1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895
//! This module implements a decision tree from the simple binary tree [gbdt::binary_tree].
//!
//! In the training process, the nodes are splited according `impurity`.
//!
//!
//! Following hyperparameters are supported:
//!
//! 1. feature_size: the size of feautures. Training data and test data should have same
//! feature_size. (default = 1)
//!
//! 2. max_depth: the max depth of the decision tree. The root node is considered to be in the layer
//! 0. (default = 2)
//!
//! 3. min_leaf_size: the minimum number of samples required to be at a leaf node during training.
//! (default = 1)
//!
//! 4. loss: the loss function type. SquaredError, LogLikelyhood and LAD are supported. See
//! [config::Loss]. (default = SquareError).
//!
//! 5. feature_sample_ratio: portion of features to be splited. When spliting a node, a subset of
//! the features (feature_size * feature_sample_ratio) will be randomly selected to calculate
//! impurity. (default = 1.0)
//!
//! [gbdt::binary_tree]: ../binary_tree/index.html
//!
//! [config::Loss]: ../config/enum.Loss.html
//!
//! # Example
//! ```
//! use gbdt::config::Loss;
//! use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
//! // set up training data
//! let data1 = Data::new_training_data(
//! vec![1.0, 2.0, 3.0],
//! 1.0,
//! 2.0,
//! None
//! );
//! let data2 = Data::new_training_data(
//! vec![1.1, 2.1, 3.1],
//! 1.0,
//! 1.0,
//! None
//! );
//! let data3 = Data::new_training_data(
//! vec![2.0, 2.0, 1.0],
//! 1.0,
//! 0.5,
//! None
//! );
//! let data4 = Data::new_training_data(
//! vec![2.0, 2.3, 1.2],
//! 1.0,
//! 3.0,
//! None,
//! );
//!
//! let mut dv = Vec::new();
//! dv.push(data1.clone());
//! dv.push(data2.clone());
//! dv.push(data3.clone());
//! dv.push(data4.clone());
//!
//!
//! // train a decision tree
//! let mut tree = DecisionTree::new();
//! tree.set_feature_size(3);
//! tree.set_max_depth(2);
//! tree.set_min_leaf_size(1);
//! tree.set_loss(Loss::SquaredError);
//! let mut cache = TrainingCache::get_cache(3, &dv, 2);
//! tree.fit(&dv, &mut cache);
//!
//!
//! // set up the test data
//! let mut dv = Vec::new();
//! dv.push(data1.clone());
//! dv.push(data2.clone());
//! dv.push(Data::new_test_data(
//! vec![2.0, 2.0, 1.0],
//! None));
//! dv.push(Data::new_test_data(
//! vec![2.0, 2.3, 1.2],
//! Some(3.0)));
//!
//!
//! // inference the test data with the decision tree
//! println!("{:?}", tree.predict(&dv));
//!
//!
//! // output:
//! // [2.0, 0.75, 0.75, 3.0]
//! ```
#[cfg(all(feature = "mesalock_sgx", not(target_env = "sgx")))]
use std::prelude::v1::*;
use crate::binary_tree::BinaryTree;
use crate::binary_tree::BinaryTreeNode;
use crate::binary_tree::TreeIndex;
use crate::config::Loss;
#[cfg(feature = "enable_training")]
use crate::fitness::almost_equal;
use std::error::Error;
#[cfg(feature = "enable_training")]
use rand::prelude::SliceRandom;
#[cfg(feature = "enable_training")]
use rand::thread_rng;
use serde_derive::{Deserialize, Serialize};
///! For now we only support std::$t using this macro.
/// We will generalize ValueType in future.
macro_rules! def_value_type {
($t: tt) => {
pub type ValueType = $t;
pub const VALUE_TYPE_MAX: ValueType = std::$t::MAX;
pub const VALUE_TYPE_MIN: ValueType = std::$t::MIN;
pub const VALUE_TYPE_UNKNOWN: ValueType = VALUE_TYPE_MIN;
};
}
// use continous variables for decision tree
def_value_type!(f32);
/// A training sample or a test sample. You can call `new_training_data` to generate a training sample, and call `new_test_data` to generate a test sample.
///
/// A training sample can be used as a test sample.
///
/// You can also directly generate a data with following guides:
///
/// 1. When using the gbdt algorithm for training, you should set the values of feature, weight and label. If Config::initial_guess_enabled is true, you should set the value of initial_guess as well. Other fields can be arbitrary value.
///
/// 2. When using the gbdt algorithm for inference, you should set the value of feature. Other fields can be arbitrary value.
///
/// 3. When directly using the decision tree for training, only "SquaredError" is supported and you should set the values of feature, weight, label and target. `label` and `target` are equal. Other fields can be arbitrary value.
///
/// 4. When directly using the decision tree for inference, only "SquaredError" is supported and you should set the values of feature.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Data {
/// the vector of features
pub feature: Vec<ValueType>,
/// the target value of the sample to be fit in one decistion tree. This value is calculated by gradient boost algorithm. If you want to use the decision tree with "SquaredError" directly, set this value with `label` value
pub target: ValueType,
/// sample's weight. Used in training.
pub weight: ValueType,
/// sample's label. Used in training. This value is the actual value of the training sample.
pub label: ValueType,
/// used by LAD loss. Calculated by gradient boost algorithm.
pub residual: ValueType,
/// used by gradient boost. Set this value if Config::initial_guess_enabled is true.
pub initial_guess: ValueType,
}
impl Data {
/// Generate a training sample.
///
/// feature: the vector of features
///
/// weight: sample's weight
///
/// label: sample's label
///
/// initial_guess: initial prediction for the sample. This value is optional. Set this value if Config::initial_guess_enabled is true.
///
/// # Example
/// ``` rust
/// use gbdt::decision_tree::Data;
/// let data1 = Data::new_training_data(vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// Some(0.5));
/// let data2 = Data::new_training_data(vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// None);
/// ```
pub fn new_training_data(
feature: Vec<ValueType>,
weight: ValueType,
label: ValueType,
initial_guess: Option<ValueType>,
) -> Self {
Data {
feature,
target: label,
weight,
label,
residual: label,
initial_guess: initial_guess.unwrap_or(0.0),
}
}
/// Generate a test sample.
///
/// label: sample's label. It's optional.
///
/// # Example
/// ``` rust
/// use gbdt::decision_tree::Data;
/// let data1 = Data::new_test_data(vec![1.0, 2.0, 3.0],
/// Some(0.5));
/// let data2 = Data::new_test_data(vec![1.0, 2.0, 3.0],
/// None);
/// ```
pub fn new_test_data(feature: Vec<ValueType>, label: Option<ValueType>) -> Self {
Data {
feature,
target: 0.0,
weight: 1.0,
label: label.unwrap_or(0.0),
residual: 0.0,
initial_guess: 0.0,
}
}
}
/// The vector of the samples
pub type DataVec = Vec<Data>;
/// The vector of the predicted values.
pub type PredVec = Vec<ValueType>;
/// Cache some values for calculating the impurity.
#[cfg(feature = "enable_training")]
struct ImpurityCache {
/// sum of target * weight
sum_s: f64,
/// sum of target * target * weight
sum_ss: f64,
/// sum of weight
sum_c: f64,
/// whether this cache is calcualted
cached: bool,
/// whether a data is in the current node
bool_vec: Vec<bool>,
}
#[cfg(feature = "enable_training")]
impl ImpurityCache {
fn new(sample_size: usize, train_data: &[usize]) -> Self {
let mut bool_vec: Vec<bool> = vec![false; sample_size];
// set bool_vec
for index in train_data.iter() {
bool_vec[*index] = true;
}
ImpurityCache {
sum_s: 0.0,
sum_ss: 0.0,
sum_c: 0.0,
cached: false, //`cached` is false
bool_vec,
}
}
}
/// These results are repeatly used together: target*weight, target*target*weight, weight
#[cfg(feature = "enable_training")]
struct CacheValue {
/// target * weight
s: f64,
/// target * target * weight
ss: f64,
/// weight
c: f64,
}
/// Cache the sort results and some calculation results
#[cfg(feature = "enable_training")]
pub struct TrainingCache {
/// Sort the training data with the feature value.
/// ordered_features[i] is the data sorted by (i+1)th feature.
/// (usize, ValueType) is the sample's index in the training set and its (i+1)th feature value.
ordered_features: Vec<Vec<(usize, ValueType)>>,
/// Sort the training data with the residual field.
/// (uisze, ValueType) is the smaple's index in the training set and its residual value.
ordered_residual: Vec<(usize, ValueType)>,
/// cache_value[i] is the (i+1)th sample's `CacheValue`
cache_value: Vec<CacheValue>, //s, ss, c
/// cache_target[i] is the (i+1)th sample's `target` value (not the label). Organizing the `target` and `CacheValue` together will have better spatial locality.
cache_target: Vec<ValueType>,
/// loigt_c[i] is the (i+1)th sample's logit value. let y = target.abs(); let logit_value = y * (2.0 - y) * weight;
logit_c: Vec<ValueType>,
/// The sample size of the training set.
sample_size: usize,
/// The feature size of the training data
feature_size: usize,
/// The prediction of the training samples.
preds: Vec<ValueType>,
/// The cache level.
/// 0: ordered_features is calculated only once. SubCache is not used.
/// 1: ordered_features is calculated in each iterations. SubCache is used.
/// 2: ordered_features is calculated only once. SubCache is used.
cache_level: u8,
}
#[cfg(feature = "enable_training")]
impl TrainingCache {
/// Allocate the training cache. Feature size, training set and cache level should be provided.
/// ``` rust
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
/// // set up training data
/// let data1 = Data::new_training_data(
/// vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// None
/// );
/// let data2 = Data::new_training_data(
/// vec![1.1, 2.1, 3.1],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data3 = Data::new_training_data(
/// vec![2.0, 2.0, 1.0],
/// 1.0,
/// 0.5,
/// None
/// );
/// let mut dv = Vec::new();
/// dv.push(data1);
/// dv.push(data2);
/// dv.push(data3);
///
/// let mut cache = TrainingCache::get_cache(3, &dv, 2);
/// ```
// Only ordered_features may be pre-computed. Other fields will be computed by calling init_one_iteration
pub fn get_cache(feature_size: usize, data: &DataVec, cache_level: u8) -> Self {
// cache_level is 0, 1 or 2.
let level = if cache_level >= 3 { 2 } else { cache_level };
let sample_size = data.len();
let logit_c = vec![0.0; data.len()];
let preds = vec![VALUE_TYPE_UNKNOWN; sample_size];
let mut cache_value = Vec::with_capacity(data.len());
for elem in data {
let item = CacheValue {
s: 0.0,
ss: 0.0,
c: f64::from(elem.weight),
};
cache_value.push(item);
}
// Calculate the ordred_features if cache_level is 0 or 2.
let ordered_features: Vec<Vec<(usize, ValueType)>> = if (level == 0) || (level == 2) {
TrainingCache::cache_features(data, feature_size)
} else {
Vec::new()
};
let ordered_residual: Vec<(usize, ValueType)> = Vec::new();
let cache_target: Vec<ValueType> = vec![0.0; data.len()];
TrainingCache {
ordered_features,
ordered_residual,
cache_value,
cache_target,
logit_c,
sample_size,
feature_size,
preds,
cache_level: level,
}
}
/// Return the training data's predictions using this decision tree. These results are computed during training and then cached.
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
/// // set up training data
/// let data1 = Data::new_training_data(
/// vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// None
/// );
/// let data2 = Data::new_training_data(
/// vec![1.1, 2.1, 3.1],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data3 = Data::new_training_data(
/// vec![2.0, 2.0, 1.0],
/// 1.0,
/// 0.5,
/// None
/// );
/// let data4 = Data::new_training_data(
/// vec![2.0, 2.3, 1.2],
/// 1.0,
/// 3.0,
/// None
/// );
///
/// let mut dv = Vec::new();
/// dv.push(data1);
/// dv.push(data2);
/// dv.push(data3);
/// dv.push(data4);
///
///
/// // train a decision tree
/// let mut tree = DecisionTree::new();
/// tree.set_feature_size(3);
/// tree.set_max_depth(2);
/// tree.set_min_leaf_size(1);
/// tree.set_loss(Loss::SquaredError);
/// let mut cache = TrainingCache::get_cache(3, &dv, 2);
/// tree.fit(&dv, &mut cache);
/// // get predictions for the training data
/// println!("{:?}", cache.get_preds());
///
///
/// // output:
/// // [2.0, 0.75, 0.75, 3.0]
/// ```
pub fn get_preds(&self) -> Vec<ValueType> {
self.preds.to_vec()
}
/// Compute the training cache in the begining of training the deceision tree.
///
/// `whole_data`: the training set.
///
/// `loss`: the loss type.
fn init_one_iteration(&mut self, whole_data: &[Data], loss: &Loss) {
// Compute the cache_target, cache_value, logit_c.
for (index, data) in whole_data.iter().enumerate() {
let target = data.target;
self.cache_target[index] = target;
let weight = f64::from(data.weight);
let target = f64::from(target);
let s = target * weight;
self.cache_value[index].s = s;
self.cache_value[index].ss = target * s;
if let Loss::LogLikelyhood = loss {
let y = target.abs();
let c = y * (2.0 - y) * weight;
self.logit_c[index] = c as ValueType;
}
}
// Compute the ordered_residual.
if let Loss::LAD = loss {
self.ordered_residual = TrainingCache::cache_residual(whole_data);
}
}
/// Compute the ordered_features.
///
/// Input the training set (`whole_data`) and feature size (`feature_size`)
///
/// Output the ordered_features.
fn cache_features(whole_data: &[Data], feature_size: usize) -> Vec<Vec<(usize, ValueType)>> {
// Allocate memory
let mut ordered_features = Vec::with_capacity(feature_size);
for _index in 0..feature_size {
let nv: Vec<(usize, ValueType)> = Vec::with_capacity(whole_data.len());
ordered_features.push(nv);
}
// Put data
for (i, item) in whole_data.iter().enumerate() {
for (index, ordered_item) in ordered_features.iter_mut().enumerate().take(feature_size)
{
ordered_item.push((i, item.feature[index]));
}
}
// Sort all the vectors
for item in ordered_features.iter_mut().take(feature_size) {
item.sort_unstable_by(|a, b| {
let v1 = a.1;
let v2 = b.1;
v1.partial_cmp(&v2).unwrap()
});
}
ordered_features
}
/// Compute the ordered_residual.
///
/// Input the training set (`whole_data`). Output the ordered_residual.
fn cache_residual(whole_data: &[Data]) -> Vec<(usize, ValueType)> {
// Allocate memory
let mut ordered_residual = Vec::with_capacity(whole_data.len());
// Put data.
for (index, elem) in whole_data.iter().enumerate() {
ordered_residual.push((index, elem.residual));
}
// Sort data
ordered_residual.sort_unstable_by(|a, b| {
let v1: ValueType = a.1;
let v2: ValueType = b.1;
v1.partial_cmp(&v2).unwrap()
});
ordered_residual
}
/// Sort data with Training Cache. Bucket sort is used.
///
/// feature_index: which feature is used to sort.
///
/// is_residual: true, sort with residual value; false, sort with feature value
///
/// to_sort: bool vector. The index is the sample's index in whole training set. The boolean value indicates whether the sample is needed to be sorted.
///
/// to_sort_size: the amount of the data.
///
/// sub_cache: SubCache.
fn sort_with_bool_vec(
&self,
feature_index: usize,
is_residual: bool,
to_sort: &[bool],
to_sort_size: usize,
sub_cache: &SubCache,
) -> Vec<(usize, ValueType)> {
// Get sorted data.
let whole_data_sorted_index = if is_residual {
if (self.cache_level == 0) || sub_cache.lazy {
&self.ordered_residual
} else {
&sub_cache.ordered_residual
}
} else if (self.cache_level == 0) || sub_cache.lazy {
&self.ordered_features[feature_index]
} else {
&sub_cache.ordered_features[feature_index]
};
// The whole_data_sorted_index.len() is greater than or equal to to_sort_size. If they are equal, then whole_data_sorted_index is what we want.
if whole_data_sorted_index.len() == to_sort_size {
return whole_data_sorted_index.to_vec();
}
// Filter the whole_data_sorted_index with the boolean vector.
let mut ret = Vec::with_capacity(to_sort_size);
for item in whole_data_sorted_index.iter() {
let (index, value) = *item;
if to_sort[index] {
ret.push((index, value));
}
}
ret
}
/// Sort data with Training Cache. Bucket sort is used.
///
/// feature_index: which feature is used to sort.
///
/// is_residual: true, sort with residual value; false, sort with feature value
///
/// to_sort: a vector containing samples' indexes.
///
/// sub_cache: SubCache.
fn sort_with_cache(
&self,
feature_index: usize,
is_residual: bool,
to_sort: &[usize],
sub_cache: &SubCache,
) -> Vec<(usize, ValueType)> {
// Allocate the boolean vector
let whole_data_sorted_index = if is_residual {
&self.ordered_residual
} else {
&self.ordered_features[feature_index]
};
let mut index_exists: Vec<bool> = vec![false; whole_data_sorted_index.len()];
// Generate the boolean vector
for index in to_sort.iter() {
index_exists[*index] = true;
}
// Call sort_with_bool_vec to get sorted data
self.sort_with_bool_vec(
feature_index,
is_residual,
&index_exists,
to_sort.len(),
sub_cache,
)
}
}
/// SubCache is used to accelerate the data sorting.
/// ordered_features and ordered_residual are used in bucket sort.
/// In TrainingCache, the two vectors contains information from the whole training set. But only the information from samples in current node are needed.
/// So SubCache only restore the information from samples in current node.
#[cfg(feature = "enable_training")]
struct SubCache {
/// Sort the samples with the feature value.
/// ordered_features[i] is the data sorted by (i+1)th feature.
/// (usize, ValueType) is the sample's index in the whole training set and its (i+1)th feature value.
ordered_features: Vec<Vec<(usize, ValueType)>>,
/// Sort the samples with the residual field.
/// (uisze, ValueType) is the smaple's index in the whole training set and its residual value.
ordered_residual: Vec<(usize, ValueType)>,
/// True means the SubCache is not computed. For the root node, the samples in current node are the whole training set. So SubCache is not needed.
lazy: bool,
}
#[cfg(feature = "enable_training")]
impl SubCache {
/// Generate the SubCache frome the TrainingCache. `data` is the whole training set. `loss` is the loss type.
fn get_cache_from_training_cache(cache: &TrainingCache, data: &[Data], loss: &Loss) -> Self {
let level = cache.cache_level;
// level 2: lazy is True, ordered_features and ordered_residual is empty.
if level == 2 {
return SubCache {
ordered_features: Vec::new(),
ordered_residual: Vec::new(),
lazy: true,
};
}
// level 0: ordered_features is empty, lazy is false.
let ordered_features = if level == 0 {
Vec::new()
} else if level == 1 {
// level 1: ordered_features is computed from the whole training set. lazy is false
TrainingCache::cache_features(data, cache.feature_size)
} else {
// Other: clone the ordered_features in the TrainingCache.
let mut ordered_features: Vec<Vec<(usize, ValueType)>> =
Vec::with_capacity(cache.feature_size);
for index in 0..cache.feature_size {
ordered_features.push(cache.ordered_features[index].to_vec());
}
ordered_features
};
// level 0: ordered_residual is empty. lazy is false.
let ordered_residual = if level == 0 {
Vec::new()
} else if level == 1 {
// level 1: ordered_residual is computed from the whole train set. lazy if alse
if let Loss::LAD = loss {
TrainingCache::cache_residual(data)
} else {
Vec::new()
}
} else {
// other: clone the ordered_features in the TrainingCache.
if let Loss::LAD = loss {
cache.ordered_residual.to_vec()
} else {
Vec::new()
}
};
SubCache {
ordered_features,
ordered_residual,
lazy: false,
}
}
/// Generate an empty SubCache.
fn get_empty() -> Self {
SubCache {
ordered_features: Vec::new(),
ordered_residual: Vec::new(),
lazy: false,
}
}
/// Split the SubCache for child nodes
///
/// left_set: the samples in left child.
///
/// right_set: the samples in right child.
///
/// cache: the TrainingCache
///
/// output: two SubCache
fn split_cache(
mut self,
left_set: &[usize],
right_set: &[usize],
cache: &TrainingCache,
) -> (Self, Self) {
// level 0: return empty SubCache
if cache.cache_level == 0 {
return (SubCache::get_empty(), SubCache::get_empty());
}
// allocate the vectors
let mut left_ordered_features: Vec<Vec<(usize, ValueType)>> =
Vec::with_capacity(cache.feature_size);
let mut right_ordered_features: Vec<Vec<(usize, ValueType)>> =
Vec::with_capacity(cache.feature_size);
let mut left_ordered_residual = Vec::with_capacity(left_set.len());
let mut right_ordered_residual = Vec::with_capacity(right_set.len());
for _ in 0..cache.feature_size {
left_ordered_features.push(Vec::with_capacity(left_set.len()));
right_ordered_features.push(Vec::with_capacity(right_set.len()));
}
// compute two boolean vectors
let mut left_bool = vec![false; cache.sample_size];
let mut right_bool = vec![false; cache.sample_size];
for index in left_set.iter() {
left_bool[*index] = true;
}
for index in right_set.iter() {
right_bool[*index] = true;
}
// If lazy is true, compute the ordered_features from the TrainingCache.
if self.lazy {
for (feature_index, feature_vec) in cache.ordered_features.iter().enumerate() {
for pair in feature_vec.iter() {
let (index, value) = *pair;
if left_bool[index] {
left_ordered_features[feature_index].push((index, value));
continue;
}
if right_bool[index] {
right_ordered_features[feature_index].push((index, value));
}
}
}
} else {
// If lazy is false, compute the ordered_features from the current SubCache
for feature_index in 0..self.ordered_features.len() {
let feature_vec = &mut self.ordered_features[feature_index];
for pair in feature_vec.iter() {
let (index, value) = *pair;
if left_bool[index] {
left_ordered_features[feature_index].push((index, value));
continue;
}
if right_bool[index] {
right_ordered_features[feature_index].push((index, value));
}
}
feature_vec.clear();
feature_vec.shrink_to_fit();
}
self.ordered_features.clear();
self.ordered_features.shrink_to_fit();
}
// If lazy is true, compute the ordered_residual from the TrainingCache
if self.lazy {
for pair in cache.ordered_residual.iter() {
let (index, value) = *pair;
if left_bool[index] {
left_ordered_residual.push((index, value));
continue;
}
if right_bool[index] {
right_ordered_residual.push((index, value));
}
}
} else {
// If lazy is false, compute the ordered_residual from the current SubCache
for pair in self.ordered_residual.into_iter() {
let (index, value) = pair;
if left_bool[index] {
left_ordered_residual.push((index, value));
continue;
}
if right_bool[index] {
right_ordered_residual.push((index, value));
}
}
}
// return result
(
SubCache {
ordered_features: left_ordered_features,
ordered_residual: left_ordered_residual,
lazy: false,
},
SubCache {
ordered_features: right_ordered_features,
ordered_residual: right_ordered_residual,
lazy: false,
},
)
}
}
/// Calculate the prediction for each leaf node.
/// data: the samples in current node
/// loss: loss type
/// cache: TrainingCache
/// sub_cache: SubCache
#[cfg(feature = "enable_training")]
fn calculate_pred(
data: &[usize],
loss: &Loss,
cache: &TrainingCache,
sub_cache: &SubCache,
) -> ValueType {
match loss {
Loss::SquaredError => average(data, cache),
Loss::LogLikelyhood => logit_optimal_value(data, cache),
Loss::LAD => lad_optimal_value(data, cache, sub_cache),
_ => average(data, cache),
}
}
/// The leaf prediction value for SquaredError loss.
#[cfg(feature = "enable_training")]
fn average(data: &[usize], cache: &TrainingCache) -> ValueType {
let mut sum: f64 = 0.0;
let mut weight: f64 = 0.0;
for index in data.iter() {
let cv: &CacheValue = &cache.cache_value[*index];
sum += cv.s;
weight += cv.c;
}
if weight.abs() < 1e-10 {
0.0
} else {
(sum / weight) as ValueType
}
}
/// The leaf prediction value for LogLikelyhood loss.
#[cfg(feature = "enable_training")]
fn logit_optimal_value(data: &[usize], cache: &TrainingCache) -> ValueType {
let mut s: f64 = 0.0;
let mut c: f64 = 0.0;
for index in data.iter() {
s += cache.cache_value[*index].s;
c += f64::from(cache.logit_c[*index]);
}
if c.abs() < 1e-10 {
0.0
} else {
(s / c) as ValueType
}
}
/// The leaf prediction value for LAD loss.
#[cfg(feature = "enable_training")]
fn lad_optimal_value(data: &[usize], cache: &TrainingCache, sub_cache: &SubCache) -> ValueType {
let sorted_data = cache.sort_with_cache(0, true, data, sub_cache);
let all_weight = sorted_data
.iter()
.fold(0.0f64, |acc, x| acc + cache.cache_value[x.0].c);
let mut weighted_median: f64 = 0.0;
let mut weight: f64 = 0.0;
for (i, pair) in sorted_data.iter().enumerate() {
weight += cache.cache_value[pair.0].c;
if (weight * 2.0) > all_weight {
if i >= 1 {
weighted_median = f64::from((pair.1 + sorted_data[i - 1].1) / 2.0);
} else {
weighted_median = f64::from(pair.1);
}
break;
}
}
weighted_median as ValueType
}
/// Return whether the data vector have same target values.
#[allow(unused)]
#[cfg(feature = "enable_training")]
fn same(iv: &[usize], cache: &TrainingCache) -> bool {
if iv.is_empty() {
return false;
}
let t: ValueType = cache.cache_target[iv[0]];
for i in iv.iter().skip(1) {
if !(almost_equal(t, cache.cache_target[*i])) {
return false;
}
}
true
}
/// The internal node of the decision tree. It's stored in the `value` of the gbdt::binary_tree::BinaryTreeNode
#[derive(Debug, Serialize, Deserialize)]
struct DTNode {
/// the feature used to split the node
feature_index: usize,
/// the feature value used to split the node
feature_value: ValueType,
/// the prediction of the leaf node
pred: ValueType,
/// how to handle missing value: -1 (left child), 0 (node prediction), 1 (right child)
missing: i8,
/// whether the node is a leaf node
is_leaf: bool,
}
impl DTNode {
/// Return an empty DTNode
pub fn new() -> Self {
DTNode {
feature_index: 0,
feature_value: 0.0,
pred: 0.0,
missing: 0,
is_leaf: false,
}
}
}
/// The decision tree.
#[derive(Debug, Serialize, Deserialize)]
pub struct DecisionTree {
/// the tree
tree: BinaryTree<DTNode>,
/// the size of feautures. Training data and test data should have same feature size.
feature_size: usize,
/// the max depth of the decision tree. The root node is considered to be in the layer 0.
max_depth: u32,
/// the minimum number of samples required to be at a leaf node during training.
min_leaf_size: usize,
/// the loss function type.
loss: Loss,
/// portion of features to be splited. When spliting a node, a subset of the features
/// (feature_size * feature_sample_ratio) will be randomly selected to calculate impurity.
feature_sample_ratio: f64,
}
impl Default for DecisionTree {
fn default() -> Self {
Self::new()
}
}
impl DecisionTree {
/// Return a new decision tree with default values (feature_size = 1, max_depth = 2,
/// min_leaf_size = 1, loss = Loss::SquaredError, feature_sample_ratio = 1.0)
///
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree};
/// let mut tree = DecisionTree::new();
/// ```
pub fn new() -> Self {
DecisionTree {
tree: BinaryTree::new(),
feature_size: 1,
max_depth: 2,
min_leaf_size: 1,
loss: Loss::SquaredError,
feature_sample_ratio: 1.0,
}
}
/// Set the size of feautures. Training data and test data should have same feature size.
///
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree};
/// let mut tree = DecisionTree::new();
/// tree.set_feature_size(3);
/// ```
pub fn set_feature_size(&mut self, size: usize) {
self.feature_size = size;
}
/// Set the max depth of the decision tree. The root node is considered to be in the layer 0.
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree};
/// let mut tree = DecisionTree::new();
/// tree.set_max_depth(2);
/// ```
pub fn set_max_depth(&mut self, max_depth: u32) {
self.max_depth = max_depth;
}
/// Set the minimum number of samples required to be at a leaf node during training.
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree};
/// let mut tree = DecisionTree::new();
/// tree.set_min_leaf_size(1);
/// ```
pub fn set_min_leaf_size(&mut self, min_leaf_size: usize) {
self.min_leaf_size = min_leaf_size;
}
/// Set the loss function type.
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree};
/// let mut tree = DecisionTree::new();
/// tree.set_loss(Loss::SquaredError);
/// ```
pub fn set_loss(&mut self, loss: Loss) {
self.loss = loss;
}
/// Set the portion of features to be splited. When spliting a node, a subset of the features
/// (feature_size * feature_sample_ratio) will be randomly selected to calculate impurity.
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree};
/// let mut tree = DecisionTree::new();
/// tree.set_feature_sample_ratio(0.9);
/// ```
pub fn set_feature_sample_ratio(&mut self, feature_sample_ratio: f64) {
self.feature_sample_ratio = feature_sample_ratio;
}
/// Use the `subset` of the `train_data` to train a decision tree
///
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
/// // set up training data
/// let data1 = Data::new_training_data(
/// vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// None
/// );
/// let data2 = Data::new_training_data(
/// vec![1.1, 2.1, 3.1],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data3 = Data::new_training_data(
/// vec![2.0, 2.0, 1.0],
/// 1.0,
/// 0.5,
/// None
/// );
/// let data4 = Data::new_training_data(
/// vec![2.0, 2.3, 1.2],
/// 1.0,
/// 3.0,
/// None
/// );
///
/// let mut dv = Vec::new();
/// dv.push(data1.clone());
/// dv.push(data2.clone());
/// dv.push(data3.clone());
/// dv.push(data4.clone());
///
///
/// // train a decision tree
/// let mut tree = DecisionTree::new();
/// tree.set_feature_size(3);
/// tree.set_max_depth(2);
/// tree.set_min_leaf_size(1);
/// tree.set_loss(Loss::SquaredError);
/// let mut cache = TrainingCache::get_cache(3, &dv, 2);
/// let subset = [0,1,2];
/// tree.fit_n(&dv, &subset, &mut cache);
/// ```
#[cfg(feature = "enable_training")]
pub fn fit_n(&mut self, train_data: &DataVec, subset: &[usize], cache: &mut TrainingCache) {
assert!(
self.feature_size == cache.feature_size,
"Decision_tree and TrainingCache should have same feature size"
);
// Compute the TrainingCache
cache.init_one_iteration(train_data, &self.loss);
let root_index = self.tree.add_root(BinaryTreeNode::new(DTNode::new()));
// Generate the SubCache
let sub_cache = SubCache::get_cache_from_training_cache(cache, train_data, &self.loss);
self.fit_node(root_index, 0, subset, cache, sub_cache);
}
/// Use the samples in `train_data` to train the decision tree.
///
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
/// // set up training data
/// let data1 = Data::new_training_data(
/// vec![1.0, 2.0, 3.0],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data2 = Data::new_training_data(
/// vec![1.1, 2.1, 3.1],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data3 = Data::new_training_data(
/// vec![2.0, 2.0, 1.0],
/// 1.0,
/// 2.0,
/// None
/// );
///
/// let mut dv = Vec::new();
/// dv.push(data1.clone());
/// dv.push(data2.clone());
/// dv.push(data3.clone());
///
///
/// // train a decision tree
/// let mut tree = DecisionTree::new();
/// tree.set_feature_size(3);
/// tree.set_max_depth(2);
/// tree.set_min_leaf_size(1);
/// tree.set_loss(Loss::SquaredError);
/// let mut cache = TrainingCache::get_cache(3, &dv, 2);
/// tree.fit(&dv, &mut cache);
///
/// ```
#[cfg(feature = "enable_training")]
pub fn fit(&mut self, train_data: &DataVec, cache: &mut TrainingCache) {
//let mut gain: Vec<ValueType> = vec![0.0; self.feature_size];
assert!(
self.feature_size == cache.feature_size,
"Decision_tree and TrainingCache should have same feature size"
);
let data_collection: Vec<usize> = (0..train_data.len()).collect();
// Compute the TrainingCache
cache.init_one_iteration(train_data, &self.loss);
let root_index = self.tree.add_root(BinaryTreeNode::new(DTNode::new()));
// Generate the SubCache
let sub_cache = SubCache::get_cache_from_training_cache(cache, train_data, &self.loss);
self.fit_node(root_index, 0, &data_collection, cache, sub_cache);
}
/// Recursively build the tree nodes. It choose a feature and a value to split the node and the data.
/// And then use the splited data to build the child nodes.
/// node: the tree index of the current node
/// depth: the deepth of the current node
/// train_data: sample data in current node
/// cache: TrainingCache
/// sub_cache: SubCache
#[cfg(feature = "enable_training")]
fn fit_node(
&mut self,
node: TreeIndex,
depth: u32,
train_data: &[usize],
cache: &mut TrainingCache,
sub_cache: SubCache,
) {
// If the node doesn't need to be splited, make this node a leaf node.
{
let node_ref = self
.tree
.get_node_mut(node)
.expect("node should not be empty!");
// compute the prediction to support unknown features
node_ref.value.pred = calculate_pred(train_data, &self.loss, cache, &sub_cache);
if (depth >= self.max_depth)
|| same(train_data, cache)
|| (train_data.len() <= self.min_leaf_size)
{
node_ref.value.is_leaf = true;
for index in train_data.iter() {
cache.preds[*index] = node_ref.value.pred;
}
return;
}
}
// Try to find a feature and a value to split the node.
let (splited_data, feature_index, feature_value) = DecisionTree::split(
train_data,
self.feature_size,
self.feature_sample_ratio,
cache,
&sub_cache,
);
{
let node_ref = self
.tree
.get_node_mut(node)
.expect("node should not be empty");
// If spliting the node is failed, make this node a leaf node
if splited_data.is_none() {
node_ref.value.is_leaf = true;
node_ref.value.pred = calculate_pred(train_data, &self.loss, cache, &sub_cache);
for index in train_data.iter() {
cache.preds[*index] = node_ref.value.pred;
}
return;
} else {
node_ref.value.feature_index = feature_index;
node_ref.value.feature_value = feature_value;
}
}
// Use the splited data to build child nodes.
if let Some((left_data, right_data, _unknown_data)) = splited_data {
let (left_sub_cache, right_sub_cache) =
sub_cache.split_cache(&left_data, &right_data, cache);
let left_index = self
.tree
.add_left_node(node, BinaryTreeNode::new(DTNode::new()));
self.fit_node(left_index, depth + 1, &left_data, cache, left_sub_cache);
let right_index = self
.tree
.add_right_node(node, BinaryTreeNode::new(DTNode::new()));
self.fit_node(right_index, depth + 1, &right_data, cache, right_sub_cache);
}
}
/// Inference the subset of the `test_data`. Return a vector of
/// predicted values. If the `i` is in the subset, then output[i] is the prediction.
/// If `i` is not in the subset, then output[i] is 0.0
///
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
/// // set up training data
/// let data1 = Data::new_training_data(
/// vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// None
/// );
/// let data2 = Data::new_training_data(
/// vec![1.1, 2.1, 3.1],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data3 = Data::new_training_data(
/// vec![2.0, 2.0, 1.0],
/// 1.0,
/// 0.5,
/// None
/// );
/// let data4 = Data::new_training_data(
/// vec![2.0, 2.3, 1.2],
/// 1.0,
/// 3.0,
/// None
/// );
///
/// let mut dv = Vec::new();
/// dv.push(data1.clone());
/// dv.push(data2.clone());
/// dv.push(data3.clone());
/// dv.push(data4.clone());
///
///
/// // train a decision tree
/// let mut tree = DecisionTree::new();
/// tree.set_feature_size(3);
/// tree.set_max_depth(2);
/// tree.set_min_leaf_size(1);
/// tree.set_loss(Loss::SquaredError);
/// let mut cache = TrainingCache::get_cache(3, &dv, 2);
/// tree.fit(&dv, &mut cache);
///
///
/// // set up the test data
/// let mut dv = Vec::new();
/// dv.push(data1.clone());
/// dv.push(data2.clone());
/// dv.push(data3.clone());
/// dv.push(data4.clone());
///
///
/// // inference the test data with the decision tree
/// let subset = [0,1,2];
/// println!("{:?}", tree.predict_n(&dv, &subset));
///
///
/// // output:
/// // [2.0, 0.75, 0.75, 0.0]
/// ```
///
/// # Panic
/// If the function is called before the decision tree is trained, it will panic.
///
/// If the test data have a smaller feature size than the tree's feature size, it will panic.
pub fn predict_n(&self, test_data: &DataVec, subset: &[usize]) -> PredVec {
let root = self
.tree
.get_node(self.tree.get_root_index())
.expect("Decision tree should have root node");
let mut ret = vec![0.0; test_data.len()];
// Inference the samples one by one.
for index in subset {
ret[*index] = self.predict_one(root, &test_data[*index]);
}
ret
}
/// Inference the values of samples in the `test_data`. Return a vector of the predicted
/// values.
///
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
/// // set up training data
/// let data1 = Data::new_training_data(
/// vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// None
/// );
/// let data2 = Data::new_training_data(
/// vec![1.1, 2.1, 3.1],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data3 = Data::new_training_data(
/// vec![2.0, 2.0, 1.0],
/// 1.0,
/// 0.5,
/// None
/// );
/// let data4 = Data::new_training_data(
/// vec![2.0, 2.3, 1.2],
/// 1.0,
/// 3.0,
/// None
/// );
///
/// let mut dv = Vec::new();
/// dv.push(data1.clone());
/// dv.push(data2.clone());
/// dv.push(data3.clone());
/// dv.push(data4.clone());
///
///
/// // train a decision tree
/// let mut tree = DecisionTree::new();
/// tree.set_feature_size(3);
/// tree.set_max_depth(2);
/// tree.set_min_leaf_size(1);
/// tree.set_loss(Loss::SquaredError);
/// let mut cache = TrainingCache::get_cache(3, &dv, 2);
/// tree.fit(&dv, &mut cache);
///
///
/// // set up the test data
/// let mut dv = Vec::new();
/// dv.push(data1.clone());
/// dv.push(data2.clone());
/// dv.push(data3.clone());
/// dv.push(data4.clone());
///
///
/// // inference the test data with the decision tree
/// println!("{:?}", tree.predict(&dv));
///
///
/// // output:
/// // [2.0, 0.75, 0.75, 3.0]
/// ```
/// # Panic
/// If the function is called before the decision tree is trained, it will panic.
///
/// If the test data have a smaller feature size than the tree's feature size, it will panic.
pub fn predict(&self, test_data: &DataVec) -> PredVec {
let root = self
.tree
.get_node(self.tree.get_root_index())
.expect("Decision tree should have root node");
// Inference the data one by one
test_data
.iter()
.map(|x| self.predict_one(root, x))
.collect()
}
/// Inference a `sample` from current `node`
/// If the current node is a leaf node, return the node's prediction. Otherwise, choose a child
/// node according to the feature and feature value of the node. Then call this function recursively.
fn predict_one(&self, node: &BinaryTreeNode<DTNode>, sample: &Data) -> ValueType {
let mut is_node_value = false;
let mut is_left_child = false;
let mut _is_right_child = false;
if node.value.is_leaf {
is_node_value = true;
} else {
assert!(
sample.feature.len() > node.value.feature_index,
"sample doesn't have the feature"
);
if sample.feature[node.value.feature_index] == VALUE_TYPE_UNKNOWN {
if node.value.missing == -1 {
is_left_child = true;
} else if node.value.missing == 0 {
is_node_value = true;
} else {
_is_right_child = true;
}
} else if sample.feature[node.value.feature_index] < node.value.feature_value {
is_left_child = true;
} else {
_is_right_child = true;
}
}
// return the node's prediction
if is_node_value {
node.value.pred
} else if is_left_child {
let left = self
.tree
.get_left_child(node)
.expect("Left child should not be None");
self.predict_one(left, sample)
} else {
let right = self
.tree
.get_right_child(node)
.expect("Right child should not be None");
self.predict_one(right, sample)
}
}
/// Split the data by calculating the impurity.
/// Step 1: Choose candidate features. If `feature_sample_ratio` < 1.0, randomly selected
/// (feature_sample_ratio * feature_size) features. Otherwise, choose all features.
///
/// Step 2: Calculate each feature's impurity and the corresponding value to split the data.
///
/// Step 3: Find the feature that has the smallest impurity.
///
/// Step 4: Use the feature and the feature value to split the data.
///
/// Output: (left set, right set, unknown set), feature index, feature value
#[cfg(feature = "enable_training")]
fn split(
train_data: &[usize],
feature_size: usize,
feature_sample_ratio: f64,
cache: &TrainingCache,
sub_cache: &SubCache,
) -> (
Option<(Vec<usize>, Vec<usize>, Vec<usize>)>,
usize,
ValueType,
) {
let mut fs = feature_size;
let mut fv: Vec<usize> = (0..).take(fs).collect();
let mut rng = thread_rng();
if feature_sample_ratio < 1.0 {
fs = (feature_sample_ratio * (feature_size as f64)) as usize;
fv.shuffle(&mut rng);
}
let mut v: ValueType = 0.0;
let mut impurity: f64 = 0.0;
let mut best_fitness: f64 = std::f64::MAX;
let mut index: usize = 0;
let mut value: ValueType = 0.0;
// Generate the ImpurityCache
let mut impurity_cache = ImpurityCache::new(cache.sample_size, train_data);
let mut find: bool = false;
let mut data_to_split: Vec<(usize, ValueType)> = Vec::new();
// Calculate each feature's impurity
for i in fv.iter().take(fs) {
let sorted_data = DecisionTree::get_impurity(
train_data,
*i,
&mut v,
&mut impurity,
cache,
&mut impurity_cache,
&sub_cache,
);
if best_fitness > impurity {
find = true;
best_fitness = impurity;
index = *i;
value = v;
data_to_split = sorted_data;
}
}
// Split the node according to the impurity
if find {
let mut left: Vec<usize> = Vec::new();
let mut right: Vec<usize> = Vec::new();
let mut unknown: Vec<usize> = Vec::new();
for pair in data_to_split.iter() {
let (item_index, feature_value) = *pair;
if feature_value == VALUE_TYPE_UNKNOWN {
unknown.push(item_index);
} else if feature_value < value {
left.push(item_index);
} else {
right.push(item_index);
}
}
let mut count: u8 = 0;
if left.is_empty() {
count += 1;
}
if right.is_empty() {
count += 1;
}
if unknown.is_empty() {
count += 1;
}
if count >= 2 {
(None, 0, 0.0)
} else {
(Some((left, right, unknown)), index, value)
}
} else {
(None, 0, 0.0)
}
}
/// Calculate the impurity.
/// train_data: samples in current node
/// feature_index: the index of the selected feature
/// value: the result of the feature value
/// impurity: the result of the impurity
/// cache: TrainingCache
/// impurity_cache: ImpurityCache
/// sub_cache: SubCache
/// output: The sorted data according to the feature
#[cfg(feature = "enable_training")]
fn get_impurity(
train_data: &[usize],
feature_index: usize,
value: &mut ValueType,
impurity: &mut f64,
cache: &TrainingCache,
impurity_cache: &mut ImpurityCache,
sub_cache: &SubCache,
) -> Vec<(usize, ValueType)> {
*impurity = std::f64::MAX;
*value = VALUE_TYPE_UNKNOWN;
// Sort the samples with the feature value
let sorted_data = cache.sort_with_bool_vec(
feature_index,
false,
&impurity_cache.bool_vec,
train_data.len(),
sub_cache,
);
let mut unknown: usize = 0;
let mut s: f64 = 0.0;
let mut ss: f64 = 0.0;
let mut c: f64 = 0.0;
for pair in sorted_data.iter() {
let (index, feature_value) = *pair;
if feature_value == VALUE_TYPE_UNKNOWN {
let cv: &CacheValue = &cache.cache_value[index];
s += cv.s;
ss += cv.ss;
c += cv.c;
unknown += 1;
} else {
break;
}
}
if unknown == sorted_data.len() {
return sorted_data;
}
let mut fitness0 = if c > 1.0 { ss - s * s / c } else { 0.0 };
if fitness0 < 0.0 {
fitness0 = 0.0;
}
if !impurity_cache.cached {
impurity_cache.sum_s = 0.0;
impurity_cache.sum_ss = 0.0;
impurity_cache.sum_c = 0.0;
for index in train_data.iter() {
let cv: &CacheValue = &cache.cache_value[*index];
impurity_cache.sum_s += cv.s;
impurity_cache.sum_ss += cv.ss;
impurity_cache.sum_c += cv.c;
}
}
s = impurity_cache.sum_s - s;
ss = impurity_cache.sum_ss - ss;
c = impurity_cache.sum_c - c;
let _fitness00: f64 = if c > 1.0 { ss - s * s / c } else { 0.0 };
let mut ls: f64 = 0.0;
let mut lss: f64 = 0.0;
let mut lc: f64 = 0.0;
let mut rs: f64 = s;
let mut rss: f64 = ss;
let mut rc: f64 = c;
for i in unknown..(sorted_data.len() - 1) {
let (index, feature_value) = sorted_data[i];
let (_next_index, next_value) = sorted_data[i + 1];
let cv: &CacheValue = &cache.cache_value[index];
s = cv.s;
ss = cv.ss;
c = cv.c;
ls += s;
lss += ss;
lc += c;
rs -= s;
rss -= ss;
rc -= c;
let f1: ValueType = feature_value;
let f2: ValueType = next_value;
if almost_equal(f1, f2) {
continue;
}
let mut fitness1: f64 = if lc > 1.0 { lss - ls * ls / lc } else { 0.0 };
if fitness1 < 0.0 {
fitness1 = 0.0;
}
let mut fitness2: f64 = if rc > 1.0 { rss - rs * rs / rc } else { 0.0 };
if fitness2 < 0.0 {
fitness2 = 0.0;
}
let fitness: f64 = fitness0 + fitness1 + fitness2;
if *impurity > fitness {
*impurity = fitness;
*value = (f1 + f2) / 2.0;
}
}
sorted_data
}
/// Print the decision tree. For debug use.
///
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
/// // set up training data
/// let data1 = Data::new_training_data(
/// vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// None
/// );
/// let data2 = Data::new_training_data(
/// vec![1.1, 2.1, 3.1],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data3 = Data::new_training_data(
/// vec![2.0, 2.0, 1.0],
/// 1.0,
/// 0.5,
/// None
/// );
///
/// let mut dv = Vec::new();
/// dv.push(data1.clone());
/// dv.push(data2.clone());
/// dv.push(data3.clone());
///
///
/// // train a decision tree
/// let mut tree = DecisionTree::new();
/// tree.set_feature_size(3);
/// tree.set_max_depth(2);
/// tree.set_min_leaf_size(1);
/// tree.set_loss(Loss::SquaredError);
/// let mut cache = TrainingCache::get_cache(3, &dv, 2);
/// let subset = [0, 1];
/// tree.fit_n(&dv, &subset, &mut cache);
///
///
/// tree.print();
///
/// // output:
///
/// // ----DTNode { feature_index: 0, feature_value: 1.05, pred: 1.5, is_leaf: false }
/// // ----DTNode { feature_index: 0, feature_value: 0.0, pred: 2.0, is_leaf: true }
/// // ----DTNode { feature_index: 0, feature_value: 0.0, pred: 1.0, is_leaf: true }
/// ```
pub fn print(&self) {
self.tree.print();
}
/// Build a decision tree from xgboost's model. xgboost can dump the model in JSON format. We used serde_json to parse a JSON string.
/// # Example
/// ``` rust
/// use serde_json::{Result, Value};
/// use gbdt::decision_tree::DecisionTree;
/// let data = r#"
/// { "nodeid": 0, "depth": 0, "split": 0, "split_condition": 750, "yes": 1, "no": 2, "missing": 2, "children": [
/// { "nodeid": 1, "leaf": 25.7333336 },
/// { "nodeid": 2, "leaf": 15.791667 }]}"#;
/// let node: Value = serde_json::from_str(data).unwrap();
/// let dt = DecisionTree::get_from_xgboost(&node);
/// ```
pub fn get_from_xgboost(node: &serde_json::Value) -> Result<Self, Box<Error>> {
// Parameters are not used in prediction process, so we use default parameters.
let mut tree = DecisionTree::new();
let index = tree.tree.add_root(BinaryTreeNode::new(DTNode::new()));
tree.add_node_from_json(index, node)?;
Ok(tree)
}
/// Recursively build the tree node from the JSON value.
fn add_node_from_json(
&mut self,
index: TreeIndex,
node: &serde_json::Value,
) -> Result<(), Box<Error>> {
{
let node_ref = self
.tree
.get_node_mut(index)
.expect("node should not be empty!");
// This is the leaf node
if let serde_json::Value::Number(pred) = &node["leaf"] {
let leaf_value = pred.as_f64().ok_or("parse 'leaf' error")?;
node_ref.value.pred = leaf_value as ValueType;
node_ref.value.is_leaf = true;
return Ok(());
} else {
// feature value
let feature_value = node["split_condition"]
.as_f64()
.ok_or("parse 'split condition' error")?;
node_ref.value.feature_value = feature_value as ValueType;
// feature index
let feature_index = match node["split"].as_i64() {
Some(v) => v,
None => {
let feature_name = node["split"].as_str().ok_or("parse 'split' error")?;
let feature_str: String = feature_name.chars().skip(3).collect();
feature_str.parse::<i64>()?
}
};
node_ref.value.feature_index = feature_index as usize;
// handle unknown feature
let missing = node["missing"].as_i64().ok_or("parse 'missing' error")?;
let left_child = node["yes"].as_i64().ok_or("parse 'yes' error")?;
let right_child = node["no"].as_i64().ok_or("parse 'no' error")?;
if missing == left_child {
node_ref.value.missing = -1;
} else if missing == right_child {
node_ref.value.missing = 1;
} else {
let err: Box<Error> = From::from("not support extra missing node".to_string());
return Err(err);
}
}
}
// ids for children
let left_child = node["yes"].as_i64().ok_or("parse 'yes' error")?;
let right_child = node["no"].as_i64().ok_or("parse 'no' error")?;
let children = node["children"]
.as_array()
.ok_or("parse 'children' error")?;
let mut find_left = false;
let mut find_right = false;
for child in children.iter() {
let node_id = child["nodeid"].as_i64().ok_or("parse 'nodeid' error")?;
// build left child
if node_id == left_child {
find_left = true;
let left_index = self
.tree
.add_left_node(index, BinaryTreeNode::new(DTNode::new()));
self.add_node_from_json(left_index, child)?;
}
// build right child
if node_id == right_child {
find_right = true;
let right_index = self
.tree
.add_right_node(index, BinaryTreeNode::new(DTNode::new()));
self.add_node_from_json(right_index, child)?;
}
}
if (!find_left) || (!find_right) {
let err: Box<Error> = From::from("children not found".to_string());
return Err(err);
}
Ok(())
}
/// For debug use. Return the number of nodes in current decision tree
///
/// # Example
/// ```
/// use gbdt::config::Loss;
/// use gbdt::decision_tree::{Data, DecisionTree, TrainingCache};
/// // set up training data
/// let data1 = Data::new_training_data(
/// vec![1.0, 2.0, 3.0],
/// 1.0,
/// 2.0,
/// None
/// );
/// let data2 = Data::new_training_data(
/// vec![1.1, 2.1, 3.1],
/// 1.0,
/// 1.0,
/// None
/// );
/// let data3 = Data::new_training_data(
/// vec![2.0, 2.0, 1.0],
/// 1.0,
/// 0.5,
/// None
/// );
///
/// let mut dv = Vec::new();
/// dv.push(data1.clone());
/// dv.push(data2.clone());
/// dv.push(data3.clone());
///
///
/// // train a decision tree
/// let mut tree = DecisionTree::new();
/// tree.set_feature_size(3);
/// tree.set_max_depth(2);
/// tree.set_min_leaf_size(1);
/// tree.set_loss(Loss::SquaredError);
/// let mut cache = TrainingCache::get_cache(3, &dv, 2);
/// let subset = [0, 1];
/// tree.fit_n(&dv, &subset, &mut cache);
///
/// assert_eq!(tree.len(), 3)
pub fn len(&self) -> usize {
self.tree.len()
}
/// Returns true if the current decision tree is empty
pub fn is_empty(&self) -> bool {
self.tree.is_empty()
}
}