 | subprob.l = prob.l - (end - begin);
|
---|
| 1578 | subprob.x = new svm_node[subprob.l][];
|
---|
| 1579 | subprob.y = new double[subprob.l];
|
---|
| 1580 |
|
---|
| 1581 | k = 0;
|
---|
| 1582 | for (j = 0; j < begin; j++) {
|
---|
| 1583 | subprob.x[k] = prob.x[perm[j]];
|
---|
| 1584 | subprob.y[k] = prob.y[perm[j]];
|
---|
| 1585 | ++k;
|
---|
| 1586 | }
|
---|
| 1587 | for (j = end; j < prob.l; j++) {
|
---|
| 1588 | subprob.x[k] = prob.x[perm[j]];
|
---|
| 1589 | subprob.y[k] = prob.y[perm[j]];
|
---|
| 1590 | ++k;
|
---|
| 1591 | }
|
---|
| 1592 | int p_count = 0, n_count = 0;
|
---|
| 1593 | for (j = 0; j < k; j++)
|
---|
| 1594 | if (subprob.y[j] > 0)
|
---|
| 1595 | p_count++;
|
---|
| 1596 | else
|
---|
| 1597 | n_count++;
|
---|
| 1598 |
|
---|
| 1599 | if (p_count == 0 && n_count == 0)
|
---|
| 1600 | for (j = begin; j < end; j++)
|
---|
| 1601 | dec_values[perm[j]] = 0;
|
---|
| 1602 | else if (p_count > 0 && n_count == 0)
|
---|
| 1603 | for (j = begin; j < end; j++)
|
---|
| 1604 | dec_values[perm[j]] = 1;
|
---|
| 1605 | else if (p_count == 0 && n_count > 0)
|
---|
| 1606 | for (j = begin; j < end; j++)
|
---|
| 1607 | dec_values[perm[j]] = -1;
|
---|
| 1608 | else {
|
---|
| 1609 | svm_parameter subparam = (svm_parameter)param.Clone();
|
---|
| 1610 | subparam.probability = 0;
|
---|
| 1611 | subparam.C = 1.0;
|
---|
| 1612 | subparam.nr_weight = 2;
|
---|
| 1613 | subparam.weight_label = new int[2];
|
---|
| 1614 | subparam.weight = new double[2];
|
---|
| 1615 | subparam.weight_label[0] = +1;
|
---|
| 1616 | subparam.weight_label[1] = -1;
|
---|
| 1617 | subparam.weight[0] = Cp;
|
---|
| 1618 | subparam.weight[1] = Cn;
|
---|
| 1619 | svm_model submodel = svm_train(subprob, subparam);
|
---|
| 1620 | for (j = begin; j < end; j++) {
|
---|
| 1621 | double[] dec_value = new double[1];
|
---|
| 1622 | svm_predict_values(submodel, prob.x[perm[j]], dec_value);
|
---|
| 1623 | dec_values[perm[j]] = dec_value[0];
|
---|
| 1624 | // ensure +1 -1 order; reason not using CV subroutine
|
---|
| 1625 | dec_values[perm[j]] *= submodel.label[0];
|
---|
| 1626 | }
|
---|
| 1627 | }
|
---|
| 1628 | }
|
---|
| 1629 | sigmoid_train(prob.l, dec_values, prob.y, probAB);
|
---|
| 1630 | }
|
---|
| 1631 |
|
---|
| 1632 | // Return parameter of a Laplace distribution
|
---|
| 1633 | private static double svm_svr_probability(svm_problem prob, svm_parameter param) {
|
---|
| 1634 | int i;
|
---|
| 1635 | int nr_fold = 5;
|
---|
| 1636 | double[] ymv = new double[prob.l];
|
---|
| 1637 | double mae = 0;
|
---|
| 1638 |
|
---|
| 1639 | svm_parameter newparam = (svm_parameter)param.Clone();
|
---|
| 1640 | newparam.probability = 0;
|
---|
| 1641 | svm_cross_validation(prob, newparam, nr_fold, ymv);
|
---|
| 1642 | for (i = 0; i < prob.l; i++) {
|
---|
| 1643 | ymv[i] = prob.y[i] - ymv[i];
|
---|
| 1644 | mae += Math.Abs(ymv[i]);
|
---|
| 1645 | }
|
---|
| 1646 | mae /= prob.l;
|
---|
| 1647 | double std = Math.Sqrt(2 * mae * mae);
|
---|
| 1648 | int count = 0;
|
---|
| 1649 | mae = 0;
|
---|
| 1650 | for (i = 0; i < prob.l; i++)
|
---|
| 1651 | if (Math.Abs(ymv[i]) > 5 * std)
|
---|
| 1652 | count = count + 1;
|
---|
| 1653 | else
|
---|
| 1654 | mae += Math.Abs(ymv[i]);
|
---|
| 1655 | mae /= (prob.l - count);
|
---|
| 1656 | svm.info("Prob. model for test data: target value = predicted value + z, " + Environment.NewLine
|
---|
| 1657 | + "z: Laplace distribution e^(-|z|/sigma)/(2sigma),sigma=" + mae + Environment.NewLine);
|
---|
| 1658 | return mae;
|
---|
| 1659 | }
|
---|
| 1660 |
|
---|
| 1661 | // label: label name, start: begin of each class, count: #data of classes, perm: indices to the original data
|
---|
| 1662 | // perm, length l, must be allocated before calling this subroutine
|
---|
| 1663 | private static void svm_group_classes(svm_problem prob, int[] nr_class_ret, int[][] label_ret, int[][] start_ret,
|
---|
| 1664 | int[][] count_ret, int[] perm) {
|
---|
| 1665 | int l = prob.l;
|
---|
| 1666 | int max_nr_class = 16;
|
---|
| 1667 | int nr_class = 0;
|
---|
| 1668 | int[] label = new int[max_nr_class];
|
---|
| 1669 | int[] count = new int[max_nr_class];
|
---|
| 1670 | int[] data_label = new int[l];
|
---|
| 1671 | int i;
|
---|
| 1672 |
|
---|
| 1673 | for (i = 0; i < l; i++) {
|
---|
| 1674 | int this_label = (int)(prob.y[i]);
|
---|
| 1675 | int j;
|
---|
| 1676 | for (j = 0; j < nr_class; j++) {
|
---|
| 1677 | if (this_label == label[j]) {
|
---|
| 1678 | ++count[j];
|
---|
| 1679 | break;
|
---|
| 1680 | }
|
---|
| 1681 | }
|
---|
| 1682 | data_label[i] = j;
|
---|
| 1683 | if (j == nr_class) {
|
---|
| 1684 | if (nr_class == max_nr_class) {
|
---|
| 1685 | max_nr_class *= 2;
|
---|
| 1686 | int[] new_data = new int[max_nr_class];
|
---|
| 1687 | Array.Copy(label, 0, new_data, 0, label.Length);
|
---|
| 1688 | label = new_data;
|
---|
| 1689 | new_data = new int[max_nr_class];
|
---|
| 1690 | Array.Copy(count, 0, new_data, 0, count.Length);
|
---|
| 1691 | count = new_data;
|
---|
| 1692 | }
|
---|
| 1693 | label[nr_class] = this_label;
|
---|
| 1694 | count[nr_class] = 1;
|
---|
| 1695 | ++nr_class;
|
---|
| 1696 | }
|
---|
| 1697 | }
|
---|
| 1698 |
|
---|
| 1699 | int[] start = new int[nr_class];
|
---|
| 1700 | start[0] = 0;
|
---|
| 1701 | for (i = 1; i < nr_class; i++)
|
---|
| 1702 | start[i] = start[i - 1] + count[i - 1];
|
---|
| 1703 | for (i = 0; i < l; i++) {
|
---|
| 1704 | perm[start[data_label[i]]] = i;
|
---|
| 1705 | ++start[data_label[i]];
|
---|
| 1706 | }
|
---|
| 1707 | start[0] = 0;
|
---|
| 1708 | for (i = 1; i < nr_class; i++)
|
---|
| 1709 | start[i] = start[i - 1] + count[i - 1];
|
---|
| 1710 |
|
---|
| 1711 | nr_class_ret[0] = nr_class;
|
---|
| 1712 | label_ret[0] = label;
|
---|
| 1713 | start_ret[0] = start;
|
---|
| 1714 | count_ret[0] = count;
|
---|
| 1715 | }
|
---|
| 1716 |
|
---|
| 1717 | //
|
---|
| 1718 | // Interface functions
|
---|
| 1719 | //
|
---|
| 1720 | public static svm_model svm_train(svm_problem prob, svm_parameter param) {
|
---|
| 1721 | svm_model model = new svm_model();
|
---|
| 1722 | model.param = param;
|
---|
| 1723 |
|
---|
| 1724 | if (param.svm_type == svm_parameter.ONE_CLASS ||
|
---|
| 1725 | param.svm_type == svm_parameter.EPSILON_SVR ||
|
---|
| 1726 | param.svm_type == svm_parameter.NU_SVR) {
|
---|
| 1727 | // regression or one-class-svm
|
---|
| 1728 | model.nr_class = 2;
|
---|
| 1729 | model.label = null;
|
---|
| 1730 | model.nSV = null;
|
---|
| 1731 | model.probA = null;
|
---|
| 1732 | model.probB = null;
|
---|
| 1733 | model.sv_coef = new double[1][];
|
---|
| 1734 |
|
---|
| 1735 | if (param.probability == 1 &&
|
---|
| 1736 | (param.svm_type == svm_parameter.EPSILON_SVR ||
|
---|
| 1737 | param.svm_type == svm_parameter.NU_SVR)) {
|
---|
| 1738 | model.probA = new double[1];
|
---|
| 1739 | model.probA[0] = svm_svr_probability(prob, param);
|
---|
| 1740 | }
|
---|
| 1741 |
|
---|
| 1742 | decision_function f = svm_train_one(prob, param, 0, 0);
|
---|
| 1743 | model.rho = new double[1];
|
---|
| 1744 | model.rho[0] = f.rho;
|
---|
| 1745 |
|
---|
| 1746 | int nSV = 0;
|
---|
| 1747 | int i;
|
---|
| 1748 | for (i = 0; i < prob.l; i++)
|
---|
| 1749 | if (Math.Abs(f.alpha[i]) > 0) ++nSV;
|
---|
| 1750 | model.l = nSV;
|
---|
| 1751 | model.SV = new svm_node[nSV][];
|
---|
| 1752 | model.sv_coef[0] = new double[nSV];
|
---|
| 1753 | int j = 0;
|
---|
| 1754 | for (i = 0; i < prob.l; i++)
|
---|
| 1755 | if (Math.Abs(f.alpha[i]) > 0) {
|
---|
| 1756 | model.SV[j] = prob.x[i];
|
---|
| 1757 | model.sv_coef[0][j] = f.alpha[i];
|
---|
| 1758 | ++j;
|
---|
| 1759 | }
|
---|
| 1760 | } else {
|
---|
| 1761 | // classification
|
---|
| 1762 | int l = prob.l;
|
---|
| 1763 | int[] tmp_nr_class = new int[1];
|
---|
| 1764 | int[][] tmp_label = new int[1][];
|
---|
| 1765 | int[][] tmp_start = new int[1][];
|
---|
| 1766 | int[][] tmp_count = new int[1][];
|
---|
| 1767 | int[] perm = new int[l];
|
---|
| 1768 |
|
---|
| 1769 | // group training data of the same class
|
---|
| 1770 | svm_group_classes(prob, tmp_nr_class, tmp_label, tmp_start, tmp_count, perm);
|
---|
| 1771 | int nr_class = tmp_nr_class[0];
|
---|
| 1772 | int[] label = tmp_label[0];
|
---|
| 1773 | int[] start = tmp_start[0];
|
---|
| 1774 | int[] count = tmp_count[0];
|
---|
| 1775 |
|
---|
| 1776 | if (nr_class == 1)
|
---|
| 1777 | svm.info("WARNING: training data in only one class. See README for details." + Environment.NewLine);
|
---|
| 1778 |
|
---|
| 1779 | svm_node[][] x = new svm_node[l][];
|
---|
| 1780 | int i;
|
---|
| 1781 | for (i = 0; i < l; i++)
|
---|
| 1782 | x[i] = prob.x[perm[i]];
|
---|
| 1783 |
|
---|
| 1784 | // calculate weighted C
|
---|
| 1785 |
|
---|
| 1786 | double[] weighted_C = new double[nr_class];
|
---|
| 1787 | for (i = 0; i < nr_class; i++)
|
---|
| 1788 | weighted_C[i] = param.C;
|
---|
| 1789 | for (i = 0; i < param.nr_weight; i++) {
|
---|
| 1790 | int j;
|
---|
| 1791 | for (j = 0; j < nr_class; j++)
|
---|
| 1792 | if (param.weight_label[i] == label[j])
|
---|
| 1793 | break;
|
---|
| 1794 | if (j == nr_class)
|
---|
| 1795 | Console.Error.WriteLine("WARNING: class label " + param.weight_label[i] +
|
---|
| 1796 | " specified in weight is not found");
|
---|
| 1797 | else
|
---|
| 1798 | weighted_C[j] *= param.weight[i];
|
---|
| 1799 | }
|
---|
| 1800 |
|
---|
| 1801 | // train k*(k-1)/2 models
|
---|
| 1802 |
|
---|
| 1803 | bool[] nonzero = new bool[l];
|
---|
| 1804 | for (i = 0; i < l; i++)
|
---|
| 1805 | nonzero[i] = false;
|
---|
| 1806 | decision_function[] f = new decision_function[nr_class * (nr_class - 1) / 2];
|
---|
| 1807 |
|
---|
| 1808 | double[] probA = null, probB = null;
|
---|
| 1809 | if (param.probability == 1) {
|
---|
| 1810 | probA = new double[nr_class * (nr_class - 1) / 2];
|
---|
| 1811 | probB = new double[nr_class * (nr_class - 1) / 2];
|
---|
| 1812 | }
|
---|
| 1813 |
|
---|
| 1814 | int p = 0;
|
---|
| 1815 | for (i = 0; i < nr_class; i++)
|
---|
| 1816 | for (int j = i + 1; j < nr_class; j++) {
|
---|
| 1817 | svm_problem sub_prob = new svm_problem();
|
---|
| 1818 | int si = start[i], sj = start[j];
|
---|
| 1819 | int ci = count[i], cj = count[j];
|
---|
| 1820 | sub_prob.l = ci + cj;
|
---|
| 1821 | sub_prob.x = new svm_node[sub_prob.l][];
|
---|
| 1822 | sub_prob.y = new double[sub_prob.l];
|
---|
| 1823 | int k;
|
---|
| 1824 | for (k = 0; k < ci; k++) {
|
---|
| 1825 | sub_prob.x[k] = x[si + k];
|
---|
| 1826 | sub_prob.y[k] = +1;
|
---|
| 1827 | }
|
---|
| 1828 | for (k = 0; k < cj; k++) {
|
---|
| 1829 | sub_prob.x[ci + k] = x[sj + k];
|
---|
| 1830 | sub_prob.y[ci + k] = -1;
|
---|
| 1831 | }
|
---|
| 1832 |
|
---|
| 1833 | if (param.probability == 1) {
|
---|
| 1834 | double[] probAB = new double[2];
|
---|
| 1835 | svm_binary_svc_probability(sub_prob, param, weighted_C[i], weighted_C[j], probAB);
|
---|
| 1836 | probA[p] = probAB[0];
|
---|
| 1837 | probB[p] = probAB[1];
|
---|
| 1838 | }
|
---|
| 1839 |
|
---|
| 1840 | f[p] = svm_train_one(sub_prob, param, weighted_C[i], weighted_C[j]);
|
---|
| 1841 | for (k = 0; k < ci; k++)
|
---|
| 1842 | if (!nonzero[si + k] && Math.Abs(f[p].alpha[k]) > 0)
|
---|
| 1843 | nonzero[si + k] = true;
|
---|
| 1844 | for (k = 0; k < cj; k++)
|
---|
| 1845 | if (!nonzero[sj + k] && Math.Abs(f[p].alpha[ci + k]) > 0)
|
---|
| 1846 | nonzero[sj + k] = true;
|
---|
| 1847 | ++p;
|
---|
| 1848 | }
|
---|
| 1849 |
|
---|
| 1850 | // build output
|
---|
| 1851 |
|
---|
| 1852 | model.nr_class = nr_class;
|
---|
| 1853 |
|
---|
| 1854 | model.label = new int[nr_class];
|
---|
| 1855 | for (i = 0; i < nr_class; i++)
|
---|
| 1856 | model.label[i] = label[i];
|
---|
| 1857 |
|
---|
| 1858 | model.rho = new double[nr_class * (nr_class - 1) / 2];
|
---|
| 1859 | for (i = 0; i < nr_class * (nr_class - 1) / 2; i++)
|
---|
| 1860 | model.rho[i] = f[i].rho;
|
---|
| 1861 |
|
---|
| 1862 | if (param.probability == 1) {
|
---|
| 1863 | model.probA = new double[nr_class * (nr_class - 1) / 2];
|
---|
| 1864 | model.probB = new double[nr_class * (nr_class - 1) / 2];
|
---|
| 1865 | for (i = 0; i < nr_class * (nr_class - 1) / 2; i++) {
|
---|
| 1866 | model.probA[i] = probA[i];
|
---|
| 1867 | model.probB[i] = probB[i];
|
---|
| 1868 | }
|
---|
| 1869 | } else {
|
---|
| 1870 | model.probA = null;
|
---|
| 1871 | model.probB = null;
|
---|
| 1872 | }
|
---|
| 1873 |
|
---|
| 1874 | int nnz = 0;
|
---|
| 1875 | int[] nz_count = new int[nr_class];
|
---|
| 1876 | model.nSV = new int[nr_class];
|
---|
| 1877 | for (i = 0; i < nr_class; i++) {
|
---|
| 1878 | int nSV = 0;
|
---|
| 1879 | for (int j = 0; j < count[i]; j++)
|
---|
| 1880 | if (nonzero[start[i] + j]) {
|
---|
| 1881 | ++nSV;
|
---|
| 1882 | ++nnz;
|
---|
| 1883 | }
|
---|
| 1884 | model.nSV[i] = nSV;
|
---|
| 1885 | nz_count[i] = nSV;
|
---|
| 1886 | }
|
---|
| 1887 |
|
---|
| 1888 | svm.info("Total nSV = " + nnz + Environment.NewLine);
|
---|
| 1889 |
|
---|
| 1890 | model.l = nnz;
|
---|
| 1891 | model.SV = new svm_node[nnz][];
|
---|
| 1892 | p = 0;
|
---|
| 1893 | for (i = 0; i < l; i++)
|
---|
| 1894 | if (nonzero[i]) model.SV[p++] = x[i];
|
---|
| 1895 |
|
---|
| 1896 | int[] nz_start = new int[nr_class];
|
---|
| 1897 | nz_start[0] = 0;
|
---|
| 1898 | for (i = 1; i < nr_class; i++)
|
---|
| 1899 | nz_start[i] = nz_start[i - 1] + nz_count[i - 1];
|
---|
| 1900 |
|
---|
| 1901 | model.sv_coef = new double[nr_class - 1][];
|
---|
| 1902 | for (i = 0; i < nr_class - 1; i++)
|
---|
| 1903 | model.sv_coef[i] = new double[nnz];
|
---|
| 1904 |
|
---|
| 1905 | p = 0;
|
---|
| 1906 | for (i = 0; i < nr_class; i++)
|
---|
| 1907 | for (int j = i + 1; j < nr_class; j++) {
|
---|
| 1908 | // classifier (i,j): coefficients with
|
---|
| 1909 | // i are in sv_coef[j-1][nz_start[i]...],
|
---|
| 1910 | // j are in sv_coef[i][nz_start[j]...]
|
---|
| 1911 |
|
---|
| 1912 | int si = start[i];
|
---|
| 1913 | int sj = start[j];
|
---|
| 1914 | int ci = count[i];
|
---|
| 1915 | int cj = count[j];
|
---|
| 1916 |
|
---|
| 1917 | int q = nz_start[i];
|
---|
| 1918 | int k;
|
---|
| 1919 | for (k = 0; k < ci; k++)
|
---|
| 1920 | if (nonzero[si + k])
|
---|
| 1921 | model.sv_coef[j - 1][q++] = f[p].alpha[k];
|
---|
| 1922 | q = nz_start[j];
|
---|
| 1923 | for (k = 0; k < cj; k++)
|
---|
| 1924 | if (nonzero[sj + k])
|
---|
| 1925 | model.sv_coef[i][q++] = f[p].alpha[ci + k];
|
---|
| 1926 | ++p;
|
---|
| 1927 | }
|
---|
| 1928 | }
|
---|
| 1929 | return model;
|
---|
| 1930 | }
|
---|
| 1931 |
|
---|
| 1932 | // Stratified cross validation
|
---|
| 1933 | public static void svm_cross_validation(svm_problem prob, svm_parameter param, int nr_fold, double[] target) {
|
---|
| 1934 | int i;
|
---|
| 1935 | int[] fold_start = new int[nr_fold + 1];
|
---|
| 1936 | int l = prob.l;
|
---|
| 1937 | int[] perm = new int[l];
|
---|
| 1938 |
|
---|
| 1939 | // stratified cv may not give leave-one-out rate
|
---|
| 1940 | // Each class to l folds -> some folds may have zero elements
|
---|
| 1941 | if ((param.svm_type == svm_parameter.C_SVC ||
|
---|
| 1942 | param.svm_type == svm_parameter.NU_SVC) && nr_fold < l) {
|
---|
| 1943 | int[] tmp_nr_class = new int[1];
|
---|
| 1944 | int[][] tmp_label = new int[1][];
|
---|
| 1945 | int[][] tmp_start = new int[1][];
|
---|
| 1946 | int[][] tmp_count = new int[1][];
|
---|
| 1947 |
|
---|
| 1948 | svm_group_classes(prob, tmp_nr_class, tmp_label, tmp_start, tmp_count, perm);
|
---|
| 1949 |
|
---|
| 1950 | int nr_class = tmp_nr_class[0];
|
---|
| 1951 | int[] start = tmp_start[0];
|
---|
| 1952 | int[] count = tmp_count[0];
|
---|
| 1953 |
|
---|
| 1954 | // random shuffle and then data grouped by fold using the array perm
|
---|
| 1955 | int[] fold_count = new int[nr_fold];
|
---|
| 1956 | int c;
|
---|
| 1957 | int[] index = new int[l];
|
---|
| 1958 | for (i = 0; i < l; i++)
|
---|
| 1959 | index[i] = perm[i];
|
---|
| 1960 | for (c = 0; c < nr_class; c++)
|
---|
| 1961 | for (i = 0; i < count[c]; i++) {
|
---|
| 1962 | int j = i + rand.Next(count[c] - i);
|
---|
| 1963 | {
|
---|
| 1964 | int _ = index[start[c] + j];
|
---|
| 1965 | index[start[c] + j] = index[start[c] + i];
|
---|
| 1966 | index[start[c] + i] = _;
|
---|
| 1967 | }
|
---|
| 1968 | }
|
---|
| 1969 | for (i = 0; i < nr_fold; i++) {
|
---|
| 1970 | fold_count[i] = 0;
|
---|
| 1971 | for (c = 0; c < nr_class; c++)
|
---|
| 1972 | fold_count[i] += (i + 1) * count[c] / nr_fold - i * count[c] / nr_fold;
|
---|
| 1973 | }
|
---|
| 1974 | fold_start[0] = 0;
|
---|
| 1975 | for (i = 1; i <= nr_fold; i++)
|
---|
| 1976 | fold_start[i] = fold_start[i - 1] + fold_count[i - 1];
|
---|
| 1977 | for (c = 0; c < nr_class; c++)
|
---|
| 1978 | for (i = 0; i < nr_fold; i++) {
|
---|
| 1979 | int begin = start[c] + i * count[c] / nr_fold;
|
---|
| 1980 | int end = start[c] + (i + 1) * count[c] / nr_fold;
|
---|
| 1981 | for (int j = begin; j < end; j++) {
|
---|
| 1982 | perm[fold_start[i]] = index[j];
|
---|
| 1983 | fold_start[i]++;
|
---|
| 1984 | }
|
---|
| 1985 | }
|
---|
| 1986 | fold_start[0] = 0;
|
---|
| 1987 | for (i = 1; i <= nr_fold; i++)
|
---|
| 1988 | fold_start[i] = fold_start[i - 1] + fold_count[i - 1];
|
---|
| 1989 | } else {
|
---|
| 1990 | for (i = 0; i < l; i++) perm[i] = i;
|
---|
| 1991 | for (i = 0; i < l; i++) {
|
---|
| 1992 | int j = i + rand.Next(l - i);
|
---|
| 1993 | {
|
---|
| 1994 | int _ = perm[i];
|
---|
| 1995 | perm[i] = perm[j];
|
---|
| 1996 | perm[j] = _;
|
---|
| 1997 | }
|
---|
| 1998 | }
|
---|
| 1999 | for (i = 0; i <= nr_fold; i++)
|
---|
| 2000 | fold_start[i] = i * l / nr_fold;
|
---|
| 2001 | }
|
---|
| 2002 |
|
---|
| 2003 | for (i = 0; i < nr_fold; i++) {
|
---|
| 2004 | int begin = fold_start[i];
|
---|
| 2005 | int end = fold_start[i + 1];
|
---|
| 2006 | int j, k;
|
---|
| 2007 | svm_problem subprob = new svm_problem();
|
---|
| 2008 |
|
---|
| 2009 | subprob.l = l - (end - begin);
|
---|
| 2010 | subprob.x = new svm_node[subprob.l][];
|
---|
| 2011 | subprob.y = new double[subprob.l];
|
---|
| 2012 |
|
---|
| 2013 | k = 0;
|
---|
| 2014 | for (j = 0; j < begin; j++) {
|
---|
| 2015 | subprob.x[k] = prob.x[perm[j]];
|
---|
| 2016 | subprob.y[k] = prob.y[perm[j]];
|
---|
| 2017 | ++k;
|
---|
| 2018 | }
|
---|
| 2019 | for (j = end; j < l; j++) {
|
---|
| 2020 | subprob.x[k] = prob.x[perm[j]];
|
---|
| 2021 | subprob.y[k] = prob.y[perm[j]];
|
---|
| 2022 | ++k;
|
---|
| 2023 | }
|
---|
| 2024 | svm_model submodel = svm_train(subprob, param);
|
---|
| 2025 | if (param.probability == 1 &&
|
---|
| 2026 | (param.svm_type == svm_parameter.C_SVC ||
|
---|
| 2027 | param.svm_type == svm_parameter.NU_SVC)) {
|
---|
| 2028 | double[] prob_estimates = new double[svm_get_nr_class(submodel)];
|
---|
| 2029 | for (j = begin; j < end; j++)
|
---|
| 2030 | target[perm[j]] = svm_predict_probability(submodel, prob.x[perm[j]], prob_estimates);
|
---|
| 2031 | } else
|
---|
| 2032 | for (j = begin; j < end; j++)
|
---|
| 2033 | target[perm[j]] = svm_predict(submodel, prob.x[perm[j]]);
|
---|
| 2034 | }
|
---|
| 2035 | }
|
---|
| 2036 |
|
---|
| 2037 | public static int svm_get_svm_type(svm_model model) {
|
---|
| 2038 | return model.param.svm_type;
|
---|
| 2039 | }
|
---|
| 2040 |
|
---|
| 2041 | public static int svm_get_nr_class(svm_model model) {
|
---|
| 2042 | return model.nr_class;
|
---|
| 2043 | }
|
---|
| 2044 |
|
---|
| 2045 | public static void svm_get_labels(svm_model model, int[] label) {
|
---|
| 2046 | if (model.label != null)
|
---|
| 2047 | for (int i = 0; i < model.nr_class; i++)
|
---|
| 2048 | label[i] = model.label[i];
|
---|
| 2049 | }
|
---|
| 2050 |
|
---|
| 2051 | public static double svm_get_svr_probability(svm_model model) {
|
---|
| 2052 | if ((model.param.svm_type == svm_parameter.EPSILON_SVR || model.param.svm_type == svm_parameter.NU_SVR) &&
|
---|
| 2053 | model.probA != null)
|
---|
| 2054 | return model.probA[0];
|
---|
| 2055 | else {
|
---|
| 2056 | Console.Error.WriteLine("Model doesn't contain information for SVR probability inference");
|
---|
| 2057 | return 0;
|
---|
| 2058 | }
|
---|
| 2059 | }
|
---|
| 2060 |
|
---|
| 2061 | public static double svm_predict_values(svm_model model, svm_node[] x, double[] dec_values) {
|
---|
| 2062 | int i;
|
---|
| 2063 | if (model.param.svm_type == svm_parameter.ONE_CLASS ||
|
---|
| 2064 | model.param.svm_type == svm_parameter.EPSILON_SVR ||
|
---|
| 2065 | model.param.svm_type == svm_parameter.NU_SVR) {
|
---|
| 2066 | double[] sv_coef = model.sv_coef[0];
|
---|
| 2067 | double sum = 0;
|
---|
| 2068 | for (i = 0; i < model.l; i++)
|
---|
| 2069 | sum += sv_coef[i] * Kernel.k_function(x, model.SV[i], model.param);
|
---|
| 2070 | sum -= model.rho[0];
|
---|
| 2071 | dec_values[0] = sum;
|
---|
| 2072 |
|
---|
| 2073 | if (model.param.svm_type == svm_parameter.ONE_CLASS)
|
---|
| 2074 | return (sum > 0) ? 1 : -1;
|
---|
| 2075 | else
|
---|
| 2076 | return sum;
|
---|
| 2077 | } else {
|
---|
| 2078 | int nr_class = model.nr_class;
|
---|
| 2079 | int l = model.l;
|
---|
| 2080 |
|
---|
| 2081 | double[] kvalue = new double[l];
|
---|
| 2082 | for (i = 0; i < l; i++)
|
---|
| 2083 | kvalue[i] = Kernel.k_function(x, model.SV[i], model.param);
|
---|
| 2084 |
|
---|
| 2085 | int[] start = new int[nr_class];
|
---|
| 2086 | start[0] = 0;
|
---|
| 2087 | for (i = 1; i < nr_class; i++)
|
---|
| 2088 | start[i] = start[i - 1] + model.nSV[i - 1];
|
---|
| 2089 |
|
---|
| 2090 | int[] vote = new int[nr_class];
|
---|
| 2091 | for (i = 0; i < nr_class; i++)
|
---|
| 2092 | vote[i] = 0;
|
---|
| 2093 |
|
---|
| 2094 | int p = 0;
|
---|
| 2095 | for (i = 0; i < nr_class; i++)
|
---|
| 2096 | for (int j = i + 1; j < nr_class; j++) {
|
---|
| 2097 | double sum = 0;
|
---|
| 2098 | int si = start[i];
|
---|
| 2099 | int sj = start[j];
|
---|
| 2100 | int ci = model.nSV[i];
|
---|
| 2101 | int cj = model.nSV[j];
|
---|
| 2102 |
|
---|
| 2103 | int k;
|
---|
| 2104 | double[] coef1 = model.sv_coef[j - 1];
|
---|
| 2105 | double[] coef2 = model.sv_coef[i];
|
---|
| 2106 | for (k = 0; k < ci; k++)
|
---|
| 2107 | sum += coef1[si + k] * kvalue[si + k];
|
---|
| 2108 | for (k = 0; k < cj; k++)
|
---|
| 2109 | sum += coef2[sj + k] * kvalue[sj + k];
|
---|
| 2110 | sum -= model.rho[p];
|
---|
| 2111 | dec_values[p] = sum;
|
---|
| 2112 |
|
---|
| 2113 | if (dec_values[p] > 0)
|
---|
| 2114 | ++vote[i];
|
---|
| 2115 | else
|
---|
| 2116 | ++vote[j];
|
---|
| 2117 | p++;
|
---|
| 2118 | }
|
---|
| 2119 |
|
---|
| 2120 | int vote_max_idx = 0;
|
---|
| 2121 | for (i = 1; i < nr_class; i++)
|
---|
| 2122 | if (vote[i] > vote[vote_max_idx])
|
---|
| 2123 | vote_max_idx = i;
|
---|
| 2124 |
|
---|
| 2125 | return model.label[vote_max_idx];
|
---|
| 2126 | }
|
---|
| 2127 | }
|
---|
| 2128 |
|
---|
| 2129 | public static double svm_predict(svm_model model, svm_node[] x) {
|
---|
| 2130 | int nr_class = model.nr_class;
|
---|
| 2131 | double[] dec_values;
|
---|
| 2132 | if (model.param.svm_type == svm_parameter.ONE_CLASS ||
|
---|
| 2133 | model.param.svm_type == svm_parameter.EPSILON_SVR ||
|
---|
| 2134 | model.param.svm_type == svm_parameter.NU_SVR)
|
---|
| 2135 | dec_values = new double[1];
|
---|
| 2136 | else
|
---|
| 2137 | dec_values = new double[nr_class * (nr_class - 1) / 2];
|
---|
| 2138 | double pred_result = svm_predict_values(model, x, dec_values);
|
---|
| 2139 | return pred_result;
|
---|
| 2140 | }
|
---|
| 2141 |
|
---|
| 2142 | public static double svm_predict_probability(svm_model model, svm_node[] x, double[] prob_estimates) {
|
---|
| 2143 | if ((model.param.svm_type == svm_parameter.C_SVC || model.param.svm_type == svm_parameter.NU_SVC) &&
|
---|
| 2144 | model.probA != null && model.probB != null) {
|
---|
| 2145 | int i;
|
---|
| 2146 | int nr_class = model.nr_class;
|
---|
| 2147 | double[] dec_values = new double[nr_class * (nr_class - 1) / 2];
|
---|
| 2148 | svm_predict_values(model, x, dec_values);
|
---|
| 2149 |
|
---|
| 2150 | double min_prob = 1e-7;
|
---|
| 2151 | double[][] pairwise_prob = new double[nr_class][];
|
---|
| 2152 |
|
---|
| 2153 | int k = 0;
|
---|
| 2154 | for (i = 0; i < nr_class; i++)
|
---|
| 2155 | pairwise_prob[i] = new double[nr_class];
|
---|
| 2156 | for (int j = i + 1; j < nr_class; j++) {
|
---|
| 2157 | pairwise_prob[i][j] =
|
---|
| 2158 | Math.Min(Math.Max(sigmoid_predict(dec_values[k], model.probA[k], model.probB[k]), min_prob), 1 - min_prob);
|
---|
| 2159 | pairwise_prob[j][i] = 1 - pairwise_prob[i][j];
|
---|
| 2160 | k++;
|
---|
| 2161 | }
|
---|
| 2162 | multiclass_probability(nr_class, pairwise_prob, prob_estimates);
|
---|
| 2163 |
|
---|
| 2164 | int prob_max_idx = 0;
|
---|
| 2165 | for (i = 1; i < nr_class; i++)
|
---|
| 2166 | if (prob_estimates[i] > prob_estimates[prob_max_idx])
|
---|
| 2167 | prob_max_idx = i;
|
---|
| 2168 | return model.label[prob_max_idx];
|
---|
| 2169 | } else
|
---|
| 2170 | return svm_predict(model, x);
|
---|
| 2171 | }
|
---|
| 2172 |
|
---|
| 2173 | private static readonly string[] svm_type_table = new string[]
|
---|
| 2174 | {
|
---|
| 2175 | "c_svc", "nu_svc", "one_class", "epsilon_svr", "nu_svr",
|
---|
| 2176 | };
|
---|
| 2177 |
|
---|
| 2178 | private static readonly string[] kernel_type_table = new string[]
|
---|
| 2179 | {
|
---|
| 2180 | "linear", "polynomial", "rbf", "sigmoid", "precomputed"
|
---|
| 2181 | };
|
---|
| 2182 |
|
---|
| 2183 |
|
---|
| 2184 | public static void svm_save_model(string model_file_name, svm_model model) {
|
---|
| 2185 | //DataOutputStream fp = new DataOutputStream(new BufferedOutputStream(new FileOutputStream(model_file_name)));
|
---|
| 2186 | var writer = new StreamWriter(model_file_name);
|
---|
| 2187 | svm_save_model(writer, model);
|
---|
| 2188 | }
|
---|
| 2189 |
|
---|
| 2190 |
|
---|
| 2191 | public static void svm_save_model(StreamWriter writer, svm_model model) {
|
---|
| 2192 |
|
---|
| 2193 | var savedCulture = Thread.CurrentThread.CurrentCulture;
|
---|
| 2194 | Thread.CurrentThread.CurrentCulture = CultureInfo.InvariantCulture;
|
---|
| 2195 | svm_parameter param = model.param;
|
---|
| 2196 |
|
---|
| 2197 | writer.Write("svm_type " + svm_type_table[param.svm_type] + Environment.NewLine);
|
---|
| 2198 | writer.Write("kernel_type " + kernel_type_table[param.kernel_type] + Environment.NewLine);
|
---|
| 2199 |
|
---|
| 2200 | if (param.kernel_type == svm_parameter.POLY)
|
---|
| 2201 | writer.Write("degree " + param.degree + Environment.NewLine);
|
---|
| 2202 |
|
---|
| 2203 | if (param.kernel_type == svm_parameter.POLY ||
|
---|
| 2204 | param.kernel_type == svm_parameter.RBF ||
|
---|
| 2205 | param.kernel_type == svm_parameter.SIGMOID)
|
---|
| 2206 | writer.Write("gamma " + param.gamma.ToString("r") + Environment.NewLine);
|
---|
| 2207 |
|
---|
| 2208 | if (param.kernel_type == svm_parameter.POLY ||
|
---|
| 2209 | param.kernel_type == svm_parameter.SIGMOID)
|
---|
| 2210 | writer.Write("coef0 " + param.coef0.ToString("r") + Environment.NewLine);
|
---|
| 2211 |
|
---|
| 2212 | int nr_class = model.nr_class;
|
---|
| 2213 | int l = model.l;
|
---|
| 2214 | writer.Write("nr_class " + nr_class + Environment.NewLine);
|
---|
| 2215 | writer.Write("total_sv " + l + Environment.NewLine);
|
---|
| 2216 |
|
---|
| 2217 | {
|
---|
| 2218 | writer.Write("rho");
|
---|
| 2219 | for (int i = 0; i < nr_class * (nr_class - 1) / 2; i++)
|
---|
| 2220 | writer.Write(" " + model.rho[i].ToString("r"));
|
---|
| 2221 | writer.Write(Environment.NewLine);
|
---|
| 2222 | }
|
---|
| 2223 |
|
---|
| 2224 | if (model.label != null) {
|
---|
| 2225 | writer.Write("label");
|
---|
| 2226 | for (int i = 0; i < nr_class; i++)
|
---|
| 2227 | writer.Write(" " + model.label[i]);
|
---|
| 2228 | writer.Write(Environment.NewLine);
|
---|
| 2229 | }
|
---|
| 2230 |
|
---|
| 2231 | if (model.probA != null) // regression has probA only
|
---|
| 2232 | {
|
---|
| 2233 | writer.Write("probA");
|
---|
| 2234 | for (int i = 0; i < nr_class * (nr_class - 1) / 2; i++)
|
---|
| 2235 | writer.Write(" " + model.probA[i].ToString("r"));
|
---|
| 2236 | writer.Write(Environment.NewLine);
|
---|
| 2237 | }
|
---|
| 2238 | if (model.probB != null) {
|
---|
| 2239 | writer.Write("probB");
|
---|
| 2240 | for (int i = 0; i < nr_class * (nr_class - 1) / 2; i++)
|
---|
| 2241 | writer.Write(" " + model.probB[i].ToString("r"));
|
---|
| 2242 | writer.Write(Environment.NewLine);
|
---|
| 2243 | }
|
---|
| 2244 |
|
---|
| 2245 | if (model.nSV != null) {
|
---|
| 2246 | writer.Write("nr_sv");
|
---|
| 2247 | for (int i = 0; i < nr_class; i++)
|
---|
| 2248 | writer.Write(" " + model.nSV[i]);
|
---|
| 2249 | writer.Write(Environment.NewLine);
|
---|
| 2250 | }
|
---|
| 2251 |
|
---|
| 2252 | writer.WriteLine("SV");
|
---|
| 2253 | double[][] sv_coef = model.sv_coef;
|
---|
| 2254 | svm_node[][] SV = model.SV;
|
---|
| 2255 |
|
---|
| 2256 | for (int i = 0; i < l; i++) {
|
---|
| 2257 | for (int j = 0; j < nr_class - 1; j++)
|
---|
| 2258 | writer.Write(sv_coef[j][i].ToString("r") + " ");
|
---|
| 2259 |
|
---|
| 2260 | svm_node[] p = SV[i];
|
---|
| 2261 | if (param.kernel_type == svm_parameter.PRECOMPUTED)
|
---|
| 2262 | writer.Write("0:" + (int)(p[0].value));
|
---|
| 2263 | else
|
---|
| 2264 | for (int j = 0; j < p.Length; j++)
|
---|
| 2265 | writer.Write(p[j].index + ":" + p[j].value.ToString("r") + " ");
|
---|
| 2266 | writer.Write(Environment.NewLine);
|
---|
| 2267 | }
|
---|
| 2268 |
|
---|
| 2269 | writer.Flush();
|
---|
| 2270 | Thread.CurrentThread.CurrentCulture = savedCulture;
|
---|
| 2271 | }
|
---|
| 2272 |
|
---|
| 2273 | private static double atof(String s) {
|
---|
| 2274 | return double.Parse(s);
|
---|
| 2275 | }
|
---|
| 2276 |
|
---|
| 2277 | private static int atoi(String s) {
|
---|
| 2278 | return int.Parse(s);
|
---|
| 2279 | }
|
---|
| 2280 |
|
---|
| 2281 |
|
---|
| 2282 | public static svm_model svm_load_model(String model_file_name) {
|
---|
| 2283 | return svm_load_model(new StreamReader(model_file_name));
|
---|
| 2284 | }
|
---|
| 2285 |
|
---|
| 2286 | public static svm_model svm_load_model(StreamReader reader) {
|
---|
| 2287 | var savedCulture = Thread.CurrentThread.CurrentCulture;
|
---|
| 2288 | Thread.CurrentThread.CurrentCulture = CultureInfo.InvariantCulture;
|
---|
| 2289 |
|
---|
| 2290 | // read parameters
|
---|
| 2291 |
|
---|
| 2292 | svm_model model = new svm_model();
|
---|
| 2293 | svm_parameter param = new svm_parameter();
|
---|
| 2294 | model.param = param;
|
---|
| 2295 | model.rho = null;
|
---|
| 2296 | model.probA = null;
|
---|
| 2297 | model.probB = null;
|
---|
| 2298 | model.label = null;
|
---|
| 2299 | model.nSV = null;
|
---|
| 2300 |
|
---|
| 2301 | while (true) {
|
---|
| 2302 | String cmd = reader.ReadLine();
|
---|
| 2303 | String arg = cmd.Substring(cmd.IndexOf(' ') + 1);
|
---|
| 2304 |
|
---|
| 2305 | if (cmd.StartsWith("svm_type")) {
|
---|
| 2306 | int i;
|
---|
| 2307 | for (i = 0; i < svm_type_table.Length; i++) {
|
---|
| 2308 | if (arg.IndexOf(svm_type_table[i], StringComparison.InvariantCultureIgnoreCase) != -1) {
|
---|
| 2309 | param.svm_type = i;
|
---|
| 2310 | break;
|
---|
| 2311 | }
|
---|
| 2312 | }
|
---|
| 2313 | if (i == svm_type_table.Length) {
|
---|
| 2314 | Console.Error.WriteLine("unknown svm type.");
|
---|
| 2315 | return null;
|
---|
| 2316 | }
|
---|
| 2317 | } else if (cmd.StartsWith("kernel_type")) {
|
---|
| 2318 | int i;
|
---|
| 2319 | for (i = 0; i < kernel_type_table.Length; i++) {
|
---|
| 2320 | if (arg.IndexOf(kernel_type_table[i], StringComparison.InvariantCultureIgnoreCase) != -1) {
|
---|
| 2321 | param.kernel_type = i;
|
---|
| 2322 | break;
|
---|
| 2323 | }
|
---|
| 2324 | }
|
---|
| 2325 | if (i == kernel_type_table.Length) {
|
---|
| 2326 | Console.Error.WriteLine("unknown kernel function.");
|
---|
| 2327 | return null;
|
---|
| 2328 | }
|
---|
| 2329 | } else if (cmd.StartsWith("degree"))
|
---|
| 2330 | param.degree = atoi(arg);
|
---|
| 2331 | else if (cmd.StartsWith("gamma"))
|
---|
| 2332 | param.gamma = atof(arg);
|
---|
| 2333 | else if (cmd.StartsWith("coef0"))
|
---|
| 2334 | param.coef0 = atof(arg);
|
---|
| 2335 | else if (cmd.StartsWith("nr_class"))
|
---|
| 2336 | model.nr_class = atoi(arg);
|
---|
| 2337 | else if (cmd.StartsWith("total_sv"))
|
---|
| 2338 | model.l = atoi(arg);
|
---|
| 2339 | else if (cmd.StartsWith("rho")) {
|
---|
| 2340 | int n = model.nr_class * (model.nr_class - 1) / 2;
|
---|
| 2341 | model.rho = new double[n];
|
---|
| 2342 | var st = arg.Split(' ', '\t', '\n', '\r', '\f');
|
---|
| 2343 | for (int i = 0; i < n; i++)
|
---|
| 2344 | model.rho[i] = atof(st[i]);
|
---|
| 2345 | } else if (cmd.StartsWith("label")) {
|
---|
| 2346 | int n = model.nr_class;
|
---|
| 2347 | model.label = new int[n];
|
---|
| 2348 | var st = arg.Split(' ', '\t', '\n', '\r', '\f');
|
---|
| 2349 | for (int i = 0; i < n; i++)
|
---|
| 2350 | model.label[i] = atoi(st[i]);
|
---|
| 2351 | } else if (cmd.StartsWith("probA")) {
|
---|
| 2352 | int n = model.nr_class * (model.nr_class - 1) / 2;
|
---|
| 2353 | model.probA = new double[n];
|
---|
| 2354 | var st = arg.Split(' ', '\t', '\n', '\r', '\f');
|
---|
| 2355 | for (int i = 0; i < n; i++)
|
---|
| 2356 | model.probA[i] = atof(st[i]);
|
---|
| 2357 | } else if (cmd.StartsWith("probB")) {
|
---|
| 2358 | int n = model.nr_class * (model.nr_class - 1) / 2;
|
---|
| 2359 | model.probB = new double[n];
|
---|
| 2360 | var st = arg.Split(' ', '\t', '\n', '\r', '\f');
|
---|
| 2361 | for (int i = 0; i < n; i++)
|
---|
| 2362 | model.probB[i] = atof(st[i]);
|
---|
| 2363 | } else if (cmd.StartsWith("nr_sv")) {
|
---|
| 2364 | int n = model.nr_class;
|
---|
| 2365 | model.nSV = new int[n];
|
---|
| 2366 | var st = arg.Split(' ', '\t', '\n', '\r', '\f');
|
---|
| 2367 | for (int i = 0; i < n; i++)
|
---|
| 2368 | model.nSV[i] = atoi(st[i]);
|
---|
| 2369 | } else if (cmd.StartsWith("SV")) {
|
---|
| 2370 | break;
|
---|
| 2371 | } else {
|
---|
| 2372 | Console.Error.WriteLine("unknown text in model file: [" + cmd + "]");
|
---|
| 2373 | return null;
|
---|
| 2374 | }
|
---|
| 2375 | }
|
---|
| 2376 |
|
---|
| 2377 | // read sv_coef and SV
|
---|
| 2378 |
|
---|
| 2379 | int m = model.nr_class - 1;
|
---|
| 2380 | int l = model.l;
|
---|
| 2381 | model.sv_coef = new double[m][];
|
---|
| 2382 | for (int k = 0; k < m; k++)
|
---|
| 2383 | model.sv_coef[k] = new double[l];
|
---|
| 2384 |
|
---|
| 2385 | model.SV = new svm_node[l][];
|
---|
| 2386 |
|
---|
| 2387 |
|
---|
| 2388 | for (int i = 0; i < l; i++) {
|
---|
| 2389 | String line = reader.ReadLine();
|
---|
| 2390 | var st = line.Split(' ', '\t', '\n', '\r', '\f', ':');
|
---|
| 2391 |
|
---|
| 2392 | for (int k = 0; k < m; k++) {
|
---|
| 2393 | model.sv_coef[k][i] = atof(st[k]);
|
---|
| 2394 | }
|
---|
| 2395 | // skip y value
|
---|
| 2396 | st = st.Skip(1).ToArray();
|
---|
| 2397 |
|
---|
| 2398 | int n = st.Length / 2;
|
---|
| 2399 | model.SV[i] = new svm_node[n];
|
---|
| 2400 | for (int j = 0; j < n; j++) {
|
---|
| 2401 | model.SV[i][j] = new svm_node();
|
---|
| 2402 | model.SV[i][j].index = atoi(st[2 * j]);
|
---|
| 2403 | model.SV[i][j].value = atof(st[2 * j + 1]);
|
---|
| 2404 | }
|
---|
| 2405 | }
|
---|
| 2406 |
|
---|
| 2407 | Thread.CurrentThread.CurrentCulture = savedCulture;
|
---|
| 2408 | return model;
|
---|
| 2409 | }
|
---|
| 2410 |
|
---|
| 2411 | public static string svm_check_parameter(svm_problem prob, svm_parameter param) {
|
---|
| 2412 | // svm_type
|
---|
| 2413 |
|
---|
| 2414 | int svm_type = param.svm_type;
|
---|
| 2415 | if (svm_type != svm_parameter.C_SVC &&
|
---|
| 2416 | svm_type != svm_parameter.NU_SVC &&
|
---|
| 2417 | svm_type != svm_parameter.ONE_CLASS &&
|
---|
| 2418 | svm_type != svm_parameter.EPSILON_SVR &&
|
---|
| 2419 | svm_type != svm_parameter.NU_SVR)
|
---|
| 2420 | return "unknown svm type";
|
---|
| 2421 |
|
---|
| 2422 | // kernel_type, degree
|
---|
| 2423 |
|
---|
| 2424 | int kernel_type = param.kernel_type;
|
---|
| 2425 | if (kernel_type != svm_parameter.LINEAR &&
|
---|
| 2426 | kernel_type != svm_parameter.POLY &&
|
---|
| 2427 | kernel_type != svm_parameter.RBF &&
|
---|
| 2428 | kernel_type != svm_parameter.SIGMOID &&
|
---|
| 2429 | kernel_type != svm_parameter.PRECOMPUTED)
|
---|
| 2430 | return "unknown kernel type";
|
---|
| 2431 |
|
---|
| 2432 | if (param.gamma < 0)
|
---|
| 2433 | return "gamma < 0";
|
---|
| 2434 |
|
---|
| 2435 | if (param.degree < 0)
|
---|
| 2436 | return "degree of polynomial kernel < 0";
|
---|
| 2437 |
|
---|
| 2438 | // cache_size,eps,C,nu,p,shrinking
|
---|
| 2439 |
|
---|
| 2440 | if (param.cache_size <= 0)
|
---|
| 2441 | return "cache_size <= 0";
|
---|
| 2442 |
|
---|
| 2443 | if (param.eps <= 0)
|
---|
| 2444 | return "eps <= 0";
|
---|
| 2445 |
|
---|
| 2446 | if (svm_type == svm_parameter.C_SVC ||
|
---|
| 2447 | svm_type == svm_parameter.EPSILON_SVR ||
|
---|
| 2448 | svm_type == svm_parameter.NU_SVR)
|
---|
| 2449 | if (param.C <= 0)
|
---|
| 2450 | return "C <= 0";
|
---|
| 2451 |
|
---|
| 2452 | if (svm_type == svm_parameter.NU_SVC ||
|
---|
| 2453 | svm_type == svm_parameter.ONE_CLASS ||
|
---|
| 2454 | svm_type == svm_parameter.NU_SVR)
|
---|
| 2455 | if (param.nu <= 0 || param.nu > 1)
|
---|
| 2456 | return "nu <= 0 or nu > 1";
|
---|
| 2457 |
|
---|
| 2458 | if (svm_type == svm_parameter.EPSILON_SVR)
|
---|
| 2459 | if (param.p < 0)
|
---|
| 2460 | return "p < 0";
|
---|
| 2461 |
|
---|
| 2462 | if (param.shrinking != 0 &&
|
---|
| 2463 | param.shrinking != 1)
|
---|
| 2464 | return "shrinking != 0 and shrinking != 1";
|
---|
| 2465 |
|
---|
| 2466 | if (param.probability != 0 &&
|
---|
| 2467 | param.probability != 1)
|
---|
| 2468 | return "probability != 0 and probability != 1";
|
---|
| 2469 |
|
---|
| 2470 | if (param.probability == 1 &&
|
---|
| 2471 | svm_type == svm_parameter.ONE_CLASS)
|
---|
| 2472 | return "one-class SVM probability output not supported yet";
|
---|
| 2473 |
|
---|
| 2474 | // check whether nu-svc is feasible
|
---|
| 2475 |
|
---|
| 2476 | if (svm_type == svm_parameter.NU_SVC) {
|
---|
| 2477 | int l = prob.l;
|
---|
| 2478 | int max_nr_class = 16;
|
---|
| 2479 | int nr_class = 0;
|
---|
| 2480 | int[] label = new int[max_nr_class];
|
---|
| 2481 | int[] count = new int[max_nr_class];
|
---|
| 2482 |
|
---|
| 2483 | int i;
|
---|
| 2484 | for (i = 0; i < l; i++) {
|
---|
| 2485 | int this_label = (int)prob.y[i];
|
---|
| 2486 | int j;
|
---|
| 2487 | for (j = 0; j < nr_class; j++)
|
---|
| 2488 | if (this_label == label[j]) {
|
---|
| 2489 | ++count[j];
|
---|
| 2490 | break;
|
---|
| 2491 | }
|
---|
| 2492 |
|
---|
| 2493 | if (j == nr_class) {
|
---|
| 2494 | if (nr_class == max_nr_class) {
|
---|
| 2495 | max_nr_class *= 2;
|
---|
| 2496 | int[] new_data = new int[max_nr_class];
|
---|
| 2497 | Array.Copy(label, 0, new_data, 0, label.Length);
|
---|
| 2498 | label = new_data;
|
---|
| 2499 |
|
---|
| 2500 | new_data = new int[max_nr_class];
|
---|
| 2501 | Array.Copy(count, 0, new_data, 0, count.Length);
|
---|
| 2502 | count = new_data;
|
---|
| 2503 | }
|
---|
| 2504 | label[nr_class] = this_label;
|
---|
| 2505 | count[nr_class] = 1;
|
---|
| 2506 | ++nr_class;
|
---|
| 2507 | }
|
---|
| 2508 | }
|
---|
| 2509 |
|
---|
| 2510 | for (i = 0; i < nr_class; i++) {
|
---|
| 2511 | int n1 = count[i];
|
---|
| 2512 | for (int j = i + 1; j < nr_class; j++) {
|
---|
| 2513 | int n2 = count[j];
|
---|
| 2514 | if (param.nu * (n1 + n2) / 2 > Math.Min(n1, n2))
|
---|
| 2515 | return "specified nu is infeasible";
|
---|
| 2516 | }
|
---|
| 2517 | }
|
---|
| 2518 | }
|
---|
| 2519 |
|
---|
| 2520 | return null;
|
---|
| 2521 | }
|
---|
| 2522 |
|
---|
| 2523 | public static int svm_check_probability_model(svm_model model) {
|
---|
| 2524 | if (((model.param.svm_type == svm_parameter.C_SVC || model.param.svm_type == svm_parameter.NU_SVC) &&
|
---|
| 2525 | model.probA != null && model.probB != null) ||
|
---|
| 2526 | ((model.param.svm_type == svm_parameter.EPSILON_SVR || model.param.svm_type == svm_parameter.NU_SVR) &&
|
---|
| 2527 | model.probA != null))
|
---|
| 2528 | return 1;
|
---|
| 2529 | else
|
---|
| 2530 | return 0;
|
---|
| 2531 | }
|
---|
| 2532 |
|
---|
| 2533 | public static void svm_set_print_string_function(Action<string> print_func) {
|
---|
| 2534 | /*if (print_func == null)
|
---|
| 2535 | svm_print_string = svm_print_stdout;
|
---|
| 2536 | else
|
---|
| 2537 | svm_print_string = print_func;
|
---|
| 2538 | */
|
---|
| 2539 | if (print_func != null) svm_print_string = print_func;
|
---|
| 2540 | }
|
---|
| 2541 | }
|
---|
| 2542 |
|
---|
| 2543 | public class svm_node : ICloneable {
|
---|
| 2544 | public int index;
|
---|
| 2545 | public double value;
|
---|
| 2546 | public object Clone() {
|
---|
| 2547 | var clone = new svm_node();
|
---|
| 2548 | clone.index = index;
|
---|
| 2549 | clone.value = value;
|
---|
| 2550 | return clone;
|
---|
| 2551 | }
|
---|
| 2552 | }
|
---|
| 2553 |
|
---|
| 2554 | public class svm_model {
|
---|
| 2555 | public svm_parameter param; // parameter
|
---|
| 2556 | public int nr_class; // number of classes, = 2 in regression/one class svm
|
---|
| 2557 | public int l; // total #SV
|
---|
| 2558 | public svm_node[][] SV; // SVs (SV[l])
|
---|
| 2559 | public double[][] sv_coef; // coefficients for SVs in decision functions (sv_coef[k-1][l])
|
---|
| 2560 | public double[] rho; // constants in decision functions (rho[k*(k-1)/2])
|
---|
| 2561 | public double[] probA; // pariwise probability information
|
---|
| 2562 | public double[] probB;
|
---|
| 2563 |
|
---|
| 2564 | // for classification only
|
---|
| 2565 |
|
---|
| 2566 | public int[] label; // label of each class (label[k])
|
---|
| 2567 | public int[] nSV; // number of SVs for each class (nSV[k])
|
---|
| 2568 | // nSV[0] + nSV[1] + ... + nSV[k-1] = l
|
---|
| 2569 | };
|
---|
| 2570 |
|
---|
| 2571 | public class svm_problem {
|
---|
| 2572 | public int l;
|
---|
| 2573 | public double[] y;
|
---|
| 2574 | public svm_node[][] x;
|
---|
| 2575 | }
|
---|
| 2576 |
|
---|
| 2577 | public interface svm_print_interface {
|
---|
| 2578 | void print(String s);
|
---|
| 2579 | }
|
---|
| 2580 |
|
---|
| 2581 | public class svm_parameter : ICloneable {
|
---|
| 2582 | /* svm_type */
|
---|
| 2583 | public const int C_SVC = 0;
|
---|
| 2584 | public const int NU_SVC = 1;
|
---|
| 2585 | public const int ONE_CLASS = 2;
|
---|
| 2586 | public const int EPSILON_SVR = 3;
|
---|
| 2587 | public const int NU_SVR = 4;
|
---|
| 2588 |
|
---|
| 2589 | /* kernel_type */
|
---|
| 2590 | public const int LINEAR = 0;
|
---|
| 2591 | public const int POLY = 1;
|
---|
| 2592 | public const int RBF = 2;
|
---|
| 2593 | public const int SIGMOID = 3;
|
---|
| 2594 | public const int PRECOMPUTED = 4;
|
---|
| 2595 |
|
---|
| 2596 | public int svm_type;
|
---|
| 2597 | public int kernel_type;
|
---|
| 2598 | public int degree; // for poly
|
---|
| 2599 | public double gamma; // for poly/rbf/sigmoid
|
---|
| 2600 | public double coef0; // for poly/sigmoid
|
---|
| 2601 |
|
---|
| 2602 | // these are for training only
|
---|
| 2603 | public double cache_size; // in MB
|
---|
| 2604 | public double eps; // stopping criteria
|
---|
| 2605 | public double C; // for C_SVC, EPSILON_SVR and NU_SVR
|
---|
| 2606 | public int nr_weight; // for C_SVC
|
---|
| 2607 | public int[] weight_label; // for C_SVC
|
---|
| 2608 | public double[] weight; // for C_SVC
|
---|
| 2609 | public double nu; // for NU_SVC, ONE_CLASS, and NU_SVR
|
---|
| 2610 | public double p; // for EPSILON_SVR
|
---|
| 2611 | public int shrinking; // use the shrinking heuristics
|
---|
| 2612 | public int probability; // do probability estimates
|
---|
| 2613 |
|
---|
| 2614 | public virtual object Clone() {
|
---|
| 2615 | var clone = new svm_parameter();
|
---|
| 2616 | clone.svm_type = svm_type;
|
---|
| 2617 | clone.kernel_type = kernel_type;
|
---|
| 2618 | clone.degree = degree;
|
---|
| 2619 | clone.gamma = gamma;
|
---|
| 2620 | clone.coef0 = coef0;
|
---|
| 2621 | clone.cache_size = cache_size;
|
---|
| 2622 | clone.eps = eps;
|
---|
| 2623 | clone.C = C;
|
---|
| 2624 | clone.nr_weight = nr_weight;
|
---|
[11306] | 2625 | if (weight_label != null) {
|
---|
| 2626 | clone.weight_label = new int[weight_label.Length];
|
---|
| 2627 | Array.Copy(weight_label, clone.weight_label, weight_label.Length);
|
---|
| 2628 | }
|
---|
| 2629 | if (weight != null) {
|
---|
| 2630 | clone.weight = new double[weight.Length];
|
---|
| 2631 | Array.Copy(weight, clone.weight, weight.Length);
|
---|
| 2632 | }
|
---|
[8607] | 2633 | clone.nu = nu;
|
---|
| 2634 | clone.p = p;
|
---|
| 2635 | clone.shrinking = shrinking;
|
---|
| 2636 | clone.probability = probability;
|
---|
| 2637 | return clone;
|
---|
| 2638 | }
|
---|
| 2639 | }
|
---|
| 2640 | }
|
---|