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
/*!
* Generic N dimensional [named tensors](http://nlp.seas.harvard.edu/NamedTensor).
*
* Tensors are generic over some type `T` and some usize `D`. If `T` is [Numeric](super::numeric)
* then the tensor can be used in a mathematical way. `D` is the number of dimensions in the tensor
* and a compile time constant. Each tensor also carries `D` dimension name and length pairs.
*/
use crate::linear_algebra;
use crate::numeric::extra::{Real, RealRef};
use crate::numeric::{Numeric, NumericRef};
use crate::tensors::indexing::{
ShapeIterator, TensorAccess, TensorIterator, TensorOwnedIterator, TensorReferenceIterator,
TensorReferenceMutIterator, TensorTranspose,
};
use crate::tensors::views::{
DataLayout, IndexRange, IndexRangeValidationError, TensorExpansion, TensorIndex, TensorMask,
TensorMut, TensorRange, TensorRef, TensorRename, TensorView,
};
use std::error::Error;
use std::fmt;
#[cfg(feature = "serde")]
use serde::Serialize;
pub mod dimensions;
mod display;
pub mod indexing;
pub mod operations;
pub mod views;
#[cfg(feature = "serde")]
pub use serde_impls::TensorDeserialize;
/**
* Dimension names are represented as static string references.
*
* This allows you to use string literals to refer to named dimensions, for example you might want
* to construct a tensor with a shape of
* `[("batch", 1000), ("height", 100), ("width", 100), ("rgba", 4)]`.
*
* Alternatively you can define the strings once as constants and refer to your dimension
* names by the constant identifiers.
*
* ```
* const BATCH: &'static str = "batch";
* const HEIGHT: &'static str = "height";
* const WIDTH: &'static str = "width";
* const RGBA: &'static str = "rgba";
* ```
*
* Although `Dimension` is interchangable with `&'static str` as it is just a type alias, Easy ML
* uses `Dimension` whenever dimension names are expected to distinguish the types from just
* strings.
*/
pub type Dimension = &'static str;
/**
* An error indicating failure to do something with a Tensor because the requested shape
* is not valid.
*/
#[derive(Clone, Debug, Eq, PartialEq)]
pub struct InvalidShapeError<const D: usize> {
shape: [(Dimension, usize); D],
}
impl<const D: usize> InvalidShapeError<D> {
/**
* Checks if this shape is valid. This is mainly for internal library use but may also be
* useful for unit testing.
*
* Note: in some functions and methods, an InvalidShapeError may be returned which is a valid
* shape, but not the right size for the quantity of data provided.
*/
pub fn is_valid(&self) -> bool {
!crate::tensors::dimensions::has_duplicates(&self.shape)
&& !self.shape.iter().any(|d| d.1 == 0)
}
/**
* Constructs an InvalidShapeError for assistance with unit testing. Note that you can
* construct an InvalidShapeError that *is* a valid shape in this way.
*/
pub fn new(shape: [(Dimension, usize); D]) -> InvalidShapeError<D> {
InvalidShapeError { shape }
}
pub fn shape(&self) -> [(Dimension, usize); D] {
self.shape
}
pub fn shape_ref(&self) -> &[(Dimension, usize); D] {
&self.shape
}
// Panics if the shape is invalid for any reason with the appropriate error message.
#[track_caller]
#[inline]
fn validate_dimensions_or_panic(shape: &[(Dimension, usize); D], data_len: usize) {
let elements = crate::tensors::dimensions::elements(shape);
if data_len != elements {
panic!(
"Product of dimension lengths must match size of data. {} != {}",
elements, data_len
);
}
if crate::tensors::dimensions::has_duplicates(shape) {
panic!("Dimension names must all be unique: {:?}", &shape);
}
if shape.iter().any(|d| d.1 == 0) {
panic!("No dimension can have 0 elements: {:?}", &shape);
}
}
// Returns true if the shape is valid and matches the data length
fn validate_dimensions(shape: &[(Dimension, usize); D], data_len: usize) -> bool {
let elements = crate::tensors::dimensions::elements(shape);
data_len == elements
&& !crate::tensors::dimensions::has_duplicates(shape)
&& !shape.iter().any(|d| d.1 == 0)
}
}
impl<const D: usize> fmt::Display for InvalidShapeError<D> {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
write!(
f,
"Dimensions must all be at least length 1 with unique names: {:?}",
self.shape
)
}
}
impl<const D: usize> Error for InvalidShapeError<D> {}
/**
* An error indicating failure to do something with a Tensor because the dimension names that
* were provided did not match with the dimension names that were valid.
*
* Typically this would be due to the same dimension name being provided multiple times, or a
* dimension name being provided that is not present in the shape of the Tensor in use.
*/
#[derive(Clone, Debug, Eq, PartialEq)]
pub struct InvalidDimensionsError<const D: usize, const P: usize> {
valid: [Dimension; D],
provided: [Dimension; P],
}
impl<const D: usize, const P: usize> InvalidDimensionsError<D, P> {
/**
* Checks if the provided dimensions have duplicate names. This is mainly for internal library
* use but may also be useful for unit testing.
*/
pub fn has_duplicates(&self) -> bool {
crate::tensors::dimensions::has_duplicates_names(&self.provided)
}
// TODO: method to check provided is a subset of valid
/**
* Constructs an InvalidDimensions for assistance with unit testing.
*/
pub fn new(provided: [Dimension; P], valid: [Dimension; D]) -> InvalidDimensionsError<D, P> {
InvalidDimensionsError { valid, provided }
}
pub fn provided_names(&self) -> [Dimension; P] {
self.provided
}
pub fn provided_names_ref(&self) -> &[Dimension; P] {
&self.provided
}
pub fn valid_names(&self) -> [Dimension; D] {
self.valid
}
pub fn valid_names_ref(&self) -> &[Dimension; D] {
&self.valid
}
}
impl<const D: usize, const P: usize> fmt::Display for InvalidDimensionsError<D, P> {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
if P > 0 {
write!(
f,
"Dimensions names {:?} were incorrect, valid dimensions in this context are: {:?}",
self.provided, self.valid
)
} else {
write!(f, "Dimensions names {:?} were incorrect", self.provided)
}
}
}
impl<const D: usize, const P: usize> Error for InvalidDimensionsError<D, P> {}
#[test]
fn test_sync() {
fn assert_sync<T: Sync>() {}
assert_sync::<InvalidShapeError<2>>();
assert_sync::<InvalidDimensionsError<2, 2>>();
}
#[test]
fn test_send() {
fn assert_send<T: Send>() {}
assert_send::<InvalidShapeError<2>>();
assert_send::<InvalidDimensionsError<2, 2>>();
}
/**
* A [named tensor](http://nlp.seas.harvard.edu/NamedTensor) of some type `T` and number of
* dimensions `D`.
*
* Tensors are a generalisation of matrices; whereas [Matrix](crate::matrices::Matrix) only
* supports 2 dimensions, and vectors are represented in Matrix by making either the rows or
* columns have a length of one, [Tensor] supports an arbitary number of dimensions,
* with 0 through 6 having full API support. A `Tensor<T, 2>` is very similar to a `Matrix<T>`
* except that this type associates each dimension with a name, and favor names to refer to
* dimensions instead of index order.
*
* Like Matrix, the type of the data in this Tensor may implement no traits, in which case the
* tensor will be rather useless. If the type implements Clone most storage and accessor methods
* are defined and if the type implements Numeric then the tensor can be used in a mathematical
* way.
*
* Like Matrix, a Tensor must always contain at least one element, and it may not not have more
* elements than `std::isize::MAX`. Concerned readers should note that on a 64 bit computer this
* maximum value is 9,223,372,036,854,775,807 so running out of memory is likely to occur first.
*
* When doing numeric operations with Tensors you should be careful to not consume a tensor by
* accidentally using it by value. All the operations are also defined on references to tensors
* so you should favor &x + &y style notation for tensors you intend to continue using.
*
* See also:
* - [indexing]
*/
#[derive(Debug)]
#[cfg_attr(feature = "serde", derive(Serialize))]
pub struct Tensor<T, const D: usize> {
data: Vec<T>,
#[cfg_attr(feature = "serde", serde(with = "serde_arrays"))]
shape: [(Dimension, usize); D],
#[cfg_attr(feature = "serde", serde(skip))]
strides: [usize; D],
}
impl<T, const D: usize> Tensor<T, D> {
/**
* Creates a Tensor with a particular number of dimensions and lengths in each dimension.
*
* The product of the dimension lengths corresponds to the number of elements the Tensor
* will store. Elements are stored in what would be row major order for a Matrix.
* Each step in memory through the N dimensions corresponds to incrementing the rightmost
* index, hence a shape of `[("row", 5), ("column", 5)]` would mean the first 6 elements
* passed in the Vec would be for (0,0), (0,1), (0,2), (0,3), (0,4), (1,0) and so on to (4,4)
* for the 25th and final element.
*
* # Panics
*
* - If the number of provided elements does not match the product of the dimension lengths.
* - If a dimension name is not unique
* - If any dimension has 0 elements
*
* Note that an empty list for dimensions is valid, and constructs a 0 dimensional tensor with
* a single element (since the product of an empty list is 1).
*/
#[track_caller]
pub fn from(shape: [(Dimension, usize); D], data: Vec<T>) -> Self {
InvalidShapeError::validate_dimensions_or_panic(&shape, data.len());
let strides = compute_strides(&shape);
Tensor {
data,
shape,
strides,
}
}
/**
* Creates a Tensor with a particular shape initialised from a function.
*
* The product of the dimension lengths corresponds to the number of elements the Tensor
* will store. Elements are stored in what would be row major order for a Matrix.
* Each step in memory through the N dimensions corresponds to incrementing the rightmost
* index, hence a shape of `[("row", 5), ("column", 5)]` would mean the first 6 elements
* passed in the Vec would be for (0,0), (0,1), (0,2), (0,3), (0,4), (1,0) and so on to (4,4)
* for the 25th and final element. These same indexes will be passed to the producer function
* to initialised the values for the Tensor.
*
* ```
* use easy_ml::tensors::Tensor;
* let tensor = Tensor::from_fn([("rows", 4), ("columns", 4)], |[r, c]| r * c);
* assert_eq!(
* tensor,
* Tensor::from([("rows", 4), ("columns", 4)], vec![
* 0, 0, 0, 0,
* 0, 1, 2, 3,
* 0, 2, 4, 6,
* 0, 3, 6, 9,
* ])
* );
* ```
*
* # Panics
*
* - If a dimension name is not unique
* - If any dimension has 0 elements
*
* Note that an empty list for dimensions is valid, and constructs a 0 dimensional tensor with
* a single element (since the product of an empty list is 1).
*/
#[track_caller]
pub fn from_fn<F>(shape: [(Dimension, usize); D], mut producer: F) -> Self
where
F: FnMut([usize; D]) -> T,
{
let length = dimensions::elements(&shape);
let mut data = Vec::with_capacity(length);
let iterator = ShapeIterator::from(shape);
for index in iterator {
data.push(producer(index));
}
Tensor::from(shape, data)
}
/**
* The shape of this tensor. Since Tensors are named Tensors, their shape is not just a
* list of lengths along each dimension, but instead a list of pairs of names and lengths.
*
* See also
* - [dimensions]
* - [indexing]
*/
pub fn shape(&self) -> [(Dimension, usize); D] {
self.shape
}
/**
* A non panicking version of [from](Tensor::from) which returns `Result::Err` if the input
* is invalid.
*
* Creates a Tensor with a particular number of dimensions and lengths in each dimension.
*
* The product of the dimension lengths corresponds to the number of elements the Tensor
* will store. Elements are stored in what would be row major order for a Matrix.
* Each step in memory through the N dimensions corresponds to incrementing the rightmost
* index, hence a shape of `[("row", 5), ("column", 5)]` would mean the first 6 elements
* passed in the Vec would be for (0,0), (0,1), (0,2), (0,3), (0,4), (1,0) and so on to (4,4)
* for the 25th and final element.
*
* Returns the Err variant if
* - If the number of provided elements does not match the product of the dimension lengths.
* - If a dimension name is not unique
* - If any dimension has 0 elements
*
* Note that an empty list for dimensions is valid, and constructs a 0 dimensional tensor with
* a single element (since the product of an empty list is 1).
*/
pub fn try_from(
shape: [(Dimension, usize); D],
data: Vec<T>,
) -> Result<Self, InvalidShapeError<D>> {
let valid = InvalidShapeError::validate_dimensions(&shape, data.len());
if !valid {
return Err(InvalidShapeError::new(shape));
}
let strides = compute_strides(&shape);
Ok(Tensor {
data,
shape,
strides,
})
}
/// Unverified constructor for interal use when we know the dimensions/data/strides are
/// unchanged and don't need reverification
pub(crate) fn direct_from(
data: Vec<T>,
shape: [(Dimension, usize); D],
strides: [usize; D],
) -> Self {
Tensor {
data,
shape,
strides,
}
}
/// Unverified constructor for interal use when we know the dimensions/data/strides are
/// the same as the existing instance and don't need reverification
#[allow(dead_code)] // pretty sure something else will want this in the future
pub(crate) fn new_with_same_shape(&self, data: Vec<T>) -> Self {
Tensor {
data,
shape: self.shape,
strides: self.strides,
}
}
}
impl<T> Tensor<T, 0> {
/**
* Creates a 0 dimensional tensor from some scalar
*/
pub fn from_scalar(value: T) -> Tensor<T, 0> {
Tensor {
data: vec![value],
shape: [],
strides: [],
}
}
/**
* Returns the sole element of the 0 dimensional tensor.
*/
pub fn into_scalar(self) -> T {
self.data
.into_iter()
.next()
.expect("Tensors always have at least 1 element")
}
}
impl<T> Tensor<T, 0>
where
T: Clone,
{
/**
* Returns a copy of the sole element in the 0 dimensional tensor.
*/
pub fn scalar(&self) -> T {
self.data
.first()
.expect("Tensors always have at least 1 element")
.clone()
}
}
impl<T> From<T> for Tensor<T, 0> {
fn from(scalar: T) -> Tensor<T, 0> {
Tensor::from_scalar(scalar)
}
}
// TODO: See if we can find a way to write the reverse Tensor<T, 0> -> T conversion using From or Into (doesn't seem like we can?)
// # Safety
//
// We promise to never implement interior mutability for Tensor.
/**
* A Tensor implements TensorRef.
*/
unsafe impl<T, const D: usize> TensorRef<T, D> for Tensor<T, D> {
fn get_reference(&self, indexes: [usize; D]) -> Option<&T> {
let i = get_index_direct(&indexes, &self.strides, &self.shape)?;
self.data.get(i)
}
fn view_shape(&self) -> [(Dimension, usize); D] {
Tensor::shape(self)
}
unsafe fn get_reference_unchecked(&self, indexes: [usize; D]) -> &T {
// The point of get_reference_unchecked is no bounds checking, and therefore
// it does not make any sense to just use `unwrap` here. The trait documents that
// it's undefind behaviour to call this method with an out of bounds index, so we
// can assume the None case will never happen.
let i = get_index_direct(&indexes, &self.strides, &self.shape).unwrap_unchecked();
self.data.get_unchecked(i)
}
fn data_layout(&self) -> DataLayout<D> {
// We always have our memory in most significant to least
DataLayout::Linear(std::array::from_fn(|i| self.shape[i].0))
}
}
// # Safety
//
// We promise to never implement interior mutability for Tensor.
unsafe impl<T, const D: usize> TensorMut<T, D> for Tensor<T, D> {
fn get_reference_mut(&mut self, indexes: [usize; D]) -> Option<&mut T> {
let i = get_index_direct(&indexes, &self.strides, &self.shape)?;
self.data.get_mut(i)
}
unsafe fn get_reference_unchecked_mut(&mut self, indexes: [usize; D]) -> &mut T {
// The point of get_reference_unchecked_mut is no bounds checking, and therefore
// it does not make any sense to just use `unwrap` here. The trait documents that
// it's undefind behaviour to call this method with an out of bounds index, so we
// can assume the None case will never happen.
let i = get_index_direct(&indexes, &self.strides, &self.shape).unwrap_unchecked();
self.data.get_unchecked_mut(i)
}
}
/**
* Any tensor of a Cloneable type implements Clone.
*/
impl<T: Clone, const D: usize> Clone for Tensor<T, D> {
fn clone(&self) -> Self {
self.map(|element| element)
}
}
/**
* Any tensor of a Displayable type implements Display
*
* You can control the precision of the formatting using format arguments, i.e.
* `format!("{:.3}", tensor)`
*/
impl<T: std::fmt::Display, const D: usize> std::fmt::Display for Tensor<T, D> {
fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
crate::tensors::display::format_view(self, f)
}
}
/**
* Any 2 dimensional tensor can be converted to a matrix with rows equal to the length of the
* first dimension in the tensor, and columns equal to the length of the second.
*/
impl<T> From<Tensor<T, 2>> for crate::matrices::Matrix<T> {
fn from(tensor: Tensor<T, 2>) -> Self {
crate::matrices::Matrix::from_flat_row_major(
(tensor.shape[0].1, tensor.shape[1].1),
tensor.data,
)
}
}
pub(crate) fn compute_strides<const D: usize>(shape: &[(Dimension, usize); D]) -> [usize; D] {
std::array::from_fn(|d| shape.iter().skip(d + 1).map(|d| d.1).product())
}
/// returns the 1 dimensional index to use to get the requested index into some tensor
#[inline]
fn get_index_direct<const D: usize>(
// indexes to use
indexes: &[usize; D],
// strides for indexing into the tensor
strides: &[usize; D],
// shape of the tensor to index into
shape: &[(Dimension, usize); D],
) -> Option<usize> {
let mut index = 0;
for d in 0..D {
let n = indexes[d];
if n >= shape[d].1 {
return None;
}
index += n * strides[d];
}
Some(index)
}
/// returns the 1 dimensional index to use to get the requested index into some tensor, without
/// checking the indexes are within bounds for the shape.
#[inline]
fn get_index_direct_unchecked<const D: usize>(
// indexes to use
indexes: &[usize; D],
// strides for indexing into the tensor
strides: &[usize; D],
) -> usize {
let mut index = 0;
for d in 0..D {
let n = indexes[d];
index += n * strides[d];
}
index
}
impl<T, const D: usize> Tensor<T, D> {
pub fn view(&self) -> TensorView<T, &Tensor<T, D>, D> {
TensorView::from(self)
}
pub fn view_mut(&mut self) -> TensorView<T, &mut Tensor<T, D>, D> {
TensorView::from(self)
}
pub fn view_owned(self) -> TensorView<T, Tensor<T, D>, D> {
TensorView::from(self)
}
/**
* Returns a TensorAccess which can be indexed in the order of the supplied dimensions
* to read values from this tensor.
*
* # Panics
*
* If the set of dimensions supplied do not match the set of dimensions in this tensor's shape.
*/
#[track_caller]
pub fn index_by(&self, dimensions: [Dimension; D]) -> TensorAccess<T, &Tensor<T, D>, D> {
TensorAccess::from(self, dimensions)
}
/**
* Returns a TensorAccess which can be indexed in the order of the supplied dimensions
* to read or write values from this tensor.
*
* # Panics
*
* If the set of dimensions supplied do not match the set of dimensions in this tensor's shape.
*/
#[track_caller]
pub fn index_by_mut(
&mut self,
dimensions: [Dimension; D],
) -> TensorAccess<T, &mut Tensor<T, D>, D> {
TensorAccess::from(self, dimensions)
}
/**
* Returns a TensorAccess which can be indexed in the order of the supplied dimensions
* to read or write values from this tensor.
*
* # Panics
*
* If the set of dimensions supplied do not match the set of dimensions in this tensor's shape.
*/
#[track_caller]
pub fn index_by_owned(self, dimensions: [Dimension; D]) -> TensorAccess<T, Tensor<T, D>, D> {
TensorAccess::from(self, dimensions)
}
/**
* Creates a TensorAccess which will index into the dimensions this Tensor was created with
* in the same order as they were provided. See [TensorAccess::from_source_order].
*/
pub fn index(&self) -> TensorAccess<T, &Tensor<T, D>, D> {
TensorAccess::from_source_order(self)
}
/**
* Creates a TensorAccess which will index into the dimensions this Tensor was
* created with in the same order as they were provided. The TensorAccess mutably borrows
* the Tensor, and can therefore mutate it. See [TensorAccess::from_source_order].
*/
pub fn index_mut(&mut self) -> TensorAccess<T, &mut Tensor<T, D>, D> {
TensorAccess::from_source_order(self)
}
/**
* Creates a TensorAccess which will index into the dimensions this Tensor was
* created with in the same order as they were provided. The TensorAccess takes ownership
* of the Tensor, and can therefore mutate it. See [TensorAccess::from_source_order].
*/
pub fn index_owned(self) -> TensorAccess<T, Tensor<T, D>, D> {
TensorAccess::from_source_order(self)
}
/**
* Returns an iterator over references to the data in this Tensor.
*/
pub fn iter_reference(&self) -> TensorReferenceIterator<T, Tensor<T, D>, D> {
TensorReferenceIterator::from(self)
}
/**
* Returns an iterator over mutable references to the data in this Tensor.
*/
pub fn iter_reference_mut(&mut self) -> TensorReferenceMutIterator<T, Tensor<T, D>, D> {
TensorReferenceMutIterator::from(self)
}
/**
* Creates an iterator over the values in this Tensor.
*/
pub fn iter_owned(self) -> TensorOwnedIterator<T, Tensor<T, D>, D>
where
T: Default,
{
TensorOwnedIterator::from(self)
}
// Non public index order reference iterator since we don't want to expose our implementation
// details to public API since then we could never change them.
pub(crate) fn direct_iter_reference(&self) -> std::slice::Iter<T> {
self.data.iter()
}
/**
* Renames the dimension names of the tensor without changing the lengths of the dimensions
* in the tensor or moving any data around.
*
* ```
* use easy_ml::tensors::Tensor;
* let mut tensor = Tensor::from([("x", 2), ("y", 3)], vec![1, 2, 3, 4, 5, 6]);
* tensor.rename(["y", "z"]);
* assert_eq!([("y", 2), ("z", 3)], tensor.shape());
* ```
*
* # Panics
*
* - If a dimension name is not unique
*/
#[track_caller]
pub fn rename(&mut self, dimensions: [Dimension; D]) {
if crate::tensors::dimensions::has_duplicates_names(&dimensions) {
panic!("Dimension names must all be unique: {:?}", &dimensions);
}
#[allow(clippy::needless_range_loop)]
for d in 0..D {
self.shape[d].0 = dimensions[d];
}
}
/**
* Renames the dimension names of the tensor and returns it without changing the lengths
* of the dimensions in the tensor or moving any data around.
*
* ```
* use easy_ml::tensors::Tensor;
* let tensor = Tensor::from([("x", 2), ("y", 3)], vec![1, 2, 3, 4, 5, 6])
* .rename_owned(["y", "z"]);
* assert_eq!([("y", 2), ("z", 3)], tensor.shape());
* ```
*
* # Panics
*
* - If a dimension name is not unique
*/
#[track_caller]
pub fn rename_owned(mut self, dimensions: [Dimension; D]) -> Tensor<T, D> {
self.rename(dimensions);
self
}
/**
* Returns a TensorView with the dimension names of the shape renamed to the provided
* dimensions. The data of this tensor and the dimension lengths and order remain unchanged.
*
* This is a shorthand for constructing the TensorView from this Tensor.
*
* ```
* use easy_ml::tensors::Tensor;
* use easy_ml::tensors::views::{TensorView, TensorRename};
* let abc = Tensor::from([("a", 3), ("b", 3), ("c", 3)], (0..27).collect());
* let xyz = abc.rename_view(["x", "y", "z"]);
* let also_xyz = TensorView::from(TensorRename::from(&abc, ["x", "y", "z"]));
* assert_eq!(xyz, also_xyz);
* assert_eq!(xyz, Tensor::from([("x", 3), ("y", 3), ("z", 3)], (0..27).collect()));
* ```
*
* # Panics
*
* If a dimension name is not unique
*/
#[track_caller]
pub fn rename_view(
&self,
dimensions: [Dimension; D],
) -> TensorView<T, TensorRename<T, &Tensor<T, D>, D>, D> {
TensorView::from(TensorRename::from(self, dimensions))
}
/**
* Changes the shape of the tensor without changing the number of dimensions or moving any
* data around.
*
* # Panics
*
* - If the number of provided elements in the new shape does not match the product of the
* dimension lengths in the existing tensor's shape.
* - If a dimension name is not unique
* - If any dimension has 0 elements
*
* ```
* use easy_ml::tensors::Tensor;
* let mut tensor = Tensor::from([("width", 2), ("height", 2)], vec![
* 1, 2,
* 3, 4
* ]);
* tensor.reshape_mut([("batch", 1), ("image", 4)]);
* assert_eq!(tensor, Tensor::from([("batch", 1), ("image", 4)], vec![ 1, 2, 3, 4 ]));
* ```
*/
#[track_caller]
pub fn reshape_mut(&mut self, shape: [(Dimension, usize); D]) {
InvalidShapeError::validate_dimensions_or_panic(&shape, self.data.len());
let strides = compute_strides(&shape);
self.shape = shape;
self.strides = strides;
}
/**
* Consumes the tensor and changes the shape of the tensor without moving any
* data around. The new Tensor may also have a different number of dimensions.
*
* # Panics
*
* - If the number of provided elements in the new shape does not match the product of the
* dimension lengths in the existing tensor's shape.
* - If a dimension name is not unique
* - If any dimension has 0 elements
*
* ```
* use easy_ml::tensors::Tensor;
* let tensor = Tensor::from([("width", 2), ("height", 2)], vec![
* 1, 2,
* 3, 4
* ]);
* let flattened = tensor.reshape_owned([("image", 4)]);
* assert_eq!(flattened, Tensor::from([("image", 4)], vec![ 1, 2, 3, 4 ]));
* ```
*/
// TODO: View version
#[track_caller]
pub fn reshape_owned<const D2: usize>(self, shape: [(Dimension, usize); D2]) -> Tensor<T, D2> {
Tensor::from(shape, self.data)
}
/**
* Returns a TensorView with a range taken in P dimensions, hiding the values **outside** the
* range from view. Error cases are documented on [TensorRange].
*
* This is a shorthand for constructing the TensorView from this Tensor.
*
* ```
* use easy_ml::tensors::Tensor;
* use easy_ml::tensors::views::{TensorView, TensorRange, IndexRange};
* # use easy_ml::tensors::views::IndexRangeValidationError;
* # fn main() -> Result<(), IndexRangeValidationError<3, 2>> {
* let samples = Tensor::from([("batch", 5), ("x", 7), ("y", 7)], (0..(5 * 7 * 7)).collect());
* let cropped = samples.range([("x", IndexRange::new(1, 5)), ("y", IndexRange::new(1, 5))])?;
* let also_cropped = TensorView::from(
* TensorRange::from(&samples, [("x", 1..6), ("y", 1..6)])?
* );
* assert_eq!(cropped, also_cropped);
* assert_eq!(
* cropped.select([("batch", 0)]),
* Tensor::from([("x", 5), ("y", 5)], vec![
* 8, 9, 10, 11, 12,
* 15, 16, 17, 18, 19,
* 22, 23, 24, 25, 26,
* 29, 30, 31, 32, 33,
* 36, 37, 38, 39, 40
* ])
* );
* # Ok(())
* # }
* ```
*/
pub fn range<R, const P: usize>(
&self,
ranges: [(Dimension, R); P],
) -> Result<TensorView<T, TensorRange<T, &Tensor<T, D>, D>, D>, IndexRangeValidationError<D, P>>
where
R: Into<IndexRange>,
{
TensorRange::from(self, ranges).map(|range| TensorView::from(range))
}
/**
* Returns a TensorView with a range taken in P dimensions, hiding the values **outside** the
* range from view. Error cases are documented on [TensorRange]. The TensorRange
* mutably borrows this Tensor, and can therefore mutate it
*
* This is a shorthand for constructing the TensorView from this Tensor.
*/
pub fn range_mut<R, const P: usize>(
&mut self,
ranges: [(Dimension, R); P],
) -> Result<
TensorView<T, TensorRange<T, &mut Tensor<T, D>, D>, D>,
IndexRangeValidationError<D, P>,
>
where
R: Into<IndexRange>,
{
TensorRange::from(self, ranges).map(|range| TensorView::from(range))
}
/**
* Returns a TensorView with a range taken in P dimensions, hiding the values **outside** the
* range from view. Error cases are documented on [TensorRange]. The TensorRange
* takes ownership of this Tensor, and can therefore mutate it
*
* This is a shorthand for constructing the TensorView from this Tensor.
*/
pub fn range_owned<R, const P: usize>(
self,
ranges: [(Dimension, R); P],
) -> Result<TensorView<T, TensorRange<T, Tensor<T, D>, D>, D>, IndexRangeValidationError<D, P>>
where
R: Into<IndexRange>,
{
TensorRange::from(self, ranges).map(|range| TensorView::from(range))
}
/**
* Returns a TensorView with a mask taken in P dimensions, hiding the values **inside** the
* range from view. Error cases are documented on [TensorMask].
*
* This is a shorthand for constructing the TensorView from this Tensor.
*
* ```
* use easy_ml::tensors::Tensor;
* use easy_ml::tensors::views::{TensorView, TensorMask, IndexRange};
* # use easy_ml::tensors::views::IndexRangeValidationError;
* # fn main() -> Result<(), IndexRangeValidationError<3, 2>> {
* let samples = Tensor::from([("batch", 5), ("x", 7), ("y", 7)], (0..(5 * 7 * 7)).collect());
* let corners = samples.mask([("x", IndexRange::new(3, 2)), ("y", IndexRange::new(3, 2))])?;
* let also_corners = TensorView::from(
* TensorMask::from(&samples, [("x", 3..5), ("y", 3..5)])?
* );
* assert_eq!(corners, also_corners);
* assert_eq!(
* corners.select([("batch", 0)]),
* Tensor::from([("x", 5), ("y", 5)], vec![
* 0, 1, 2, 5, 6,
* 7, 8, 9, 12, 13,
* 14, 15, 16, 19, 20,
*
* 35, 36, 37, 40, 41,
* 42, 43, 44, 47, 48
* ])
* );
* # Ok(())
* # }
* ```
*/
pub fn mask<R, const P: usize>(
&self,
masks: [(Dimension, R); P],
) -> Result<TensorView<T, TensorMask<T, &Tensor<T, D>, D>, D>, IndexRangeValidationError<D, P>>
where
R: Into<IndexRange>,
{
TensorMask::from(self, masks).map(|mask| TensorView::from(mask))
}
/**
* Returns a TensorView with a mask taken in P dimensions, hiding the values **inside** the
* range from view. Error cases are documented on [TensorMask]. The TensorMask
* mutably borrows this Tensor, and can therefore mutate it
*
* This is a shorthand for constructing the TensorView from this Tensor.
*/
pub fn mask_mut<R, const P: usize>(
&mut self,
masks: [(Dimension, R); P],
) -> Result<
TensorView<T, TensorMask<T, &mut Tensor<T, D>, D>, D>,
IndexRangeValidationError<D, P>,
>
where
R: Into<IndexRange>,
{
TensorMask::from(self, masks).map(|mask| TensorView::from(mask))
}
/**
* Returns a TensorView with a mask taken in P dimensions, hiding the values **inside** the
* range from view. Error cases are documented on [TensorMask]. The TensorMask
* takes ownership of this Tensor, and can therefore mutate it
*
* This is a shorthand for constructing the TensorView from this Tensor.
*/
pub fn mask_owned<R, const P: usize>(
self,
masks: [(Dimension, R); P],
) -> Result<TensorView<T, TensorMask<T, Tensor<T, D>, D>, D>, IndexRangeValidationError<D, P>>
where
R: Into<IndexRange>,
{
TensorMask::from(self, masks).map(|mask| TensorView::from(mask))
}
/**
* Creates and returns a new tensor with all value pairs of two tensors with the same shape
* mapped by a function. The value pairs are not copied for you, if you're using `Copy` types
* or need to clone the values anyway, you can use
* [`Tensor::elementwise`](Tensor::elementwise) instead.
*
* # Generics
*
* This method can be called with any right hand side that can be converted to a TensorView,
* which includes `Tensor`, `&Tensor`, `&mut Tensor` as well as references to a `TensorView`.
*
* # Panics
*
* If the two tensors have different shapes.
*/
#[track_caller]
pub fn elementwise_reference<S, I, M>(&self, rhs: I, mapping_function: M) -> Tensor<T, D>
where
I: Into<TensorView<T, S, D>>,
S: TensorRef<T, D>,
M: Fn(&T, &T) -> T,
{
self.elementwise_reference_less_generic(rhs.into(), mapping_function)
}
/**
* Creates and returns a new tensor with all value pairs of two tensors with the same shape
* mapped by a function. The mapping function also receives each index corresponding to the
* value pairs. The value pairs are not copied for you, if you're using `Copy` types
* or need to clone the values anyway, you can use
* [`Tensor::elementwise_with_index`](Tensor::elementwise_with_index) instead.
*
* # Generics
*
* This method can be called with any right hand side that can be converted to a TensorView,
* which includes `Tensor`, `&Tensor`, `&mut Tensor` as well as references to a `TensorView`.
*
* # Panics
*
* If the two tensors have different shapes.
*/
#[track_caller]
pub fn elementwise_reference_with_index<S, I, M>(
&self,
rhs: I,
mapping_function: M,
) -> Tensor<T, D>
where
I: Into<TensorView<T, S, D>>,
S: TensorRef<T, D>,
M: Fn([usize; D], &T, &T) -> T,
{
self.elementwise_reference_less_generic_with_index(rhs.into(), mapping_function)
}
#[track_caller]
fn elementwise_reference_less_generic<S, M>(
&self,
rhs: TensorView<T, S, D>,
mapping_function: M,
) -> Tensor<T, D>
where
S: TensorRef<T, D>,
M: Fn(&T, &T) -> T,
{
let left_shape = self.shape();
let right_shape = rhs.shape();
if left_shape != right_shape {
panic!(
"Dimensions of left and right tensors are not the same: (left: {:?}, right: {:?})",
left_shape, right_shape
);
}
let mapped = self
.direct_iter_reference()
.zip(rhs.iter_reference())
.map(|(x, y)| mapping_function(x, y))
.collect();
// We're not changing the shape of the Tensor, so don't need to revalidate
Tensor::direct_from(mapped, self.shape, self.strides)
}
#[track_caller]
fn elementwise_reference_less_generic_with_index<S, M>(
&self,
rhs: TensorView<T, S, D>,
mapping_function: M,
) -> Tensor<T, D>
where
S: TensorRef<T, D>,
M: Fn([usize; D], &T, &T) -> T,
{
let left_shape = self.shape();
let right_shape = rhs.shape();
if left_shape != right_shape {
panic!(
"Dimensions of left and right tensors are not the same: (left: {:?}, right: {:?})",
left_shape, right_shape
);
}
// we just checked both shapes were the same, so we don't need to propagate indexes
// for both tensors because they'll be identical
let mapped = self
.direct_iter_reference()
.zip(rhs.iter_reference().with_index())
.map(|(x, (i, y))| mapping_function(i, x, y))
.collect();
// We're not changing the shape of the Tensor, so don't need to revalidate
Tensor::direct_from(mapped, self.shape, self.strides)
}
/**
* Returns a TensorView which makes the order of the data in this tensor appear to be in
* a different order. The order of the dimension names is unchanged, although their lengths
* may swap.
*
* This is a shorthand for constructing the TensorView from this Tensor.
*
* See also: [transpose](Tensor::transpose), [TensorTranspose]
*
* # Panics
*
* If the set of dimensions in the tensor does not match the set of dimensions provided. The
* order need not match.
*/
pub fn transpose_view(
&self,
dimensions: [Dimension; D],
) -> TensorView<T, TensorTranspose<T, &Tensor<T, D>, D>, D> {
TensorView::from(TensorTranspose::from(self, dimensions))
}
}
impl<T, const D: usize> Tensor<T, D>
where
T: Clone,
{
/**
* Creates a tensor with a particular number of dimensions and length in each dimension
* with all elements initialised to the provided value.
*
* # Panics
*
* - If a dimension name is not unique
* - If any dimension has 0 elements
*/
#[track_caller]
pub fn empty(shape: [(Dimension, usize); D], value: T) -> Self {
let elements = crate::tensors::dimensions::elements(&shape);
Tensor::from(shape, vec![value; elements])
}
/**
* Gets a copy of the first value in this tensor.
* For 0 dimensional tensors this is the only index `[]`, for 1 dimensional tensors this
* is `[0]`, for 2 dimensional tensors `[0,0]`, etcetera.
*/
pub fn first(&self) -> T {
self.data
.first()
.expect("Tensors always have at least 1 element")
.clone()
}
/**
* Returns a new Tensor which has the same data as this tensor, but with the order of data
* changed. The order of the dimension names is unchanged, although their lengths may swap.
*
* For example, with a `[("x", x), ("y", y)]` tensor you could call
* `transpose(["y", "x"])` which would return a new tensor with a shape of
* `[("x", y), ("y", x)]` where every (x,y) of its data corresponds to (y,x) in the original.
*
* This method need not shift *all* the dimensions though, you could also swap the width
* and height of images in a tensor with a shape of
* `[("batch", b), ("h", h), ("w", w), ("c", c)]` via `transpose(["batch", "w", "h", "c"])`
* which would return a new tensor where all the images have been swapped over the diagonal.
*
* See also: [TensorAccess], [reorder](Tensor::reorder)
*
* # Panics
*
* If the set of dimensions in the tensor does not match the set of dimensions provided. The
* order need not match (and if the order does match, this function is just an expensive
* clone).
*/
#[track_caller]
pub fn transpose(&self, dimensions: [Dimension; D]) -> Tensor<T, D> {
let shape = self.shape;
let mut reordered = self.reorder(dimensions);
// Transposition is essentially reordering, but we retain the dimension name ordering
// of the original order, this means we may swap dimension lengths, but the dimensions
// will not change order.
#[allow(clippy::needless_range_loop)]
for d in 0..D {
reordered.shape[d].0 = shape[d].0;
}
reordered
}
/**
* Modifies this tensor to have the same data as before, but with the order of data changed.
* The order of the dimension names is unchanged, although their lengths may swap.
*
* For example, with a `[("x", x), ("y", y)]` tensor you could call
* `transpose_mut(["y", "x"])` which would edit the tensor, updating its shape to
* `[("x", y), ("y", x)]`, so every (x,y) of its data corresponds to (y,x) before the
* transposition.
*
* The order swapping will try to be in place, but this is currently only supported for
* square tensors with 2 dimensions. Other types of tensors will not be transposed in place.
*
* # Panics
*
* If the set of dimensions in the tensor does not match the set of dimensions provided. The
* order need not match (and if the order does match, this function is just an expensive
* clone).
*/
#[track_caller]
pub fn transpose_mut(&mut self, dimensions: [Dimension; D]) {
let shape = self.shape;
self.reorder_mut(dimensions);
// Transposition is essentially reordering, but we retain the dimension name ordering
// we had before, this means we may swap dimension lengths, but the dimensions
// will not change order.
#[allow(clippy::needless_range_loop)]
for d in 0..D {
self.shape[d].0 = shape[d].0;
}
}
/**
* Returns a new Tensor which has the same data as this tensor, but with the order of the
* dimensions and corresponding order of data changed.
*
* For example, with a `[("x", x), ("y", y)]` tensor you could call
* `reorder(["y", "x"])` which would return a new tensor with a shape of
* `[("y", y), ("x", x)]` where every (y,x) of its data corresponds to (x,y) in the original.
*
* This method need not shift *all* the dimensions though, you could also swap the width
* and height of images in a tensor with a shape of
* `[("batch", b), ("h", h), ("w", w), ("c", c)]` via `reorder(["batch", "w", "h", "c"])`
* which would return a new tensor where every (b,w,h,c) of its data corresponds to (b,h,w,c)
* in the original.
*
* See also: [TensorAccess], [transpose](Tensor::transpose)
*
* # Panics
*
* If the set of dimensions in the tensor does not match the set of dimensions provided. The
* order need not match (and if the order does match, this function is just an expensive
* clone).
*/
#[track_caller]
pub fn reorder(&self, dimensions: [Dimension; D]) -> Tensor<T, D> {
let reorderd = match TensorAccess::try_from(&self, dimensions) {
Ok(reordered) => reordered,
Err(_error) => panic!(
"Dimension names provided {:?} must be the same set of dimension names in the tensor: {:?}",
dimensions,
&self.shape,
),
};
let reorderd_shape = reorderd.shape();
Tensor::from(reorderd_shape, reorderd.iter().collect())
}
/**
* Modifies this tensor to have the same data as before, but with the order of the
* dimensions and corresponding order of data changed.
*
* For example, with a `[("x", x), ("y", y)]` tensor you could call
* `reorder_mut(["y", "x"])` which would edit the tensor, updating its shape to
* `[("y", y), ("x", x)]`, so every (y,x) of its data corresponds to (x,y) before the
* transposition.
*
* The order swapping will try to be in place, but this is currently only supported for
* square tensors with 2 dimensions. Other types of tensors will not be reordered in place.
*
* # Panics
*
* If the set of dimensions in the tensor does not match the set of dimensions provided. The
* order need not match (and if the order does match, this function is just an expensive
* clone).
*/
#[track_caller]
pub fn reorder_mut(&mut self, dimensions: [Dimension; D]) {
use crate::tensors::dimensions::DimensionMappings;
if D == 2 && crate::tensors::dimensions::is_square(&self.shape) {
let dimension_mapping = match DimensionMappings::new(&self.shape, &dimensions) {
Some(dimension_mapping) => dimension_mapping,
None => panic!(
"Dimension names provided {:?} must be the same set of dimension names in the tensor: {:?}",
dimensions,
&self.shape,
),
};
let shape = dimension_mapping.map_shape_to_requested(&self.shape);
let shape_iterator = ShapeIterator::from(shape);
for index in shape_iterator {
let i = index[0];
let j = index[1];
if j >= i {
let mapped_index = dimension_mapping.map_dimensions_to_source(&index);
// Swap elements from the upper triangle (using index order of the actual tensor's
// shape)
let temp = self.get_reference(index).unwrap().clone();
// tensor[i,j] becomes tensor[mapping(i,j)]
*self.get_reference_mut(index).unwrap() =
self.get_reference(mapped_index).unwrap().clone();
// tensor[mapping(i,j)] becomes tensor[i,j]
*self.get_reference_mut(mapped_index).unwrap() = temp;
// If the mapping is a noop we've assigned i,j to i,j
// If the mapping is i,j -> j,i we've assigned i,j to j,i and j,i to i,j
}
}
// now update our shape and strides to match
self.shape = shape;
self.strides = compute_strides(&shape);
} else {
// fallback to allocating a new reordered tensor
let reordered = self.reorder(dimensions);
self.data = reordered.data;
self.shape = reordered.shape;
self.strides = reordered.strides;
}
}
/**
* Returns an iterator over copies of the data in this Tensor.
*/
pub fn iter(&self) -> TensorIterator<T, Tensor<T, D>, D> {
TensorIterator::from(self)
}
/**
* Creates and returns a new tensor with all values from the original with the
* function applied to each. This can be used to change the type of the tensor
* such as creating a mask:
* ```
* use easy_ml::tensors::Tensor;
* let x = Tensor::from([("a", 2), ("b", 2)], vec![
* 0.0, 1.2,
* 5.8, 6.9
* ]);
* let y = x.map(|element| element > 2.0);
* let result = Tensor::from([("a", 2), ("b", 2)], vec![
* false, false,
* true, true
* ]);
* assert_eq!(&y, &result);
* ```
*/
pub fn map<U>(&self, mapping_function: impl Fn(T) -> U) -> Tensor<U, D> {
let mapped = self
.data
.iter()
.map(|x| mapping_function(x.clone()))
.collect();
// We're not changing the shape of the Tensor, so don't need to revalidate
Tensor::direct_from(mapped, self.shape, self.strides)
}
/**
* Creates and returns a new tensor with all values from the original and
* the index of each value mapped by a function.
*/
pub fn map_with_index<U>(&self, mapping_function: impl Fn([usize; D], T) -> U) -> Tensor<U, D> {
let mapped = self
.iter()
.with_index()
.map(|(i, x)| mapping_function(i, x))
.collect();
// We're not changing the shape of the Tensor, so don't need to revalidate
Tensor::direct_from(mapped, self.shape, self.strides)
}
/**
* Applies a function to all values in the tensor, modifying
* the tensor in place.
*/
pub fn map_mut(&mut self, mapping_function: impl Fn(T) -> T) {
for value in self.data.iter_mut() {
*value = mapping_function(value.clone());
}
}
/**
* Applies a function to all values and each value's index in the tensor, modifying
* the tensor in place.
*/
pub fn map_mut_with_index(&mut self, mapping_function: impl Fn([usize; D], T) -> T) {
self.iter_reference_mut()
.with_index()
.for_each(|(i, x)| *x = mapping_function(i, x.clone()));
}
/**
* Creates and returns a new tensor with all value pairs of two tensors with the same shape
* mapped by a function.
*
* ```
* use easy_ml::tensors::Tensor;
* let lhs = Tensor::from([("a", 4)], vec![1, 2, 3, 4]);
* let rhs = Tensor::from([("a", 4)], vec![0, 1, 2, 3]);
* let multiplied = lhs.elementwise(&rhs, |l, r| l * r);
* assert_eq!(
* multiplied,
* Tensor::from([("a", 4)], vec![0, 2, 6, 12])
* );
* ```
*
* # Generics
*
* This method can be called with any right hand side that can be converted to a TensorView,
* which includes `Tensor`, `&Tensor`, `&mut Tensor` as well as references to a `TensorView`.
*
* # Panics
*
* If the two tensors have different shapes.
*/
#[track_caller]
pub fn elementwise<S, I, M>(&self, rhs: I, mapping_function: M) -> Tensor<T, D>
where
I: Into<TensorView<T, S, D>>,
S: TensorRef<T, D>,
M: Fn(T, T) -> T,
{
self.elementwise_reference_less_generic(rhs.into(), |lhs, rhs| {
mapping_function(lhs.clone(), rhs.clone())
})
}
/**
* Creates and returns a new tensor with all value pairs of two tensors with the same shape
* mapped by a function. The mapping function also receives each index corresponding to the
* value pairs.
*
* # Generics
*
* This method can be called with any right hand side that can be converted to a TensorView,
* which includes `Tensor`, `&Tensor`, `&mut Tensor` as well as references to a `TensorView`.
*
* # Panics
*
* If the two tensors have different shapes.
*/
#[track_caller]
pub fn elementwise_with_index<S, I, M>(&self, rhs: I, mapping_function: M) -> Tensor<T, D>
where
I: Into<TensorView<T, S, D>>,
S: TensorRef<T, D>,
M: Fn([usize; D], T, T) -> T,
{
self.elementwise_reference_less_generic_with_index(rhs.into(), |i, lhs, rhs| {
mapping_function(i, lhs.clone(), rhs.clone())
})
}
}
impl<T> Tensor<T, 1>
where
T: Numeric,
for<'a> &'a T: NumericRef<T>,
{
/**
* Computes the scalar product of two equal length vectors. For two vectors `[a,b,c]` and
* `[d,e,f]`, returns `a*d + b*e + c*f`. This is also known as the dot product.
*
* ```
* use easy_ml::tensors::Tensor;
* let tensor = Tensor::from([("sequence", 5)], vec![3, 4, 5, 6, 7]);
* assert_eq!(tensor.scalar_product(&tensor), 3*3 + 4*4 + 5*5 + 6*6 + 7*7);
* ```
*
* # Generics
*
* This method can be called with any right hand side that can be converted to a TensorView,
* which includes `Tensor`, `&Tensor`, `&mut Tensor` as well as references to a `TensorView`.
*
* # Panics
*
* If the two vectors are not of equal length or their dimension names do not match.
*/
// Would like this impl block to be in operations.rs too but then it would show first in the
// Tensor docs which isn't ideal
pub fn scalar_product<S, I>(&self, rhs: I) -> T
where
I: Into<TensorView<T, S, 1>>,
S: TensorRef<T, 1>,
{
self.scalar_product_less_generic(rhs.into())
}
}
impl<T> Tensor<T, 2>
where
T: Numeric,
for<'a> &'a T: NumericRef<T>,
{
/**
* Returns the determinant of this square matrix, or None if the matrix
* does not have a determinant. See [`linear_algebra`](super::linear_algebra::determinant_tensor())
*/
pub fn determinant(&self) -> Option<T> {
linear_algebra::determinant_tensor::<T, _, _>(self)
}
/**
* Computes the inverse of a matrix provided that it exists. To have an inverse a
* matrix must be square (same number of rows and columns) and it must also have a
* non zero determinant. See [`linear_algebra`](super::linear_algebra::inverse_tensor())
*/
pub fn inverse(&self) -> Option<Tensor<T, 2>> {
linear_algebra::inverse_tensor::<T, _, _>(self)
}
/**
* Computes the covariance matrix for this feature matrix along the specified feature
* dimension in this matrix. See [`linear_algebra`](crate::linear_algebra::covariance()).
*/
pub fn covariance(&self, feature_dimension: Dimension) -> Tensor<T, 2> {
linear_algebra::covariance::<T, _, _>(self, feature_dimension)
}
}
// FIXME: want this to be callable in the main numeric impl block
impl<T> Tensor<T, 2>
where
T: Numeric,
{
/**
* Creates a diagonal matrix of the provided size with the diagonal elements
* set to the provided value and all other elements in the tensor set to 0.
* A diagonal matrix is always square.
*
* The size is still taken as a shape to facilitate creating a diagonal matrix
* from the dimensionality of an existing one. If the provided value is 1 then
* this will create an identity matrix.
*
* A 3 x 3 identity matrix:
* ```ignore
* [
* 1, 0, 0
* 0, 1, 0
* 0, 0, 1
* ]
* ```
*
* # Panics
*
* - If the shape is not square.
* - If a dimension name is not unique
* - If any dimension has 0 elements
*/
#[track_caller]
pub fn diagonal(shape: [(Dimension, usize); 2], value: T) -> Tensor<T, 2> {
if !crate::tensors::dimensions::is_square(&shape) {
panic!("Shape must be square: {:?}", shape);
}
let mut tensor = Tensor::empty(shape, T::zero());
for ([r, c], x) in tensor.iter_reference_mut().with_index() {
if r == c {
*x = value.clone();
}
}
tensor
}
}
impl<T> Tensor<T, 2> {
/**
* Converts this 2 dimensional Tensor into a Matrix.
*
* This is a wrapper around the `From<Tensor<T, 2>>` implementation.
*
* The Matrix will have the data in the same order, with rows equal to the length of
* the first dimension in the tensor, and columns equal to the length of the second.
*/
pub fn into_matrix(self) -> crate::matrices::Matrix<T> {
self.into()
}
}
/**
* Methods for tensors with numerical real valued types, such as f32 or f64.
*
* This excludes signed and unsigned integers as they do not support decimal
* precision and hence can't be used for operations like square roots.
*
* Third party fixed precision and infinite precision decimal types should
* be able to implement all of the methods for [Real] and then utilise these functions.
*/
impl<T: Numeric + Real> Tensor<T, 1>
where
for<'a> &'a T: NumericRef<T> + RealRef<T>,
{
/**
* Computes the [L2 norm](https://en.wikipedia.org/wiki/Euclidean_vector#Length)
* of this vector, also referred to as the length or magnitude,
* and written as ||x||, or sometimes |x|.
*
* ||**a**|| = sqrt(a<sub>1</sub><sup>2</sup> + a<sub>2</sub><sup>2</sup> + a<sub>3</sub><sup>2</sup>...) = sqrt(**a**<sup>T</sup> * **a**)
*
* This is a shorthand for `(x.iter().map(|x| x * x).sum().sqrt()`, ie
* the square root of the dot product of a vector with itself.
*
* The euclidean length can be used to compute a
* [unit vector](https://en.wikipedia.org/wiki/Unit_vector), that is, a
* vector with length of 1. This should not be confused with a unit matrix,
* which is another name for an identity matrix.
*
* ```
* use easy_ml::tensors::Tensor;
* let a = Tensor::from([("data", 3)], vec![ 1.0, 2.0, 3.0 ]);
* let length = a.euclidean_length(); // (1^2 + 2^2 + 3^2)^0.5
* let unit = a.map(|x| x / length);
* assert_eq!(unit.euclidean_length(), 1.0);
* ```
*/
// TODO: Scalar ops for tensors
pub fn euclidean_length(&self) -> T {
self.direct_iter_reference()
.map(|x| x * x)
.sum::<T>()
.sqrt()
}
}
#[cfg(feature = "serde")]
mod serde_impls {
use crate::tensors::{Dimension, InvalidShapeError, Tensor};
use serde::Deserialize;
use std::convert::TryFrom;
/**
* Deserialised data for a Tensor. Can be converted into a Tensor by providing `&'static str`
* dimension names.
*/
#[derive(Deserialize)]
#[serde(rename = "Tensor")]
pub struct TensorDeserialize<'a, T, const D: usize> {
data: Vec<T>,
#[serde(with = "serde_arrays")]
#[serde(borrow)]
shape: [(&'a str, usize); D],
}
impl<'a, T, const D: usize> TensorDeserialize<'a, T, D> {
/**
* Converts this deserialised Tensor data to a Tensor, using the provided `&'static str`
* dimension names in place of what was serialised (which wouldn't necessarily live
* long enough).
*/
pub fn into_tensor(
self,
dimensions: [Dimension; D],
) -> Result<Tensor<T, D>, InvalidShapeError<D>> {
let shape = std::array::from_fn(|d| (dimensions[d], self.shape[d].1));
// Safety: Use the normal constructor that performs validation to prevent invalid
// serialized data being created as a Tensor, which would then break all the
// code that's relying on these invariants.
// By never serialising the strides in the first place, we reduce the possibility
// of creating invalid serialised represenations at the slight increase in
// serialisation work.
Tensor::try_from(shape, self.data)
}
}
/**
* Converts this deserialised Tensor data which has a static lifetime for the dimension
* names to a Tensor, using the serialised data.
*/
impl<T, const D: usize> TryFrom<TensorDeserialize<'static, T, D>> for Tensor<T, D> {
type Error = InvalidShapeError<D>;
fn try_from(value: TensorDeserialize<'static, T, D>) -> Result<Self, Self::Error> {
Tensor::try_from(value.shape, value.data)
}
}
}
#[cfg(feature = "serde")]
#[test]
fn test_serialize() {
fn assert_serialize<T: Serialize>() {}
assert_serialize::<Tensor<f64, 3>>();
assert_serialize::<Tensor<f64, 2>>();
assert_serialize::<Tensor<f64, 1>>();
assert_serialize::<Tensor<f64, 0>>();
}
#[cfg(feature = "serde")]
#[test]
fn test_deserialize() {
use serde::Deserialize;
fn assert_deserialize<'de, T: Deserialize<'de>>() {}
assert_deserialize::<TensorDeserialize<f64, 3>>();
assert_deserialize::<TensorDeserialize<f64, 2>>();
assert_deserialize::<TensorDeserialize<f64, 1>>();
assert_deserialize::<TensorDeserialize<f64, 0>>();
}
#[cfg(feature = "serde")]
#[test]
fn test_serialization_deserialization_loop() {
#[rustfmt::skip]
let tensor = Tensor::from(
[("rows", 3), ("columns", 4)],
vec![
1, 2, 3, 4,
5, 6, 7, 8,
9, 10, 11, 12
],
);
let encoded = toml::to_string(&tensor).unwrap();
assert_eq!(
encoded,
r#"data = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]
shape = [["rows", 3], ["columns", 4]]
"#,
);
let parsed: Result<TensorDeserialize<i32, 2>, _> = toml::from_str(&encoded);
assert!(parsed.is_ok());
let result = parsed.unwrap().into_tensor(["rows", "columns"]);
assert!(result.is_ok());
assert_eq!(result.unwrap(), tensor);
}
#[cfg(feature = "serde")]
#[test]
fn test_deserialization_validation() {
let parsed: Result<TensorDeserialize<i32, 2>, _> = toml::from_str(
r#"data = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]
shape = [["rows", 4], ["columns", 4]]
"#,
);
assert!(parsed.is_ok());
let result = parsed.unwrap().into_tensor(["rows", "columns"]);
assert!(result.is_err());
}
macro_rules! tensor_select_impl {
(impl Tensor $d:literal 1) => {
impl<T> Tensor<T, $d> {
/**
* Selects the provided dimension name and index pairs in this Tensor, returning a
* TensorView which has fewer dimensions than this Tensor, with the removed dimensions
* always indexed as the provided values.
*
* This is a shorthand for manually constructing the TensorView and
* [TensorIndex]
*
* Note: due to limitations in Rust's const generics support, this method is only
* implemented for `provided_indexes` of length 1 and `D` from 1 to 6. You can fall
* back to manual construction to create `TensorIndex`es with multiple provided
* indexes if you need to reduce dimensionality by more than 1 dimension at a time.
*/
#[track_caller]
pub fn select(
&self,
provided_indexes: [(Dimension, usize); 1],
) -> TensorView<T, TensorIndex<T, &Tensor<T, $d>, $d, 1>, { $d - 1 }> {
TensorView::from(TensorIndex::from(self, provided_indexes))
}
/**
* Selects the provided dimension name and index pairs in this Tensor, returning a
* TensorView which has fewer dimensions than this Tensor, with the removed dimensions
* always indexed as the provided values. The TensorIndex mutably borrows this
* Tensor, and can therefore mutate it
*
* See [select](Tensor::select)
*/
#[track_caller]
pub fn select_mut(
&mut self,
provided_indexes: [(Dimension, usize); 1],
) -> TensorView<T, TensorIndex<T, &mut Tensor<T, $d>, $d, 1>, { $d - 1 }> {
TensorView::from(TensorIndex::from(self, provided_indexes))
}
/**
* Selects the provided dimension name and index pairs in this Tensor, returning a
* TensorView which has fewer dimensions than this Tensor, with the removed dimensions
* always indexed as the provided values. The TensorIndex takes ownership of this
* Tensor, and can therefore mutate it
*
* See [select](Tensor::select)
*/
#[track_caller]
pub fn select_owned(
self,
provided_indexes: [(Dimension, usize); 1],
) -> TensorView<T, TensorIndex<T, Tensor<T, $d>, $d, 1>, { $d - 1 }> {
TensorView::from(TensorIndex::from(self, provided_indexes))
}
}
};
}
tensor_select_impl!(impl Tensor 6 1);
tensor_select_impl!(impl Tensor 5 1);
tensor_select_impl!(impl Tensor 4 1);
tensor_select_impl!(impl Tensor 3 1);
tensor_select_impl!(impl Tensor 2 1);
tensor_select_impl!(impl Tensor 1 1);
macro_rules! tensor_expand_impl {
(impl Tensor $d:literal 1) => {
impl<T> Tensor<T, $d> {
/**
* Expands the dimensionality of this tensor by adding dimensions of length 1 at
* a particular position within the shape, returning a TensorView which has more
* dimensions than this Tensor.
*
* This is a shorthand for manually constructing the TensorView and
* [TensorExpansion]
*
* Note: due to limitations in Rust's const generics support, this method is only
* implemented for `extra_dimension_names` of length 1 and `D` from 0 to 5. You can
* fall back to manual construction to create `TensorExpansion`s with multiple provided
* indexes if you need to increase dimensionality by more than 1 dimension at a time.
*/
#[track_caller]
pub fn expand(
&self,
extra_dimension_names: [(usize, Dimension); 1],
) -> TensorView<T, TensorExpansion<T, &Tensor<T, $d>, $d, 1>, { $d + 1 }> {
TensorView::from(TensorExpansion::from(self, extra_dimension_names))
}
/**
* Expands the dimensionality of this tensor by adding dimensions of length 1 at
* a particular position within the shape, returning a TensorView which has more
* dimensions than this Tensor. The TensorIndex mutably borrows this
* Tensor, and can therefore mutate it
*
* See [expand](Tensor::expand)
*/
#[track_caller]
pub fn expand_mut(
&mut self,
extra_dimension_names: [(usize, Dimension); 1],
) -> TensorView<T, TensorExpansion<T, &mut Tensor<T, $d>, $d, 1>, { $d + 1 }> {
TensorView::from(TensorExpansion::from(self, extra_dimension_names))
}
/**
* Expands the dimensionality of this tensor by adding dimensions of length 1 at
* a particular position within the shape, returning a TensorView which has more
* dimensions than this Tensor. The TensorIndex takes ownership of this
* Tensor, and can therefore mutate it
*
* See [expand](Tensor::expand)
*/
#[track_caller]
pub fn expand_owned(
self,
extra_dimension_names: [(usize, Dimension); 1],
) -> TensorView<T, TensorExpansion<T, Tensor<T, $d>, $d, 1>, { $d + 1 }> {
TensorView::from(TensorExpansion::from(self, extra_dimension_names))
}
}
};
}
tensor_expand_impl!(impl Tensor 0 1);
tensor_expand_impl!(impl Tensor 1 1);
tensor_expand_impl!(impl Tensor 2 1);
tensor_expand_impl!(impl Tensor 3 1);
tensor_expand_impl!(impl Tensor 4 1);
tensor_expand_impl!(impl Tensor 5 1);