File size: 126,007 Bytes
9375c9a |
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 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 |
// Copyright (C) 2015 Davis E. King ([email protected])
// License: Boost Software License See LICENSE.txt for the full license.
#undef DLIB_DNn_LAYERS_ABSTRACT_H_
#ifdef DLIB_DNn_LAYERS_ABSTRACT_H_
#include "../cuda/tensor_abstract.h"
#include "core_abstract.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
class SUBNET
{
/*!
WHAT THIS OBJECT REPRESENTS
This object represents a deep neural network. In particular, it is
the simplified interface through which layer objects interact with their
subnetworks. A layer's two important tasks are to (1) take outputs from its
subnetwork and forward propagate them through itself and (2) to backwards
propagate an error gradient through itself and onto its subnetwork.
The idea of a subnetwork is illustrated in the following diagram:
+---------------------------------------------------------+
| loss <-- layer1 <-- layer2 <-- ... <-- layern <-- input |
+---------------------------------------------------------+
^ ^
\__ subnetwork for layer1 __/
Therefore, by "subnetwork" we mean the part of the network closer to the
input.
Note that there is no dlib::SUBNET type. It is shown here purely to
document the interface layer objects expect to see when they interact
with a network.
!*/
public:
// You aren't allowed to copy subnetworks from inside a layer.
SUBNET(const SUBNET&) = delete;
SUBNET& operator=(const SUBNET&) = delete;
const tensor& get_output(
) const;
/*!
ensures
- returns the output of this subnetwork. This is the data that the next
layer in the network will take as input.
- have_same_dimensions(#get_gradient_input(), get_output()) == true
!*/
tensor& get_gradient_input(
);
/*!
ensures
- returns the error gradient for this subnetwork. That is, this is the
error gradient that this network will use to update itself. Therefore,
when performing back propagation, layers that sit on top of this
subnetwork write their back propagated error gradients into
get_gradient_input(). Or to put it another way, during back propagation,
layers take the contents of their get_gradient_input() and back propagate
it through themselves and store the results into their subnetwork's
get_gradient_input().
!*/
const NEXT_SUBNET& subnet(
) const;
/*!
ensures
- returns the subnetwork of *this network. With respect to the diagram
above, if *this was layer1 then subnet() would return the network that
begins with layer2.
!*/
NEXT_SUBNET& subnet(
);
/*!
ensures
- returns the subnetwork of *this network. With respect to the diagram
above, if *this was layer1 then subnet() would return the network that
begins with layer2.
!*/
const layer_details_type& layer_details(
) const;
/*!
ensures
- returns the layer_details_type instance that defines the behavior of the
layer at the top of this network. I.e. returns the layer details that
defines the behavior of the layer nearest to the network output rather
than the input layer. For computational layers, this is the object
implementing the EXAMPLE_COMPUTATIONAL_LAYER_ interface that defines the
layer's behavior.
!*/
unsigned int sample_expansion_factor (
) const;
/*!
ensures
- When to_tensor() is invoked on this network's input layer it converts N
input objects into M samples, all stored inside a resizable_tensor. It
is always the case that M is some integer multiple of N.
sample_expansion_factor() returns the value of this multiplier. To be
very specific, it is always true that M==I*N where I is some integer.
This integer I is what is returned by sample_expansion_factor().
It should be noted that computational layers likely do not care about the
sample expansion factor. It is only really of concern inside a loss
layer where you need to know its value so that tensor samples can be
matched against truth objects. Moreover, in most cases the sample
expansion factor is 1.
!*/
};
// ----------------------------------------------------------------------------------------
class EXAMPLE_COMPUTATIONAL_LAYER_
{
/*!
WHAT THIS OBJECT REPRESENTS
Each computational layer in a deep neural network can be thought of as a
function, f(data,parameters), that takes in a data tensor, some parameters,
and produces an output tensor. You create an entire deep network by
composing these functions. Importantly, you are able to use a wide range
of different functions to accommodate the task you are trying to
accomplish. Therefore, dlib includes a number of common layer types but if
you want to define your own then you simply implement a class with the same
interface as EXAMPLE_COMPUTATIONAL_LAYER_.
Note that there is no dlib::EXAMPLE_COMPUTATIONAL_LAYER_ type. It is shown
here purely to document the interface that a layer object must implement.
The central work of defining a layer is implementing the forward and backward
methods. When you do this you have four options:
- Implement the forward() and backward() methods according to the
specification shown below. Do not implement forward_inplace() and
backward_inplace().
- Implement the forward() and backward() methods according to the
specification shown below, except exclude the computed_output
parameter from backward(). Doing this will allow dlib to make some
layers execute in-place and therefore run a little faster and use
less memory. Do not implement forward_inplace() and
backward_inplace().
- Implement the forward_inplace() and backward_inplace() methods
according to the specification shown below. Do not implement
forward() and backward(). These in-place methods allow some types of
layers to be implemented more efficiently.
- Implement the forward_inplace() and backward_inplace() methods
according to the specification shown below, except exclude the
computed_output parameter from backward_inplace(). Doing this will
allow dlib to make some layers execute in-place and therefore run a
little faster and use less memory. Do not implement forward() and
backward().
It should also be noted that layers may define additional layer specific
fields and the solvers can use these fields as they see fit. For example,
some layers define get_learning_rate_multiplier() and
get_weight_decay_multiplier() methods. The solvers that come with dlib
look at these methods, if they exist, and adjust the learning rate or
weight decay for that layer according to the multiplier. Therefore, you
can add these methods to your layer types if you want, or even define new
fields and new solvers that use those fields in some way.
!*/
public:
EXAMPLE_COMPUTATIONAL_LAYER_(
);
/*!
ensures
- Default constructs this object. This function is not required to do
anything in particular but it must exist, that is, it is required that
layer objects be default constructable.
!*/
EXAMPLE_COMPUTATIONAL_LAYER_ (
const EXAMPLE_COMPUTATIONAL_LAYER_& item
);
/*!
ensures
- EXAMPLE_COMPUTATIONAL_LAYER_ objects are copy constructable
!*/
EXAMPLE_COMPUTATIONAL_LAYER_(
const some_other_layer_type& item
);
/*!
ensures
- Constructs this object from item. This form of constructor is optional
but it allows you to provide a conversion from one layer type to another.
For example, the following code is valid only if my_layer2 can be
constructed from my_layer1:
relu<fc<my_layer1<fc<input<matrix<float>>>>>> my_dnn1;
relu<fc<my_layer2<fc<input<matrix<float>>>>>> my_dnn2(my_dnn1);
This kind of pattern is useful if you want to use one type of layer
during training but a different type of layer during testing since it
allows you to easily convert between related deep neural network types.
Additionally, if you provide a constructor to build a layer from another
layer type you should also write your layer's deserialize() routine such
that it can read that other layer's serialized data in addition to your
own serialized data.
!*/
template <typename SUBNET>
void setup (
const SUBNET& sub
);
/*!
requires
- SUBNET implements the SUBNET interface defined at the top of this file.
ensures
- performs any necessary initial memory allocations and/or sets parameters
to their initial values prior to learning. Therefore, calling setup
destroys any previously learned parameters. Also, typically setup()
would look at the dimensions of the outputs of sub and configure the
number of parameters in *this accordingly.
!*/
template <typename SUBNET>
void forward(
const SUBNET& sub,
resizable_tensor& data_output
);
/*!
requires
- SUBNET implements the SUBNET interface defined at the top of this file.
- setup() has been called.
ensures
- Runs the output of the subnetwork through this layer and stores the
results into #data_output. In particular, forward() can use any of the
outputs in sub (e.g. sub.get_output(), sub.subnet().get_output(), etc.)
to compute whatever it wants.
!*/
template <typename SUBNET>
void backward(
const tensor& computed_output, // this parameter is optional
const tensor& gradient_input,
SUBNET& sub,
tensor& params_grad
);
/*!
requires
- SUBNET implements the SUBNET interface defined at the top of this file.
- setup() has been called.
- computed_output is the tensor resulting from calling forward(sub,computed_output).
Moreover, this was the most recent call to forward(). This means that
forward() is allowed to cache intermediate results so they can be used
during the backward computation.
- have_same_dimensions(gradient_input, computed_output) == true
- have_same_dimensions(sub.get_gradient_input(), sub.get_output()) == true
- have_same_dimensions(params_grad, get_layer_params()) == true
ensures
- This function outputs the gradients of this layer with respect to the
input data from sub and also with respect to this layer's parameters.
These gradients are stored into #sub and #params_grad, respectively. To be
precise, the gradients are taken of a function f(sub,get_layer_params())
which is defined thusly:
- Recalling that computed_output is a function of both sub and get_layer_params(),
since it is the result of calling forward(sub,computed_output):
let f(sub,get_layer_params()) == dot(computed_output, gradient_input)
Then we define the following gradient vectors:
- PARAMETER_GRADIENT == gradient of f(sub,get_layer_params()) with
respect to get_layer_params().
- for all valid I:
- DATA_GRADIENT_I == gradient of f(sub,get_layer_params()) with
respect to layer<I>(sub).get_output() (recall that forward() can
draw inputs from the immediate sub layer, sub.subnet(), or
any earlier layer. So you must consider the gradients with
respect to all inputs drawn from sub)
Finally, backward() outputs these gradients by performing:
- params_grad = PARAMETER_GRADIENT
- for all valid I:
- layer<I>(sub).get_gradient_input() += DATA_GRADIENT_I
!*/
void forward_inplace(
const tensor& data_input,
tensor& data_output
);
/*!
requires
- have_same_dimensions(data_input,data_output) == true
- setup() has been called.
ensures
- Runs the data_input tensor through this layer and stores the output into
#data_output.
- This function supports in-place operation, i.e. having
is_same_object(data_input, data_output)==true
!*/
void backward_inplace(
const tensor& computed_output, // this parameter is optional
const tensor& gradient_input,
tensor& data_grad,
tensor& params_grad
);
/*!
requires
- setup() has been called.
- computed_output is the tensor resulting from the most recent call to
forward_inplace(). This means that forward_inplace() is allowed to cache
intermediate results so they can be used during the backward computation.
- have_same_dimensions(gradient_input, data_grad) == true
- have_same_dimensions(gradient_input, computed_output) == true
- have_same_dimensions(params_grad, get_layer_params()) == true
ensures
- This function supports in-place operation, i.e. having
is_same_object(gradient_input, data_grad)==true
- This function outputs the gradients of this layer with respect to the
input data from a sublayer and also with respect to this layer's parameters.
These gradients are stored into #data_grad and #params_grad, respectively. To be
precise, the gradients are taken of a function f(data_input,get_layer_params())
which is defined thusly:
- Recalling that computed_output is a function of both the input to
forward_inplace() and get_layer_params(), since it is the result of
calling forward_inplace(data_input,computed_output):
let f(data_input,get_layer_params()) == dot(computed_output, gradient_input)
Then we define the following gradient vectors:
- PARAMETER_GRADIENT == gradient of f(data_input,get_layer_params()) with
respect to get_layer_params().
- DATA_GRADIENT == gradient of f(data_input,get_layer_params()) with respect
to data_input.
Finally, backward_inplace() outputs these gradients by performing:
- params_grad = PARAMETER_GRADIENT
- if (is_same_object(gradient_input, data_grad)) then
- data_grad = DATA_GRADIENT
- else
- data_grad += DATA_GRADIENT
!*/
const tensor& get_layer_params(
) const;
/*!
ensures
- returns the parameters that define the behavior of forward().
!*/
tensor& get_layer_params(
);
/*!
ensures
- returns the parameters that define the behavior of forward().
!*/
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
/*!
These two functions are optional. If provided, they should map between
(column,row) coordinates in input and output tensors of forward(). Providing
these functions allows you to use global utility functions like
input_tensor_to_output_tensor().
!*/
void clean (
);
/*!
Implementing this function is optional. If you don't need it then you don't
have to provide a clean(). But if you do provide it then it must behave as
follows:
ensures
- calling clean() causes this object to forget about everything except its
parameters. This is useful if your layer caches information between
forward and backward passes and you want to clean out that cache
information before saving the network to disk.
!*/
};
std::ostream& operator<<(std::ostream& out, const EXAMPLE_COMPUTATIONAL_LAYER_& item);
/*!
print a string describing this layer.
!*/
void to_xml(const EXAMPLE_COMPUTATIONAL_LAYER_& item, std::ostream& out);
/*!
This function is optional, but required if you want to print your networks with
net_to_xml(). Therefore, to_xml() prints a layer as XML.
!*/
void serialize(const EXAMPLE_COMPUTATIONAL_LAYER_& item, std::ostream& out);
void deserialize(EXAMPLE_COMPUTATIONAL_LAYER_& item, std::istream& in);
/*!
provides serialization support
!*/
// For each layer you define, always define an add_layer template so that layers can be
// easily composed. Moreover, the convention is that the layer class ends with an _
// while the add_layer template has the same name but without the trailing _.
template <typename SUBNET>
using EXAMPLE_COMPUTATIONAL_LAYER = add_layer<EXAMPLE_COMPUTATIONAL_LAYER_, SUBNET>;
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
enum fc_bias_mode
{
FC_HAS_BIAS = 0,
FC_NO_BIAS = 1
};
struct num_fc_outputs
{
num_fc_outputs(unsigned long n) : num_outputs(n) {}
unsigned long num_outputs;
};
template <
unsigned long num_outputs,
fc_bias_mode bias_mode
>
class fc_
{
/*!
REQUIREMENTS ON num_outputs
num_outputs > 0
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a fully connected layer that
takes an input tensor and multiplies it by a weight matrix and outputs the
results.
The dimensions of the tensors output by this layer are as follows (letting
IN be the input tensor and OUT the output tensor):
- OUT.num_samples() == IN.num_samples()
- OUT.k() == get_num_outputs()
- OUT.nr() == 1
- OUT.nc() == 1
!*/
public:
fc_(
);
/*!
ensures
- #get_num_outputs() == num_outputs
- #get_bias_mode() == bias_mode
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 1
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 0
!*/
fc_(
num_fc_outputs o
);
/*!
ensures
- #get_num_outputs() == o.num_outputs
- #get_bias_mode() == bias_mode
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 1
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 0
!*/
unsigned long get_num_outputs (
) const;
/*!
ensures
- This layer outputs column vectors that contain get_num_outputs()
elements. That is, the output tensor T from forward() will be such that:
- T.num_samples() == however many samples were given to forward().
- T.k() == get_num_outputs()
- The rest of the dimensions of T will be 1.
!*/
void set_num_outputs(
long num
);
/*!
requires
- num > 0
- get_layer_params().size() == 0 || get_num_outputs() == num
(i.e. You can't change the number of outputs in fc_ if the parameter
tensor has already been allocated.)
ensures
- #get_num_outputs() == num
!*/
fc_bias_mode get_bias_mode (
) const;
/*!
ensures
- returns the bias mode which determines if this layer includes bias terms.
That is, if the bias mode is FC_HAS_BIAS then a different constant scalar
is added to each of the outputs of this layer.
!*/
double get_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its parameters be
multiplied by get_learning_rate_multiplier().
!*/
double get_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its parameters be
multiplied by get_weight_decay_multiplier().
!*/
void set_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_learning_rate_multiplier() == val
!*/
void set_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_weight_decay_multiplier() == val
!*/
double get_bias_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its bias parameters be
multiplied by get_learning_rate_multiplier()*get_bias_learning_rate_multiplier().
!*/
double get_bias_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its bias parameters be
multiplied by get_weight_decay_multiplier()*get_bias_weight_decay_multiplier().
!*/
void set_bias_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_learning_rate_multiplier() == val
!*/
void set_bias_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_weight_decay_multiplier() == val
!*/
void disable_bias(
);
/*!
ensures
- bias_is_disabled() returns true
!*/
bool bias_is_disabled(
) const;
/*!
ensures
- returns true if bias learning is disabled for this layer. This means the biases will
not be learned during the training and they will not be used in the forward or backward
methods either.
!*/
alias_tensor_const_instance get_weights(
) const;
/*!
ensures
- returns an alias of get_layer_params(), containing the weights matrix of
the fully connected layer.
- #get_weights().num_samples() is the number of elements in input sample,
i.e. sublayer's output's k * nc * nr.
- #get_bias().k() == #get_num_outputs()
- if get_bias_mode() == FC_HAS_BIAS:
- #get_layer_params().size() == (#get_weights().size() + #get_biases().size())
- else:
- #get_layer_params().size() == #get_weights().size()
!*/
alias_tensor_instance get_weights(
);
/*!
ensures
- returns an alias of get_layer_params(), containing the weights matrix of
the fully connected layer.
- #get_weights().num_samples() is the number of elements in input sample,
i.e. sublayer's output's k * nc * nr.
- #get_bias().k() == #get_num_outputs()
- if get_bias_mode() == FC_HAS_BIAS:
- #get_layer_params().size() == (#get_weights().size() + #get_biases().size())
- else:
- #get_layer_params().size() == #get_weights().size()
!*/
alias_tensor_const_instance get_biases(
) const;
/*!
requires
- #get_bias_mode() == FC_HAS_BIAS
ensures
- returns an alias of get_layer_params(), containing the bias vector of
the fully connected layer.
- #get_bias().num_samples() == 1
- #get_bias().k() == #get_num_outputs()
- #get_layer_params().size() == (#get_weights().size() + #get_biases().size())
!*/
alias_tensor_instance get_biases(
);
/*!
requires
- #get_bias_mode() == FC_HAS_BIAS
ensures
- returns an alias of get_layer_params(), containing the bias vector of
the fully connected layer.
- #get_bias().num_samples() == 1
- #get_bias().k() == #get_num_outputs()
- #get_layer_params().size() == (#get_weights().size() + #get_biases().size())
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
unsigned long num_outputs,
typename SUBNET
>
using fc = add_layer<fc_<num_outputs,FC_HAS_BIAS>, SUBNET>;
template <
unsigned long num_outputs,
typename SUBNET
>
using fc_no_bias = add_layer<fc_<num_outputs,FC_NO_BIAS>, SUBNET>;
// ----------------------------------------------------------------------------------------
struct num_con_outputs
{
num_con_outputs(unsigned long n) : num_outputs(n) {}
unsigned long num_outputs;
};
template <
long _num_filters,
long _nr,
long _nc,
int _stride_y,
int _stride_x,
int _padding_y = _stride_y!=1? 0 : _nr/2,
int _padding_x = _stride_x!=1? 0 : _nc/2
>
class con_
{
/*!
REQUIREMENTS ON TEMPLATE ARGUMENTS
- _num_filters > 0
- _nr >= 0
- _nc >= 0
- _stride_y > 0
- _stride_x > 0
- _padding_y >= 0
- _padding_x >= 0
- Also, we require that:
- if (_nr == 0) then
- _padding_y == 0
- else
- _padding_y < _nr
- if (_nc == 0) then
- _padding_x == 0
- else
- _padding_x < _nc
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a convolution layer that takes an
input tensor (nominally representing an image) and convolves it with a set
of filters and then outputs the results.
The dimensions of the tensors output by this layer are as follows (letting
IN be the input tensor and OUT the output tensor):
- OUT.num_samples() == IN.num_samples()
- OUT.k() == num_filters()
- OUT.nr() == 1+(IN.nr() + 2*padding_y() - nr())/stride_y()
- OUT.nc() == 1+(IN.nc() + 2*padding_x() - nc())/stride_x()
Note also that setting _nr or _nc to 0 has a special meaning of "set the
filter size equal to the input image size". Specifically, it means:
- if (_nr == 0) then
- nr() == IN.nr()
- OUT.nr() == 1
- if (_nc == 0) then
- nc() == IN.nc()
- OUT.nc() == 1
!*/
public:
con_(
);
/*!
ensures
- #num_filters() == _num_filters
- #nr() == _nr
- #nc() == _nc
- #stride_y() == _stride_y
- #stride_x() == _stride_x
- #padding_y() == _padding_y
- #padding_x() == _padding_x
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 1
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 0
!*/
con_(
num_con_outputs o
);
/*!
ensures
- #num_filters() == o.num_outputs
- #nr() == _nr
- #nc() == _nc
- #stride_y() == _stride_y
- #stride_x() == _stride_x
- #padding_y() == _padding_y
- #padding_x() == _padding_x
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 1
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 0
!*/
long num_filters(
) const;
/*!
ensures
- returns the number of filters contained in this layer. The k dimension
of the output tensors produced by this layer will be equal to the number
of filters.
!*/
void set_num_filters(
long num
);
/*!
requires
- num > 0
- get_layer_params().size() == 0 || num_filters() == num
(i.e. You can't change the number of filters in con_ if the parameter
tensor has already been allocated.)
ensures
- #num_filters() == num
!*/
long nr(
) const;
/*!
ensures
- returns the number of rows in the filters in this layer. Note that if
nr()==0 then it means the size of the filter is not yet assigned, but
once setup() is called nr() will be set to the input tensor's nr().
Therefore, nr()==0 has the special interpretation of "be the same size as
the input tensor".
!*/
long nc(
) const;
/*!
ensures
- returns the number of columns in the filters in this layer. Note that if
nc()==0 then it means the size of the filter is not yet assigned, but
once setup() is called nc() will be set to the input tensor's nc().
Therefore, nc()==0 has the special interpretation of "be the same size as
the input tensor".
!*/
long stride_y(
) const;
/*!
ensures
- returns the vertical stride used when convolving the filters over an
image. That is, each filter will be moved stride_y() pixels down at a
time when it moves over the image.
!*/
long stride_x(
) const;
/*!
ensures
- returns the horizontal stride used when convolving the filters over an
image. That is, each filter will be moved stride_x() pixels right at a
time when it moves over the image.
!*/
long padding_y(
) const;
/*!
ensures
- returns the number of pixels of zero padding added to the top and bottom
sides of the image.
!*/
long padding_x(
) const;
/*!
ensures
- returns the number of pixels of zero padding added to the left and right
sides of the image.
!*/
double get_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its parameters be
multiplied by get_learning_rate_multiplier().
!*/
double get_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its parameters be
multiplied by get_weight_decay_multiplier().
!*/
void set_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_learning_rate_multiplier() == val
!*/
void set_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_weight_decay_multiplier() == val
!*/
double get_bias_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its bias parameters be
multiplied by get_learning_rate_multiplier()*get_bias_learning_rate_multiplier().
!*/
double get_bias_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its bias parameters be
multiplied by get_weight_decay_multiplier()*get_bias_weight_decay_multiplier().
!*/
void set_bias_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_learning_rate_multiplier() == val
!*/
void set_bias_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_weight_decay_multiplier() == val
!*/
void disable_bias(
);
/*!
ensures
- bias_is_disabled() returns true
!*/
bool bias_is_disabled(
) const;
/*!
ensures
- returns true if bias learning is disabled for this layer. This means the biases will
not be learned during the training and they will not be used in the forward or backward
methods either.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
long num_filters,
long nr,
long nc,
int stride_y,
int stride_x,
typename SUBNET
>
using con = add_layer<con_<num_filters,nr,nc,stride_y,stride_x>, SUBNET>;
// ----------------------------------------------------------------------------------------
template <
long _num_filters,
long _nr,
long _nc,
int _stride_y,
int _stride_x,
int _padding_y = _stride_y!=1? 0 : _nr/2,
int _padding_x = _stride_x!=1? 0 : _nc/2
>
class cont_
{
/*!
REQUIREMENTS ON TEMPLATE ARGUMENTS
All of them must be > 0.
Also, we require that:
- 0 <= _padding_y && _padding_y < _nr
- 0 <= _padding_x && _padding_x < _nc
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a transposed convolution layer
that takes an input tensor and transpose convolves (sometimes called
"deconvolution") it with a set of filters and then outputs the results.
This is essentially a convolutional layer that allows fractional strides.
Therefore, you can make output tensors that are larger than the input
tensors using this layer type.
The dimensions of the tensors output by this layer are as follows (letting
IN be the input tensor and OUT the output tensor):
- OUT.num_samples() == IN.num_samples()
- OUT.k() == num_filters()
- OUT.nr() == stride_y()*(IN.nr()-1) + nr() - 2*padding_y()
- OUT.nc() == stride_x()*(IN.nc()-1) + nc() - 2*padding_x()
!*/
public:
cont_(
);
/*!
ensures
- #num_filters() == _num_filters
- #nr() == _nr
- #nc() == _nc
- #stride_y() == _stride_y
- #stride_x() == _stride_x
- #padding_y() == _padding_y
- #padding_x() == _padding_x
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 1
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 0
!*/
cont_(
num_con_outputs o
);
/*!
ensures
- #num_filters() == o.num_outputs
- #nr() == _nr
- #nc() == _nc
- #stride_y() == _stride_y
- #stride_x() == _stride_x
- #padding_y() == _padding_y
- #padding_x() == _padding_x
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 1
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 0
!*/
long num_filters(
) const;
/*!
ensures
- returns the number of filters contained in this layer. The k dimension
of the output tensors produced by this layer will be equal to the number
of filters.
!*/
void set_num_filters(
long num
);
/*!
requires
- num > 0
- get_layer_params().size() == 0 || num_filters() == num
(i.e. You can't change the number of filters in cont_ if the parameter
tensor has already been allocated.)
ensures
- #num_filters() == num
!*/
long nr(
) const;
/*!
ensures
- returns the number of rows in the filters in this layer.
!*/
long nc(
) const;
/*!
ensures
- returns the number of columns in the filters in this layer.
!*/
long stride_y(
) const;
/*!
ensures
- returns the vertical stride used when convolving the filters over an
image. That is, each filter will be moved 1.0/stride_y() pixels down at
a time when it moves over the image.
!*/
long stride_x(
) const;
/*!
ensures
- returns the horizontal stride used when convolving the filters over an
image. That is, each filter will be moved 1.0/stride_x() pixels right at
a time when it moves over the image.
!*/
long padding_y(
) const;
/*!
ensures
- returns the number of pixels of zero padding added to the top and bottom
sides of the image.
!*/
long padding_x(
) const;
/*!
ensures
- returns the number of pixels of zero padding added to the left and right
sides of the image.
!*/
double get_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its parameters be
multiplied by get_learning_rate_multiplier().
!*/
double get_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its parameters be
multiplied by get_weight_decay_multiplier().
!*/
void set_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_learning_rate_multiplier() == val
!*/
void set_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_weight_decay_multiplier() == val
!*/
double get_bias_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its bias parameters be
multiplied by get_learning_rate_multiplier()*get_bias_learning_rate_multiplier().
!*/
double get_bias_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its bias parameters be
multiplied by get_weight_decay_multiplier()*get_bias_weight_decay_multiplier().
!*/
void set_bias_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_learning_rate_multiplier() == val
!*/
void set_bias_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_weight_decay_multiplier() == val
!*/
void disable_bias(
);
/*!
ensures
- bias_is_disabled() returns true
!*/
bool bias_is_disabled(
) const;
/*!
ensures
- returns true if bias learning is disabled for this layer. This means the biases will
not be learned during the training and they will not be used in the forward or backward
methods either.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
long num_filters,
long nr,
long nc,
int stride_y,
int stride_x,
typename SUBNET
>
using cont = add_layer<cont_<num_filters,nr,nc,stride_y,stride_x>, SUBNET>;
// ----------------------------------------------------------------------------------------
template <
int scale_y,
int scale_x
>
class upsample_
{
/*!
REQUIREMENTS ON TEMPLATE ARGUMENTS
All of them must be >= 1.
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it allows you to upsample a layer using
bilinear interpolation. To be very specific, it upsamples each of the
channels in an input tensor. Therefore, if IN is the input tensor to this
layer and OUT the output tensor, then we will have:
- OUT.num_samples() == IN.num_samples()
- OUT.k() == IN.k()
- OUT.nr() == IN.nr()*scale_y
- OUT.nc() == IN.nc()*scale_x
- for all valid i,k: image_plane(OUT,i,k) is a copy of
image_plane(IN,i,k) that has been bilinearly interpolated to fit into
the shape of image_plane(OUT,i,k).
!*/
public:
upsample_(
);
/*!
ensures
- This object has no state, so the constructor does nothing, aside from
providing default constructability.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
int scale,
typename SUBNET
>
using upsample = add_layer<upsample_<scale,scale>, SUBNET>;
// ----------------------------------------------------------------------------------------
template <
long NR_,
long NC_
>
class resize_to_
{
/*!
REQUIREMENTS ON THE INPUT ARGUMENTS
- NR_ >= 1
- NC_ >= 1
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it allows you to resize a layer using
bilinear interpolation. To be very specific, it resizes each of the
channels in an input tensor. Therefore, if IN is the input tensor to this
layer and OUT the output tensor, then we will have:
- OUT.num_samples() == IN.num_samples()
- OUT.k() == IN.k()
- OUT.nr() == NR_
- OUT.nc() == NC_
- for all valid i,k: image_plane(OUT,i,k) is a copy of
image_plane(IN,i,k) that has been bilinearly interpolated to fit into
the shape of image_plane(OUT,i,k).
!*/
public:
resize_to_(
);
/*!
ensures
- This object has no state, so the constructor does nothing, aside from
providing default constructability.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
long NR,
long NC,
typename SUBNET
>
using resize_to = add_layer<resize_to_<NR,NC>, SUBNET>;
// ----------------------------------------------------------------------------------------
class dropout_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a dropout layer. Therefore, it
passes its inputs through the stochastic function f(x) which outputs either
0 or x. The probability of 0 being output is given by the drop_rate
argument to this object's constructor.
Note that, after you finish training a network with dropout, it is a good
idea to replace each dropout_ layer with a multiply_ layer because the
multiply_ layer is faster and deterministic.
!*/
public:
explicit dropout_(
float drop_rate = 0.5
);
/*!
requires
- 0 <= drop_rate <= 1
ensures
- #get_drop_rate() == drop_rate
!*/
float get_drop_rate (
) const;
/*!
ensures
- returns the probability that an individual input value to this layer will
be replaced with 0.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <typename SUBNET>
using dropout = add_layer<dropout_, SUBNET>;
// ----------------------------------------------------------------------------------------
class multiply_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a basic layer that just
multiplies its input tensor with a constant value and returns the result.
It therefore has no learnable parameters.
!*/
public:
explicit multiply_(
float val = 0.5
);
/*!
ensures
- #get_multiply_value() == val
!*/
multiply_ (
const dropout_& item
);
/*!
ensures
- #get_multiply_value() == 1-item.get_drop_rate()
(i.e. We construct the multiply_ layer so that it is essentially a
deterministic version of the given dropout_ layer)
!*/
float get_multiply_value (
) const;
/*!
ensures
- this layer simply multiplies its input tensor by get_multiply_value() and
produces the result as output.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <typename SUBNET>
using multiply = add_layer<multiply_, SUBNET>;
// ----------------------------------------------------------------------------------------
const double DEFAULT_LAYER_NORM_EPS = 1e-5;
class layer_norm_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a batch normalization layer that
implements the method described in the paper:
Layer Normalization by Jimmy Lei Ba, Jamie Ryan Kiros, Geoffrey E. Hinton
In particular, this layer produces output tensors with the same
dimensionality as the input tensors, except that the mean and variances of
the elements in each sample have been standardized to 0 and 1 respectively.
This is different from batch normalization, since this layer learns one scaling
factor and one bias for each sample in the batch, independently. As a result,
this layer is batch-size independent.
!*/
public:
layer_norm_(
);
/*!
ensures
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 0
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 1
- #get_eps() == DEFAULT_LAYER_NORM_EPS
!*/
explicit layer_norm_(
double eps_ = DEFAULT_LAYER_NORM_EPS
)
/*!
requires
- eps > 0
ensures
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 0
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 1
- #get_eps() == eps
!*/
double get_eps(
) const;
/*!
ensures
- When doing layer normalization, we are dividing by the standard
deviation. This epsilon value returned by this function is added to the
variance to prevent the division from dividing by zero.
!*/
double get_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its parameters be
multiplied by get_learning_rate_multiplier().
!*/
double get_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its parameters be
multiplied by get_weight_decay_multiplier().
!*/
void set_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_learning_rate_multiplier() == val
!*/
void set_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_weight_decay_multiplier() == val
!*/
double get_bias_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its bias parameters be
multiplied by get_learning_rate_multiplier()*get_bias_learning_rate_multiplier().
!*/
double get_bias_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its bias parameters be
multiplied by get_weight_decay_multiplier()*get_bias_weight_decay_multiplier().
!*/
void set_bias_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_learning_rate_multiplier() == val
!*/
void set_bias_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_weight_decay_multiplier() == val
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
// ----------------------------------------------------------------------------------------
enum layer_mode
{
CONV_MODE = 0, // convolutional mode
FC_MODE = 1 // fully connected mode
};
const double DEFAULT_BATCH_NORM_EPS = 0.0001;
template <
layer_mode mode
>
class bn_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a batch normalization layer that
implements the method described in the paper:
Batch Normalization: Accelerating Deep Network Training by Reducing
Internal Covariate Shift by Sergey Ioffe and Christian Szegedy
In particular, this layer produces output tensors with the same
dimensionality as the input tensors, except that the mean and variances of
the elements have been standardized to 0 and 1 respectively.
It should also be noted that when tensors with a num_samples() dimension of
1 are passed to this layer it doesn't perform batch normalization.
Instead, it runs in "inference mode" where the learned linear normalizing
transformation is used to transform the tensor.
Finally, after you finish training a batch normalized network, it is a good
idea to replace each bn_ layer with an affine_ layer because the affine_
layer is faster and will never surprise you by performing batch
normalization on tensors that have a num_samples() dimension > 1. This allows
you to run large mini-batches of samples through your final network without
batch normalization executing at all.
!*/
public:
bn_(
);
/*!
ensures
- #get_mode() == mode
- #get_running_stats_window_size() == 100
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 0
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 1
- #get_eps() == tt::DEFAULT_BATCH_NORM_EPS
!*/
explicit bn_(
unsigned long window_size,
double eps = tt::DEFAULT_BATCH_NORM_EPS
);
/*!
requires
- eps > 0
- window_size > 0
ensures
- #get_mode() == mode
- #get_running_stats_window_size() == window_size
- #get_learning_rate_multiplier() == 1
- #get_weight_decay_multiplier() == 0
- #get_bias_learning_rate_multiplier() == 1
- #get_bias_weight_decay_multiplier() == 1
- #get_eps() == eps
!*/
layer_mode get_mode(
) const;
/*!
ensures
- returns the mode of this layer, either CONV_MODE or FC_MODE.
If the mode is FC_MODE then the normalization is applied across the
samples in a tensor (i.e. k()*nr()*nc() different things will be
normalized). Otherwise, normalization is applied across everything
except for the k() dimension, resulting in there being only k()
normalization equations that are applied spatially over the tensor.
Therefore, if you are putting batch normalization after a fully connected
layer you should use FC_MODE. Otherwise, if you are putting batch
normalization after a convolutional layer you should use CONV_MODE.
!*/
double get_eps(
) const;
/*!
ensures
- When doing batch normalization, we are dividing by the standard
deviation. This epsilon value returned by this function is added to the
variance to prevent the division from dividing by zero.
!*/
unsigned long get_running_stats_window_size (
) const;
/*!
ensures
- Just as recommended in the batch normalization paper, this object keeps a
running average of the mean and standard deviations of the features.
These averages are used during "inference mode" so you can run a single
object through a batch normalized network. They are also what is used to
initialize an affine_ layer that is constructed from a bn_ layer. This
function returns the effective number of recent samples used to compute
the running average.
!*/
void set_running_stats_window_size (
unsigned long new_window_size
);
/*!
requires
- new_window_size > 0
ensures
- #get_running_stats_window_size() == new_window_size
!*/
double get_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its parameters be
multiplied by get_learning_rate_multiplier().
!*/
double get_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its parameters be
multiplied by get_weight_decay_multiplier().
!*/
void set_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_learning_rate_multiplier() == val
!*/
void set_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_weight_decay_multiplier() == val
!*/
double get_bias_learning_rate_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the learning rate used to optimize its bias parameters be
multiplied by get_learning_rate_multiplier()*get_bias_learning_rate_multiplier().
!*/
double get_bias_weight_decay_multiplier(
) const;
/*!
ensures
- returns a multiplier number. The interpretation is that this object is
requesting that the weight decay used to optimize its bias parameters be
multiplied by get_weight_decay_multiplier()*get_bias_weight_decay_multiplier().
!*/
void set_bias_learning_rate_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_learning_rate_multiplier() == val
!*/
void set_bias_weight_decay_multiplier(
double val
);
/*!
requires
- val >= 0
ensures
- #get_bias_weight_decay_multiplier() == val
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <typename SUBNET>
using bn_con = add_layer<bn_<CONV_MODE>, SUBNET>;
template <typename SUBNET>
using bn_fc = add_layer<bn_<FC_MODE>, SUBNET>;
// ----------------------------------------------------------------------------------------
template <typename net_type>
void set_all_bn_running_stats_window_sizes (
const net_type& net,
unsigned long new_window_size
);
/*!
requires
- new_window_size > 0
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
ensures
- Sets the get_running_stats_window_size() field of all bn_ layers in net to
new_window_size.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
void disable_duplicative_biases (
const net_type& net
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
ensures
- Disables bias for all bn_ and layer_norm_ inputs.
- Sets the get_bias_learning_rate_multiplier() and get_bias_weight_decay_multiplier()
to zero of all bn_ and layer_norm_ inputs.
!*/
// ----------------------------------------------------------------------------------------
class affine_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it applies a simple pointwise linear
transformation to an input tensor. You can think of it as having two
parameter tensors, A and B. If the input tensor is called INPUT then the
output of this layer is:
A*INPUT+B
where all operations are performed element wise and each sample in the
INPUT tensor is processed separately.
Moreover, this object has two modes that affect the dimensionalities of A
and B and how they are applied to compute A*INPUT+B. If
get_mode()==FC_MODE then A and B each have the same dimensionality as the
input tensor, except their num_samples() dimensions are 1. If
get_mode()==CONV_MODE then A and B have all their dimensions set to 1
except for k(), which is equal to INPUT.k().
In either case, the computation of A*INPUT+B is performed pointwise over all
the elements of INPUT using either:
OUTPUT(n,k,r,c) == A(1,k,r,c)*INPUT(n,k,r,c)+B(1,k,r,c)
or
OUTPUT(n,k,r,c) == A(1,k,1,1)*INPUT(n,k,r,c)+B(1,k,1,1)
as appropriate.
Finally, note that the parameters of this layer are not learnable and
therefore not modified during network updates. Instead, the layer will
perform the identity transformation unless it is initialized with a bn_
layer, in which case it will perform whatever transformation the bn_ layer
has learned.
!*/
public:
affine_(
);
/*!
ensures
- #get_mode() == FC_MODE
!*/
affine_(
layer_mode mode
);
/*!
ensures
- #get_mode() == mode
!*/
template <
layer_mode mode
>
affine_(
const bn_<mode>& layer
);
/*!
ensures
- Constructs affine_ so that it performs the same transformation as the
supplied batch normalization layer. You would want to do this after you
finish training a network with bn_ layers because the affine_ layer will
execute faster.
- #get_mode() == layer.get_mode()
!*/
layer_mode get_mode(
) const;
/*!
ensures
- returns the mode of this layer, either CONV_MODE or FC_MODE.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the
EXAMPLE_COMPUTATIONAL_LAYER_ interface. Also note that get_layer_params()
always returns an empty tensor since there are no learnable parameters in this
object.
!*/
};
template <typename SUBNET>
using affine = add_layer<affine_, SUBNET>;
// ----------------------------------------------------------------------------------------
template <
long _nr,
long _nc,
int _stride_y,
int _stride_x,
int _padding_y = _stride_y!=1? 0 : _nr/2,
int _padding_x = _stride_x!=1? 0 : _nc/2
>
class max_pool_
{
/*!
REQUIREMENTS ON TEMPLATE ARGUMENTS
- _nr >= 0
- _nc >= 0
- _stride_y > 0
- _stride_x > 0
- _padding_y >= 0
- _padding_x >= 0
- if (_nr != 0) then
- _padding_y < _nr
- else
- _padding_y == 0
- if (_nc != 0) then
- _padding_x < _nr
- else
- _padding_x == 0
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a max pooling layer that takes an
input tensor and downsamples it. It does this by sliding a window over the
images in an input tensor and outputting, for each channel, the maximum
element within the window.
If _nr == 0 then it means the filter size covers all the rows in the input
tensor, similarly for the _nc parameter. To be precise, if we call the
input tensor IN and the output tensor OUT, then OUT is defined as follows:
- let FILT_NR == (nr()==0) ? IN.nr() : nr()
- let FILT_NC == (nc()==0) ? IN.nc() : nc()
- OUT.num_samples() == IN.num_samples()
- OUT.k() == IN.k()
- OUT.nr() == 1+(IN.nr() + 2*padding_y() - FILT_NR)/stride_y()
- OUT.nc() == 1+(IN.nc() + 2*padding_x() - FILT_NC)/stride_x()
- for all valid s, k, r, and c:
- image_plane(OUT,s,k)(r,c) == max(subm_clipped(image_plane(IN,s,k),
centered_rect(x*stride_x() + FILT_NC/2 - padding_x(),
y*stride_y() + FILT_NR/2 - padding_y(),
FILT_NC,
FILT_NR)))
!*/
public:
max_pool_ (
);
/*!
ensures
- #nr() == _nr
- #nc() == _nc
- #stride_y() == _stride_y
- #stride_x() == _stride_x
- #padding_y() == _padding_y
- #padding_x() == _padding_x
!*/
long nr(
) const;
/*!
ensures
- returns the number of rows in the pooling window or 0 if the window size
is "the entire input tensor".
!*/
long nc(
) const;
/*!
ensures
- returns the number of rows in the pooling window or 0 if the window size
is "the entire input tensor".
!*/
long stride_y(
) const;
/*!
ensures
- returns the vertical stride used when scanning the max pooling window
over an image. That is, each window will be moved stride_y() pixels down
at a time when it moves over the image.
!*/
long stride_x(
) const;
/*!
ensures
- returns the horizontal stride used when scanning the max pooling window
over an image. That is, each window will be moved stride_x() pixels down
at a time when it moves over the image.
!*/
long padding_y(
) const;
/*!
ensures
- returns the number of pixels of zero padding added to the top and bottom
sides of the image.
!*/
long padding_x(
) const;
/*!
ensures
- returns the number of pixels of zero padding added to the left and right
sides of the image.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& computed_output, const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <
long nr,
long nc,
int stride_y,
int stride_x,
typename SUBNET
>
using max_pool = add_layer<max_pool_<nr,nc,stride_y,stride_x>, SUBNET>;
template <
typename SUBNET
>
using max_pool_everything = add_layer<max_pool_<0,0,1,1>, SUBNET>;
// ----------------------------------------------------------------------------------------
template <
long _nr,
long _nc,
int _stride_y,
int _stride_x,
int _padding_y = _stride_y!=1? 0 : _nr/2,
int _padding_x = _stride_x!=1? 0 : _nc/2
>
class avg_pool_
{
/*!
REQUIREMENTS ON TEMPLATE ARGUMENTS
- _nr >= 0
- _nc >= 0
- _stride_y > 0
- _stride_x > 0
- _padding_y >= 0
- _padding_x >= 0
- if (_nr != 0) then
- _padding_y < _nr
- else
- _padding_y == 0
- if (_nc != 0) then
- _padding_x < _nr
- else
- _padding_x == 0
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines an average pooling layer that
takes an input tensor and downsamples it. It does this by sliding a window
over the images in an input tensor and outputting, for each channel, the
average element within the window.
If _nr == 0 then it means the filter size covers all the rows in the input
tensor, similarly for the _nc parameter. To be precise, if we call the
input tensor IN and the output tensor OUT, then OUT is defined as follows:
- let FILT_NR == (nr()==0) ? IN.nr() : nr()
- let FILT_NC == (nc()==0) ? IN.nc() : nc()
- OUT.num_samples() == IN.num_samples()
- OUT.k() == IN.k()
- OUT.nr() == 1+(IN.nr() + 2*padding_y() - FILT_NR)/stride_y()
- OUT.nc() == 1+(IN.nc() + 2*padding_x() - FILT_NC)/stride_x()
- for all valid s, k, r, and c:
- image_plane(OUT,s,k)(r,c) == mean(subm_clipped(image_plane(IN,s,k),
centered_rect(x*stride_x() + FILT_NC/2 - padding_x(),
y*stride_y() + FILT_NR/2 - padding_y(),
FILT_NC,
FILT_NR)))
!*/
public:
avg_pool_ (
);
/*!
ensures
- #nr() == _nr
- #nc() == _nc
- #stride_y() == _stride_y
- #stride_x() == _stride_x
- #padding_y() == _padding_y
- #padding_x() == _padding_x
!*/
long nr(
) const;
/*!
ensures
- returns the number of rows in the pooling window or 0 if the window size
is "the entire input tensor".
!*/
long nc(
) const;
/*!
ensures
- returns the number of rows in the pooling window or 0 if the window size
is "the entire input tensor".
!*/
long stride_y(
) const;
/*!
ensures
- returns the vertical stride used when scanning the pooling window
over an image. That is, each window will be moved stride_y() pixels down
at a time when it moves over the image.
!*/
long stride_x(
) const;
/*!
ensures
- returns the horizontal stride used when scanning the pooling window
over an image. That is, each window will be moved stride_x() pixels down
at a time when it moves over the image.
!*/
long padding_y(
) const;
/*!
ensures
- returns the number of pixels of zero padding added to the top and bottom
sides of the image.
!*/
long padding_x(
) const;
/*!
ensures
- returns the number of pixels of zero padding added to the left and right
sides of the image.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& computed_output, const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <
long nr,
long nc,
int stride_y,
int stride_x,
typename SUBNET
>
using avg_pool = add_layer<avg_pool_<nr,nc,stride_y,stride_x>, SUBNET>;
template <
typename SUBNET
>
using avg_pool_everything = add_layer<avg_pool_<0,0,1,1>, SUBNET>;
// ----------------------------------------------------------------------------------------
class relu_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a rectified linear layer.
Therefore, it passes its inputs through the function
f(x)=max(x,0)
where f() is applied pointwise across the input tensor.
!*/
public:
relu_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <typename SUBNET>
using relu = add_layer<relu_, SUBNET>;
// ----------------------------------------------------------------------------------------
class prelu_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a parametric rectified linear
layer. Therefore, it passes its inputs through the function
f(x) = x>0 ? x : p*x
where f() is applied pointwise across the input tensor and p is a scalar
parameter learned by this layer.
This is the layer type introduced in the paper:
He, Kaiming, et al. "Delving deep into rectifiers: Surpassing
human-level performance on imagenet classification." Proceedings of the
IEEE International Conference on Computer Vision. 2015.
!*/
public:
explicit prelu_(
float initial_param_value = 0.25
);
/*!
ensures
- The p parameter will be initialized with initial_param_value.
- #get_initial_param_value() == initial_param_value.
!*/
float get_initial_param_value (
) const;
/*!
ensures
- returns the initial value of the prelu parameter.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <typename SUBNET>
using prelu = add_layer<prelu_, SUBNET>;
// ----------------------------------------------------------------------------------------
class leaky_relu_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a leaky rectified linear
layer. Therefore, it passes its inputs through the function
f(x) = x>0 ? x : alpha*x
where f() is applied pointwise across the input tensor and alpha is a
non-learned scalar.
This is the layer type introduced in the paper:
A. L. Maas, A. Y. Hannun, and A. Y. Ng. "Rectifier nonlinearities improve
neural network acoustic models". In ICML, 2013.
!*/
public:
explicit leaky_relu_(
float alpha = 0.01f
);
/*!
ensures
- the alpha parameter will be initialized with the alpha value
!*/
float get_alpha(
) const;
/*!
ensures
- returns the alpha parameter of the leaky_relu
!*/
template <typename SUBNET> void setup(const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <typename SUBNET>
using leaky_relu = add_layer<prelu_, SUBNET>;
// ----------------------------------------------------------------------------------------
class sig_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a sigmoid layer. Therefore, it
passes its inputs through the function
f(x)=1/(1+exp(-x))
where f() is applied pointwise across the input tensor.
!*/
public:
sig_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <typename SUBNET>
using sig = add_layer<sig_, SUBNET>;
// ----------------------------------------------------------------------------------------
class mish_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a mish layer. Therefore, it
passes its inputs through the function
f(x)= x*tanh(log(1+exp(x)))
where f() is applied pointwise across the input tensor.
!*/
public:
mish_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& data_output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor&);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <typename SUBNET>
using mish = add_layer<mish_, SUBNET>;
// ----------------------------------------------------------------------------------------
class htan_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a hyperbolic tangent layer.
Therefore, it passes its inputs through the function
f(x)=std::tanh(x)
where f() is applied pointwise across the input tensor.
!*/
public:
htan_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <typename SUBNET>
using htan = add_layer<htan_, SUBNET>;
// ----------------------------------------------------------------------------------------
class gelu_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a gelu layer. Therefore, it
passes its inputs through the function
f(x)= x/2 * (1 + erf(x/sqrt(2))
where f() is applied pointwise across the input tensor.
This is the layer type introduced in the paper:
Dan Hendrycks, Kevin Gimpel. "Gaussian Error Linear Units (GELUs)".
!*/
public:
gelu_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& data_output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor&);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <typename SUBNET>
using gelu = add_layer<gelu_, SUBNET>;
// ----------------------------------------------------------------------------------------
class softmax_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a softmax layer. To be precise,
we define the softmax function s(x) as:
s(x) == exp(x)/sum(exp(x))
where x is a vector. Then this layer treats its input tensor as a
collection of multi-channel images and applies s() to each spatial location
in each image. In each application, the tensor::k() channel elements at
each position are input to s() and then replaced by the outputs of s().
This means that, for example, if you collapsed each output image to a 1
channel image by adding the channels then you would end up with images
where each pixel value was 1. This is because the sum of the outputs of
s() will always be equal to 1.
!*/
public:
softmax_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <typename SUBNET>
using softmax = add_layer<softmax_, SUBNET>;
// ----------------------------------------------------------------------------------------
class softmax_all_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, it defines a softmax layer. To be precise,
we define the softmax function s(x) as:
s(x) == exp(x)/sum(exp(x))
where x is a vector. Then this layer treats its input tensor as a
collection of tensor::num_samples() vectors and applies s() to each vector
in the tensor. Therefore, there are logically tensor::num_samples()
invocations of s().
!*/
public:
softmax_all_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_
interface. Note that this layer doesn't have any parameters, so the tensor
returned by get_layer_params() is always empty.
!*/
};
template <typename SUBNET>
using softmax_all = add_layer<softmax_all_, SUBNET>;
// ----------------------------------------------------------------------------------------
template <
template<typename> class tag
>
class add_prev_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. This layer simply adds the output of two previous layers.
In particular, it adds the tensor from its immediate predecessor layer,
sub.get_output(), with the tensor from a deeper layer,
layer<tag>(sub).get_output().
Therefore, you supply a tag via add_prev_'s template argument that tells it
what layer to add to the output of the previous layer. The result of this
addition is output by add_prev_. Finally, the addition happens pointwise
according to 4D tensor arithmetic. If the dimensions don't match then
missing elements are presumed to be equal to 0. Moreover, each dimension
of the output tensor is equal to the maximum dimension of either of the
inputs. That is, if the tensors A and B are being added to produce C then:
- C.num_samples() == max(A.num_samples(), B.num_samples())
- C.k() == max(A.k(), B.k())
- C.nr() == max(A.nr(), B.nr())
- C.nc() == max(A.nc(), B.nc())
!*/
public:
add_prev_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
template<typename> class tag,
typename SUBNET
>
using add_prev = add_layer<add_prev_<tag>, SUBNET>;
// Here we add some convenient aliases for using add_prev_ with the tag layers.
template <typename SUBNET> using add_prev1 = add_prev<tag1, SUBNET>;
template <typename SUBNET> using add_prev2 = add_prev<tag2, SUBNET>;
template <typename SUBNET> using add_prev3 = add_prev<tag3, SUBNET>;
template <typename SUBNET> using add_prev4 = add_prev<tag4, SUBNET>;
template <typename SUBNET> using add_prev5 = add_prev<tag5, SUBNET>;
template <typename SUBNET> using add_prev6 = add_prev<tag6, SUBNET>;
template <typename SUBNET> using add_prev7 = add_prev<tag7, SUBNET>;
template <typename SUBNET> using add_prev8 = add_prev<tag8, SUBNET>;
template <typename SUBNET> using add_prev9 = add_prev<tag9, SUBNET>;
template <typename SUBNET> using add_prev10 = add_prev<tag10, SUBNET>;
using add_prev1_ = add_prev_<tag1>;
using add_prev2_ = add_prev_<tag2>;
using add_prev3_ = add_prev_<tag3>;
using add_prev4_ = add_prev_<tag4>;
using add_prev5_ = add_prev_<tag5>;
using add_prev6_ = add_prev_<tag6>;
using add_prev7_ = add_prev_<tag7>;
using add_prev8_ = add_prev_<tag8>;
using add_prev9_ = add_prev_<tag9>;
using add_prev10_ = add_prev_<tag10>;
// ----------------------------------------------------------------------------------------
template <
template<typename> class tag
>
class mult_prev_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. This layer simply multiplies the output of two previous
layers. In particular, it multiplies the tensor from its immediate
predecessor layer, sub.get_output(), with the tensor from a deeper layer,
layer<tag>(sub).get_output().
Therefore, you supply a tag via mult_prev_'s template argument that tells
it what layer to multiply with the output of the previous layer. The
result of this multiplication is output by mult_prev_. Finally, the
multiplication happens pointwise according to 4D tensor arithmetic. If the
dimensions don't match then missing elements are presumed to be equal to 0.
Moreover, each dimension of the output tensor is equal to the maximum
dimension of either of the inputs. That is, if the tensors A and B are
being multiplied to produce C then:
- C.num_samples() == max(A.num_samples(), B.num_samples())
- C.k() == max(A.k(), B.k())
- C.nr() == max(A.nr(), B.nr())
- C.nc() == max(A.nc(), B.nc())
!*/
public:
mult_prev_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
template<typename> class tag,
typename SUBNET
>
using mult_prev = add_layer<mult_prev_<tag>, SUBNET>;
// Here we add some convenient aliases for using mult_prev_ with the tag layers.
template <typename SUBNET> using mult_prev1 = mult_prev<tag1, SUBNET>;
template <typename SUBNET> using mult_prev2 = mult_prev<tag2, SUBNET>;
template <typename SUBNET> using mult_prev3 = mult_prev<tag3, SUBNET>;
template <typename SUBNET> using mult_prev4 = mult_prev<tag4, SUBNET>;
template <typename SUBNET> using mult_prev5 = mult_prev<tag5, SUBNET>;
template <typename SUBNET> using mult_prev6 = mult_prev<tag6, SUBNET>;
template <typename SUBNET> using mult_prev7 = mult_prev<tag7, SUBNET>;
template <typename SUBNET> using mult_prev8 = mult_prev<tag8, SUBNET>;
template <typename SUBNET> using mult_prev9 = mult_prev<tag9, SUBNET>;
template <typename SUBNET> using mult_prev10 = mult_prev<tag10, SUBNET>;
using mult_prev1_ = mult_prev_<tag1>;
using mult_prev2_ = mult_prev_<tag2>;
using mult_prev3_ = mult_prev_<tag3>;
using mult_prev4_ = mult_prev_<tag4>;
using mult_prev5_ = mult_prev_<tag5>;
using mult_prev6_ = mult_prev_<tag6>;
using mult_prev7_ = mult_prev_<tag7>;
using mult_prev8_ = mult_prev_<tag8>;
using mult_prev9_ = mult_prev_<tag9>;
using mult_prev10_ = mult_prev_<tag10>;
// ----------------------------------------------------------------------------------------
template <
template<typename> class tag
>
class resize_prev_to_tagged_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. This layer resizes the output channels of the previous layer
to have the same number of rows and columns as the output of the tagged layer.
This layer uses bilinear interpolation. If the sizes match already, then it
simply copies the data.
Therefore, you supply a tag via resize_prev_to_tagged's template argument that
tells it what layer to use for the target size.
If tensor PREV is resized to size of tensor TAGGED, then a tensor OUT is
produced such that:
- OUT.num_samples() == PREV.num_samples()
- OUT.k() == PREV.k()
- OUT.nr() == TAGGED.nr()
- OUT.nc() == TAGGED.nc()
!*/
public:
resize_prev_to_tagged_(
);
template <typename SUBNET> void setup(const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
template<typename> class tag,
typename SUBNET
>
using resize_prev_to_tagged = add_layer<resize_prev_to_tagged_<tag>, SUBNET>;
// ----------------------------------------------------------------------------------------
template <
template<typename> class tag
>
class scale_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. This layer scales the output channels of the tagged layer
by multiplying it with the output of the previous layer. To be specific:
- Let INPUT == layer<tag>(sub).get_output()
- Let SCALES == sub.get_output()
- This layer takes INPUT and SCALES as input.
- The output of this layer has the same dimensions as INPUT.
- This layer requires:
- SCALES.num_samples() == INPUT.num_samples()
- SCALES.k() == INPUT.k()
- SCALES.nr() == 1
- SCALES.nc() == 1
- The output tensor is produced by pointwise multiplying SCALES with
INPUT at each spatial location. Therefore, if OUT is the output of
this layer then we would have:
OUT(n,k,r,c) == INPUT(n,k,r,c)*SCALES(n,k)
!*/
public:
scale_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
template<typename> class tag,
typename SUBNET
>
using scale = add_layer<scale_<tag>, SUBNET>;
// Here we add some convenient aliases for using scale_ with the tag layers.
template <typename SUBNET> using scale1 = scale<tag1, SUBNET>;
template <typename SUBNET> using scale2 = scale<tag2, SUBNET>;
template <typename SUBNET> using scale3 = scale<tag3, SUBNET>;
template <typename SUBNET> using scale4 = scale<tag4, SUBNET>;
template <typename SUBNET> using scale5 = scale<tag5, SUBNET>;
template <typename SUBNET> using scale6 = scale<tag6, SUBNET>;
template <typename SUBNET> using scale7 = scale<tag7, SUBNET>;
template <typename SUBNET> using scale8 = scale<tag8, SUBNET>;
template <typename SUBNET> using scale9 = scale<tag9, SUBNET>;
template <typename SUBNET> using scale10 = scale<tag10, SUBNET>;
using scale1_ = scale_<tag1>;
using scale2_ = scale_<tag2>;
using scale3_ = scale_<tag3>;
using scale4_ = scale_<tag4>;
using scale5_ = scale_<tag5>;
using scale6_ = scale_<tag6>;
using scale7_ = scale_<tag7>;
using scale8_ = scale_<tag8>;
using scale9_ = scale_<tag9>;
using scale10_ = scale_<tag10>;
// ----------------------------------------------------------------------------------------
template <
template<typename> class tag
>
class scale_prev_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. This layer scales the output channels of the tagged layer
by multiplying it with the output of the previous layer. It is excatly the
same as the scale_ layer, but with the inputs swapped, which is useful since
it allows mapping between inputs and outputs of this layer. To be specific:
- Let INPUT == sub.get_output()
- Let SCALES == layer<tag>(sub).get_output()
- This layer takes INPUT and SCALES as input.
- The output of this layer has the same dimensions as INPUT.
- This layer requires:
- SCALES.num_samples() == INPUT.num_samples()
- SCALES.k() == INPUT.k()
- SCALES.nr() == 1
- SCALES.nc() == 1
- The output tensor is produced by pointwise multiplying SCALES with
INPUT at each spatial location. Therefore, if OUT is the output of
this layer then we would have:
OUT(n,k,r,c) == INPUT(n,k,r,c)*SCALES(n,k)
!*/
public:
scale_prev_(
);
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
template<typename> class tag,
typename SUBNET
>
using scale_prev = add_layer<scale_prev_<tag>, SUBNET>;
// Here we add some convenient aliases for using scale_prev_ with the tag layers.
template <typename SUBNET> using scale_prev1 = scale_prev<tag1, SUBNET>;
template <typename SUBNET> using scale_prev2 = scale_prev<tag2, SUBNET>;
template <typename SUBNET> using scale_prev3 = scale_prev<tag3, SUBNET>;
template <typename SUBNET> using scale_prev4 = scale_prev<tag4, SUBNET>;
template <typename SUBNET> using scale_prev5 = scale_prev<tag5, SUBNET>;
template <typename SUBNET> using scale_prev6 = scale_prev<tag6, SUBNET>;
template <typename SUBNET> using scale_prev7 = scale_prev<tag7, SUBNET>;
template <typename SUBNET> using scale_prev8 = scale_prev<tag8, SUBNET>;
template <typename SUBNET> using scale_prev9 = scale_prev<tag9, SUBNET>;
template <typename SUBNET> using scale_prev10 = scale_prev<tag10, SUBNET>;
using scale_prev1_ = scale_prev_<tag1>;
using scale_prev2_ = scale_prev_<tag2>;
using scale_prev3_ = scale_prev_<tag3>;
using scale_prev4_ = scale_prev_<tag4>;
using scale_prev5_ = scale_prev_<tag5>;
using scale_prev6_ = scale_prev_<tag6>;
using scale_prev7_ = scale_prev_<tag7>;
using scale_prev8_ = scale_prev_<tag8>;
using scale_prev9_ = scale_prev_<tag9>;
using scale_prev10_ = scale_prev_<tag10>;
// ----------------------------------------------------------------------------------------
template<
template<typename> class... TAG_TYPES
>
class concat_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. This layer simply concatenates the output of tagged layers.
Importantly, each input layer must have the same dimensions (i.e.
num_samples, nr, and nc) except for the k channel, which may vary. This is
because the concatenation happens along the k dimension. That is, the
output of this network is a tensor, OUT, that is the concatenation of the
tensors:
for each (tag in TAG_TYPES)
layer<tag>(subnet).get_output()
Therefore, out.num_samples(), out.nr(), and out.nc() match the dimensions
of the input tensors while OUT.k() is the sum of the input layer's k()
dimensions.
!*/
public:
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
dpoint map_input_to_output(dpoint p) const;
dpoint map_output_to_input(dpoint p) const;
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
// concat layer definitions
template <template<typename> class TAG1,
template<typename> class TAG2,
typename SUBNET>
using concat2 = add_layer<concat_<TAG1, TAG2>, SUBNET>;
template <template<typename> class TAG1,
template<typename> class TAG2,
template<typename> class TAG3,
typename SUBNET>
using concat3 = add_layer<concat_<TAG1, TAG2, TAG3>, SUBNET>;
template <template<typename> class TAG1,
template<typename> class TAG2,
template<typename> class TAG3,
template<typename> class TAG4,
typename SUBNET>
using concat4 = add_layer<concat_<TAG1, TAG2, TAG3, TAG4>, SUBNET>;
template <template<typename> class TAG1,
template<typename> class TAG2,
template<typename> class TAG3,
template<typename> class TAG4,
template<typename> class TAG5,
typename SUBNET>
using concat5 = add_layer<concat_<TAG1, TAG2, TAG3, TAG4, TAG5>, SUBNET>;
// ----------------------------------------------------------------------------------------
/*!A inception layer definitions !*/
// Now define inception layer tag types. These layer aliases allow creating
// the networks described in the paper:
// Szegedy, Christian, et al. "Going deeper with convolutions." Proceedings of
// the IEEE Conference on Computer Vision and Pattern Recognition. 2015.
// See the dnn_inception_ex.cpp example for a complete example of their use. Note also
// that we use tag ID numbers >= 1000 to avoid conflict with user's tag layers.
template <typename SUBNET> using itag0 = add_tag_layer< 1000 + 0, SUBNET>;
template <typename SUBNET> using itag1 = add_tag_layer< 1000 + 1, SUBNET>;
template <typename SUBNET> using itag2 = add_tag_layer< 1000 + 2, SUBNET>;
template <typename SUBNET> using itag3 = add_tag_layer< 1000 + 3, SUBNET>;
template <typename SUBNET> using itag4 = add_tag_layer< 1000 + 4, SUBNET>;
template <typename SUBNET> using itag5 = add_tag_layer< 1000 + 5, SUBNET>;
// skip to inception input
template <typename SUBNET> using iskip = add_skip_layer< itag0, SUBNET>;
// here are some templates to be used for creating inception layer groups
template <template<typename>class B1,
template<typename>class B2,
typename SUBNET>
using inception2 = concat2<itag1, itag2, itag1<B1<iskip< itag2<B2< itag0<SUBNET>>>>>>>;
template <template<typename>class B1,
template<typename>class B2,
template<typename>class B3,
typename SUBNET>
using inception3 = concat3<itag1, itag2, itag3, itag1<B1<iskip< itag2<B2<iskip< itag3<B3< itag0<SUBNET>>>>>>>>>>;
template <template<typename>class B1,
template<typename>class B2,
template<typename>class B3,
template<typename>class B4,
typename SUBNET>
using inception4 = concat4<itag1, itag2, itag3, itag4,
itag1<B1<iskip< itag2<B2<iskip< itag3<B3<iskip< itag4<B4< itag0<SUBNET>>>>>>>>>>>>>;
template <template<typename>class B1,
template<typename>class B2,
template<typename>class B3,
template<typename>class B4,
template<typename>class B5,
typename SUBNET>
using inception5 = concat5<itag1, itag2, itag3, itag4, itag5,
itag1<B1<iskip< itag2<B2<iskip< itag3<B3<iskip< itag4<B4<iskip< itag5<B5< itag0<SUBNET>>>>>>>>>>>>>>>>;
// ----------------------------------------------------------------------------------------
const double DEFAULT_L2_NORM_EPS = 1e-5;
class l2normalize_
{
/*!
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. It takes tensors as input and L2 normalizes them. In particular,
it has the following properties:
- The output tensors from this layer have the same dimensions as the
input tensors.
- If you think of each input tensor as a set of tensor::num_samples()
vectors, then the output tensor contains the same vectors except they
have been length normalized so that their L2 norms are all 1. I.e.
for each vector v we will have ||v||==1.
!*/
public:
explicit l2normalize_(
double eps = tt::DEFAULT_L2_NORM_EPS
);
/*!
requires
- eps > 0
ensures
- #get_eps() == eps
!*/
double get_eps(
) const;
/*!
ensures
- When we normalize a vector we divide it by its L2 norm. However, the
get_eps() value is added to the squared norm prior to division to avoid
ever dividing by zero.
!*/
template <typename SUBNET> void setup (const SUBNET& sub);
void forward_inplace(const tensor& input, tensor& output);
void backward_inplace(const tensor& computed_output, const tensor& gradient_input, tensor& data_grad, tensor& params_grad);
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
// ----------------------------------------------------------------------------------------
template <
long _offset,
long _k,
long _nr,
long _nc
>
class extract_
{
/*!
REQUIREMENTS ON TEMPLATE ARGUMENTS
- 0 <= _offset
- 0 < _k
- 0 < _nr
- 0 < _nc
WHAT THIS OBJECT REPRESENTS
This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
defined above. In particular, the output of this layer is simply a copy of
the input tensor. However, you can configure the extract layer to output
only some subset of the input tensor and also to reshape it. Therefore,
the dimensions of the tensor output by this layer are as follows (letting
IN be the input tensor and OUT the output tensor):
- OUT.num_samples() == IN.num_samples()
- OUT.k() == _k
- OUT.nr() == _nr
- OUT.nc() == _nc
So the output will always have the same number of samples as the input, but
within each sample (the k,nr,nc part) we will copy only a subset of the
values. Moreover, the _offset parameter controls which part of each sample
we take. To be very precise, we will have:
- let IN_SIZE = IN.k()*IN.nr()*IN.nc()
- let OUT_SIZE = _k*_nr*_nc
- for i in range[0,IN.num_samples()) and j in range[0,OUT_SIZE):
- OUT.host()[i*OUT_SIZE+j] == IN.host()[i*IN_SIZE+_offset+j]
Finally, all this means that the input tensor to this layer must have a big
enough size to accommodate taking a _k*_nr*_nc slice from each of its
samples.
!*/
public:
template <typename SUBNET> void setup (const SUBNET& sub);
template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
const tensor& get_layer_params() const;
tensor& get_layer_params();
/*!
These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
!*/
};
template <
long offset,
long k,
long nr,
long nc,
typename SUBNET
>
using extract = add_layer<extract_<offset,k,nr,nc>, SUBNET>;
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_DNn_LAYERS_ABSTRACT_H_
|