Skip to content

P
PaddleOCR

s920243400 / PaddleOCR
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleOCR

通知 1
Star 1
Fork 0
P
PaddleOCR
切换分支/标签
查找文件 普通视图历史永久链接Permalink
文本识别实践部分.ipynb 205.8 KB
Notebook
Newer Older
L
reset  
LDOUBLEV 已提交
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 
{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "# 文本识别实战\n",
    "\n",
    "上一章理论部分，介绍了文本识别领域的主要方法，其中CRNN是较早被提出也是目前工业界应用较多的方法。本章将详细介绍如何基于PaddleOCR完成CRNN文本识别模型的搭建、训练、评估和预测。数据集采用 icdar 2015，其中训练集有4468张，测试集有2077张。\n",
    "\n",
    "\n",
    "通过本章的学习，你可以掌握：\n",
    "\n",
    "1. 如何使用paddleocr whl 包快速完成文本识别预测\n",
    "\n",
    "2. CRNN的基本原理和网络结构\n",
    "\n",
    "3. 模型训练的必须步骤和调参方式\n",
    "\n",
    "4. 使用自定义的数据集训练网络\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "## 1. 快速体验\n",
    "\n",
    "### 1.1 安装相关的依赖及whl包\n",
    "\n",
    "首先确认安装了 paddle 以及 paddleocr，如果已经安装过，忽略该步骤。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple\n",
      "Requirement already satisfied: paddlepaddle-gpu in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (2.1.2.post101)\n",
      "Requirement already satisfied: protobuf>=3.1.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlepaddle-gpu) (3.14.0)\n",
      "Requirement already satisfied: numpy>=1.13; python_version >= \"3.5\" and platform_system != \"Windows\" in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlepaddle-gpu) (1.20.3)\n",
      "Requirement already satisfied: Pillow in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlepaddle-gpu) (7.1.2)\n",
      "Requirement already satisfied: six in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlepaddle-gpu) (1.15.0)\n",
      "Requirement already satisfied: requests>=2.20.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlepaddle-gpu) (2.22.0)\n",
      "Requirement already satisfied: astor in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlepaddle-gpu) (0.8.1)\n",
      "Requirement already satisfied: decorator in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlepaddle-gpu) (4.4.2)\n",
      "Requirement already satisfied: gast<=0.4.0,>=0.3.3; platform_system != \"Windows\" in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlepaddle-gpu) (0.3.3)\n",
      "Requirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests>=2.20.0->paddlepaddle-gpu) (1.25.6)\n",
      "Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests>=2.20.0->paddlepaddle-gpu) (2019.9.11)\n",
      "Requirement already satisfied: idna<2.9,>=2.5 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests>=2.20.0->paddlepaddle-gpu) (2.8)\n",
      "Requirement already satisfied: chardet<3.1.0,>=3.0.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests>=2.20.0->paddlepaddle-gpu) (3.0.4)\n",
      "Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple\n",
      "Collecting pip\n",
      "\u001b[?25l  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/a4/6d/6463d49a933f547439d6b5b98b46af8742cc03ae83543e4d7688c2420f8b/pip-21.3.1-py3-none-any.whl (1.7MB)\n",
      "\u001b[K     |████████████████████████████████| 1.7MB 8.4MB/s eta 0:00:01\n",
      "\u001b[?25hInstalling collected packages: pip\n",
      "  Found existing installation: pip 19.2.3\n",
      "    Uninstalling pip-19.2.3:\n",
      "      Successfully uninstalled pip-19.2.3\n",
      "Successfully installed pip-21.3.1\n",
      "Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple\n",
      "Collecting paddleocr\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/e1/b6/5486e674ce096667dff247b58bf0fb789c2ce17a10e546c2686a2bb07aec/paddleocr-2.3.0.2-py3-none-any.whl (250 kB)\n",
      "     |████████████████████████████████| 250 kB 3.3 MB/s            \n",
      "\u001b[?25hCollecting lmdb\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/2e/dd/ada2fd91cd7832979069c556607903f274470c3d3d2274e0a848908272e8/lmdb-1.2.1-cp37-cp37m-manylinux2010_x86_64.whl (299 kB)\n",
      "     |████████████████████████████████| 299 kB 12.8 MB/s            \n",
      "\u001b[?25hCollecting lxml\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/7b/01/16a9b80c8ce4339294bb944f08e157dbfcfbb09ba9031bde4ddf7e3e5499/lxml-4.7.1-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_24_x86_64.whl (6.4 MB)\n",
      "     |████████████████████████████████| 6.4 MB 52.4 MB/s            \n",
      "\u001b[?25hCollecting python-Levenshtein\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/2a/dc/97f2b63ef0fa1fd78dcb7195aca577804f6b2b51e712516cc0e902a9a201/python-Levenshtein-0.12.2.tar.gz (50 kB)\n",
      "     |████████████████████████████████| 50 kB 1.6 MB/s             \n",
      "\u001b[?25h  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
      "\u001b[?25hCollecting scikit-image\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/9a/44/8f8c7f9c9de7fde70587a656d7df7d056e6f05192a74491f7bc074a724d0/scikit_image-0.19.1-cp37-cp37m-manylinux_2_12_x86_64.manylinux2010_x86_64.whl (13.3 MB)\n",
      "     |████████████████████████████████| 13.3 MB 56.1 MB/s            \n",
      "\u001b[?25hRequirement already satisfied: numpy in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddleocr) (1.20.3)\n",
      "Collecting imgaug==0.4.0\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/66/b1/af3142c4a85cba6da9f4ebb5ff4e21e2616309552caca5e8acefe9840622/imgaug-0.4.0-py2.py3-none-any.whl (948 kB)\n",
      "     |████████████████████████████████| 948 kB 62.9 MB/s            \n",
      "\u001b[?25hCollecting opencv-contrib-python==4.4.0.46\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/08/51/1e0a206dd5c70fea91084e6f43979dc13e8eb175760cc7a105083ec3eb68/opencv_contrib_python-4.4.0.46-cp37-cp37m-manylinux2014_x86_64.whl (55.7 MB)\n",
      "     |████████████████████████████████| 55.7 MB 44 kB/s              0:01\n",
      "\u001b[?25hCollecting premailer\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/b1/07/4e8d94f94c7d41ca5ddf8a9695ad87b888104e2fd41a35546c1dc9ca74ac/premailer-3.10.0-py2.py3-none-any.whl (19 kB)\n",
      "Collecting shapely\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/ae/20/33ce377bd24d122a4d54e22ae2c445b9b1be8240edb50040b40add950cd9/Shapely-1.8.0-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.whl (1.1 MB)\n",
      "     |████████████████████████████████| 1.1 MB 14.5 MB/s            \n",
      "\u001b[?25hRequirement already satisfied: visualdl in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddleocr) (2.2.0)\n",
      "Collecting fasttext==0.9.1\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/10/61/2e01f1397ec533756c1d893c22d9d5ed3fce3a6e4af1976e0d86bb13ea97/fasttext-0.9.1.tar.gz (57 kB)\n",
      "     |████████████████████████████████| 57 kB 9.0 MB/s             \n",
      "\u001b[?25h  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
      "\u001b[?25hRequirement already satisfied: cython in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddleocr) (0.29)\n",
      "Requirement already satisfied: openpyxl in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddleocr) (3.0.5)\n",
      "Collecting pyclipper\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/c5/fa/2c294127e4f88967149a68ad5b3e43636e94e3721109572f8f17ab15b772/pyclipper-1.3.0.post2-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.whl (603 kB)\n",
      "     |████████████████████████████████| 603 kB 7.6 MB/s            \n",
      "\u001b[?25hRequirement already satisfied: tqdm in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddleocr) (4.36.1)\n",
      "Collecting pybind11>=2.2\n",
      "  Using cached https://pypi.tuna.tsinghua.edu.cn/packages/a8/3b/fc246e1d4c7547a7a07df830128e93c6215e9b93dcb118b2a47a70726153/pybind11-2.8.1-py2.py3-none-any.whl (208 kB)\n",
      "Requirement already satisfied: setuptools>=0.7.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from fasttext==0.9.1->paddleocr) (56.2.0)\n",
      "Requirement already satisfied: Pillow in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from imgaug==0.4.0->paddleocr) (7.1.2)\n",
      "Requirement already satisfied: imageio in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from imgaug==0.4.0->paddleocr) (2.6.1)\n",
      "Requirement already satisfied: scipy in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from imgaug==0.4.0->paddleocr) (1.6.3)\n",
      "Requirement already satisfied: opencv-python in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from imgaug==0.4.0->paddleocr) (4.1.1.26)\n",
      "Requirement already satisfied: matplotlib in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from imgaug==0.4.0->paddleocr) (2.2.3)\n",
      "Requirement already satisfied: six in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from imgaug==0.4.0->paddleocr) (1.15.0)\n",
      "Collecting PyWavelets>=1.1.1\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/a1/9c/564511b6e1c4e1d835ed2d146670436036960d09339a8fa2921fe42dad08/PyWavelets-1.2.0-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.whl (6.1 MB)\n",
      "     |████████████████████████████████| 6.1 MB 3.8 MB/s            \n",
      "\u001b[?25hRequirement already satisfied: packaging>=20.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image->paddleocr) (20.9)\n",
      "Requirement already satisfied: networkx>=2.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image->paddleocr) (2.4)\n",
      "Collecting tifffile>=2019.7.26\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/d8/38/85ae5ed77598ca90558c17a2f79ddaba33173b31cf8d8f545d34d9134f0d/tifffile-2021.11.2-py3-none-any.whl (178 kB)\n",
      "     |████████████████████████████████| 178 kB 7.1 MB/s            \n",
      "\u001b[?25hRequirement already satisfied: et-xmlfile in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from openpyxl->paddleocr) (1.0.1)\n",
      "Requirement already satisfied: jdcal in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from openpyxl->paddleocr) (1.4.1)\n",
      "Requirement already satisfied: requests in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from premailer->paddleocr) (2.22.0)\n",
      "Collecting cssselect\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/3b/d4/3b5c17f00cce85b9a1e6f91096e1cc8e8ede2e1be8e96b87ce1ed09e92c5/cssselect-1.1.0-py2.py3-none-any.whl (16 kB)\n",
      "Collecting cssutils\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/24/c4/9db28fe567612896d360ab28ad02ee8ae107d0e92a22db39affd3fba6212/cssutils-2.3.0-py3-none-any.whl (404 kB)\n",
      "     |████████████████████████████████| 404 kB 134 kB/s            \n",
      "\u001b[?25hRequirement already satisfied: cachetools in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from premailer->paddleocr) (4.0.0)\n",
      "Requirement already satisfied: pre-commit in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddleocr) (1.21.0)\n",
      "Requirement already satisfied: Flask-Babel>=1.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddleocr) (1.0.0)\n",
      "Requirement already satisfied: flask>=1.1.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddleocr) (1.1.1)\n",
      "Requirement already satisfied: flake8>=3.7.9 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddleocr) (3.8.2)\n",
      "Requirement already satisfied: shellcheck-py in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddleocr) (0.7.1.1)\n",
      "Requirement already satisfied: pandas in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddleocr) (1.1.5)\n",
      "Requirement already satisfied: bce-python-sdk in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddleocr) (0.8.53)\n",
      "Requirement already satisfied: protobuf>=3.11.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddleocr) (3.14.0)\n",
      "Requirement already satisfied: pyflakes<2.3.0,>=2.2.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->paddleocr) (2.2.0)\n",
      "Requirement already satisfied: pycodestyle<2.7.0,>=2.6.0a1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->paddleocr) (2.6.0)\n",
      "Requirement already satisfied: importlib-metadata in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->paddleocr) (0.23)\n",
      "Requirement already satisfied: mccabe<0.7.0,>=0.6.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->paddleocr) (0.6.1)\n",
      "Requirement already satisfied: itsdangerous>=0.24 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->paddleocr) (1.1.0)\n",
      "Requirement already satisfied: Jinja2>=2.10.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->paddleocr) (2.11.0)\n",
      "Requirement already satisfied: click>=5.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->paddleocr) (7.0)\n",
      "Requirement already satisfied: Werkzeug>=0.15 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->paddleocr) (0.16.0)\n",
      "Requirement already satisfied: pytz in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Flask-Babel>=1.0.0->visualdl->paddleocr) (2019.3)\n",
      "Requirement already satisfied: Babel>=2.3 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Flask-Babel>=1.0.0->visualdl->paddleocr) (2.8.0)\n",
      "Requirement already satisfied: decorator>=4.3.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from networkx>=2.2->scikit-image->paddleocr) (4.4.2)\n",
      "Requirement already satisfied: pyparsing>=2.0.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from packaging>=20.0->scikit-image->paddleocr) (2.4.2)\n",
      "Requirement already satisfied: pycryptodome>=3.8.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from bce-python-sdk->visualdl->paddleocr) (3.9.9)\n",
      "Requirement already satisfied: future>=0.6.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from bce-python-sdk->visualdl->paddleocr) (0.18.0)\n",
      "Requirement already satisfied: kiwisolver>=1.0.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from matplotlib->imgaug==0.4.0->paddleocr) (1.1.0)\n",
      "Requirement already satisfied: python-dateutil>=2.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from matplotlib->imgaug==0.4.0->paddleocr) (2.8.0)\n",
      "Requirement already satisfied: cycler>=0.10 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from matplotlib->imgaug==0.4.0->paddleocr) (0.10.0)\n",
      "Requirement already satisfied: aspy.yaml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddleocr) (1.3.0)\n",
      "Requirement already satisfied: virtualenv>=15.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddleocr) (16.7.9)\n",
      "Requirement already satisfied: nodeenv>=0.11.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddleocr) (1.3.4)\n",
      "Requirement already satisfied: pyyaml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddleocr) (5.1.2)\n",
      "Requirement already satisfied: cfgv>=2.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddleocr) (2.0.1)\n",
      "Requirement already satisfied: identify>=1.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddleocr) (1.4.10)\n",
      "Requirement already satisfied: toml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddleocr) (0.10.0)\n",
      "Requirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->premailer->paddleocr) (1.25.6)\n",
      "Requirement already satisfied: chardet<3.1.0,>=3.0.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->premailer->paddleocr) (3.0.4)\n",
      "Requirement already satisfied: idna<2.9,>=2.5 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->premailer->paddleocr) (2.8)\n",
      "Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->premailer->paddleocr) (2019.9.11)\n",
      "Requirement already satisfied: MarkupSafe>=0.23 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Jinja2>=2.10.1->flask>=1.1.1->visualdl->paddleocr) (1.1.1)\n",
      "Requirement already satisfied: zipp>=0.5 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from importlib-metadata->flake8>=3.7.9->visualdl->paddleocr) (3.6.0)\n",
      "Building wheels for collected packages: fasttext, python-Levenshtein\n",
      "  Building wheel for fasttext (setup.py) ... \u001b[?25ldone\n",
      "\u001b[?25h  Created wheel for fasttext: filename=fasttext-0.9.1-cp37-cp37m-linux_x86_64.whl size=2584156 sha256=acb4d4fde73d31c7dfdd2ae3de0da25a558c34c672d4904e6a5c4279185fe5af\n",
      "  Stored in directory: /home/aistudio/.cache/pip/wheels/a1/cb/b3/a25a8ce16c1a4ff102c1e40d6eaa4dfc9d5695b92d57331b36\n",
      "  Building wheel for python-Levenshtein (setup.py) ... \u001b[?25ldone\n",
      "\u001b[?25h  Created wheel for python-Levenshtein: filename=python_Levenshtein-0.12.2-cp37-cp37m-linux_x86_64.whl size=171687 sha256=56b4a2de4349a05004121050df68b488ffd253dcc59187ca07b89b62d40c0218\n",
      "  Stored in directory: /home/aistudio/.cache/pip/wheels/38/b9/a4/3729726160fb103833de468adb5ce019b58543ae41d0b0e446\n",
      "Successfully built fasttext python-Levenshtein\n",
      "Installing collected packages: tifffile, PyWavelets, shapely, scikit-image, pybind11, lxml, cssutils, cssselect, python-Levenshtein, pyclipper, premailer, opencv-contrib-python, lmdb, imgaug, fasttext, paddleocr\n",
      "Successfully installed PyWavelets-1.2.0 cssselect-1.1.0 cssutils-2.3.0 fasttext-0.9.1 imgaug-0.4.0 lmdb-1.2.1 lxml-4.7.1 opencv-contrib-python-4.4.0.46 paddleocr-2.3.0.2 premailer-3.10.0 pybind11-2.8.1 pyclipper-1.3.0.post2 python-Levenshtein-0.12.2 scikit-image-0.19.1 shapely-1.8.0 tifffile-2021.11.2\n"
     ]
    }
   ],
   "source": [
    "# 安装 PaddlePaddle GPU 版本\n",
    "!pip install paddlepaddle-gpu\n",
    "# 安装 paddleocr whl包\n",
    "! pip install -U pip\n",
    "! pip install paddleocr"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "### 1.2 快速预测文字内容\n",
    "\n",
    "paddleocr whl包会自动下载ppocr轻量级模型作为默认模型\n",
    "\n",
    "下面展示如何使用whl包进行识别预测：\n",
    "\n",
    "测试图片：\n",
    "\n",
    "![](https://ai-studio-static-online.cdn.bcebos.com/531d9b3aff45449893b33bcb5dd13971057fcb4038f045578b3abd99fa3a96f2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[2021/12/23 20:28:44] root WARNING: version 2.1 not support cls models, use version 2.0 instead\n",
      "download https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar to /home/aistudio/.paddleocr/2.2.1/ocr/det/ch/ch_PP-OCRv2_det_infer/ch_PP-OCRv2_det_infer.tar\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/skimage/morphology/_skeletonize.py:241: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.\n",
      "Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations\n",
      "  0, 1, 1, 0, 0, 1, 0, 0, 0], dtype=np.bool)\n",
      "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/skimage/morphology/_skeletonize.py:256: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.\n",
      "Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations\n",
      "  0, 0, 0, 0, 0, 0, 0, 0, 0], dtype=np.bool)\n",
      "  0%|          | 0.00/3.19M [00:00<?, ?iB/s]100%|██████████| 3.19M/3.19M [00:00<00:00, 7.80MiB/s]\n",
      " 14%|█▎        | 1.20M/8.88M [00:00<00:00, 11.7MiB/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "download https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar to /home/aistudio/.paddleocr/2.2.1/ocr/rec/ch/ch_PP-OCRv2_rec_infer/ch_PP-OCRv2_rec_infer.tar\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 24%|██▍       | 2.15M/8.88M [00:00<00:00, 10.8MiB/s]100%|██████████| 8.88M/8.88M [00:01<00:00, 6.38MiB/s]\n",
      " 17%|█▋        | 249k/1.45M [00:00<00:00, 2.42MiB/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "download https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar to /home/aistudio/.paddleocr/2.2.1/ocr/cls/ch_ppocr_mobile_v2.0_cls_infer/ch_ppocr_mobile_v2.0_cls_infer.tar\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 90%|█████████ | 1.31M/1.45M [00:00<00:00, 3.32MiB/s]100%|██████████| 1.45M/1.45M [00:00<00:00, 4.53MiB/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Namespace(benchmark=False, cls_batch_num=6, cls_image_shape='3, 48, 192', cls_model_dir='/home/aistudio/.paddleocr/2.2.1/ocr/cls/ch_ppocr_mobile_v2.0_cls_infer', cls_thresh=0.9, cpu_threads=10, det=True, det_algorithm='DB', det_db_box_thresh=0.6, det_db_score_mode='fast', det_db_thresh=0.3, det_db_unclip_ratio=1.5, det_east_cover_thresh=0.1, det_east_nms_thresh=0.2, det_east_score_thresh=0.8, det_limit_side_len=960, det_limit_type='max', det_model_dir='/home/aistudio/.paddleocr/2.2.1/ocr/det/ch/ch_PP-OCRv2_det_infer', det_sast_nms_thresh=0.2, det_sast_polygon=False, det_sast_score_thresh=0.5, drop_score=0.5, e2e_algorithm='PGNet', e2e_char_dict_path='./ppocr/utils/ic15_dict.txt', e2e_limit_side_len=768, e2e_limit_type='max', e2e_model_dir=None, e2e_pgnet_mode='fast', e2e_pgnet_polygon=True, e2e_pgnet_score_thresh=0.5, e2e_pgnet_valid_set='totaltext', enable_mkldnn=False, gpu_mem=500, help='==SUPPRESS==', image_dir=None, ir_optim=True, label_list=['0', '180'], lang='ch', layout_path_model='lp://PubLayNet/ppyolov2_r50vd_dcn_365e_publaynet/config', max_batch_size=10, max_text_length=25, min_subgraph_size=15, output='./output/table', precision='fp32', process_id=0, rec=True, rec_algorithm='CRNN', rec_batch_num=6, rec_char_dict_path='/home/aistudio/PaddleOCR/ppocr/utils/ppocr_keys_v1.txt', rec_char_type='ch', rec_image_shape='3, 32, 320', rec_model_dir='/home/aistudio/.paddleocr/2.2.1/ocr/rec/ch/ch_PP-OCRv2_rec_infer', save_log_path='./log_output/', show_log=True, table_char_dict_path=None, table_char_type='en', table_max_len=488, table_model_dir=None, total_process_num=1, type='ocr', use_angle_cls=False, use_dilation=False, use_gpu=True, use_mp=False, use_pdserving=False, use_space_char=True, use_tensorrt=False, version='2.1', vis_font_path='./doc/fonts/simfang.ttf', warmup=True)\n",
      "[2021/12/23 20:28:48] root WARNING: Since the angle classifier is not initialized, the angle classifier will not be uesd during the forward process\n",
      "('SLOW', 0.9776376)\n"
     ]
    }
   ],
   "source": [
    "from paddleocr import PaddleOCR\n",
    "\n",
    "ocr = PaddleOCR()  # need to run only once to download and load model into memory\n",
    "img_path = '/home/aistudio/work/word_19.png'\n",
    "result = ocr.ocr(img_path, det=False)\n",
    "for line in result:\n",
    "    print(line)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "执行完上述代码块，将返回识别结果和识别置信度\n",
    "\n",
    "```\n",
    "('SLOW', 0.9776376)\n",
    "```\n",
    "\n",
    "至此，你掌握了如何使用 paddleocr whl 包进行预测。`./work/` 路径下有更多测试图片，可以尝试其他图片结果。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "## 2. 预测原理详解\n",
    "\n",
    "第一节中 paddleocr 加载训练好的 CRNN 识别模型进行预测，本节将详细介绍 CRNN 的原理及流程。\n",
    "\n",
    "### 2.1 所属类别\n",
    "\n",
    "CRNN 是基于CTC的算法，在理论部分介绍的分类图中，处在如下位置。可以看出CRNN主要用于解决规则文本，基于CTC的算法有较快的预测速度并且很好的适用长文本。因此CRNN是PPOCR选择的中文识别算法。\n",
    "<center><img src=https://ai-studio-static-online.cdn.bcebos.com/0e74d46918e5423684e06df8e6eb992cf9f7c485d6c142fb98cc263714559898 width=\"600\"></center>\n",
    "\n",
    "\n",
    "### 2.2 算法详解\n",
    "\n",
    "CRNN 的网络结构体系如下所示，从下往上分别为卷积层、递归层和转录层三部分：\n",
    "\n",
    "<center><img src=https://ai-studio-static-online.cdn.bcebos.com/f6fae3ff66bd413fa182d75782034a2af6aab1994fa148a08e6565f3fb75b18d width=\"600\"></center>\n",
    "\n",
    "1)backbone：\n",
    "\n",
    "卷积网络作为底层的骨干网络，用于从输入图像中提取特征序列。由于 `conv`、`max-pooling`、`elementwise` 和激活函数都作用在局部区域上，所以它们是平移不变的。因此，特征映射的每一列对应于原始图像的一个矩形区域(称为感受野)，并且这些矩形区域与它们在特征映射上对应的列从左到右的顺序相同。由于CNN需要将输入的图像缩放到固定的尺寸以满足其固定的输入维数，因此它不适合长度变化很大的序列对象。为了更好的支持变长序列，CRNN将backbone最后一层输出的特征向量送到了RNN层，转换为序列特征。\n",
    "\n",
    "<center><img src=https://ai-studio-static-online.cdn.bcebos.com/6694818123724b0d92d05b63dc9dfb08c7ced6c47c3b4f4d9b110ae9ccfe941d width=\"600\"></center>\n",
    "\n",
    "2)neck: \n",
    "\n",
    "递归层，在卷积网络的基础上，构建递归网络，将图像特征转换为序列特征，预测每个帧的标签分布。\n",
    "RNN具有很强的捕获序列上下文信息的能力。使用上下文线索进行基于图像的序列识别比单独处理每个像素更有效。以场景文本识别为例，宽字符可能需要几个连续的帧来充分描述。此外，有些歧义字符在观察其上下文时更容易区分。其次，RNN可以将误差差分反向传播回卷积层，使网络可以统一训练。第三，RNN能够对任意长度的序列进行操作，解决了文本图片变长的问题。CRNN使用双层LSTM作为递归层，解决了长序列训练过程中的梯度消失和梯度爆炸问题。\n",
    "\n",
    "<center><img src=https://ai-studio-static-online.cdn.bcebos.com/41cdb7fb08fb4b55923b0baf66b783e46fd063223d05416fa952369ad20ac83c width=\"600\"></center>\n",
    "\n",
    "\n",
    "3)head: \n",
    "\n",
    "转录层，通过全连接网络和softmax激活函数，将每帧的预测转换为最终的标签序列。最后使用 CTC Loss 在无需序列对齐的情况下，完成CNN和RNN的联合训练。CTC 有一套特别的合并序列机制，LSTM输出序列后，需要在时序上分类得到预测结果。可能存在多个时间步对应同一个类别，因此需要对相同结果进行合并。为避免合并本身存在的重复字符，CTC 引入了一个 `blank` 字符插入在重复字符之间。\n",
    "\n",
    "<center><img src=https://ai-studio-static-online.cdn.bcebos.com/bea6be2f5d9d4ac791118737c3f2f140a2a261e0d8a540a3b0ef239b6bcb2c43 width=\"600\"></center>\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "### 2.3 代码实现\n",
    "\n",
    "整个网络结构非常简洁，代码实现也相对简单，可以跟随预测流程依次搭建模块。本节需要完成：数据输入、backbone搭建、neck搭建、head搭建。\n",
    "\n",
    "**【数据输入】**\n",
    "\n",
    "数据送入网络前需要缩放到统一尺寸(3,32,320)，并完成归一化处理。这里省略掉训练时需要的数据增强部分，以单张图为例展示预处理的必须步骤（[源码位置](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/data/imaug/rec_img_aug.py#L126)）：\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import cv2\n",
    "import math\n",
    "import numpy as np\n",
    "\n",
    "def resize_norm_img(img):\n",
    "    \"\"\"\n",
    "    数据缩放和归一化\n",
    "    :param img: 输入图片\n",
    "    \"\"\"\n",
    "\n",
    "    # 默认输入尺寸\n",
    "    imgC = 3\n",
    "    imgH = 32\n",
    "    imgW = 320\n",
    "\n",
    "    # 图片的真实高宽\n",
    "    h, w = img.shape[:2]\n",
    "    # 图片真实长宽比\n",
    "    ratio = w / float(h)\n",
    "\n",
    "    # 按比例缩放\n",
    "    if math.ceil(imgH * ratio) > imgW:\n",
    "        # 如大于默认宽度，则宽度为imgW\n",
    "        resized_w = imgW\n",
    "    else:\n",
    "        # 如小于默认宽度则以图片真实宽为准\n",
    "        resized_w = int(math.ceil(imgH * ratio))\n",
    "    # 缩放\n",
    "    resized_image = cv2.resize(img, (resized_w, imgH))\n",
    "    resized_image = resized_image.astype('float32')\n",
    "    # 归一化\n",
    "    resized_image = resized_image.transpose((2, 0, 1)) / 255\n",
    "    resized_image -= 0.5\n",
    "    resized_image /= 0.5\n",
    "    # 对宽度不足的位置，补0\n",
    "    padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32)\n",
    "    padding_im[:, :, 0:resized_w] = resized_image\n",
    "    # 转置 padding 后的图片用于可视化\n",
    "    draw_img = padding_im.transpose((1,2,0))\n",
    "    return padding_im, draw_img\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAADbCAYAAAB9XmrcAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDIuMi4zLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvIxREBQAAIABJREFUeJztvWmMJdl1JvadePGWfLlUVlZW1tq19UKySZFNuqdFLRBkamhTHMM9HghjcowRfxDogS3CkiHDwxkBhgQY8MiQJXsAQYP2iB7OQBDFoehhQ6BHpqjWaCRITTa3Xtnd1dVLVXXtWbm8fPkyXry4/nFvvPNFVUQuXVmZlU/nA7LqZmTEjbtEnLjnu2cR5xwMBoPBsPcR7XYDDAaDwbA9MIFuMBgMIwIT6AaDwTAiMIFuMBgMIwIT6AaDwTAiMIFuMBgMIwIT6AaDwTAi2BaBLiKfEJFXROSsiHx+O+o0GAwGw9Ygd+pYJCI1AK8C+DiACwC+DeDTzrmX7rx5BoPBYNgs4m2o4zEAZ51z5wBARL4E4HEAlQJ930TdHZ5plfyFPy5SfrgUUlrkX4TqEOHz6Rw631W0JQrnS6THBmlWcX8FfzeLH9GKPlO7nMtKj5chElW6Cv2ke2aFxpTfnk/hO3K/3fBcPZnvGRXqc1TOSo8zpHJO3x1kg3Hzbam8muqpOn/9h3Rz9y+fjKpLy56jwrk8zoWaK57zqi5UNqDi/GH7Nj55M+NSaGJVnfwe5+8ovQuFd6jYgvK28PtC9xFHY1fV9OIk3NZWfkf5eFYy/1fmu1jsrG36DdgOgX4MwHn6/QKAH731JBF5AsATAHBofxNP/vJHwh/oRRcadEeTUTIXmatx3VQmgQY9p0bnx7WGlmMdAq4ny2hwI62z2WwCABoNrWPh5pKeWysXqIPBYFhO+1TOtHMx3cfF2t5ev19af+Ty5lH7Ym1XvV7X+ug+a2tr2q6UHjSqZ5CVC+l66D+gAiUZpMNjjbqe24y1vmyg90ySZFhOU72WUfgwROXMIAu0/JwqAcF1SOFjqXUMCvKx/Pyqa7Ps9oe0qg+bqWMz/ednKnN+HBsxHaMx5+eZ6yv0Py0XurWaPossxQof6ZJ3dEADym3lvsU1fUar+N9MLy3UU1hIUBsbDf+M8vPPz7yj+9ciff+5nzwX/HxLRu80ywtavcRNEqt5452+w02SHSz/1ui9yOXb537jz7AV7NimqHPuSefco865R/dN1De+wGAwGAxbwnas0C8CuI9+Px6ObQ4FvWUz35fN699CX/A6rcR5MdVZ0S/3zeXOsDy/sDgsr/X065qvLvjrPzczPSy3Wkoltalci3llLaXlwgqRfmm1xvQ4ra7ykiNNxJGG0lnVlcX8/PywfGN+YVg+f/6CXkur0l6vNyyz1jM2Pj4s52PQHpsYHps9sG9YPn360LBcozZKTcclAq0oaWKywupPy7WIV7R6bSPML6/gBjz/TZqLSOcuHeg9azV9Ror9p/kiFZmnK29i1SqbV7C84uTnKBvQKp5ehWadtKKBPospaW45FZb2qT+krdVAK86+9o3BmqvQu9inMYqpLYOU5o4anPcpprEYsFZY0Ep1zGN6/njl3hjTuYsbtBikd6fPGnCgQPnY1KQ+l91ud1guaJ90z7VU29sg7b7Z0v6zVrBKYxFFtNIPWpJEOrbjdX7+dA4HTJEFTXyrbON2rNC/DeBBETktIg0AnwLw1DbUazAYDIYt4I5X6M65VEQ+B+CPAdQAfME59+Idt8xgMBgMW8J2UC5wzn0dwNe3dE2JMlG5EV2yQVXj3elNWMSwKhyRmrecrA7L//HZ7w/Lr72htARp68j38GJq6+FpVcP+wd//e8Myq5yrKyvD8sTUpN5/Ve8vFVQAEqIc6kR/TEwBALqrqh6+cvbNYfn1N5VOuXZNKZf5JaWW1hLaLE2UonEZqb91VRcjLGu7goJXJ9V+fFz7cN/Zt4blY0f2D8snT54elg8cUFqG6YTuirYxog2lRoM2Wvt0Tl4Ht47m+a033xmWf/iatqtHG4EDunp8QmmuWqRj9MiPvH9YnppU+qlWC5RHTze2WmPtYfkczUXS1/vcXNA+MM0Q11VtT3tKEcxM67Nz4ujhYbkRzl/p6hw2mTZZ0zEcb2u7mX9kKiImmi8lyuHmTX2O37pwWe/f1PPbY77+5SWlLacmdSyOHdF283O+r6XtunbpujZxoM9cOzzzQFGGEOOBOGyKRrG+lys8zwOiooijWCO6yhH9NejTe0HUUb2h9Sf0HrXaKjDGQhuErVzWiIoiOpfYR7Ta/lmItki6mKeowWAwjAhMoBsMBsOIYFsoly3DSYFG0eMVfEnBEqaEqqlwbGEUrR+03B/o+WtUXiXGIyXb09yet05/H4gOI6t+NVKz4gpbYra4KTgF0U752KSq2RlZv1wJNMoLL782PPbCy68My4srSuf0iUJZ6qialxWcL8hunyw+aqR+shVF0gv1O1VDx5a1fQvzl4bli5dU5b50Te32T59W+uXo4SPD8sSYqtbiyD53oH2qk8VFHNrryMlrNdF+3pjXe56/qFTBIk10WqDltDwxqVTI0aNHh+VGS1X3Vm65QDp8QnPVIzrrP/zlt4blzgrNBT3+dZr/Bln2HJ5V6qpPdZ457Q3NGg0d5xr5BLCNOVtFRdA6IiF780zLffKb+OGrSle99EN97ro6RRgfb4W69Z5nTh4bltsTahU1VtM5fOGC1v36q68Py/PzN7VPRKMwRdIlqgth7CKiMJvkP1Er2KxruU5zV2P/GLBVDp1DlB6xMqhTG3OrrFpN6xuje0bkN7D/gL7nx07652xQyUOXw1boBoPBMCIwgW4wGAwjgt2hXO4AOb1SRa04omf4FHZmYKcNVmiywvlajstiX5B2ys4+CXEuZKhQsGAZsPswO7+QKujqRHnUddf8yk3d/X/5lbMAgO+9oFai5y+p08i4shZojakVQkQqdI3oooToioQcKwaJ1pmShUaulTbJ2YPdntdIC75xQy0o+skbwzJbH3E4hfYxtYSoE/3TJ4ufmNYjLqcLqD/8AHDfVla1YV012kBK88/OP1FNaRF2shF+fcL9M3aNp3v21lQnX+1pmVgDkCEKIqIF2ZP8yoJaX83d1PLp++/3hZjd4ak/YPqFrDn4JSEKwQk7x1BoB5ojbu+CsiLorfrnpU3+cDwuDXqea2RCdnFFLYFevHZtWH7nkt6oPaGUW0QUSa9H1Fled40o1Kv03NB4MhNaiPdUHvmjEJ+Iz+EwF/WCQ9kgHNNz2bKGy2fOHBiWxw54Woqp4s3AVugGg8EwIjCBbjAYDCOCe4py2VQo1ZJzK+kX5lOiuLQsNY7BwOdQPRxtL+ygC0eGJA4n5WhwHEuDdC6Ot8J6cUyqKFMxlyn2ynef16jEL778MgBgcZmsTNSAAHWKgcH8z9iEHmcqKCariYQsRAoxVlgDDMedK7/uyAF1FOmvqaq8Ss4Zly5f1Xa1tfExRZs8RPFh6g0ao57yJUPLDR5z0ok5BglbEEXkIdYkykvIyoXjd9RE6x+Q5VDucJbRsYiooojipHCMlSZRaxlxVDE9IzGp80tdHcc3LqgV0dwRHz7pxFG1FCrGFCGLmwFTbhzWmKMgMv1CJhzUrgL9EJWU+R1y5ZRPj96/LlV4g2ieeY6DQ68oO0glNYrgGWixjBtAIWBSppP42abIn3y8xhFOWRYQLeaIdmTqrh7uxXVkFEuH2DesUp+zcL6rkG1VsBW6wWAwjAjuqRV6VcR4DsKvq3G2nwb9Xcu8QC+s6GnFn9HXesBlurbPG63Du0vp39do02xQsCvnjRKKx0yrNcT6lV9Z1k3El147Nyx/+7kXhuXLl/0Xfb/upUBoJXBtXlcC9YaW9x/Q3dIzZ+4flotx1bUthbjqNI5Li961++JFDa554/KNYbm3pn3IaDOV84HcXFS37jff1k0xKcSsVlv140e0s31a0Q7C88Ab2xnHo6dd7B7tRHZWefeL47rr4ZhshXtr5RudcP4CXrVLg1bopH0t06ZskmkdvUTvMzamWsE4bUT2Ul2hX7immtv06z4lwez+2eGxJoVsqJG9d5rqszCgVWkh8qDoALC2styh1SVtehd061CNI7+OlEOJ8iqfNJfWlEYtrVN4DCyqDwFPV7+nY9HhjeagJLJhA49nOiiPwV+niIgSU9tJi+EVMNu5X1zUZ52mHe1QHqPXvMaCiQTNSkqhLxKvxWZmh24wGAx/M2EC3WAwGEYE9xTlsqmNzqEd+sZ1uEK6KD2H6RK2j03J95pMteHYnz+cE0Xl6iRvxAhtsg5ItWZ3XqnxOeTWf12pixd/qG7Q1+ZJRQ6XcmS8VYreyHbgswfVJfx9731oWH7Pe7RcI8olpqiODWojuz7n0fmuHlc1/+pltR9+hTZwOe0eu7inNBecYEQuaHTEMYpeN0HlmFziJVA6hdSBUj4XvBE9oA1dtivu6b4tyGu8sOnpCnbo9dvqc0Qz9chou0mbeT2iMAb8LFLdfaYXKcJfKjqm74QwEFdvaITDCYr2WKeQAPxu9VKlCpg3KdjT0/2bLa1nrK1t6ZHfQr5byu/FoLDhqteNtbW+QV/raBPnRYEvi2l3aVNyokkJMSZ8/Y5TTVJ5uaOb6Rw+gelPHiNOmcjhOZrjHGWUU9nRpnPgfVpE4bToQatR2IzxSTUKKIuMshnYCt1gMBhGBCbQDQaDYUSwe5RLhUXL8M8F+qW2wd85wUV5vcJ5N6mckdrKFi+cAL0QETHUT9o0BhwagGx/QQkL+mSfzeEBoqaqbQmp5RcuXtHyeY0OWGM73BDtL00KNgZDPHhaI/N9gBIz3H/mhJ7E6h/ZG2cUPq5PbtUgm+jxoBafOqqUy5EZtRknL3Scf0ttpq+Ry3qNVOGMKLKFRVWLL76jtuqTlNP06GG1eMkj4gnZNnHShVqD3c35ONmykzpdI1osqrENO6vlZAkS1kZMwxTc5zkhC1lHJAOlPNKC5ZS2ZYHy3jItwhTR+cuecnn1rIZVmJ5QOmNuVi1IJNaxyECRLOlBL9BJNKbc/4yic6acpyG8PMT4obuq1NJKV/sTNZUifOSBB4blg2TlstLVa6Ma5xSl+aI5nWj7Z1Do2XrjDY3kePZ1HaN3LilFmBUS5VCCC6pn//TMsHz0pIanOPOgRpN0EeUJDdxdLdNjY0QVRpn2bXJMx/zonL9PnazNNgNboRsMBsOIwAS6wWAwjAjuKSuXKmzG+qUcZOVAXIWQGhNxpgqmS4oeBNqWcLorOCfxuaRykyUMR+lb61OOzJidRlRJvXZDc4AuUtKIfdOqWsbBiWhxQWkDMnjBA6dODcsPndZyg3bckzW9dozonzpZGQxYnyZrnTyqXNykPKcU3P/973vPsNzpaN+uUsKClBxbYrLgWCNaanFFaYkb5GRy4sR92q5AnWWcyIT90SO2bNHDHIWR81uOkUs+J/VgaxWmyGphrpnaaxLNw0kd2OInobqjBs0tRSFkK4sa0zxEF64seguh1986Pzx28rjSAFNT6kzWJAsmppky4nNcxdjxGPVp7Nj9pR6erzpFOxwj65g6hVLgCKODRJ/FuTF6R4k64iQcqzT+TBHFgTrkHLnvowQbS1eUwrzeZ6swSmRBMmKC2v7gcQ2t8MEPfXBYHp+m9taIugymZkLWT216/7KC8x3l6x3m161I+lOBTa/QReQLInJVRF6gYzMi8g0ReS38v3+9OgwGg8Fw97AVyuVfAfjELcc+D+CbzrkHAXwz/G4wGAyGXcCmKRfn3J+LyKlbDj8O4KdD+YsA/gzAP96oLpFi3BBFxfdlA4uYYuV0WcGygJJQEOXBUQ2ZWmHHojrdvxmiAGZUB1stLHdUhTpAjgJ12oV3tFOfkkp16bKqgpeuqJUL+W9gktS/pWVPXfA++IeI5niYrAbGOZYMUSgtihMSUWAJToLRKET7p2JQC6OC35U+UpOkKt93QlXeRXIgunBJ+8ltYSrmyhVN6nF47uCwfJWcr44dnfP3p8ZwTk+huCYRRV4cOFVzmy2KTkkOShyzg+PQtCeUxnDBsYafa3by4v40OfPDTaWQOGxHxFkY6Nq1AdN1eq9W27eRc+Q+96I6dk3tU6uR+08qVZV1dC5SzrXLOXVp0scpv22dPK6aLUogsuJphsYEOaFR4JdCjCXqT5OeP459kmZ6LTsLNjiCKdcq/vwxygzSWVGHqzblqB2nwCrEBBWSZ6QUP8f19P2eIuoqpiQwklFymDB3UcZxfyhmUCFqK8mrnKrJ7hLlUoFDzrncHu0ygEN3WJ/BYDAY3iW2zcrF+Z3Lys+JiDwhIs+KyLMLlHXeYDAYDNuDO7VyuSIiR5xzl0TkCICrVSc6554E8CQAvPfE1Nb0iG0AW8cUElZQmdVljqXA5+QKEgf3zyoYIbYaSAtOG4o+WWWw9QvnEmxyGFaikVrBomRqTI8d3K/70lOU4KJODeavuJAqGJHVRI2dSeieEauI+THuUIE30DvNzqhDxoED6hD0zhV17FijxBeNQswSrefGgqrOa+z8FOgVTqRQyBfLFk9STsXx/PM4VzmuFY6Hsqvw/KlxKN9+Qpfxc0mxT3j+idqYmdD5rZP1yfJNP47drlIo80tavnxV6alDh1SRbo6po1YGpRbYgqVOVNigML7lyTZzQxRX4hB4G6jPQpRLzHFl6Zljy7FCvBk6J3f4GlDdnOul2aJ5iSn2TiGqLluocYhjpejWelqeIIos4orC+5URtRTTM8LvHMfSdcPjO0u5PAXgM6H8GQBfu8P6DAaDwfAusRWzxd8H8FcA3iMiF0TkswD+GYCPi8hrAP52+N1gMBgMu4CtWLl8uuJPP7P12wpKvyUVOUXLG7S1XHusflfRLMUynb/BZy8j2oLPHZCqNsjYCULB1g+9Xq/0OGcPEhqjPO/kfnIaObhfqY1xoi1cpup0RNYMINW+QLmwNu2YiqG8myVOXoU0rvTbvn3UxjmN/VJ/TfvWo3i/LQoxK+R8skA0QmdFxyt3IolqrIZzqFtuGMUAobpjUpszCubDar5U0HUiefheHauCvxk77XBcD8r1usY0Hs1/m+b3+PHj2l7q37lVr/4vLWqcnMVlreP8RY2lc/SYOsccOaT0V8T0H1lCNSnbEtMcGT2LBSIubxYxLpmUU458XUq/ZRRXyBGn5wphfRVlc81xmhxZpAwoA9WArFwoGRUyUfqPAgyjQxYsnD0qypSWapCDYJbLAM5RyhnYQP3k9ob+y1ZkIsz132AwGEYGJtANBoNhRLArsVycc4X4LDt+/woLFlbXSRMvJF6W3KSFVciKvnA4UrZaiKjyPmV6WSWnBU5kG0dKnXBcj8EgOFC0VN3jjD41UlXX1tiygmgB3pHnfhQSb5dbYuS0VyRFcmH4Z+pznWLWTJHDEbeXY2/wPQtUCI3pcldjf/SC5ciYEM0Ulc9LIQE1zQWXbzF50MNs/SO3W7Rw+F4ew5ieLQrlgVaL3MJSirFDljCtWNt1+ODcsFyniq6+8zYA4CaxaRy+9vINdc46T0m9p/cptRVT/J4+xYdOiP5hLq6QkJvulVu8sAVRVcJ2tjhjHxqeOzasGuB2K6vb6gnPN89+j2iuNaJN1ugkpryY0e2TvFijbEQJJRXPChQwZ88ehD4w5cQWZHycaNmC9cvmYSt0g8FgGBGYQDcYDIYRwZ4In7tVi5YyZGSpURWOt9LKhZ0fohLLDlKnOAJvRmo7Wz/EjXIKgR1rOEsSZwlKOTxs3iemh+hcVk+Z8qnFZE3AY8GWGGSjwcqfFBxr8pPLrUlqFFZ3jXb+YxrDcaJfFlco9gk5YmSFwdDO5kmqAaAfnLLGyLKHUT3neo4U5pyvpaxWbNFUQr8M6NyYqZ2IqRgKu0sjVmMLjgGfo9dOjWscmHGiq2ZnfEaiyw2aH6p7mWK2vHn+7WH5wIxa0LD1S4M8cdYoZG6tziGmq5y1iv8DtzyLhSeKkqSz9QezWVn5s8hwRZsXf664W47c3pZ+4TjVQG1PqZw4toohxza6NqLfslArO03VCqGR3G3n+vbmtJFZuRgMBsPfSJhANxgMhhHBrlEu5ZYhd06tVIEpByflmU6KyXD1aC263cql0HpS7QrxYHg3nx2CiJdxpMJlKVuiUP2F5DGc7Nq33RHNIgXHGqqk6jihYD/AFEVV/JIwFqzacqLjel2tJrI1dQJip6UWxSnheCd9jmtD5YjuxY5YRVrmdhQSIFfMS2WZVGtXCIPKdQbVmi2IaOLIrwWOLJiylGkxPYfpFxBdF1F5kuiX40ePAgAuvK3JkFc7GmuEnbYuX9WQS2/Q+e1JtXiZPaDxXgZEucQU10XIyqbMcqwQS4fGn8sFqoaolVrG70g5dcJgijA/h99szlLGWZJqFSKHLWvY4Imfy1XKNrS2T6lDkEVN7lAlVCG3K6qiPF0em6i8fVWwFbrBYDCMCEygGwwGw4hg96xctsFyZStgOmNAYT2zrPybVoiMyupkrhYVrCDYgoIrKQ+lGtfL46EUaAMyJuBMMqDkxbVAaQg5nhTijpBKzM5RacE6g/pBTS86StBfOAtQPjBshcD0E3EIWYWjDluNsJNVo6kqrCO9uEdqLse7yeeAKS+2moBjKoTGk+ORMC1A7WUqpBiThywXSqxcCs5RNC9N6k+XMyPRkHOclmygz25vVa1V4kidjOZCJqczZ84Mj517/bVheYEol+6q9u3tCxeG5f37p4flqUkN08sJqxtEuRQcsQg5zVBFswxQTrnERLPEFfNSc+XvVIFGDM9ARscalLGqJfoOUVTdQvJwfr9jfi0pefmAnr+E49CQ81FursYxWQY05yyA2ZqJqditwFboBoPBMCLYpRW63LICXB8bfXWqauLrOBdkxgu0im0WvjbiTcFQrmq/0IZnIQA/22rTCoE3ZQcZ2/VSbkraRKuRe3ae4MBFvPpWOOrFoLhDNSzGhZ6WR8QrjAZ1ezDcuKHVFC+KKUxB1aYlbzL1KWFFa5w0ESlfofPqfnjPQnfK55ZXP4VclLy6LmyKsoZCq8usZBOLtBVKr1pYQTYKG4sU4oEXdhw9kN3WKYGFkNY51fZ1njqptuSXLp0bljsahLGw0XZjXsMnXLysSTCO3KcbzpMNzY2b1bTtTmjTuxDtsOSNrVhlF1aoNHacYCQqvEflER75bdT3jq/jTWZ6/2jMo4oKWQ+J6QGL6X3lNhYewmF7yzd2s2LD6fbvjsGwFbrBYDCMCEygGwwGw4hgVyiXKIrRnvSB9ZeX1VZ2bEztahvs7s4RBkMUukaDqA3OLZmoqhjVtHtpn2gGsn3OKKr9eF034lqRqpZJV+vP25hGqm46UuEWVrStE+PqVu3IJnt+STf/0oFeOzauyQaWV94clgdjtFlElE4SNss6ND7dNVXP0wFtOIGSZLA2R4kMeEORwXRJn+mH4MNcoIFo86lLduI1GvN0maLdEZ3kaON2jaIN8mbpGIUKyEjpTgb+nP5A78mJMWoNUo/ZKLx2u7s1AMQcKZE20WVAzyVFJMw3MeucVIQGOiGqJmrqGGVOKSQex5SoKKFQAWP0xq5156ntvv6jh9SW/NABHavuotIpFNQTy0tafvuC1jdzSMuzqb6X8x0doxUKQ7iS6Fg0o9BI2mWMqP/7W2SznVCIg0K0Sy6SfwY/fsL0y+2bjj3aZC1c15wclnv0jmRMXSba3iY1a9CjxB/0XDT7Wo6IinF5rluic4R+4fARmWMfkxJ/l03AVugGg8EwIjCBbjAYDCOCXaFcMpdhZdWrxpeuXhsej+j7wrbafXIbR6BXpvfpzvu+KVWh6g21LE1p57nGqjXvlJMtKduVUkrDQtS83FWXkzdwfQXagkMvsv0JnROT+h3XymmRHkVhbNA5UaArVlbUUqGzTPr0LFkWEP006Gl9TW5jVdIOzsHJ+R0DzTAgFZIpL97Br5NlT5X98oDyeLKNOdNlrTqp9kTj5OPOlifCCS4KLtZZ6XFGRLYNBQsGTuDBiRfyw4MCJ0C3KbfO4Ah/PCq80mL6h0MLxBydMISzYLf2U6ePDcsrFG3xwts39Z6UX2O5o3P3xluagzRqq306Yn2/YvIV4PfOBYpsQPRXn57hpKcUSoP4jIyoMEd+ANWrThpTNmIJ8zUouNJzxEa954BpDrZOKVi80PvNMoVDNbCFTsZ+LoPwP9XBFiyFhMVsfne7pc5msOkVuojcJyJPi8hLIvKiiPxiOD4jIt8QkdfC//s3qstgMBgM24+tUC4pgF92zj0M4KMAfkFEHgbweQDfdM49COCb4XeDwWAw7DA2Tbk45y4BuBTKyyLyMoBjAB4H8NPhtC8C+DMA/3i9ugaDDJ1AE1x4R1W7t85p4P2Y3Nn7CUWbC/+fOK5uzw+/9z3D8v79qiAknJczLlf565HuWtdJu2nSyPRp17oR5eocO0qoCtmMKI9mjSgMUpU5euN4S9sy0dabkhEPaqTbc7ty54N+TxNDrHbUbIEjvI21lJ7oJUrLsEs6O1llFS7WtYJqG1y8B7xTz8kj2GpD+8/5TQc0txzhjum3OqmlbHEy1VbarRUcXiIOTcDJI5iKSfm43jMuSV6yHorOUqHMneCxqnBsK5xecFQqv2dB5S/kepXbbnTi+PFheZkciG7eUPolIyunZaLrzp8/PyzvO3hwWG5Pz+g9abxqZN2UBcqFnfkSnn/6wwQ9l4UojOxMxGNXoJn4nJKxq3BIqgKfwhRO1ap3q1E7dwLvalNURE4B+DCAZwAcCsIeAC4DOFRxmcFgMBjuIrYs0EVkAsAfAvgl59wS/835z1HpJ0lEnhCRZ0Xk2aWVpOwUg8FgMNwBtmTlIiJ1eGH+e865r4bDV0TkiHPukogcAXC17Frn3JMAngSA+49NuTSoSMsdpQsWVBMsRMRLOWRHODy5T3fQBzV1fGhNqHNOSpHpslW9T0x0SYOoGC5zIoWI80gGNa5W2Pkm5wi2QqBySs4REcXAiMnhZZy8Rph+WaXoeM3odl0w40h6S4t6TzremFKHkz5RPpzrtJhTk9VVTsLBySxuj3BYo0okVnX6Js3ztavXh+XlJZp0QjMZwzfLAAAgAElEQVRWFT4tRKfTdu2bVOumVjDXKMTmoPpipoKY8mDjo4JVSgUVVZEoYziMFUukrCKUn6swuGEDnc1kOcjP76/pOLfaaoVy+JDSJlcOqtPQyvLFYZkeF9TI++jaNbVEa6yqtcoqvVOFaIqhLXWix1LHMY7IyYscsbKUrNmqEqyAQLxIVohamt+nPMJhwYClItxoMYLqxg53G1Eu1Uk6NqJqtkbZbMXKRQD8LoCXnXO/SX96CsBnQvkzAL62pRYYDAaDYVuwlRX6TwD4hwCeF5Hvh2P/FMA/A/BlEfksgLcA/P3tbaLBYDAYNoOtWLn8Baqt3H9mKzcViRDXvSNCo6V0yfikUhED2gmvkWNJEmJGrCSqwq0m2o21TFW45VWl+JM1sj5pkfpNDi+cVICtBdKULVduj80QUdzNYvIEVU+RkTpJ59dq2t7xttIMB2Y0Dsz8eVWRU1KFJdfzycpk8ebN0vL0mDp+CMes4FguBYWNVFiyFuqTtUKu5tbIJKdByRDiMbVCeeuS0iyXrqgK31kuj71S5fzlUj1nuq1j1AgOUmy1IaTyc9/YsgQFioZpAY7xUk65cJ7afN457C7TBhxW906Su3D9BauYnAIkCq27onGS9u/bNyyfPKEORxcuKkNar1GcEoo9dOXKlWE5WtQ6O+SsFBWogeDkRXk8Hc1Fn2kmsoQa0PsiJSGrgVvCyrIjXCEOrb8/02ZsWcTEBOd9ZQc5ro8TxTDYcikthAdmq68gLwrUCnWBu1N6l63BXP8NBoNhRGAC3WAwGEYEuxLLRaII9eBQML1frVJ++Io6FlEU2IIyl2vU86SqX1vU8r5VVXdWE/1etSjuREzWF6yLObbQADsfUejd3BKG1G1WTyPO+jPg/JPsHUGmFRR6d3xcKZejhw8PyxeuaroZDo+bO0LVSCVcIcuDK9c0ZOrMlKrcLQ4fTKGHb0mIqm0nqxyO6zJMKcrxYMiaoUehRq/PK/0zP6/9IQYHMdXdWyXnI+rzgWntB4dbzi1t+jQvNC2I6VGPXDn9wk47rEKDrJUKeTLZWiZUFG1ChWaHoEJmLM6Aw9mrSjIj3dqWvCNsqbNG2Z2mKOzwkUPqlHdoVq1fVntKrbDj3M15fY4a42t0DsetKcnSE1H4ajq1SyFz1xKyICs4lhWS9uo5BeszPYVz2Q4zaXGu26jcWsQVrLm2ZlGyJSuXghQrv6dRLgaDwWAYwgS6wWAwjAh2h3KBqsj7aPed44dm5ORQJ905t9DoZ3qsl3KcVorvMcmxUVTNG9B9uj3dtV/u6k3TQmYeigMTduUHoqpnq6WWHYyqxMisZrGlRLOlzj9Hjmiy3+m31BKhd41Cn4ax4DCy3VXt59sXNE7O1KSO89E5VbPrlAA4JYqo0HRyuKrXtZwFHZn70O2pCn327TeH5VfOaXmenImkwdl4ia6hLEzjNL4nT58elhvN2x/fghrMFBLRQpzEOGNrHqqHKZeIuIAC5VII8Zo7kGgdUmE1xHQeO7YUnFwIBbXcVZjoBP6Hab46zZsjmm98TJ+X+8+cHJZ57s5fpWxINL+c1JszLDHysctIvPA4J2TBxvFeagXLFqqQrVXYQoiHt8QpqJgXvfxdrELRsah8jhxZQmXFBpTUU+5YVDHl7xq2QjcYDIYRgQl0g8FgGBHsCuXioCpQneiCVlOtFpI+JXumuB5JSMy7Sll35pd0N3+xq8dZVY4bWkePMiBdW9CwovMUTKbTJdW1zqF3vVpKEWtRI0qIQ4qCY1aQY0VB+SP9q1HXsZia1jDA4+NKi8h1jdWS+xM5msbFjlJIg74680ztU2eeWl3H+djRo9ou4rkSyrAUk7VCneiPKKj8PXIw6fR0PJ978eVh+dxFtZRgWmKMnKnY+WQQ6Ty2pzVmy/0PnhmW2REpCzr6oCr2BRth0BwNWD0uhPXIysvscFSIz5GXObsV375cba/x+azalze9aIkxYHop0AxEibTGdZ4dUyWiz9nJY0rtvfOOPiNvnFear9kg6oQoH3YmYopE6Q0eW46NQ+8FZ6/ihM0F559yVMVHyemd8vkpjn/lilbKKRrZMPZKRVu5fcwy8jnbEGrXVugGg8EwIjCBbjAYDCOCXaFc4BwkqK4H9qv1xYFZpRk6XXVyWFxWNb6WJ6MlJ6BXzr0+LPdInzlxXOmEQV9pgcUltRQ5/46qlhxjxBHN48hpYhD08kZbjx05qplhauS0VKe4Jmy1w0ly1yh5riOao9HUOCUPv/+Dw/L1eUr2e8FTKhQlFdMzasFyjWJw/PV3nhuWL9/QGDcnrqiTzzEar4OUpYZxs6P01o0Fbwlx7ty54bGz517Tcyl8MfkbFcL0Doj+SCiTUp2ShB8hWqBJzlcJJaROQ4hhR6pyRHX0iMKTWBvQpgTjKVEYzTGi6LqUeJucuJptmt++5+BYbebMWAWapcKZhc9pt/T+HBKXE6ZP79O2Lyz5Z6FOmb4GfX3O+Pkbo7g++/drHSdOEP1C78KVBX1eVgfkWLRGGbHIyasdxo6f7VlqK4fdXe7qu90kq506Z/iqYCIytj6i4zmlw2GdWdA1m0qRcswejhnUpkxKHD8nJeurCaK0ChZK9IBHga6UijjJBQqrJGOWu1tJog0Gg8Fwb8MEusFgMIwIdodyEQy3xUkrwtysxnU5e+6dYZmigOLAEa9SNlpq+bFIVi7Pv/CDYfm1s68My2OUjDkhtWmFrGVS9nJgiwtyMuqFuCI1UvJmKB5Ns6HqbEoqXD/htEuU4YfUrJSsYthZZ3ZGE/O+//3vH5br9VcBAIvzavmSrOl9WhNK23DGnHNv69i+/oYmAx6neB8NsgrifhQca4KzCqvQix3KrqRaNvbvJ3qC+j+/oHPHGufJU0pjnTyp4V6JUUBCPMLQcYbVWUpH1CfHmtVE29uleCesNtc5Tgj57/T7+rwkVJZwUq1gKUPJzelB7xKF019jyoUstMiypEHlOGZrEYqDEso8V3GdKR+9f0IUTkZc2IF9+rwcO6zxXpa6en6TnotVCsTjaC6yLFjUpHRsoOUaUV5TZMHV6ygVylQE0xVS9DjSkrDFjYRz9Tqmf5hCiSuck9jJrsFOaQNOqq2C6eD0rNZfiEPj/2NrqqjCmqXMsWyrfke2QjcYDIYRgQl0g8FgGBHsUiwXhzjoIrzLfOrkfcPy629pKN2lRaUUlpf8rniDqAXOv8uJYZc7ahGQpGS1UrBE0CGQGsWH6dG1lD13qu3PObh/enhs7qCqW22yjnCkkvPuuJAFTZ2cdgYUY9RRlqSZaaVcHjqjsUzS4N30AxqfRXKOGp9kCkV37QeU1afW1LZ0qJ/pitIC7FjDDjJ5KOH6mMagObxP/77cVQuahKwzupS8iS0YHrhf5/9vfUQtew7N6Fh3lymWDanxWUjUzZYlFMoEMceMIQsKjvzLzmecdaZB1A3TBYOMsj3lzzFRNRz3JCNVPaY6aFoKdTcpC1Sb6Zc6q/McY3pwW5tqxP+w01JCNCNHct5PoYlPnVKa69K8OqhdodDHa0QXNcYoVHUYRw4vnK6pNQvPYaej94yZzqOB5CxghTJnBCfk9JKj+po00C2KQdMmS6VCnBpa6jYpkXujQNFQG/n4FsLGFOPElBzfIudiK3SDwWAYEeySHbpuYvEmB9vEfviDt2/+AcArr/nVwuUrusqYUvN1TExqHQUbb95AK3wKOcSjfqFTsuHljZP7jvvEE488/ODwWLupdTTp3IQ3gtjelJYuGX3aebVacDfv66ZUo64nnQhtcZTz89ybbw3LV6+ru/0q5SJtjZHhOmHAyzWalwaFPuBNx7WwuZbQxmK9SSt4Gmbe2OYgiWceUtv3H/mAzvlh8k9wqS7pY6dj2oxpAzTYkPMmc8wr8ez21eyt5YjOH9CKusbdp43IPj0j+cqtkIylEL1Sr+ONywFvvvZu3+QEgBaFahgMuB56vsL9eV+f79OKOWKontRPKMTChLb+8KxqhccOq09ClzZUB5QnN6aHd6wehbZqW4TGVjLaoOTnnJ8/6j+/C7zRybIjo3IenqEQPoGu41yrXQqVQQv0QkTIAY2RcGBV0pAKURjpGSxTIgrZVzkMQcSr9bx8l+zQRaQlIt8SkR+IyIsi8mvh+GkReUZEzorIH4hIY6O6DAaDwbD92ArlsgbgY865DwF4BMAnROSjAH4dwG855x4AcBPAZ7e/mQaDwWDYCJumXJzfScx33OrhxwH4GIB/EI5/EcCvAvid9WvLgOBCnJIdbJ02tI4dUdvuuPnwsNwaewMAcPZNtZ9eI7WGXfyXl8n2k4O60fm1WFW7yUndLDl2SFX+GaKCPvCg35R834Mnhsf6K7pRxPp5TBs4jYhUTqJZOJFGRIk6WnXKr7mkm1Ix1X/skN8sPHxQNw0PHVb+6QWKdnj+wuVheY3olwa7mBOFUEjUQO11xBHk5uesVo719cJjR7St983p2B4/rhu7DzzwgF5L7uOrC+p6HlPYwoPTek5G6r8b5Go+bawRFZSR7XlM9AvTQkyFsZofUx4H3sRPUq2zHjaXhezNOTIj24QztcDjTAwJaswQOaZuyPWeqKha4Fw4dypv8vGb3qBNwQHRNmurFGKDnrPjR9QmvUvPjnMaWmKF8seuhg31AW1+tyh6ZsQhLmhu0woqjOkaUFKRLOKNU6ZR/XGh8JnjFO4goz6zWftUm3xCaMOX/SMmxnTspqfUht5VbtDmm5sVUTULG6G3X3dX7dBFpCYi3wdwFcA3ALwOYMG54ShfAHCs6nqDwWAw3D1sSaA75wbOuUcAHAfwGID3bvZaEXlCRJ4VkWcXO/2NLzAYDAbDlvCurFyccwsi8jSAHwMwLSJxWKUfB3Cx4ponATwJAA8dn3TZ0Baa1E9K9rlGduDTk6pmP/a3PgQAOHlKKY9L15SSWFpR1a5g5UK2xGxjHkdqnzo3p6rloUOHhuX9k6paTQUTjYxykbo1ypEpqpLFhV17Nk5l9ZsizGUcEkDrYWsJpo4i1wv3pAiTR9QmfmryPxmWz1N+0fPvEP1CY9RZJZd0slXmYP+sxrZbXi1tkgo/Na7j+cj71BLoINmSzxygnKaU4KRPbYnJFKbGLvTsbp+yGu/buFbw0ycLDqKw7jumc3uwT0kyyPjYkTrP9uEH59T6o0FR+2qBu+H8mxy9kX0fHnjg/mH5ENn7r7J7Or0XkxNKi3GkRMeWU6EbNU4SwckjCjlS6Tmjc9hXwNHzN0PUwtycPl8c54Cfo9yiJKP3bJLymE7QM9InC66okJCiPMFIreD6X0655BEKB/TcLi8pLbpKER45Lep4W/0p+nV6/mmcx4iiLFJnhWSy2qf83aSYFYXokRX9ebfYipXLQRGZDuUxAB8H8DKApwH8XDjtMwC+dsetMhgMBsOWsZUV+hEAXxSRGvyH4MvOuT8SkZcAfElE/hcA3wPwu3ehnQaDwWDYAFuxcnkOwIdLjp+D59O3hpJd4ZTUrwFbJTRVFZqa9hHh6qT6TB9Qy44B0RkRqYR9Sp7AEQMjcr3nnKbjZHHBKr8LOTN7K0qzjLEFCzuwkPpdc+W73C7iiHh0PohmIct+VyPqKIxRjyI51hrqNHRkVqPnTe9TtfkIUQ51yoeakFVEP+HoeeUqbz24pI+RW/VES8dzktT5MbLy4MhzyQqpv2QhMjFBzjT0LKx2Fuh8siIK9TuygqiRM9l+iiT5njOn9JyGWjD1iYro18hahuocp3406rdTGsVolDTndOpD71HLHnZgKTgcEV1Ui3ReJtqcKYScX8J92Tms0eAEK3ou5+uNKdzFgNqbFJyWdE5PHdckGLMH9b3j56iehy0gh5w6zflEk2lJ7U5aCDGB0jK75RRymhbOCE5mNJ8z5Kj2/offMywfnNN3YWJKaUGmHFs0zzWiwmYPqiXeMj2XHAZgaK1CpjKca5bpJPY4yimcrWYZNdd/g8FgGBGYQDcYDIYRwe7EcoFDbt3CVhOrXVUtx1qqcjcp9shasMToUdD9ep1Uftq2Luy8E20ysU8pnAZdy9YvKzfVEiSiXe6JoE6Ok4NBjaIkcvD8PnuKEOqi9+S4GgNW1Th5AVnCCKm/k8HiYWyMxpDol5VlVQMd6YEHJ1UVX6MkBeNEOdVrStGwas/qYs6QFPIlkqruiLbp97WNA7DzDecU1fN7q9qPOpkFjI/r3LGVS07/9DmYCdEPNYpIMUVRKMfaSrl06do10WdxQNZHfE+O8aHKsfaHHXhaNR3zLkW1bNLzN0aWIJFoG4Xi19SEnb+oHOg9zos51tb7r5BD0BpZMzWJZuA8rv0VjkNEllhNpiL1WSxYXwVaLqOHpVGgHHRse5RfN6axKNh+0PNX9LQpj4MyjNVElVy7ou9zc1z7fHBGaaPxSaUomXKZmNB3YXVZ86t2lihAEaqchUJ+U1470zvPr06fDWUqkmBsBFuhGwwGw4jABLrBYDCMCMS9y6X9Hd1U5BqAFQDXNzp3BDAL6+eo4W9KX62fu4+TzrmDG5/msSsCHQBE5Fnn3KO7cvMdhPVz9PA3pa/Wz70Ho1wMBoNhRGAC3WAwGEYEuynQn9zFe+8krJ+jh78pfbV+7jHsGoduMBgMhu2FUS4Gg8EwIjCBbjAYDCOCXRHoIvIJEXlFRM6KyOd3ow13AyJyn4g8LSIviciLIvKL4fiMiHxDRF4L/+/fqK69gJCS8Hsi8kfh99Mi8kyY1z8QIX/7PQoRmRaRr4jID0XkZRH5sVGcTxH5H8Iz+4KI/L6ItEZhPkXkCyJyVUReoGOl8yce/zz09zkR+cjutfzdYccFeoin/tsAfhbAwwA+LSIPr3/VnkEK4Jedcw8D+CiAXwh9+zyAbzrnHgTwzfD7KOAX4ZOc5Ph1AL/lnHsAwE0An92VVm0v/k8A/945914AH4Lv70jNp4gcA/DfA3jUOfcB+JRan8JozOe/AvCJW45Vzd/PAngw/DyBDZPd33vYjRX6YwDOOufOOR9d6EsAHt+Fdmw7nHOXnHPfDeVl+Jf/GHz/vhhO+yKAv7s7Ldw+iMhxAH8HwL8MvwuAjwH4Sjhlz/dTRPYB+CmEpC3OucQ5t4ARnE/4QH1jIhIDaAO4hBGYT+fcnwOYv+Vw1fw9DuBfO4+/hk+veQR7CLsh0I8BOE+/XwjHRgoicgo+IcgzAA455/KknpcBHKq4bC/h/wDwP0ETIR4AsOA0w8QozOtpANcA/N+BWvqXIjKOEZtP59xFAL8B4G14Qb4I4DsYvfnMUTV/e1422aboXYCITAD4QwC/5Jxb4r85bye6p21FReS/AHDVOfed3W7LXUYM4CMAfsc592H4+EMFemVE5nM//Or0NICjAMZxO00xkhiF+WPshkC/COA++v14ODYSEJE6vDD/PefcV8PhK7nqFv6/ulvt2yb8BID/UkTehKfMPgbPNU8HlR0YjXm9AOCCc+6Z8PtX4AX8qM3n3wbwhnPumnOuD+Cr8HM8avOZo2r+9rxs2g2B/m0AD4Yd9Ab85stTu9CObUfgkX8XwMvOud+kPz0F4DOh/BkAX9vptm0nnHP/xDl33Dl3Cn7+/tQ5998AeBrAz4XTRqGflwGcF5E8CeXPAHgJIzaf8FTLR0WkHZ7hvJ8jNZ+Eqvl7CsDPB2uXjwJYJGpmb8A5t+M/AD4J4FUArwP4ld1ow13q10/Cq2/PAfh++PkkPL/8TQCvAfgTADO73dZt7PNPA/ijUD4D4FsAzgL4twCau92+bejfIwCeDXP67wDsH8X5BPBrAH4I4AUA/wZAcxTmE8Dvw+8L9OE1rs9WzR982qHfDnLpeXirn13vw1Z+zPXfYDAYRgS2KWowGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjAhPoBoPBMCIwgW4wGAwjgjsS6CLyCRF5RUTOisjnt6tRBoPBYNg6xDn37i4UqQF4FcDHAVwA8G0An3bOvbR9zTMYDAbDZhHfwbWPATjrnDsHACLyJQCPA6gU6OM1cdN1YPgJccDweyKAIPyIPxSJLzvnr6nVgFoEJAmGJ7sMyDL/dxHACZA5rS8Sfw0ADAb+3HC7YTuiGt0j8tc4BwwyABm1LxyXUPetx5HX6bQP/L0UQQGhmcN+uNBeOP1bfrnL9L75PfL/+FqhyvM+iQBRBEgEZANt8/BUmodM/Hn5n6MwdvkYX1mDwWDYeVx3zh3c6KQ7EejHAJyn3y8A+NFbTxKRJwA8AQD7asA/OgpkEYAgiDP4/+PI8z+tGGgEIdLpAmcvAi8AmPeXoAXgwwBO7QNmp4CoAaQJ0EuBLPYnxC0gin19SQZ0loA33wB+CODlsp70/X+TAD4I4LHTQDv2wizLgDT1/zfC7+2GPz9JgcaEb1gvCXU1gE7H92ei5ducJECWAoj0vJzrarf9uUnP3y8G0GgBSIFeT09sRKGuMGa9HhDH/qfX9fW3G8BUDDTgy+0pf2038W1IUqA9ASwt+etyYd3rAI0YaLWBbgakcRjsyI9xmgLzS0DWAH6tdAANBsNdxlubOelOBPqm4Jx7EsCTAHC0KS5DEG5ZsQG5gMuyIPABLPSALwO4dVH4CoD/dQZopUBnAZia9UIJEZAlvtJGwwuyv3geeHqTbV0G8JcA/vIN4DSAH98PnDgOIAHSnhd6cUqDFgHPPg98raSuOoCfPwIcPRwEdgrMLwDfugKcA7ACoBnOPQrgp5rAqRPAUgeYiHx/ZmdCfwAsLAFRG7i6AHx/0X+cgrxGG8CjdeDwNDCVAa9eAhbC3zMAPXghH4dxzsLv+Zin4f+W7yogwPSEF/oL80C3B/QyYPboJgfSYDDsCu5EoF8EcB/9fjwcWx9MY4RVIOAFeRR5Yd8Lf++mtwvzHL3EC6VGHFbRmV+VN1p+Ffr2O8C/Pw+8/i46BgBvAHjjJvCjN4GPnAZmpoEoAaLMfyyyxGsQb1Zc3wdw7hIwPeO1jjQFlrrA83RO3rc3AHTXgL+34O/TaPkVe5qp0I3g7/nOIvCdcGyF6pruA40u8NzK5j9glXDwX7eAGoCHAHzs6p1WbDAY7ibuxMrl2wAeFJHTItIA8CkAT2366iC0c047L6eZXzGm0JV6GTq98D2I/TUZPGWAFvDmdeDf3YEwZzwD4Nzb/gOSQu+FCEAcVrQVyP+WBtqm16s+9wqAywtAktebeSqll/ofROuPRwy/gj+7hb5tFgN4qurtxbtQucFg2Da8a4HunEsBfA7AH8O/7192zr24mWtz7px/94WwgM+F1zr6QzcB0giIWv7cuOH59HeuA199fTOqwubxpwPgpVe95pDF4WPT8AJ0PRUnRaArwmp7I33o633g6rynZ5ABSz3Px3c6vq4oLo4bI4EX/N2td2/TuJt1GwyGO8cdcejOua8D+PpWr8sy5XIj+H9yoZ6QxFpv9dsNm6AJ/Oo3agCdBPjWW361uxEEnnueCG24tM65fQB/PgCmrwNHp/3GYRQ0iSoBC3gB2O0CWSus0jf4fK4BOHcNmMtpl67/cAGeXmpFyneXodXyfbqx/m3eNZbuUr0Gg2F7cNc3RQtwGHLoGQJnjuKGaG79AqwvvNKwgu8FCxc0gMtXgR9s0IQP7/f/T034+zcibxHS6wLvnPc2l/2S61YAPHsDeDQL1jVRsBRZ515DagZ+1d1d7wsV8B8BnLoKHD/sJycNy+IsBeKs+n49+DadGgPOr258n62iDuDwPgBGuxgM9yx2VqDfiiDN8xV7/nuUt6qxzqWx/+l0vKCMG8Cb6yzNmwB+/KAXwgDQDgI5S7ygnGsDhx8ETlwH/uQmUCYTXwNwagmYmPIr9NYGAr0Hby4Yx75PrRbKvxa34IVFYHbWXzf8qJH1TxlS+L48NAdMXAWWVr1des/5/6cmvFbz0ipwc5173wdg5hCQ9fwm8HQMRD0/XifmYALdYLiHYbFcDAaDYUSwoyv03GMxzh1XcvBnJdaV6Hr89FInbIqGupY63ja7Cj85Dsy29PdW7DlpNPz1reDcFE8Djy0B/2FQXs87A+B4BLQafgW+HnoIWkBKWscm8D0AJy4AH3hIefcsW39PAfBj206BMxNA2g5tSNQaqAu/kn+m4vqDAB69z2sSWTBKnwAQt/2Fke2KGgz3NHaccomi4KGY74oScl49od+r0B0EOiLQHkvz1ZuhH4LfzGxF2uE4AxqptifKgCj19zw6B+y/VE5NvA3gA0nYtMzWbyMANNqeCsk9TjeL764Bp3JPVPjN1SQFoiZKjfPzqmP4PYEk9R+SVkO5/FYMLCzffi3X0U6Cw1LPt7uRBSekCMhMoBsM9zR2h0O/VQpmtxTDqnSjVW0SNkTTzDv5VOFUM3DB8Lww4FfjjcgL9jj/yAQBPdXyjjRlK9lleK/NRjCXXE9GRwgfC/iP1Fb4rfMAzl0ATh33v/cyz4FHMUoFeoZgJx95zSCLwj5DXg5/i2vwhuUluAkvxKfgG5wL9DjSegwGw72LHRfo2UZLWqDoHrnOKWka4qwA6FQEjdwHb5WSeytFmV6fx3vJhW4e2ySKgLkDqLT/u74MTE17U8n1kCEIw8iv/rc62N9aBubyD1Ab6HaKZp2M3IQyd8hih6v8w5MC6GxwzyRYGsVZoIpyE9NcqzIYDPcsdt/KpeRQmgvd9QR6TQV6tA6f3QC852XqV5ssEOPUr1rzsANxsMNIod8AAA7iSURBVLLJzRHrKDdKWUAQdo3SLhT6Mvw2ZRp0rAz3NYHzt6y8rwB4Kbi7PvSgF+pphSaSC/RcE4hSPZZrPQk25uE78Fx7/uFrZP4h4Rg7BoPh3sSOCnTnNl6h50IaWP/cOMRwyYL54HrURxY2TjNaoWfwwj0KwhyBT08bPlYLUu+kU2al14MK8nXvCz0xt3mvwtFZ4PLF2z8gfxX+n74OHD3u47UUgrgE5II6AtBIoLb+UKHeioCpBsptMgMWUmCppR/TVqT275vSrgwGw65h51foGwiFLPWRDYH1PSujnMoIAbnWXTw2vOBOUBTouSaARIV6CiBO1hdeKdbXHnKwkN1ohR4DeHgf8IMKO+/nbgKzcz7wGK6VtwmRj0DZChuijdzOPzhhpTEws4FA76bAQuy1jyjzVjIt+B8T6AbDvY2dXaGD6BQUhXAeuiXLQgILhLjcFcjjmuRUQNVKOYH/W5KFELj0t6F8ioLFSljFZ10NElaGzW4QRvA3H8Z9j4pJKwroAkdngOuL5XFoLgK4MB9MCEsQQsF7iiRoHlnoQ4TAp28UqwDe8ocjYuabqUlqFLrBcK9jRwV6HPs4KDPTnibpdeCpjZbnhntZMPELUjdLfejWMqOMt5eBw3NeEL+9TljXRXgLmLk8cQZZuQCe3umG0AGNCf9/p+cDb3VLqA3AUy7dxAcG22gAkxBHvdHyH5SqhH+9Jc97f/AkcLEilP3T14ADFdfPAWgFj9n8Y5drBSm8NdBS6vu2HqYGwFSYlyyEC44BpG5jE02DwbC72FGB3k+Bv74CpFf8jXvwjivvBXDikOeuczNCAN50rqKufLMT8EKsYuEKAOguAY05v0ptkyVLlvk6GsGZKUmD/XakySDKPiYt6Cp4o1gubHueriMRW/BtazV8KqiqaJFVgbeW4D8Y6S3xcQDSUqCx5qvQgF+lZymQDkI9sv41BoPh3sCOC/S/Kjl+GcAnloDp2bB5F1bo7cwL6rKF8oT4NHFZ5D8Ec+MVJ8JHLZxoe4EX5yngoiBgg/BOA82Qe1UuLVWaa2MOgdq4hcIpw1Cgb+BYlNvCZwnw6IPAxdc2qPgWhEW1bsSmt2yK5tTPRpvSUOrJYDDsLdwTtOg7ALK2pzmSnHZJvGCfqLimgeAYBG+JMTPl6ZkyzDsAadEyJgvSLk29wO/1/Kq73fCC9c11kiHPwps8Rsn6Aj1FqD/k80zXsxkMK/6k49vxI+ucWoYEABqB7w7mmWnJT1LF+XCbMyBzZNvuCrS6wWC4R7GjAp2zyTNW4LP1XO0ACx0VuPlGaRnyDcdc6rRawImKc88CuHpV7bQTqMDLUkriHBIlZ/Bp4aow3fTCPErXF+gJfN1p+ECtu+odYOj4NH8dOHXQJ63eLNbg6ZQyQT78iKXrm1kCKtCZprHFusGwN7CjAj2qVQvdCwOgkwLxhN+cbEz41WaVR3+EICQB79kI4FTFuSsAXrjpk06nwXwvifxP2gDiltcOUgDXl4DvrhOG9yC85pCGD8lmVui5N+t6yGkRZKqdfGCDa27FUjd8QPLVeC7UQQJ+gzqGK3TYitxg2GvYUKCLyH0i8rSIvCQiL4rIL4bjvyoiF0Xk++Hnk5u54VzF8asAerEX6EnD/1yeXyd2d+DAhyZ5ETB3yHt3luEZAK9eA5JYf5Yyf5/WLNAKmYh+eMPHPa/CKcDbdAPDGClVGJpVhl82GuzcBn+q5YX67KS3aqmybLkVvCIvrNDTgjKzfhvgKZZbrBdNuBsMewCb2RRNAfyyc+67IjIJ4Dsi8o3wt99yzv3GVm54uMKf/g0A6UUv8HPLuvUSlEbBozPPBJQEn/czAF6puOZpAJdDrrlTdWBm1tezlPpcpN9dWz/5gwCYGPdmjkkGTGyQJBrw5ooINtxxldkMgImaOivFEZB2/Ebuo01/7I/X4fRzdAbAVJDEMdmR5x+W9SisHDEACBAR127C3GDYG9hQoDvnLiGk3HTOLYvIy/CWdQaDwWC4h7Als0UROQXgw/AMxk8A+JyI/DyAZ+FX8bctcEXkCQBPAMBkDZg7jEoj6/PhZyMcQnCBD7RLmvgNwTTyeS/PLlabHL6c/9/H+pmhS3AKnmbpdOGXrfH60Qsdih6l0Xor9Lby3/kGapYAU8HAft/axtnf8giUecTECH5M0kh/38xqO48+mWW+E0O79ko3V4PBcC9g05uiIjIB4A8B/JJzbgnA7wC4H8Aj8KLxfy+7zjn3pHPuUefco+2a9xB9zx02+mEAs9PeC7PX8QKoFWKxTDSAH9+KecgmcQDA7Lg3cUwG3qwvamw8gL0kbIwm69uh5/HYkz6Gwndp2ZsxJp3NbZD2oA5BqQs/wSY9CkJ+oy94HhMG8H0ctk/I4ctgMNyT2NQrKiJ1eGH+e865rwKAc+6Kc27gnMsA/F8AHttMXQ0Aj+5/l60NOHUMmG6p6WAcBFUcrE5mp4H339ktCjgI4AMHQ5CqLPwfPk4bORb1+l4wJqlP2FyFJAQIyxDCBYSZSfr+5+i+jXmuBCU24y4IZud58Y3aGwG3pQGM4G344x11QzMYDFvFZqxcBMDvAnjZOfebdPwInfZfAXhhw7s5v1KdnQb+Tg0YexcN/q8PBi/NkE1nIvIepXHiowzGPSBdAE7tA/7T5uYtRKrwfgDvbQKYB2Zjn8pubgyYjoCotzGF0QOQycbBvJKeF5itmtcAogho12hTswcc3+BeLdw+oRwGIBa/+boeOIRCJPQTmUA3GO51bOYV/QkA/xDA8yLy/XDsnwL4tIg8As+qvgngH21Ukcu8YIpib2f90xnwzhrwg0004j4A7637RM/JPIYJkdm1PcrtrDNvWz7VBtpd4NyicuebwTh8fJkHDgDtto8Fk/S8EG8ECd5Lgagb0rVVYD+AhvhYM1ni27Yft1vSTEJXwY1IaZJGBESBc++ueYF9BNXU/62TyREtI/iV+kLV5gIhQ+DRGyGfKDYXLthgMOwuNmPl8hcod/D8+pbvFlzwe10gXQWm9wOHjwNH571DTwbP/3JAwOk6cHgWmGh5Adq5DkxFnm5BiNWdx/xOMv+xmIj9Jun8ghdGHzgEnEmBV2+oY003/CTwQqsN79J/vA6cmAKm2xpydzqEJejM+/u2G4HiSYATdeDDZIbZhbb/qPgPQpYoFfLecM9OOC8GcBghGmPmefZ2yEyUuNA/qHnk0dDeDkK43PAzB2CGxi0X5vkEZwDadeCBKSC5AVxAMazCBLwG0B739E8ch1U5tAKLh24w3NvYeSU69VYiM5PezX5hHj4W+JQXQN1usGABMD3jV8bz14H5xK/OZ6d8mNgoA7o9L3QmWt5RKKcXssw750xMAEs9YL7jheVD+/wmZd7zdlh+xlkQYIG6iXsh/gq8gO2kfkNwouU1i1aQqFnmN2NPiW8rAFxPQuTDEOsl7QbhL8B0sJvvDYCZpqa6y1fBWc8L8bkWcH3F13M0fEobEdAdeMHbFs/HN4LVSQyvKeTp50LCpQLVgshz8VdvAIdrwOEIaAf1Is58/Z0FoLfiPXqjLGyCxl776WYYJgcxGAz3JsS5nbNDE5FlVPv97FXMAri+243YRlh/7n2MWp+sPxvjpHPu4EYn7fQK/RXn3KM7fM+7ChF5dpT6ZP259zFqfbL+bB9sq8tgMBhGBCbQDQaDYUSw0wL9yR2+305g1Ppk/bn3MWp9sv5sE3Z0U9RgMBgMdw9GuRgMBsOIYMcEuoh8QkReEZGzIvL5nbrvdkJE3hSR50NCj2fDsRkR+YaIvBb+v8NINXcXIvIFEbkqIi/QsdI+iMc/D3P2nIh8ZPdaXo6K/lQmXxGRfxL684qI/Oe70+pqrJNQZk/O0btJkLMH5qglIt8SkR+EPv1aOH5aRJ4Jbf8DEWmE483w+9nw91N3rXHOubv+A5+/+XX4/BMNeG//h3fi3tvcjzcBzN5y7H8D8PlQ/jyAX9/tdm7Qh58C8BEAL2zUBwCfBPD/wnsKfxTAM7vd/k3251cB/I8l5z4cnr0mgNPhmaztdh9uaeMRAB8J5UkAr4Z278k5Wqc/e3mOBMBEKNfhw4l/FMCXAXwqHP8XAP7bUP7vAPyLUP4UgD+4W23bqRX6YwDOOufOOecSAF8C8PgO3ftu43EAXwzlLwL4u7vYlg3hnPtzAPO3HK7qw+MA/rXz+GsA07cEZdt1VPSnCo8D+JJzbs059wZ8/vBNRQndKTjnLjnnvhvKy/BhiI5hj87ROv2pwl6YI+ecy1Mh1MOPA/AxAF8Jx2+do3zuvgLgZ0LQw23HTgn0YyjmrriAvZn1yAH4/0TkOyFxBwAccj6rEwBchs+/sddQ1Ye9PG+fCxTEF4gG21P9uSWhzJ6fo1v6A+zhORKRWghWeBXAN+A1iQXnXB4uits97FP4+yLuPBBsKWxTdGv4SefcRwD8LIBfEJGf4j86r1PtabOhUegDNpl85V5GSUKZIfbiHL3bBDn3KpzPBfEIfEy7x+Dj7u06dkqgX4SPgJvjOCoT0d27cM5dDP9fBfD/wE/klVzFDf9f3b0WvmtU9WFPzpurTr6yJ/pTllAGe3iOtpgg557vD8M5twCff/7H4OmuPJwKt3vYp/D3fQBu3I327JRA/zaAB8MucAN+Y+CpHbr3tkBExkVkMi8D+M/gk3o8BeAz4bTPAPja7rTwjlDVh6cA/HywpPgogEVS++9ZSHXylacAfCpYHZwG8CCAb+10+9ZD4FZvSyiDPTpHVf3Z43N0UESmQ3kMwMfh9waeBvBz4bRb5yifu58D8KdBy9p+7ODO8Cfhd7hfB/ArO3XfbWz/Gfjd9x8AeDHvAzwX9k0ArwH4EwAzu93WDfrx+/Aqbh+e5/tsVR/gd/N/O8zZ8wAe3e32b7I//ya09zn4l+kInf8roT+vAPjZ3W5/SX9+Ep5OeQ7A98PPJ/fqHK3Tn708Rx8E8L3Q9hcA/M/h+Bn4j89ZAP8WQDMcb4Xfz4a/n7lbbTNPUYPBYBgR2KaowWAwjAhMoBsMBsOIwAS6wWAwjAhMoBsMBsOIwAS6wWAwjAhMoBsMBsOIwAS6wWAwjAhMoBsMBsOI4P8HQw6G7YaZIk4AAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "# 读图\n",
    "raw_img = cv2.imread(\"/home/aistudio/work/word_1.png\")\n",
    "plt.figure()\n",
    "plt.subplot(2,1,1)\n",
    "# 可视化原图\n",
    "plt.imshow(raw_img)\n",
    "# 缩放并归一化\n",
    "padding_im, draw_img = resize_norm_img(raw_img)\n",
    "plt.subplot(2,1,2)\n",
    "# 可视化网络输入图\n",
    "plt.imshow(draw_img)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "**【网络结构】**\n",
    "\n",
    "* backbone\n",
    "\n",
    "PaddleOCR 使用 MobileNetV3 作为骨干网络，组网顺序与网络结构一致，首先定义网络中的公共模块([源码位置](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/modeling/backbones/rec_mobilenet_v3.py))：ConvBNLayer、ResidualUnit、make_divisible"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import paddle\n",
    "import paddle.nn as nn\n",
    "import paddle.nn.functional as F\n",
    "\n",
    "class ConvBNLayer(nn.Layer):\n",
    "    def __init__(self,\n",
    "                 in_channels,\n",
    "                 out_channels,\n",
    "                 kernel_size,\n",
    "                 stride,\n",
    "                 padding,\n",
    "                 groups=1,\n",
    "                 if_act=True,\n",
    "                 act=None):\n",
    "        \"\"\"\n",
    "        卷积BN层\n",
    "        :param in_channels: 输入通道数\n",
    "        :param out_channels: 输出通道数\n",
    "        :param kernel_size: 卷积核尺寸\n",
    "        :parma stride: 步长大小\n",
    "        :param padding: 填充大小\n",
    "        :param groups: 二维卷积层的组数\n",
    "        :param if_act: 是否添加激活函数\n",
    "        :param act: 激活函数\n",
    "        \"\"\"\n",
    "        super(ConvBNLayer, self).__init__()\n",
    "        self.if_act = if_act\n",
    "        self.act = act\n",
    "        self.conv = nn.Conv2D(\n",
    "            in_channels=in_channels,\n",
    "            out_channels=out_channels,\n",
    "            kernel_size=kernel_size,\n",
    "            stride=stride,\n",
    "            padding=padding,\n",
    "            groups=groups,\n",
    "            bias_attr=False)\n",
    "\n",
    "        self.bn = nn.BatchNorm(num_channels=out_channels, act=None)\n",
    "\n",
    "    def forward(self, x):\n",
    "        # conv层\n",
    "        x = self.conv(x)\n",
    "        # batchnorm层\n",
    "        x = self.bn(x)\n",
    "        # 是否使用激活函数\n",
    "        if self.if_act:\n",
    "            if self.act == \"relu\":\n",
    "                x = F.relu(x)\n",
    "            elif self.act == \"hardswish\":\n",
    "                x = F.hardswish(x)\n",
    "            else:\n",
    "                print(\"The activation function({}) is selected incorrectly.\".\n",
    "                      format(self.act))\n",
    "                exit()\n",
    "        return x\n",
    "\n",
    "class SEModule(nn.Layer):\n",
    "    def __init__(self, in_channels, reduction=4):\n",
    "        \"\"\"\n",
    "        SE模块\n",
    "        :param in_channels: 输入通道数\n",
    "        :param reduction: 通道缩放率\n",
    "        \"\"\"        \n",
    "        super(SEModule, self).__init__()\n",
    "        self.avg_pool = nn.AdaptiveAvgPool2D(1)\n",
    "        self.conv1 = nn.Conv2D(\n",
    "            in_channels=in_channels,\n",
    "            out_channels=in_channels // reduction,\n",
    "            kernel_size=1,\n",
    "            stride=1,\n",
    "            padding=0)\n",
    "        self.conv2 = nn.Conv2D(\n",
    "            in_channels=in_channels // reduction,\n",
    "            out_channels=in_channels,\n",
    "            kernel_size=1,\n",
    "            stride=1,\n",
    "            padding=0)\n",
    "\n",
    "    def forward(self, inputs):\n",
    "        # 平均池化\n",
    "        outputs = self.avg_pool(inputs)\n",
    "        # 第一个卷积层\n",
    "        outputs = self.conv1(outputs)\n",
    "        # relu激活函数\n",
    "        outputs = F.relu(outputs)\n",
    "        # 第二个卷积层\n",
    "        outputs = self.conv2(outputs)\n",
    "        # hardsigmoid 激活函数\n",
    "        outputs = F.hardsigmoid(outputs, slope=0.2, offset=0.5)\n",
    "        return inputs * outputs\n",
    "\n",
    "\n",
    "class ResidualUnit(nn.Layer):\n",
    "    def __init__(self,\n",
    "                 in_channels,\n",
    "                 mid_channels,\n",
    "                 out_channels,\n",
    "                 kernel_size,\n",
    "                 stride,\n",
    "                 use_se,\n",
    "                 act=None):\n",
    "        \"\"\"\n",
    "        残差层\n",
    "        :param in_channels: 输入通道数\n",
    "        :param mid_channels: 中间通道数\n",
    "        :param out_channels: 输出通道数\n",
    "        :param kernel_size: 卷积核尺寸\n",
    "        :parma stride: 步长大小\n",
    "        :param use_se: 是否使用se模块\n",
    "        :param act: 激活函数\n",
    "        \"\"\" \n",
    "        super(ResidualUnit, self).__init__()\n",
    "        self.if_shortcut = stride == 1 and in_channels == out_channels\n",
    "        self.if_se = use_se\n",
    "\n",
    "        self.expand_conv = ConvBNLayer(\n",
    "            in_channels=in_channels,\n",
    "            out_channels=mid_channels,\n",
    "            kernel_size=1,\n",
    "            stride=1,\n",
    "            padding=0,\n",
    "            if_act=True,\n",
    "            act=act)\n",
    "        self.bottleneck_conv = ConvBNLayer(\n",
    "            in_channels=mid_channels,\n",
    "            out_channels=mid_channels,\n",
    "            kernel_size=kernel_size,\n",
    "            stride=stride,\n",
    "            padding=int((kernel_size - 1) // 2),\n",
    "            groups=mid_channels,\n",
    "            if_act=True,\n",
    "            act=act)\n",
    "        if self.if_se:\n",
    "            self.mid_se = SEModule(mid_channels)\n",
    "        self.linear_conv = ConvBNLayer(\n",
    "            in_channels=mid_channels,\n",
    "            out_channels=out_channels,\n",
    "            kernel_size=1,\n",
    "            stride=1,\n",
    "            padding=0,\n",
    "            if_act=False,\n",
    "            act=None)\n",
    "\n",
    "    def forward(self, inputs):\n",
    "        x = self.expand_conv(inputs)\n",
    "        x = self.bottleneck_conv(x)\n",
    "        if self.if_se:\n",
    "            x = self.mid_se(x)\n",
    "        x = self.linear_conv(x)\n",
    "        if self.if_shortcut:\n",
    "            x = paddle.add(inputs, x)\n",
    "        return x\n",
    "\n",
    "\n",
    "def make_divisible(v, divisor=8, min_value=None):\n",
    "    \"\"\"\n",
    "    确保被8整除\n",
    "    \"\"\" \n",
    "    if min_value is None:\n",
    "        min_value = divisor\n",
    "    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)\n",
    "    if new_v < 0.9 * v:\n",
    "        new_v += divisor\n",
    "    return new_v\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "利用公共模块搭建骨干网络"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "class MobileNetV3(nn.Layer):\n",
    "    def __init__(self,\n",
    "                 in_channels=3,\n",
    "                 model_name='small',\n",
    "                 scale=0.5,\n",
    "                 small_stride=None,\n",
    "                 disable_se=False,\n",
    "                 **kwargs):\n",
    "        super(MobileNetV3, self).__init__()\n",
    "        self.disable_se = disable_se\n",
    "        \n",
    "        small_stride = [1, 2, 2, 2]\n",
    "\n",
    "        if model_name == \"small\":\n",
    "            cfg = [\n",
    "                # k, exp, c,  se,     nl,  s,\n",
    "                [3, 16, 16, True, 'relu', (small_stride[0], 1)],\n",
    "                [3, 72, 24, False, 'relu', (small_stride[1], 1)],\n",
    "                [3, 88, 24, False, 'relu', 1],\n",
    "                [5, 96, 40, True, 'hardswish', (small_stride[2], 1)],\n",
    "                [5, 240, 40, True, 'hardswish', 1],\n",
    "                [5, 240, 40, True, 'hardswish', 1],\n",
    "                [5, 120, 48, True, 'hardswish', 1],\n",
    "                [5, 144, 48, True, 'hardswish', 1],\n",
    "                [5, 288, 96, True, 'hardswish', (small_stride[3], 1)],\n",
    "                [5, 576, 96, True, 'hardswish', 1],\n",
    "                [5, 576, 96, True, 'hardswish', 1],\n",
    "            ]\n",
    "            cls_ch_squeeze = 576\n",
    "        else:\n",
    "            raise NotImplementedError(\"mode[\" + model_name +\n",
    "                                      \"_model] is not implemented!\")\n",
    "\n",
    "        supported_scale = [0.35, 0.5, 0.75, 1.0, 1.25]\n",
    "        assert scale in supported_scale, \\\n",
    "            \"supported scales are {} but input scale is {}\".format(supported_scale, scale)\n",
    "\n",
    "        inplanes = 16\n",
    "        # conv1\n",
    "        self.conv1 = ConvBNLayer(\n",
    "            in_channels=in_channels,\n",
    "            out_channels=make_divisible(inplanes * scale),\n",
    "            kernel_size=3,\n",
    "            stride=2,\n",
    "            padding=1,\n",
    "            groups=1,\n",
    "            if_act=True,\n",
    "            act='hardswish')\n",
    "        i = 0\n",
    "        block_list = []\n",
    "        inplanes = make_divisible(inplanes * scale)\n",
    "        for (k, exp, c, se, nl, s) in cfg:\n",
    "            se = se and not self.disable_se\n",
    "            block_list.append(\n",
    "                ResidualUnit(\n",
    "                    in_channels=inplanes,\n",
    "                    mid_channels=make_divisible(scale * exp),\n",
    "                    out_channels=make_divisible(scale * c),\n",
    "                    kernel_size=k,\n",
    "                    stride=s,\n",
    "                    use_se=se,\n",
    "                    act=nl))\n",
    "            inplanes = make_divisible(scale * c)\n",
    "            i += 1\n",
    "        self.blocks = nn.Sequential(*block_list)\n",
    "\n",
    "        self.conv2 = ConvBNLayer(\n",
    "            in_channels=inplanes,\n",
    "            out_channels=make_divisible(scale * cls_ch_squeeze),\n",
    "            kernel_size=1,\n",
    "            stride=1,\n",
    "            padding=0,\n",
    "            groups=1,\n",
    "            if_act=True,\n",
    "            act='hardswish')\n",
    "\n",
    "        self.pool = nn.MaxPool2D(kernel_size=2, stride=2, padding=0)\n",
    "        self.out_channels = make_divisible(scale * cls_ch_squeeze)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.conv1(x)\n",
    "        x = self.blocks(x)\n",
    "        x = self.conv2(x)\n",
    "        x = self.pool(x)\n",
    "        return x\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "至此就完成了骨干网络的定义，可通过 paddle.summary 结构可视化整个网络结构："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "-------------------------------------------------------------------------------\n",
      "   Layer (type)         Input Shape          Output Shape         Param #    \n",
      "===============================================================================\n",
      "     Conv2D-1        [[1, 3, 32, 320]]     [1, 8, 16, 160]          216      \n",
      "    BatchNorm-1      [[1, 8, 16, 160]]     [1, 8, 16, 160]          32       \n",
      "   ConvBNLayer-1     [[1, 3, 32, 320]]     [1, 8, 16, 160]           0       \n",
      "     Conv2D-2        [[1, 8, 16, 160]]     [1, 8, 16, 160]          64       \n",
      "    BatchNorm-2      [[1, 8, 16, 160]]     [1, 8, 16, 160]          32       \n",
      "   ConvBNLayer-2     [[1, 8, 16, 160]]     [1, 8, 16, 160]           0       \n",
      "     Conv2D-3        [[1, 8, 16, 160]]     [1, 8, 16, 160]          72       \n",
      "    BatchNorm-3      [[1, 8, 16, 160]]     [1, 8, 16, 160]          32       \n",
      "   ConvBNLayer-3     [[1, 8, 16, 160]]     [1, 8, 16, 160]           0       \n",
      "AdaptiveAvgPool2D-1  [[1, 8, 16, 160]]       [1, 8, 1, 1]            0       \n",
      "     Conv2D-4          [[1, 8, 1, 1]]        [1, 2, 1, 1]           18       \n",
      "     Conv2D-5          [[1, 2, 1, 1]]        [1, 8, 1, 1]           24       \n",
      "    SEModule-1       [[1, 8, 16, 160]]     [1, 8, 16, 160]           0       \n",
      "     Conv2D-6        [[1, 8, 16, 160]]     [1, 8, 16, 160]          64       \n",
      "    BatchNorm-4      [[1, 8, 16, 160]]     [1, 8, 16, 160]          32       \n",
      "   ConvBNLayer-4     [[1, 8, 16, 160]]     [1, 8, 16, 160]           0       \n",
      "  ResidualUnit-1     [[1, 8, 16, 160]]     [1, 8, 16, 160]           0       \n",
      "     Conv2D-7        [[1, 8, 16, 160]]     [1, 40, 16, 160]         320      \n",
      "    BatchNorm-5      [[1, 40, 16, 160]]    [1, 40, 16, 160]         160      \n",
      "   ConvBNLayer-5     [[1, 8, 16, 160]]     [1, 40, 16, 160]          0       \n",
      "     Conv2D-8        [[1, 40, 16, 160]]    [1, 40, 8, 160]          360      \n",
      "    BatchNorm-6      [[1, 40, 8, 160]]     [1, 40, 8, 160]          160      \n",
      "   ConvBNLayer-6     [[1, 40, 16, 160]]    [1, 40, 8, 160]           0       \n",
      "     Conv2D-9        [[1, 40, 8, 160]]     [1, 16, 8, 160]          640      \n",
      "    BatchNorm-7      [[1, 16, 8, 160]]     [1, 16, 8, 160]          64       \n",
      "   ConvBNLayer-7     [[1, 40, 8, 160]]     [1, 16, 8, 160]           0       \n",
      "  ResidualUnit-2     [[1, 8, 16, 160]]     [1, 16, 8, 160]           0       \n",
      "     Conv2D-10       [[1, 16, 8, 160]]     [1, 48, 8, 160]          768      \n",
      "    BatchNorm-8      [[1, 48, 8, 160]]     [1, 48, 8, 160]          192      \n",
      "   ConvBNLayer-8     [[1, 16, 8, 160]]     [1, 48, 8, 160]           0       \n",
      "     Conv2D-11       [[1, 48, 8, 160]]     [1, 48, 8, 160]          432      \n",
      "    BatchNorm-9      [[1, 48, 8, 160]]     [1, 48, 8, 160]          192      \n",
      "   ConvBNLayer-9     [[1, 48, 8, 160]]     [1, 48, 8, 160]           0       \n",
      "     Conv2D-12       [[1, 48, 8, 160]]     [1, 16, 8, 160]          768      \n",
      "   BatchNorm-10      [[1, 16, 8, 160]]     [1, 16, 8, 160]          64       \n",
      "  ConvBNLayer-10     [[1, 48, 8, 160]]     [1, 16, 8, 160]           0       \n",
      "  ResidualUnit-3     [[1, 16, 8, 160]]     [1, 16, 8, 160]           0       \n",
      "     Conv2D-13       [[1, 16, 8, 160]]     [1, 48, 8, 160]          768      \n",
      "   BatchNorm-11      [[1, 48, 8, 160]]     [1, 48, 8, 160]          192      \n",
      "  ConvBNLayer-11     [[1, 16, 8, 160]]     [1, 48, 8, 160]           0       \n",
      "     Conv2D-14       [[1, 48, 8, 160]]     [1, 48, 4, 160]         1,200     \n",
      "   BatchNorm-12      [[1, 48, 4, 160]]     [1, 48, 4, 160]          192      \n",
      "  ConvBNLayer-12     [[1, 48, 8, 160]]     [1, 48, 4, 160]           0       \n",
      "AdaptiveAvgPool2D-2  [[1, 48, 4, 160]]      [1, 48, 1, 1]            0       \n",
      "     Conv2D-15        [[1, 48, 1, 1]]       [1, 12, 1, 1]           588      \n",
      "     Conv2D-16        [[1, 12, 1, 1]]       [1, 48, 1, 1]           624      \n",
      "    SEModule-2       [[1, 48, 4, 160]]     [1, 48, 4, 160]           0       \n",
      "     Conv2D-17       [[1, 48, 4, 160]]     [1, 24, 4, 160]         1,152     \n",
      "   BatchNorm-13      [[1, 24, 4, 160]]     [1, 24, 4, 160]          96       \n",
      "  ConvBNLayer-13     [[1, 48, 4, 160]]     [1, 24, 4, 160]           0       \n",
      "  ResidualUnit-4     [[1, 16, 8, 160]]     [1, 24, 4, 160]           0       \n",
      "     Conv2D-18       [[1, 24, 4, 160]]     [1, 120, 4, 160]        2,880     \n",
      "   BatchNorm-14      [[1, 120, 4, 160]]    [1, 120, 4, 160]         480      \n",
      "  ConvBNLayer-14     [[1, 24, 4, 160]]     [1, 120, 4, 160]          0       \n",
      "     Conv2D-19       [[1, 120, 4, 160]]    [1, 120, 4, 160]        3,000     \n",
      "   BatchNorm-15      [[1, 120, 4, 160]]    [1, 120, 4, 160]         480      \n",
      "  ConvBNLayer-15     [[1, 120, 4, 160]]    [1, 120, 4, 160]          0       \n",
      "AdaptiveAvgPool2D-3  [[1, 120, 4, 160]]     [1, 120, 1, 1]           0       \n",
      "     Conv2D-20        [[1, 120, 1, 1]]      [1, 30, 1, 1]          3,630     \n",
      "     Conv2D-21        [[1, 30, 1, 1]]       [1, 120, 1, 1]         3,720     \n",
      "    SEModule-3       [[1, 120, 4, 160]]    [1, 120, 4, 160]          0       \n",
      "     Conv2D-22       [[1, 120, 4, 160]]    [1, 24, 4, 160]         2,880     \n",
      "   BatchNorm-16      [[1, 24, 4, 160]]     [1, 24, 4, 160]          96       \n",
      "  ConvBNLayer-16     [[1, 120, 4, 160]]    [1, 24, 4, 160]           0       \n",
      "  ResidualUnit-5     [[1, 24, 4, 160]]     [1, 24, 4, 160]           0       \n",
      "     Conv2D-23       [[1, 24, 4, 160]]     [1, 120, 4, 160]        2,880     \n",
      "   BatchNorm-17      [[1, 120, 4, 160]]    [1, 120, 4, 160]         480      \n",
      "  ConvBNLayer-17     [[1, 24, 4, 160]]     [1, 120, 4, 160]          0       \n",
      "     Conv2D-24       [[1, 120, 4, 160]]    [1, 120, 4, 160]        3,000     \n",
      "   BatchNorm-18      [[1, 120, 4, 160]]    [1, 120, 4, 160]         480      \n",
      "  ConvBNLayer-18     [[1, 120, 4, 160]]    [1, 120, 4, 160]          0       \n",
      "AdaptiveAvgPool2D-4  [[1, 120, 4, 160]]     [1, 120, 1, 1]           0       \n",
      "     Conv2D-25        [[1, 120, 1, 1]]      [1, 30, 1, 1]          3,630     \n",
      "     Conv2D-26        [[1, 30, 1, 1]]       [1, 120, 1, 1]         3,720     \n",
      "    SEModule-4       [[1, 120, 4, 160]]    [1, 120, 4, 160]          0       \n",
      "     Conv2D-27       [[1, 120, 4, 160]]    [1, 24, 4, 160]         2,880     \n",
      "   BatchNorm-19      [[1, 24, 4, 160]]     [1, 24, 4, 160]          96       \n",
      "  ConvBNLayer-19     [[1, 120, 4, 160]]    [1, 24, 4, 160]           0       \n",
      "  ResidualUnit-6     [[1, 24, 4, 160]]     [1, 24, 4, 160]           0       \n",
      "     Conv2D-28       [[1, 24, 4, 160]]     [1, 64, 4, 160]         1,536     \n",
      "   BatchNorm-20      [[1, 64, 4, 160]]     [1, 64, 4, 160]          256      \n",
      "  ConvBNLayer-20     [[1, 24, 4, 160]]     [1, 64, 4, 160]           0       \n",
      "     Conv2D-29       [[1, 64, 4, 160]]     [1, 64, 4, 160]         1,600     \n",
      "   BatchNorm-21      [[1, 64, 4, 160]]     [1, 64, 4, 160]          256      \n",
      "  ConvBNLayer-21     [[1, 64, 4, 160]]     [1, 64, 4, 160]           0       \n",
      "AdaptiveAvgPool2D-5  [[1, 64, 4, 160]]      [1, 64, 1, 1]            0       \n",
      "     Conv2D-30        [[1, 64, 1, 1]]       [1, 16, 1, 1]          1,040     \n",
      "     Conv2D-31        [[1, 16, 1, 1]]       [1, 64, 1, 1]          1,088     \n",
      "    SEModule-5       [[1, 64, 4, 160]]     [1, 64, 4, 160]           0       \n",
      "     Conv2D-32       [[1, 64, 4, 160]]     [1, 24, 4, 160]         1,536     \n",
      "   BatchNorm-22      [[1, 24, 4, 160]]     [1, 24, 4, 160]          96       \n",
      "  ConvBNLayer-22     [[1, 64, 4, 160]]     [1, 24, 4, 160]           0       \n",
      "  ResidualUnit-7     [[1, 24, 4, 160]]     [1, 24, 4, 160]           0       \n",
      "     Conv2D-33       [[1, 24, 4, 160]]     [1, 72, 4, 160]         1,728     \n",
      "   BatchNorm-23      [[1, 72, 4, 160]]     [1, 72, 4, 160]          288      \n",
      "  ConvBNLayer-23     [[1, 24, 4, 160]]     [1, 72, 4, 160]           0       \n",
      "     Conv2D-34       [[1, 72, 4, 160]]     [1, 72, 4, 160]         1,800     \n",
      "   BatchNorm-24      [[1, 72, 4, 160]]     [1, 72, 4, 160]          288      \n",
      "  ConvBNLayer-24     [[1, 72, 4, 160]]     [1, 72, 4, 160]           0       \n",
      "AdaptiveAvgPool2D-6  [[1, 72, 4, 160]]      [1, 72, 1, 1]            0       \n",
      "     Conv2D-35        [[1, 72, 1, 1]]       [1, 18, 1, 1]          1,314     \n",
      "     Conv2D-36        [[1, 18, 1, 1]]       [1, 72, 1, 1]          1,368     \n",
      "    SEModule-6       [[1, 72, 4, 160]]     [1, 72, 4, 160]           0       \n",
      "     Conv2D-37       [[1, 72, 4, 160]]     [1, 24, 4, 160]         1,728     \n",
      "   BatchNorm-25      [[1, 24, 4, 160]]     [1, 24, 4, 160]          96       \n",
      "  ConvBNLayer-25     [[1, 72, 4, 160]]     [1, 24, 4, 160]           0       \n",
      "  ResidualUnit-8     [[1, 24, 4, 160]]     [1, 24, 4, 160]           0       \n",
      "     Conv2D-38       [[1, 24, 4, 160]]     [1, 144, 4, 160]        3,456     \n",
      "   BatchNorm-26      [[1, 144, 4, 160]]    [1, 144, 4, 160]         576      \n",
      "  ConvBNLayer-26     [[1, 24, 4, 160]]     [1, 144, 4, 160]          0       \n",
      "     Conv2D-39       [[1, 144, 4, 160]]    [1, 144, 2, 160]        3,600     \n",
      "   BatchNorm-27      [[1, 144, 2, 160]]    [1, 144, 2, 160]         576      \n",
      "  ConvBNLayer-27     [[1, 144, 4, 160]]    [1, 144, 2, 160]          0       \n",
      "AdaptiveAvgPool2D-7  [[1, 144, 2, 160]]     [1, 144, 1, 1]           0       \n",
      "     Conv2D-40        [[1, 144, 1, 1]]      [1, 36, 1, 1]          5,220     \n",
      "     Conv2D-41        [[1, 36, 1, 1]]       [1, 144, 1, 1]         5,328     \n",
      "    SEModule-7       [[1, 144, 2, 160]]    [1, 144, 2, 160]          0       \n",
      "     Conv2D-42       [[1, 144, 2, 160]]    [1, 48, 2, 160]         6,912     \n",
      "   BatchNorm-28      [[1, 48, 2, 160]]     [1, 48, 2, 160]          192      \n",
      "  ConvBNLayer-28     [[1, 144, 2, 160]]    [1, 48, 2, 160]           0       \n",
      "  ResidualUnit-9     [[1, 24, 4, 160]]     [1, 48, 2, 160]           0       \n",
      "     Conv2D-43       [[1, 48, 2, 160]]     [1, 288, 2, 160]       13,824     \n",
      "   BatchNorm-29      [[1, 288, 2, 160]]    [1, 288, 2, 160]        1,152     \n",
      "  ConvBNLayer-29     [[1, 48, 2, 160]]     [1, 288, 2, 160]          0       \n",
      "     Conv2D-44       [[1, 288, 2, 160]]    [1, 288, 2, 160]        7,200     \n",
      "   BatchNorm-30      [[1, 288, 2, 160]]    [1, 288, 2, 160]        1,152     \n",
      "  ConvBNLayer-30     [[1, 288, 2, 160]]    [1, 288, 2, 160]          0       \n",
      "AdaptiveAvgPool2D-8  [[1, 288, 2, 160]]     [1, 288, 1, 1]           0       \n",
      "     Conv2D-45        [[1, 288, 1, 1]]      [1, 72, 1, 1]         20,808     \n",
      "     Conv2D-46        [[1, 72, 1, 1]]       [1, 288, 1, 1]        21,024     \n",
      "    SEModule-8       [[1, 288, 2, 160]]    [1, 288, 2, 160]          0       \n",
      "     Conv2D-47       [[1, 288, 2, 160]]    [1, 48, 2, 160]        13,824     \n",
      "   BatchNorm-31      [[1, 48, 2, 160]]     [1, 48, 2, 160]          192      \n",
      "  ConvBNLayer-31     [[1, 288, 2, 160]]    [1, 48, 2, 160]           0       \n",
      "  ResidualUnit-10    [[1, 48, 2, 160]]     [1, 48, 2, 160]           0       \n",
      "     Conv2D-48       [[1, 48, 2, 160]]     [1, 288, 2, 160]       13,824     \n",
      "   BatchNorm-32      [[1, 288, 2, 160]]    [1, 288, 2, 160]        1,152     \n",
      "  ConvBNLayer-32     [[1, 48, 2, 160]]     [1, 288, 2, 160]          0       \n",
      "     Conv2D-49       [[1, 288, 2, 160]]    [1, 288, 2, 160]        7,200     \n",
      "   BatchNorm-33      [[1, 288, 2, 160]]    [1, 288, 2, 160]        1,152     \n",
      "  ConvBNLayer-33     [[1, 288, 2, 160]]    [1, 288, 2, 160]          0       \n",
      "AdaptiveAvgPool2D-9  [[1, 288, 2, 160]]     [1, 288, 1, 1]           0       \n",
      "     Conv2D-50        [[1, 288, 1, 1]]      [1, 72, 1, 1]         20,808     \n",
      "     Conv2D-51        [[1, 72, 1, 1]]       [1, 288, 1, 1]        21,024     \n",
      "    SEModule-9       [[1, 288, 2, 160]]    [1, 288, 2, 160]          0       \n",
      "     Conv2D-52       [[1, 288, 2, 160]]    [1, 48, 2, 160]        13,824     \n",
      "   BatchNorm-34      [[1, 48, 2, 160]]     [1, 48, 2, 160]          192      \n",
      "  ConvBNLayer-34     [[1, 288, 2, 160]]    [1, 48, 2, 160]           0       \n",
      "  ResidualUnit-11    [[1, 48, 2, 160]]     [1, 48, 2, 160]           0       \n",
      "     Conv2D-53       [[1, 48, 2, 160]]     [1, 288, 2, 160]       13,824     \n",
      "   BatchNorm-35      [[1, 288, 2, 160]]    [1, 288, 2, 160]        1,152     \n",
      "  ConvBNLayer-35     [[1, 48, 2, 160]]     [1, 288, 2, 160]          0       \n",
      "    MaxPool2D-1      [[1, 288, 2, 160]]    [1, 288, 1, 80]           0       \n",
      "===============================================================================\n",
      "Total params: 259,056\n",
      "Trainable params: 246,736\n",
      "Non-trainable params: 12,320\n",
      "-------------------------------------------------------------------------------\n",
      "Input size (MB): 0.12\n",
      "Forward/backward pass size (MB): 44.38\n",
      "Params size (MB): 0.99\n",
      "Estimated Total Size (MB): 45.48\n",
      "-------------------------------------------------------------------------------\n",
      "\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "{'total_params': 259056, 'trainable_params': 246736}"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 定义网络输入shape\n",
    "IMAGE_SHAPE_C = 3\n",
    "IMAGE_SHAPE_H = 32\n",
    "IMAGE_SHAPE_W = 320\n",
    "\n",
    "\n",
    "# 可视化网络结构\n",
    "paddle.summary(MobileNetV3(),[(1, IMAGE_SHAPE_C, IMAGE_SHAPE_H, IMAGE_SHAPE_W)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "backbone output: [1, 288, 1, 80]\n"
     ]
    }
   ],
   "source": [
    "# 图片输入骨干网络\n",
    "backbone = MobileNetV3()\n",
    "# 将numpy数据转换为Tensor\n",
    "input_data = paddle.to_tensor([padding_im])\n",
    "# 骨干网络输出\n",
    "feature = backbone(input_data)\n",
    "# 查看feature map的纬度\n",
    "print(\"backbone output:\", feature.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "* neck\n",
    "\n",
    "neck 部分将backbone输出的视觉特征图转换为1维向量输入送到 LSTM 网络中，输出序列特征（ [源码位置](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/modeling/necks/rnn.py) ）："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "class Im2Seq(nn.Layer):\n",
    "    def __init__(self, in_channels, **kwargs):\n",
    "        \"\"\"\n",
    "        图像特征转换为序列特征\n",
    "        :param in_channels: 输入通道数\n",
    "        \"\"\" \n",
    "        super().__init__()\n",
    "        self.out_channels = in_channels\n",
    "\n",
    "    def forward(self, x):\n",
    "        B, C, H, W = x.shape\n",
    "        assert H == 1\n",
    "        x = x.squeeze(axis=2)\n",
    "        x = x.transpose([0, 2, 1])  # (NWC)(batch, width, channels)\n",
    "        return x\n",
    "\n",
    "class EncoderWithRNN(nn.Layer):\n",
    "    def __init__(self, in_channels, hidden_size):\n",
    "        super(EncoderWithRNN, self).__init__()\n",
    "        self.out_channels = hidden_size * 2\n",
    "        self.lstm = nn.LSTM(\n",
    "            in_channels, hidden_size, direction='bidirectional', num_layers=2)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x, _ = self.lstm(x)\n",
    "        return x\n",
    "\n",
    "\n",
    "class SequenceEncoder(nn.Layer):\n",
    "    def __init__(self, in_channels, hidden_size=48, **kwargs):\n",
    "        \"\"\"\n",
    "        序列编码\n",
    "        :param in_channels: 输入通道数\n",
    "        :param hidden_size: 隐藏层size\n",
    "        \"\"\" \n",
    "        super(SequenceEncoder, self).__init__()\n",
    "        self.encoder_reshape = Im2Seq(in_channels)\n",
    "\n",
    "        self.encoder = EncoderWithRNN(\n",
    "            self.encoder_reshape.out_channels, hidden_size)\n",
    "        self.out_channels = self.encoder.out_channels\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.encoder_reshape(x)\n",
    "        x = self.encoder(x)\n",
    "        return x\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "sequence shape: [1, 80, 96]\n"
     ]
    }
   ],
   "source": [
    "neck = SequenceEncoder(in_channels=288)\n",
    "sequence = neck(feature)\n",
    "print(\"sequence shape:\", sequence.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "* head\n",
    "\n",
    "预测头部分由全连接层和softmax组成，用于计算序列特征时间步上的标签概率分布，本示例仅支持模型识别小写英文字母和数字（26+10）36个类别（[源码位置](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/modeling/heads/rec_ctc_head.py)）:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "class CTCHead(nn.Layer):\n",
    "    def __init__(self,\n",
    "                 in_channels,\n",
    "                 out_channels,\n",
    "                 **kwargs):\n",
    "        \"\"\"\n",
    "        CTC 预测层\n",
    "        :param in_channels: 输入通道数\n",
    "        :param out_channels: 输出通道数\n",
    "        \"\"\" \n",
    "        super(CTCHead, self).__init__()\n",
    "        self.fc = nn.Linear(\n",
    "            in_channels,\n",
    "            out_channels)\n",
    "        \n",
    "        # 思考：out_channels 应该等于多少？\n",
    "        self.out_channels = out_channels\n",
    "\n",
    "    def forward(self, x):\n",
    "        predicts = self.fc(x)\n",
    "        result = predicts\n",
    "\n",
    "        if not self.training:\n",
    "            predicts = F.softmax(predicts, axis=2)\n",
    "            result = predicts\n",
    "\n",
    "        return result"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "在网络随机初始化的情况下，输出结果是无序的，经过SoftMax之后，可以得到各时间步上的概率最大的预测结果，其中：`pred_id` 代表预测的标签ID，`pre_scores` 代表预测结果的置信度："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "predict shape: [1, 80, 37]\n",
      "pred_id: Tensor(shape=[1, 80], dtype=int64, place=CUDAPlace(0), stop_gradient=False,\n",
      "       [[23, 28, 23, 23, 23, 23, 23, 23, 23, 23, 23, 30, 30, 30, 31, 23, 23, 23, 23, 23, 23, 23, 31, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 5 ]])\n",
      "pred_scores: Tensor(shape=[1, 80], dtype=float32, place=CUDAPlace(0), stop_gradient=False,\n",
      "       [[0.03683758, 0.03368053, 0.03604801, 0.03504696, 0.03696444, 0.03597261, 0.03925638, 0.03650934, 0.03873367, 0.03572492, 0.03543066, 0.03618268, 0.03805700, 0.03496549, 0.03329032, 0.03565763, 0.03846950, 0.03922413, 0.03970327, 0.03638541, 0.03572393, 0.03618102, 0.03565401, 0.03636984, 0.03691722, 0.03718850, 0.03623354, 0.03877943, 0.03731697, 0.03563465, 0.03447339, 0.03365586, 0.03312979, 0.03285240, 0.03273271, 0.03269565, 0.03269779, 0.03271412, 0.03273287, 0.03274929, 0.03276210, 0.03277146, 0.03277802, 0.03278249, 0.03278547, 0.03278742, 0.03278869, 0.03278949, 0.03279000, 0.03279032, 0.03279052, 0.03279064, 0.03279071, 0.03279077, 0.03279081, 0.03279087, 0.03279094, 0.03279106, 0.03279124, 0.03279152, 0.03279196, 0.03279264, 0.03279363, 0.03279509, 0.03279718, 0.03280006, 0.03280392, 0.03280888, 0.03281487, 0.03282148, 0.03282760, 0.03283087, 0.03282646, 0.03280647, 0.03275031, 0.03263619, 0.03242587, 0.03194289, 0.03122442, 0.02986610]])\n"
     ]
    }
   ],
   "source": [
    "ctc_head = CTCHead(in_channels=96, out_channels=37)\n",
    "predict = ctc_head(sequence)\n",
    "print(\"predict shape:\", predict.shape)\n",
    "result = F.softmax(predict, axis=2)\n",
    "pred_id = paddle.argmax(result, axis=2)\n",
    "pred_socres = paddle.max(result, axis=2)\n",
    "print(\"pred_id:\", pred_id)\n",
    "print(\"pred_scores:\", pred_socres)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "* 后处理\n",
    "\n",
    "识别网络最终返回的结果是各个时间步上的最大索引值，最终期望的输出是对应的文字结果，因此CRNN的后处理是一个解码过程，主要逻辑如下："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def decode(text_index, text_prob=None, is_remove_duplicate=False):\n",
    "    \"\"\" convert text-index into text-label. \"\"\"\n",
    "    character = \"-0123456789abcdefghijklmnopqrstuvwxyz\"\n",
    "    result_list = []\n",
    "    # 忽略tokens [0] 代表ctc中的blank位\n",
    "    ignored_tokens = [0]\n",
    "    batch_size = len(text_index)\n",
    "    for batch_idx in range(batch_size):\n",
    "        char_list = []\n",
    "        conf_list = []\n",
    "        for idx in range(len(text_index[batch_idx])):\n",
    "            if text_index[batch_idx][idx] in ignored_tokens:\n",
    "                continue\n",
    "            # 合并blank之间相同的字符\n",
    "            if is_remove_duplicate:\n",
    "                # only for predict\n",
    "                if idx > 0 and text_index[batch_idx][idx - 1] == text_index[\n",
    "                        batch_idx][idx]:\n",
    "                    continue\n",
    "            # 将解码结果存在char_list内\n",
    "            char_list.append(character[int(text_index[batch_idx][\n",
    "                idx])])\n",
    "            # 记录置信度\n",
    "            if text_prob is not None:\n",
    "                conf_list.append(text_prob[batch_idx][idx])\n",
    "            else:\n",
    "                conf_list.append(1)\n",
    "        text = ''.join(char_list)\n",
    "        # 输出结果\n",
    "        result_list.append((text, np.mean(conf_list)))\n",
    "    return result_list"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "以 head 部分随机初始化预测出的结果为例，进行解码得到："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tensor(shape=[1, 80], dtype=int64, place=CUDAPlace(0), stop_gradient=False,\n",
      "       [[23, 28, 23, 23, 23, 23, 23, 23, 23, 23, 23, 30, 30, 30, 31, 23, 23, 23, 23, 23, 23, 23, 31, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 5 ]])\n",
      "decode out: [('mrmmmmmmmmmtttummmmmmmummmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm4', 0.034180813)]\n"
     ]
    }
   ],
   "source": [
    "pred_id = paddle.argmax(result, axis=2)\n",
    "pred_socres = paddle.max(result, axis=2)\n",
    "print(pred_id)\n",
    "decode_out = decode(pred_id, pred_socres)\n",
    "print(\"decode out:\", decode_out)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "**小测试：** 如果输入模型训练好的index，解码结果是否正确呢？"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "out: [('pain', 1.0)]\n"
     ]
    }
   ],
   "source": [
    "# 替换模型预测好的结果\n",
    "right_pred_id = paddle.to_tensor([['xxxxxxxxxxxxx']])\n",
    "tmp_scores = paddle.ones(shape=right_pred_id.shape)\n",
    "out = decode(right_pred_id, tmp_scores)\n",
    "print(\"out:\",out)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "上述步骤完成了网络的搭建，也实现了一个简单的前向预测过程。\n",
    "\n",
    "没有经过训练的网络无法正确预测结果，因此需要定义损失函数、优化策略，将整个网络run起来，下面将详细介绍网络训练原理。\n",
    "\n",
    "\n",
    "## 3. 训练原理详解\n",
    "### 3.1 准备训练数据\n",
    "PaddleOCR 支持两种数据格式:\n",
    " - `lmdb` 用于训练以lmdb格式存储的数据集(LMDBDataSet);\n",
    " - `通用数据` 用于训练以文本文件存储的数据集(SimpleDataSet);\n",
    " \n",
    " 本次只介绍通用数据格式读取\n",
    "\n",
    "训练数据的默认存储路径是 `./train_data`, 执行以下命令解压数据："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "!cd /home/aistudio/work/train_data/ && tar xf ic15_data.tar "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "解压完成后，训练图片都在同一个文件夹内，并有一个txt文件（rec_gt_train.txt）记录图片路径和标签，txt文件里的内容如下:\n",
    "\n",
    "```\n",
    "\" 图像文件名         图像标注信息 \"\n",
    "\n",
    "train/word_1.png\tGenaxis Theatre\n",
    "train/word_2.png\t[06]\n",
    "...\n",
    "```\n",
    "\n",
    "**注意：** txt文件中默认将图片路径和图片标签用 \\t 分割，如用其他方式分割将造成训练报错。\n",
    "\n",
    "\n",
    "数据集应有如下文件结构：\n",
    "```\n",
    "|-train_data\n",
    "  |-ic15_data\n",
    "    |- rec_gt_train.txt\n",
    "    |- train\n",
    "        |- word_001.png\n",
    "        |- word_002.jpg\n",
    "        |- word_003.jpg\n",
    "        | ...\n",
    "    |- rec_gt_test.txt\n",
    "    |- test\n",
    "        |- word_001.png\n",
    "        |- word_002.jpg\n",
    "        |- word_003.jpg\n",
    "        | ...\n",
    "```\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "确认配置文件中的数据路径是否正确，以 [rec_icdar15_train.yml](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/configs/rec/rec_icdar15_train.yml)为例：\n",
    "\n",
    "```\n",
    "Train:\n",
    "  dataset:\n",
    "    name: SimpleDataSet\n",
    "    # 训练数据根目录\n",
    "    data_dir: ./train_data/ic15_data/\n",
    "    # 训练数据标签\n",
    "    label_file_list: [\"./train_data/ic15_data/rec_gt_train.txt\"]\n",
    "    transforms:\n",
    "      - DecodeImage: # load image\n",
    "          img_mode: BGR\n",
    "          channel_first: False\n",
    "      - CTCLabelEncode: # Class handling label\n",
    "      - RecResizeImg:\n",
    "          image_shape: [3, 32, 100]  # [3,32,320]\n",
    "      - KeepKeys:\n",
    "          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order\n",
    "  loader:\n",
    "    shuffle: True\n",
    "    batch_size_per_card: 256\n",
    "    drop_last: True\n",
    "    num_workers: 8\n",
    "    use_shared_memory: False\n",
    "\n",
    "Eval:\n",
    "  dataset:\n",
    "    name: SimpleDataSet\n",
    "    # 评估数据根目录\n",
    "    data_dir: ./train_data/ic15_data\n",
    "    # 评估数据标签\n",
    "    label_file_list: [\"./train_data/ic15_data/rec_gt_test.txt\"]\n",
    "    transforms:\n",
    "      - DecodeImage: # load image\n",
    "          img_mode: BGR\n",
    "          channel_first: False\n",
    "      - CTCLabelEncode: # Class handling label\n",
    "      - RecResizeImg:\n",
    "          image_shape: [3, 32, 100]\n",
    "      - KeepKeys:\n",
    "          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order\n",
    "  loader:\n",
    "    shuffle: False\n",
    "    drop_last: False\n",
    "    batch_size_per_card: 256\n",
    "    num_workers: 4\n",
    "    use_shared_memory: False\n",
    "    ```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "### 3.2 数据预处理\n",
    "\n",
    "送入网络的训练数据，需要保证一个batch内维度一致，同时为了不同维度之间的特征在数值上有一定的比较性，需要对数据做统一尺度**缩放**和**归一化**。\n",
    "\n",
    "为了增加模型的鲁棒性，抑制过拟合提升泛化性能，需要实现一定的**数据增广**。\n",
    "\n",
    "* 缩放和归一化\n",
    "\n",
    "第二节中已经介绍了相关内容，这是图片送入网络之前的最后一步操作。调用 `resize_norm_img` 完成图片缩放、padding和归一化。\n",
    "\n",
    "* 数据增广\n",
    "\n",
    "PaddleOCR中实现了多种数据增广方式，如：颜色反转、随机切割、仿射变化、随机噪声等等，这里以简单的随机切割为例，更多增广方式可参考：[rec_img_aug.py](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/data/imaug/rec_img_aug.py)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def get_crop(image):\n",
    "    \"\"\"\n",
    "    random crop\n",
    "    \"\"\"\n",
    "    import random\n",
    "    h, w, _ = image.shape\n",
    "    top_min = 1\n",
    "    top_max = 8\n",
    "    top_crop = int(random.randint(top_min, top_max))\n",
    "    top_crop = min(top_crop, h - 1)\n",
    "    crop_img = image.copy()\n",
    "    ratio = random.randint(0, 1)\n",
    "    if ratio:\n",
    "        crop_img = crop_img[top_crop:h, :, :]\n",
    "    else:\n",
    "        crop_img = crop_img[0:h - top_crop, :, :]\n",
    "    return crop_img\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# 读图\n",
    "raw_img = cv2.imread(\"/home/aistudio/work/word_1.png\")\n",
    "plt.figure()\n",
    "plt.subplot(2,1,1)\n",
    "# 可视化原图\n",
    "plt.imshow(raw_img)\n",
    "# 随机切割\n",
    "crop_img = get_crop(raw_img)\n",
    "plt.subplot(2,1,2)\n",
    "# 可视化增广图\n",
    "plt.imshow(crop_img)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "### 3.3 训练主程序\n",
    "\n",
    "模型训练的入口代码是 [train.py](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/tools/train.py)，它展示了训练中所需的各个模块： `build dataloader`, `build post process`, `build model` , `build loss`, `build optim`, `build metric`，将各部分串联后即可开始训练：\n",
    "\n",
    "* 构建 dataloader\n",
    "\n",
    "训练模型需要将数据组成指定数目的 batch ，并在训练过程中依次 yield 出来，本例中调用了 PaddleOCR 中实现的 [SimpleDataSet](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/data/simple_dataset.py)\n",
    "\n",
    "基于原始代码稍作修改，其返回单条数据的主要逻辑如下"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def __getitem__(data_line, data_dir):\n",
    "    import os\n",
    "    mode = \"train\"\n",
    "    delimiter = '\\t'\n",
    "    try:\n",
    "        substr = data_line.strip(\"\\n\").split(delimiter)\n",
    "        file_name = substr[0]\n",
    "        label = substr[1]\n",
    "        img_path = os.path.join(data_dir, file_name)\n",
    "        data = {'img_path': img_path, 'label': label}\n",
    "        if not os.path.exists(img_path):\n",
    "            raise Exception(\"{} does not exist!\".format(img_path))\n",
    "        with open(data['img_path'], 'rb') as f:\n",
    "            img = f.read()\n",
    "            data['image'] = img\n",
    "        # 预处理操作，先注释掉\n",
    "        # outs = transform(data, self.ops)\n",
    "        outs = data\n",
    "    except Exception as e:\n",
    "        print(\"When parsing line {}, error happened with msg: {}\".format(\n",
    "                data_line, e))\n",
    "        outs = None\n",
    "    return outs"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "假设当前输入的标签为 `train/word_1.png\tGenaxis Theatre`, 训练数据的路径为 `/home/aistudio/work/train_data/ic15_data/`, 解析出的结果是一个字典，里面包含 `img_path` `label` `image` 三个字段："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'img_path': '/home/aistudio/work/train_data/ic15_data/train/word_1.png', 'label': 'Genaxis Theatre', 'image': b'\\x89PNG\\r\\n\\x1a\\n\\x00\\x00\\x00\\rIHDR\\x00\\x00\\x00Y\\x00\\x00\\x00\\x0e\\x08\\x02\\x00\\x00\\x00\\xcb\\xe2\\'\\xb7\\x00\\x00\\x00\\x01sRGB\\x00\\xae\\xce\\x1c\\xe9\\x00\\x00\\x00\\x04gAMA\\x00\\x00\\xb1\\x8f\\x0b\\xfca\\x05\\x00\\x00\\x00 cHRM\\x00\\x00z&\\x00\\x00\\x80\\x84\\x00\\x00\\xfa\\x00\\x00\\x00\\x80\\xe8\\x00\\x00u0\\x00\\x00\\xea`\\x00\\x00:\\x98\\x00\\x00\\x17p\\x9c\\xbaQ<\\x00\\x00\\x0bmIDATHK\\x8d\\x96\\xf9S[\\xd7\\x15\\x80\\x01\\xa7\\x93\\xa4\\xfd1\\x99L\\xea\\x80\\xc4\\xa2]B\\x0bb\\xdf\\x84\\x04\\x18\\x8c\\x01\\xb3\\x8aE\\xec\\x12\\x02\\t\\xb4KhC\\xfb\\xbe=\\xed\\xbb\\x04\\xc2l&N\\xd2\\xb4\\x93i\\x9bv\\xa6\\x7fL\\xdb\\xe9d\\xe2N\\xd3d<u6C\\x8f\\xc0I\\x9c\\xb1\\x7f\\xc8\\x9d\\x8f3\\xf7\\xdd\\xf7\\xee\\x9d{\\xbfw\\xceC\\x95\\xd3\\xc3\\xe3\\x04\\x02\\xe1\\xad\\xb7\\xde\\xbazv\\xf9\\xcd7\\xdf<~\\xfc\\xf8\\xef\\xff\\xfc\\xc7\\xe7\\x9f\\xff\\xfb\\x7fO\\x9f>y\\xf2\\xe4\\xeb\\xaf\\xbf\\xbe\\xbc\\xbc|\\xedWU\\xaf]\\xb7\\xab\\xab\\xab\\xca\\xca\\xab\\x8a\\x8a\\x8a\\xca\\xca\\xca\\xeb\\x08\\x97\\xb7\\xa0sY\\xf1\\xda\\xf7\\xdf?\\x83vUYQUU\\x05wa\\xd6w\\xcf\\x9e\\xdd\\xbau\\xeb\\xf5\\xd7_G\\xa1Po\\xfe\\xe6\\xd7\\x9f}\\xf6\\x19,\\x08+|\\xf9\\xe5\\x97U\\xe5\\xd9\\xcf\\xdb\\x1bo\\xbc\\xf1\\xf6\\xdbo\\xdf\\xbe}\\x1b\\xf6\\xf0\\xce;\\xef\\xc0\"\\xdf~\\xfb\\xedw\\xdf}\\xf7\\xf4\\xe9\\xd3\\xaf\\xbe\\xfa\\xea\\x8b/\\xbex\\xf2\\xe4\\xbf\\xffy\\xfc\\xaf[\\x95\\xcf`eh\\xcf\\xa7]\\x96\\xb7\\x01\\x8f\\x95\\xb7q\\xf5\\xe3~*\\xe16l\\xa6\\xb2\\xf2\\xcd\\xab\\x8a[\\xb0\\rh\\xd7\\xfb)?\\x00\\rV\\xae\\xac*O\\xfc\\xe1\\x08\\x97\\x15\\x15U\\x95W\\x97\\xf0W\\x05\\xf3LJ]&\\x14\\x07r\\xa1D\\xdc\\x1d\\xd2\\x8a\\x15\\x0b\\xe3\\\\J-\\x89\\x88&\\xe0\\xaa\\xb1\\xb8j\\x0c\\xbe\\x06K@\\xe1\\x08h\\x0c\\x1e\\xd5@@\\xd7\\x03xT\\x1d\\x1e\\x85\\xc6\\xd5\\xa0p\\xa8jlu\\r\\xb6\\x1aM\\xc1R\\xb0\\xb5x\\xf4\\xed\\xba\\x9aw\\xd1\\x10\\xd1\\xd5\\r5\\xb7\\xeb\\xde{\\x17M\\xc0R\\x18\\xf4\\xd6\\xd9\\x99\\xc5\\x95UAw\\x0f\\xa7\\x91\\xca\\xa4\\xd2\\x9a\\xd1\\xb5\\xd8\\xea\\xf7jQ\\xa82ht\\x1d\\x99\\xdc8<<\\xa2P\\xa8\\xc2\\xe1\\xe8\\xe9\\xe9\\xf9\\xe1\\xe1Q:\\x9d\\x8dF\\xe3\\x1e\\x8f\\x0f\\x06gg\\xe7\\xfazY4\\n\\x9eN\\xc6\\xd2H\\x18\\x80Jl\\x00h\\x84z\\x80\\xd4P\\x03P\\x1aP@#\\x06\\rP\\x1b\\x80\\xba\\x16\\x12\\x9d\\x81\\xa7Q\\xeaI\\xb0\\xf3\\xf2\\xfe\\xd1\\x18r=\\xbe\\x11C\\xa4bIT,\\x81\\x86#\\xde@\\xc7\\x13\\xe8x\\x12\\x03Oh&\\x90\\x80\\n\\x97\\xc1\\x1a\\xb0\\xba5b\\xb9f[\\xa6\\x12J\\x16F\\xa79m\\xacNz\\x07\\xa7\\xbd\\x7f|\\xe0>orq}nmy\\x9a73256p\\xf7\\x1egh\\x84=0\\xd4\\xcb\\x1e\\xe8\\xeaf\\xb7w\\xf4\\xb40\\xbb\\x98\\x8c\\xf6\\xa6\\xa6\\xb1\\xe1\\xf1;\\xec\\xe1\\xde\\x8e\\xbe\\xae\\xd6\\x9e\\xaevVO\\'\\xbb\\xbb\\xa3\\xaf\\xb3\\xad\\xb7\\xbd\\xad\\xa7\\x8f5\\xb8\\xc1\\xdf\\xde\\x16I9\\xfd\\xc3\\xadm\\xdd}\\xec;\\x10{{8,\\x16\\xbb\\xb3\\xb3\\x9b\\xc1`R\\xa9\\xf4\\xbe>\\xce\\xc6\\x86\\xc0d\\xb2 H\\x04\\x14X\\xadv\\x8b\\xc5\\x06\\x972\\x99\\x82\\xcb\\x9d\\xe7@c\\xc3\\x84vVw[w\\x1b\\xa3\\x9dIke\\x90\\x9bi\\xe4\\x16*\\x81\\x8ak\\xa0\\xe2\\xeb\\xe88\\x0c\\x1d_\\x0f\\x91Ah`\\xe0\\xb1\\x0c<N\"\\xd8\\x12\\xf0\\xd6f\\xee\\x8d\\x0fv\\xb3:\\xe8L&\\xa9\\xb1\\x89H\\x02\\xa8X\\\\\\x19\\x1c\\x86\\x86\\xc7\\x02t\\x02\\xee\\x06&\\x11\\\\P*\\xfc\\x16\\x97\\xdbh\\xdb\\\\\\\\\\x1b\\xeb\\x1bj\\'7\\xb5\\x91\\x18c\\x9c\\x11\\xadDg\\xd5\\xb9\\x10w<\\x1d\\xce\\x05\\x1c\\x88tK&\\x11J\\xbd6o\\xc8\\x13\\xc8\\xc4\\x12>\\x87K\\xabP\\xcdO\\xcd\\x88\\xf8|\\xbdJ\\x15\\xf2\\xf8b\\xa1\\x84\\xdb\\xea\\x97l)\\xf6\\xf7\\xac\\xd2\\x1d\\xb5\\x90\\xbf;9>\\'X\\x17\\xabU\\xc6M\\xc1\\xae\\\\\\xae\\xdb\\xda\\x92\\t\\x85R\\x91Ha\\xb7\\xfbuz\\xb3i\\xdff1;dR\\xd5\\xf8\\xd8T#\\x85\\x81\\xaa\\xa9\\xef\\xec\\xe8\\x15\\x8b\\xa4>oH.S//\\xadK%J$\\x14\\x8b\\xc7\\xd20\\xe2\\xf3\\x05\\x90P`\\x89\\xc7\\x9d\\x9d\\x990\\x194\\xbb\\xdb\\x9bF\\x9d\\xca\\xa8U\\x8b\\x04k3\\xf7GW\\x17\\xb9\\x06\\xa5\\xdc\\xaa\\xd3l\\xaf\\xadl,\\xce9\\x8c\\xfa\\xb3\\x83\\xc2I\\xbe\\xe0w:\\xb5J\\x99d{sG\\xc8WJ\\xc4R\\xd1&\\x8f;\\x05S\\xb4J\\xa9A#S\\xcbD\"\\xc1\\xca\\xf4\\xf8p?\\xab\\xf5No\\x17\\xb8`\\x82\\x0b\\xcf\\xbe\\x1d\\\\\\xec\\xac\\t\\xf9\\xdc\\xe5\\xd5\\xa9\\x85\\xe5\\x899!\\x8f\\xef4\\x94E\\xe4\\xe3\\xa5\\x93\\xc2E.V4i\\xacz\\xc5~,\\x10?H\\x1f\\x94\\xb2\\xa5\\\\\"\\x93\\x8a$T\\x12\\x85Eo\\n\\xfb\\x02\\xa5\\xdc\\xc1\\xf1\\xc1\\xe9\\xc9\\xc1C\\xc4\\x1b\\x8b\\x87\\xd2\\x01W8\\xe0\\x8d\\x1a\\xb4V\\x9b\\xd9c6:\\x95\\n\\xbd\\xd3\\x110\\x1a\\x1dz\\xbdM\\xa7\\xb3\\xc6\\xe3\\xf9P(\\x19\\xf0G\\xcc&\\xfb\\x96pgrb\\xb6\\xa5\\xb9\\xa3\\xa6\\xba\\x8eAoY[\\x15\\xd8mn\\x10\\x04.\\x14rM&]8*\\x9d\\xa6S\\xf9L&\\x97\\xcf\\xe6\\xe42\\x89T\\xb2\\x93I\\'\\x11\\xbf\\xaf\\x90Ig\\x12q\\xa7\\xd5bP\\xab=6[>\\x99<+\\x1e\\xc4\\xfd\\x01\\x9dL\\xae\\x95\\xca\\x02\\x0e\\xdb\\x87\\x0fO\\x83\\x1e\\xe7\\xbeV\\xe5uX\\xd21$\\x9f\\x8a\\xa5\\xa2\\x81\\xa0\\xc7\\x96\\x8e\\x05\\x8f\\n\\xa9\\xd3\\xa3\\xccI)\\x9dM\\x06\\xf7\\x94\\xa2\\xd9\\xc9;c\\xc3\\x1cH\\x19&\\x81Zv\\x818|z\\xa9\\xda\\xb1g\\x8a\\xb9\\x82~\\x93\\xd3\\xa2\\xde\\xcf\"\\x85\\x83\\xd4\\xe9i\\xf1\\xfd\\x8b\\xa3\\x8f\\x8e\\xb2g\\x01G$\\xe8\\x8c\\x9e\\x1f\\xbe\\x0f9\\xb2\\xaf19M\\xae|\\xa2`T\\x1b\\xdd\\x16w.\\x9e;/\\x9d?:\\xfe\\xe8\\xcf\\x1f\\xff\\xf5({\\x02:\\x82\\xee\\xc8\\x83\\xc2Y!UJ\\xc5\\x0b\\xfbz\\xbbVm\\x8a\\x84R~_\\xd4a\\xf7[\\xcc\\xee|\\xeeA6S\\x8aG\\xd2&\\x83M\\xb0\\xbe=39\\x0f\\xd5\\x04\\xdf\\x17\\n\\x91\\xbe\\xc0]v\\xd9}\\n\\xa9fyq].Q\\xa7\\xe2\\xb9D4c\\xb7\\xb8\\xddN_<\\x9eT\\xa9T:\\x9d\\xee\\xe4\\xe4\\xa4\\x98/\\\\\\x9c?<:,\\x85C\\x88`m]\\xbe+1\\x1b\\x8c\\xc9p\\xf4(_\\x0c\\xb8<6\\xa3)\\x1b\\x8f?:;\\xd6\\xab\\xe5\\xfc\\x95E\\xbbI\\x9fKF\\x11\\x9f\\xcb\\xb8\\'\\xd7\\xa9$`$\\x9b@\\x8a\\xd9\\xc8\\xc5i\\xe1\\xfc8g7kxs\\xa3\\xf7G\\x06\\x9a\\x88\\x94&\\xe2\\xb5\\x8b\\xa0\\xcd\\xa3\\x10\\xee\\x82\\x8b\\x07\\xc9\\xc2a,\\x8b8\\x02\\xb9p1\\x1b-%C\\xb9\\x98?\\xed\\xb7\\x87M\\x1a\\xbb\\xd7\\x1a:?\\xfc\\xc0m\\x0e\\xac/\\x08w\\x05\\x8a\\\\\\xec\\xd0at\\x07\\x9d\\xe1\\xc3\\xf4\\x83\\x8b\\xa3\\x0f\\xc0Q1Yr\\x9b}|\\x1e\\xdc\\x95\\xa5\"9H\\x93b\\xfaH.V\\xef\\x8a\\x94>w\\xd8i\\xf3C\\xa6\\xc8eZ$\\x90\\x005\\xf1H\\xd6\\xebB\\xd4R\\xdd\\x02w\\xb5\\xa7\\x8d]\\x87\\xc2\\xd3\\x88\\xcc\\xd9)\\x9e\\xdf\\x89@\\x89-L/+\\xa4\\xdad8\\x0bSV\\x17\\xf8\\xbb\"y*\\x99\\xb3Y]n\\x97\\xff\\xe2\\xe1\\x87\\xb9L\\xb1\\x98/E\\x90\\xb8\\xdd\\xea\\xda\\\\\\x13*%*\\x9dJ\\x1f\\xf6G\\xe0\\x95 \\x1eD\\xaf\\xd4G\\xfc\\xc8\\xe9\\xe1!\\x94\\xc6\\xd4\\xe8\\xa8A\\xa3LE\\x11\\x9dJ>?=\\xb18;Y\\xcc$R\\xd1P4\\xe8\\xcd%\\xc3\\x80F.\\x86\\xa4`w\\xb7\\x97]\\x10\\xe8\\x15P ^\\x93C\\xbc\\xbai\\x94j\\n\\xe1\\x14\\xb8\\x88{#p\\xfe\\xa0;\\xe1\\xb1\\x85\\xad\\x06\\xafFf\\x12\\xf1\\x15*\\x891\\x16\\xcc\\x01K\\\\\\xc1\\xea\\xfcv\"Tt\\xec\\x07=\\xd6H,\\x90/$\\x8e\\xddF\\x1f\\xa4\\x92\\xcf\\x1c\\x14\\xf2D:\\x89\\xf1A\\xee\\xfc\\xd3\\xdf\\xff\\xed\\x93\\x0f>U\\x8a\\xb5\\xbbB\\xa5\\xc7\\x16\\xb2\\xef{\\xb5\\n\\x93xK\\xe9s\\x84C\\x9ex1}|z\\xf8\\x08D\\x0bV\\xc4\\xdd\\xcdl\\\\-\\xa5\\x99\\xd21yo\\x0e\\xb2O\\xb1\\xa3]\\x9e\\xdd\\xb0\\x1a\\xdcg\\x07\\x8fLZ\\x07\\xf4\\x8d\\x1a\\x1b\\xa4R \\x18\\xf5\\x07\"\\xa5\\xa33\\x9f\\x17q\\xd8\\xbd*\\xa5nG,\\x97\\x88\\x15\\x0e\\xab\\'\\x12\\x88\\x9d\\x1c\\x9e}\\xf2\\xbb?F\\x83q\\xc9\\xb6L\\xab\\xd4\\xe7S\\x05\\xa9H2?\\xbd`1\\x98\\x0f\\xb2E\\x93n\\x7f~zne\\x91\\x97\\x8a\\xc5\\xa3\\xa1`\\xd0\\xebID\\xc2\\xf10\\xa2\\x94JF\\x87\\x87\\xd8=\\xddt\\x12\\x95N\\xa4W\\x98U\\xfa\\x98\\'\\x04:R\\xbeH)\\x9eCl\\xde}\\xa8\\xf0}_\\xc4\\x9fI\\x86\\x8ba_\\xdaf\\xf4I\\xb64\\n\\xb1\\xc1g\\x8f\\x85\\xdc\\xa9\\xb9\\x89\\x8d\\xa9{\\xcb\\x1eK\\xc2\\xbc\\x170i\\xfcvC8\\xee/\\xfa-\\xe1\\x847\\x15\\xb4\"R\\xbe\\xc2\\xa4\\xb4\\x06m\\xe1|\\xe4 \\x85\\xe4\\x1dF\\xafFb\\x00\\xb3N\\xb3\\x1f\\x8e\\x04\\xe7\\x0c\\xfb\\x92\\xf1P\\xf6A\\xee\\xe2\\xe3\\x87\\x7f\\xca\\xc5\\x8e\\xf8<q;\\x9dE\\xa8\\xa51\\xc9\\x9dC\\xac\\xf1\\x80#f\\xd28E\\xebr\\xb79\\xf4\\xe8\\xf8\\x0f.Spsy\\xd7k\\x8f\\x04|\\t\\x87#`\\xb3\\x052 \\xc5\\x97\\x08\\xf8cF\\xbdS\\xbbgVH\\xf5\\x90_\\xd9\\xf8\\xe1\\xc5\\xf1\\x87\\x9f|\\xf4\\x97R\\xee\\x04\\xbc\\xfb\\x1c\\xa1B\\xea\\x01D\\xad\\xc2\\x08\\x11\\xd2\\xd3\\xa4\\xb3p\\'\\xe7\\xe7\\xa6\\x16\\x82\\x1e$\\x19I\\xe5\\x92\\xf9B:\\x0f\\x1d\\x95T9~\\xf7>\\xab\\xb3\\x0fD\\x94]\\xf0\\xe7W@\\x07\\xfcg\\x85\\x8f\\x85\\x07jxn\\xe5\\xfe\\xc0\\xa8`I\\xa4W\\xd9 /\\xdcVD+\\xb7n\\xf0v\\xf8KR\\xa3\\xda\\xad\\x91Z\\xc7\\x06\\x17\\xef\\xf4Nk$N\\xd9\\x96y{M+\\xde\\xd0{\\xcc\\xf1<r\\x98\\xf4\\xa4\\xc2\\xb6\\x88^bth]J\\xa1Z\\xb7k\\x00)F\\xb9\\x99\\xcf\\xdb\\x96okv\\xf8\\xf2\\xad\\xd5]\\xa8/\\x8b\\xd6\\t\\x82\\xc2\\xceD\\xc4\\x95\\xda\\xdb\\xd9\\x1f\\xe3L7\\xe1\\xdb\\xf1\\xd5\\x8d4L\\x0b\\xbbuH\\xb5\\xad7\\xc8\\xac\\xeb\\xdc-\\xc1\\x82\\xd8\\xa9\\xf7K\\xf9\\xea\\xa9\\xa1\\xf9]\\x81J.5\\x8aD*\\xa1P\\xe9r\\x85\\x83\\xc1L:}\\xe4vF\\xadf\\x9flGo\\xd9\\xf7\\x04\\xbdP\\x07\\xf0\\x1d(@\\xc6\\x99\\xb4.\\xb9Xg7\\x06\"\\xbe\\xac\\xcb\\x8c\\x98\\xb5\\x1e\\xb7%,\\xdd\\xde\\x1b\\x19\\x98\\xba\\xd37n\\xd1{R\\x91\\xc2a\\xf6\\x14\\n\\x10R\\x15\\n\\xb0\\xbb\\x95\\xd3\\x88g\\xd0\\x89e*\\xfa\\xdb{GX\\x83\\xa3\\xec\\xa1\\xf9\\xb1\\xe9\\xf1\\xfe\\xbb\\xad\\xc4&\\x1a\\x86B\\xc1\\xd0\\x06zG\\xb8\\x13K\\xc0\\xdd\\xfe\\xc9\\x0e&\\xbb\\xafs\\x987#\\x98\\x1cYl\\xa1\\xf6\\x911\\xad\\x93#KC\\xac\\x99v\\xfa`+\\xb5\\x7fiZ\\xb8\\xb5\\xb4;3<3\\xce\\x19\\xe7\\xb4rF9c4,\\xbd\\x83\\xd6y\\x8f=\\xba8\\xc1\\xebnf\\rt\\x0f\\xf5\\xb5\\xf7\\x03=-}K3\\xab\\xabs\\x02\\xf9\\xa6fcn\\xebN\\xd7=\\x06\\xae\\x95\\x84\\xa6a~K\\x84\\xd8\\xcb\\xec\\x1f\\xee\\x19[\\xe7\\n\\x07;G`|bpv\\x94=\\xd9L\\xec\\x18\\xec\\x19\\xbd\\xd37\\xda\\xd5\\xde\\xcf\\xa4wq\\xa7\\x97\\x84|\\xa9Fi\\\\\\x99\\x17L\\x8dr\\xe9\\xa4\\xe6\\x81\\xde\\xe1\\x15\\xee\\xfa\\xd6\\xaa\\x18>U\\xcb\\xb3kC\\xac\\x11\\x1c\\x8a\\xdc\\xc9\\xe0\\xac\\xcf\\xef\\x00Sw\\x17a{\\xd3#\\xbc\\x0e:\\xbb\\x85\\xd2C\\xc5\\xb6\\xc2\\x88pE\\xb2\\xbe\\xb0=w\\x7f\\x052\\xb1\\x95\\xdaC\\'\\xb44\\x11\\x98M\\x04F\\x05\\xa9\\x96\\xf0\"d4\\x11.\\xf1\\xb5\\xf8k\\xa0\\xf3\\x02h\\x12\\xbe\\x0c\\xe5\\x07\\xa8x\\xf4\\r\\x14\\x1c\\x8a\\x88C\\xe3_\\t\\xfc\\x1e}\\x19b\\x1d\\x85P\\xdb\\x08\\xdc,u\\xd3\\'\\xd6Qo:7\\xfd\\x1bH\\xf5\\xf0\\xf3\\xb1\\x91\\x8e\\xa1\\xd2\\xb1\\x8d\\xd7\\x90_\\x8c\\x0c\\x1c\\x05h\\xc2\\x91^\\x84\\x81\\xa3\\xd21\\xf0F[\\xaei\\xfe!B\\xa7\\x99\\x8emy\\x15\\xcc&\\\\S\\x13\\x8e^A\\xa8\\xfb\\xd1\\x05Xx\\xce\\xab]\\xfc\\xe8\\xe5\\xb9\\x94\\x1b5e0h\\x1c\\x06\\x8d}9^[\\xc0\\xbd\\x14\\xf1\\x84:\"X~n\\xf0\\xda#,\\x02\\x83\\x84Z2\\xb1\\x9e\\x04\\xa6 \\x92\\xea\\x1bI\\rd\\x88\\x94z2\\x03\\xd3\\xc8\\xc4\\x90\\x99\\x18\\xe2\\xab\\xc031?\\x83\\x81%2\\xb04\\x06\\x96\\xc1\\xc02\\x7fadb\\x9b\\x98/\\xba\\x80\\x8c\\xb8\\x01\\x8c\\xfc\\xfc\\xfd\\xdf$\\xc2O\\'\\x7f!5\\x9e\\xdf\\xc2\\xa0\\x08e\\x11/E,\\x9aX\\x16\\xf1R,\\x8b@\\xe3\\xb0(\\x00{\\x13\\xcb:j\\xf1e\\x11\\xf5\\xc4\\xb2\\x88\\x06\\x80L\\xc6\\x00\\x94\\xc6\\x06\\xc8\\x05\\xd0\\xf1\\n\\xe8\\r\\xa4\\x17\\xf9\\xe9\\x99\\xeb|\\x81\\x04\\xf9e\\x91\\n\"\\x80\\xff\\x03\\x99\\xa0+\\x94\\xbd\\xf0X\\xa1\\x00\\x00\\x00\\x00IEND\\xaeB`\\x82'}\n"
     ]
    }
   ],
   "source": [
    "data_line = \"train/word_1.png\tGenaxis Theatre\"\n",
    "data_dir = \"/home/aistudio/work/train_data/ic15_data/\"\n",
    "\n",
    "item = __getitem__(data_line, data_dir)\n",
    "print(item)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "实现完单条数据返回逻辑后，调用 `padde.io.Dataloader` 即可把数据组合成batch，具体可参考 [build_dataloader]()\n",
    "\n",
    "\n",
    "* build model\n",
    "\n",
    "  build model 即搭建主要网络结构，具体细节如《2.3 代码实现》所述，本节不做过多介绍，各模块代码可参考[modeling](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.3/ppocr/modeling)\n",
    "\n",
    "* build loss\n",
    "  \n",
    "  CRNN 模型的损失函数为 CTC loss, 飞桨集成了常用的 Loss 函数，只需调用实现即可："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import paddle.nn as nn\n",
    "class CTCLoss(nn.Layer):\n",
    "    def __init__(self, use_focal_loss=False, **kwargs):\n",
    "        super(CTCLoss, self).__init__()\n",
    "        # blank 是 ctc 的无意义连接符\n",
    "        self.loss_func = nn.CTCLoss(blank=0, reduction='none')\n",
    "\n",
    "    def forward(self, predicts, batch):\n",
    "        if isinstance(predicts, (list, tuple)):\n",
    "            predicts = predicts[-1]\n",
    "        # 转置模型 head 层的预测结果，沿channel层排列\n",
    "        predicts = predicts.transpose((1, 0, 2)) #[80,1,37]\n",
    "        N, B, _ = predicts.shape\n",
    "        preds_lengths = paddle.to_tensor([N] * B, dtype='int64')\n",
    "        labels = batch[1].astype(\"int32\")\n",
    "        label_lengths = batch[2].astype('int64')\n",
    "        # 计算损失函数\n",
    "        loss = self.loss_func(predicts, labels, preds_lengths, label_lengths)\n",
    "        loss = loss.mean()\n",
    "        return {'loss': loss}"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "* build post process\n",
    "\n",
    " 具体细节同样在《2.3 代码实现》有详细介绍，实现逻辑与之前一致。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "* build optim\n",
    "\n",
    "优化器使用 `Adam` , 同样调用飞桨API： `paddle.optimizer.Adam`\n",
    "\n",
    "* build metric\n",
    "\n",
    "metric 部分用于计算模型指标，PaddleOCR的文本识别中，将整句预测正确判断为预测正确，因此准确率计算主要逻辑如下："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def metric(preds, labels):    \n",
    "    correct_num = 0\n",
    "    all_num = 0\n",
    "    norm_edit_dis = 0.0\n",
    "    for (pred), (target) in zip(preds, labels):\n",
    "        pred = pred.replace(\" \", \"\")\n",
    "        target = target.replace(\" \", \"\")\n",
    "        if pred == target:\n",
    "            correct_num += 1\n",
    "        all_num += 1\n",
    "    correct_num += correct_num\n",
    "    all_num += all_num\n",
    "    return {\n",
    "        'acc': correct_num / all_num,\n",
    "    }"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "acc: {'acc': 0.6}\n"
     ]
    }
   ],
   "source": [
    "preds = [\"aaa\", \"bbb\", \"ccc\", \"123\", \"456\"]\n",
    "labels = [\"aaa\", \"bbb\", \"ddd\", \"123\", \"444\"]\n",
    "acc = metric(preds, labels)\n",
    "print(\"acc:\", acc)\n",
    "# 五个预测结果中,完全正确的有3个，因此准确率应为0.6"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "将以上各部分组合起来，即是完整的训练流程："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "\n",
    "def main(config, device, logger, vdl_writer):\n",
    "    # init dist environment\n",
    "    if config['Global']['distributed']:\n",
    "        dist.init_parallel_env()\n",
    "\n",
    "    global_config = config['Global']\n",
    "\n",
    "    # build dataloader\n",
    "    train_dataloader = build_dataloader(config, 'Train', device, logger)\n",
    "    if len(train_dataloader) == 0:\n",
    "        logger.error(\n",
    "            \"No Images in train dataset, please ensure\\n\" +\n",
    "            \"\\t1. The images num in the train label_file_list should be larger than or equal with batch size.\\n\"\n",
    "            +\n",
    "            \"\\t2. The annotation file and path in the configuration file are provided normally.\"\n",
    "        )\n",
    "        return\n",
    "\n",
    "    if config['Eval']:\n",
    "        valid_dataloader = build_dataloader(config, 'Eval', device, logger)\n",
    "    else:\n",
    "        valid_dataloader = None\n",
    "\n",
    "    # build post process\n",
    "    post_process_class = build_post_process(config['PostProcess'],\n",
    "                                            global_config)\n",
    "\n",
    "    # build model\n",
    "    # for rec algorithm\n",
    "    if hasattr(post_process_class, 'character'):\n",
    "        char_num = len(getattr(post_process_class, 'character'))\n",
    "        if config['Architecture'][\"algorithm\"] in [\"Distillation\",\n",
    "                                                   ]:  # distillation model\n",
    "            for key in config['Architecture'][\"Models\"]:\n",
    "                config['Architecture'][\"Models\"][key][\"Head\"][\n",
    "                    'out_channels'] = char_num\n",
    "        else:  # base rec model\n",
    "            config['Architecture'][\"Head\"]['out_channels'] = char_num\n",
    "\n",
    "    model = build_model(config['Architecture'])\n",
    "    if config['Global']['distributed']:\n",
    "        model = paddle.DataParallel(model)\n",
    "\n",
    "    # build loss\n",
    "    loss_class = build_loss(config['Loss'])\n",
    "\n",
    "    # build optim\n",
    "    optimizer, lr_scheduler = build_optimizer(\n",
    "        config['Optimizer'],\n",
    "        epochs=config['Global']['epoch_num'],\n",
    "        step_each_epoch=len(train_dataloader),\n",
    "        parameters=model.parameters())\n",
    "\n",
    "    # build metric\n",
    "    eval_class = build_metric(config['Metric'])\n",
    "    # load pretrain model\n",
    "    pre_best_model_dict = load_model(config, model, optimizer)\n",
    "    logger.info('train dataloader has {} iters'.format(len(train_dataloader)))\n",
    "    if valid_dataloader is not None:\n",
    "        logger.info('valid dataloader has {} iters'.format(\n",
    "            len(valid_dataloader)))\n",
    "\n",
    "    use_amp = config[\"Global\"].get(\"use_amp\", False)\n",
    "    if use_amp:\n",
    "        AMP_RELATED_FLAGS_SETTING = {\n",
    "            'FLAGS_cudnn_batchnorm_spatial_persistent': 1,\n",
    "            'FLAGS_max_inplace_grad_add': 8,\n",
    "        }\n",
    "        paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING)\n",
    "        scale_loss = config[\"Global\"].get(\"scale_loss\", 1.0)\n",
    "        use_dynamic_loss_scaling = config[\"Global\"].get(\n",
    "            \"use_dynamic_loss_scaling\", False)\n",
    "        scaler = paddle.amp.GradScaler(\n",
    "            init_loss_scaling=scale_loss,\n",
    "            use_dynamic_loss_scaling=use_dynamic_loss_scaling)\n",
    "    else:\n",
    "        scaler = None\n",
    "\n",
    "    # start train\n",
    "    program.train(config, train_dataloader, valid_dataloader, device, model,\n",
    "                  loss_class, optimizer, lr_scheduler, post_process_class,\n",
    "                  eval_class, pre_best_model_dict, logger, vdl_writer, scaler)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "## 4. 完整训练任务\n",
    "\n",
    "### 4.1 启动训练\n",
    "\n",
    "PaddleOCR 识别任务与检测任务类似，是通过配置文件传输参数的。\n",
    "\n",
    "要进行完整的模型训练，首先需要下载整个项目并安装相关依赖："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple\n",
      "Requirement already satisfied: shapely in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 1)) (1.8.0)\n",
      "Collecting scikit-image==0.17.2\n",
      "  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/d7/ee/753ea56fda5bc2a5516a1becb631bf5ada593a2dd44f21971a13a762d4db/scikit_image-0.17.2-cp37-cp37m-manylinux1_x86_64.whl (12.5 MB)\n",
      "     |████████████████████████████████| 12.5 MB 8.4 MB/s            \n",
      "\u001b[?25hRequirement already satisfied: imgaug==0.4.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 3)) (0.4.0)\n",
      "Requirement already satisfied: pyclipper in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 4)) (1.3.0.post2)\n",
      "Requirement already satisfied: lmdb in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 5)) (1.2.1)\n",
      "Requirement already satisfied: tqdm in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 6)) (4.36.1)\n",
      "Requirement already satisfied: numpy in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 7)) (1.20.3)\n",
      "Requirement already satisfied: visualdl in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 8)) (2.2.0)\n",
      "Requirement already satisfied: python-Levenshtein in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 9)) (0.12.2)\n",
      "Requirement already satisfied: opencv-contrib-python==4.4.0.46 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 10)) (4.4.0.46)\n",
      "Requirement already satisfied: lxml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 11)) (4.7.1)\n",
      "Requirement already satisfied: premailer in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 12)) (3.10.0)\n",
      "Requirement already satisfied: openpyxl in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from -r requirements.txt (line 13)) (3.0.5)\n",
      "Requirement already satisfied: imageio>=2.3.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image==0.17.2->-r requirements.txt (line 2)) (2.6.1)\n",
      "Requirement already satisfied: matplotlib!=3.0.0,>=2.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image==0.17.2->-r requirements.txt (line 2)) (2.2.3)\n",
      "Requirement already satisfied: tifffile>=2019.7.26 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image==0.17.2->-r requirements.txt (line 2)) (2021.11.2)\n",
      "Requirement already satisfied: PyWavelets>=1.1.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image==0.17.2->-r requirements.txt (line 2)) (1.2.0)\n",
      "Requirement already satisfied: pillow!=7.1.0,!=7.1.1,>=4.3.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image==0.17.2->-r requirements.txt (line 2)) (7.1.2)\n",
      "Requirement already satisfied: networkx>=2.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image==0.17.2->-r requirements.txt (line 2)) (2.4)\n",
      "Requirement already satisfied: scipy>=1.0.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-image==0.17.2->-r requirements.txt (line 2)) (1.6.3)\n",
      "Requirement already satisfied: six in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from imgaug==0.4.0->-r requirements.txt (line 3)) (1.15.0)\n",
      "Requirement already satisfied: opencv-python in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from imgaug==0.4.0->-r requirements.txt (line 3)) (4.1.1.26)\n",
      "Requirement already satisfied: flask>=1.1.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (1.1.1)\n",
      "Requirement already satisfied: requests in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (2.22.0)\n",
      "Requirement already satisfied: pre-commit in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (1.21.0)\n",
      "Requirement already satisfied: shellcheck-py in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (0.7.1.1)\n",
      "Requirement already satisfied: pandas in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (1.1.5)\n",
      "Requirement already satisfied: Flask-Babel>=1.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (1.0.0)\n",
      "Requirement already satisfied: bce-python-sdk in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (0.8.53)\n",
      "Requirement already satisfied: flake8>=3.7.9 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (3.8.2)\n",
      "Requirement already satisfied: protobuf>=3.11.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->-r requirements.txt (line 8)) (3.14.0)\n",
      "Requirement already satisfied: setuptools in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from python-Levenshtein->-r requirements.txt (line 9)) (56.2.0)\n",
      "Requirement already satisfied: cssutils in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from premailer->-r requirements.txt (line 12)) (2.3.0)\n",
      "Requirement already satisfied: cachetools in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from premailer->-r requirements.txt (line 12)) (4.0.0)\n",
      "Requirement already satisfied: cssselect in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from premailer->-r requirements.txt (line 12)) (1.1.0)\n",
      "Requirement already satisfied: jdcal in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from openpyxl->-r requirements.txt (line 13)) (1.4.1)\n",
      "Requirement already satisfied: et-xmlfile in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from openpyxl->-r requirements.txt (line 13)) (1.0.1)\n",
      "Requirement already satisfied: pycodestyle<2.7.0,>=2.6.0a1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->-r requirements.txt (line 8)) (2.6.0)\n",
      "Requirement already satisfied: pyflakes<2.3.0,>=2.2.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->-r requirements.txt (line 8)) (2.2.0)\n",
      "Requirement already satisfied: importlib-metadata in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->-r requirements.txt (line 8)) (0.23)\n",
      "Requirement already satisfied: mccabe<0.7.0,>=0.6.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->-r requirements.txt (line 8)) (0.6.1)\n",
      "Requirement already satisfied: click>=5.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->-r requirements.txt (line 8)) (7.0)\n",
      "Requirement already satisfied: Jinja2>=2.10.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->-r requirements.txt (line 8)) (2.11.0)\n",
      "Requirement already satisfied: Werkzeug>=0.15 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->-r requirements.txt (line 8)) (0.16.0)\n",
      "Requirement already satisfied: itsdangerous>=0.24 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->-r requirements.txt (line 8)) (1.1.0)\n",
      "Requirement already satisfied: pytz in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Flask-Babel>=1.0.0->visualdl->-r requirements.txt (line 8)) (2019.3)\n",
      "Requirement already satisfied: Babel>=2.3 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Flask-Babel>=1.0.0->visualdl->-r requirements.txt (line 8)) (2.8.0)\n",
      "Requirement already satisfied: cycler>=0.10 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image==0.17.2->-r requirements.txt (line 2)) (0.10.0)\n",
      "Requirement already satisfied: pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image==0.17.2->-r requirements.txt (line 2)) (2.4.2)\n",
      "Requirement already satisfied: python-dateutil>=2.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image==0.17.2->-r requirements.txt (line 2)) (2.8.0)\n",
      "Requirement already satisfied: kiwisolver>=1.0.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image==0.17.2->-r requirements.txt (line 2)) (1.1.0)\n",
      "Requirement already satisfied: decorator>=4.3.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from networkx>=2.0->scikit-image==0.17.2->-r requirements.txt (line 2)) (4.4.2)\n",
      "Requirement already satisfied: pycryptodome>=3.8.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from bce-python-sdk->visualdl->-r requirements.txt (line 8)) (3.9.9)\n",
      "Requirement already satisfied: future>=0.6.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from bce-python-sdk->visualdl->-r requirements.txt (line 8)) (0.18.0)\n",
      "Requirement already satisfied: cfgv>=2.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->-r requirements.txt (line 8)) (2.0.1)\n",
      "Requirement already satisfied: virtualenv>=15.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->-r requirements.txt (line 8)) (16.7.9)\n",
      "Requirement already satisfied: nodeenv>=0.11.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->-r requirements.txt (line 8)) (1.3.4)\n",
      "Requirement already satisfied: toml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->-r requirements.txt (line 8)) (0.10.0)\n",
      "Requirement already satisfied: pyyaml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->-r requirements.txt (line 8)) (5.1.2)\n",
      "Requirement already satisfied: aspy.yaml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->-r requirements.txt (line 8)) (1.3.0)\n",
      "Requirement already satisfied: identify>=1.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->-r requirements.txt (line 8)) (1.4.10)\n",
      "Requirement already satisfied: idna<2.9,>=2.5 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->visualdl->-r requirements.txt (line 8)) (2.8)\n",
      "Requirement already satisfied: chardet<3.1.0,>=3.0.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->visualdl->-r requirements.txt (line 8)) (3.0.4)\n",
      "Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->visualdl->-r requirements.txt (line 8)) (2019.9.11)\n",
      "Requirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->visualdl->-r requirements.txt (line 8)) (1.25.6)\n",
      "Requirement already satisfied: MarkupSafe>=0.23 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Jinja2>=2.10.1->flask>=1.1.1->visualdl->-r requirements.txt (line 8)) (1.1.1)\n",
      "Requirement already satisfied: zipp>=0.5 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from importlib-metadata->flake8>=3.7.9->visualdl->-r requirements.txt (line 8)) (3.6.0)\n",
      "Installing collected packages: scikit-image\n",
      "  Attempting uninstall: scikit-image\n",
      "    Found existing installation: scikit-image 0.19.1\n",
      "    Uninstalling scikit-image-0.19.1:\n",
      "      Successfully uninstalled scikit-image-0.19.1\n",
      "Successfully installed scikit-image-0.17.2\n"
     ]
    }
   ],
   "source": [
    "# 克隆PaddleOCR代码\n",
    "#!git clone https://gitee.com/paddlepaddle/PaddleOCR\n",
    "# 修改代码运行的默认目录为 /home/aistudio/PaddleOCR\n",
    "import os\n",
    "os.chdir(\"/home/aistudio/PaddleOCR\")\n",
    "# 安装PaddleOCR第三方依赖\n",
    "!pip install -r requirements.txt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "创建软链，将训练数据放在PaddleOCR项目下："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "!ln -s /home/aistudio/work/train_data/ /home/aistudio/PaddleOCR/"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "下载预训练模型：\n",
    "\n",
    "为了加快收敛速度，建议下载训练好的模型在 icdar2015 数据上进行 finetune"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "--2021-12-22 15:39:39--  https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_none_bilstm_ctc_v2.0_train.tar\n",
      "Resolving paddleocr.bj.bcebos.com (paddleocr.bj.bcebos.com)... 182.61.200.195, 182.61.200.229, 2409:8c04:1001:1002:0:ff:b001:368a\n",
      "Connecting to paddleocr.bj.bcebos.com (paddleocr.bj.bcebos.com)|182.61.200.195|:443... connected.\n",
      "HTTP request sent, awaiting response... 200 OK\n",
      "Length: 51200000 (49M) [application/x-tar]\n",
      "Saving to: ‘./pretrain_models/rec_mv3_none_bilstm_ctc_v2.0_train.tar’\n",
      "\n",
      "rec_mv3_none_bilstm 100%[===================>]  48.83M  15.5MB/s    in 3.6s    \n",
      "\n",
      "2021-12-22 15:39:42 (13.7 MB/s) - ‘./pretrain_models/rec_mv3_none_bilstm_ctc_v2.0_train.tar’ saved [51200000/51200000]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "!cd PaddleOCR/\n",
    "# 下载MobileNetV3的预训练模型\n",
    "!wget -nc -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_none_bilstm_ctc_v2.0_train.tar\n",
    "# 解压模型参数\n",
    "!tar -xf pretrain_models/rec_mv3_none_bilstm_ctc_v2.0_train.tar && rm -rf pretrain_models/rec_mv3_none_bilstm_ctc_v2.0_train.tar"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "启动训练命令很简单，指定好配置文件即可。另外在命令行中可以通过 `-o` 修改配置文件中的参数值。启动训练命令如下所示\n",
    "\n",
    "其中：\n",
    "\n",
    "* `Global.pretrained_model`: 加载的预训练模型路径\n",
    "* `Global.character_dict_path` ： 字典路径（这里只支持26个小写字母+数字）\n",
    "* `Global.eval_batch_step` ： 评估频率\n",
    "* `Global.epoch_num`： 总训练轮数\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/skimage/morphology/_skeletonize.py:241: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.\n",
      "Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations\n",
      "  0, 1, 1, 0, 0, 1, 0, 0, 0], dtype=np.bool)\n",
      "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/skimage/morphology/_skeletonize.py:256: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.\n",
      "Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations\n",
      "  0, 0, 0, 0, 0, 0, 0, 0, 0], dtype=np.bool)\n",
      "[2021/12/23 20:28:15] root INFO: Architecture : \n",
      "[2021/12/23 20:28:15] root INFO:     Backbone : \n",
      "[2021/12/23 20:28:15] root INFO:         model_name : large\n",
      "[2021/12/23 20:28:15] root INFO:         name : MobileNetV3\n",
      "[2021/12/23 20:28:15] root INFO:         scale : 0.5\n",
      "[2021/12/23 20:28:15] root INFO:     Head : \n",
      "[2021/12/23 20:28:15] root INFO:         fc_decay : 0\n",
      "[2021/12/23 20:28:15] root INFO:         name : CTCHead\n",
      "[2021/12/23 20:28:15] root INFO:     Neck : \n",
      "[2021/12/23 20:28:15] root INFO:         encoder_type : rnn\n",
      "[2021/12/23 20:28:15] root INFO:         hidden_size : 96\n",
      "[2021/12/23 20:28:15] root INFO:         name : SequenceEncoder\n",
      "[2021/12/23 20:28:15] root INFO:     Transform : None\n",
      "[2021/12/23 20:28:15] root INFO:     algorithm : CRNN\n",
      "[2021/12/23 20:28:15] root INFO:     model_type : rec\n",
      "[2021/12/23 20:28:15] root INFO: Eval : \n",
      "[2021/12/23 20:28:15] root INFO:     dataset : \n",
      "[2021/12/23 20:28:15] root INFO:         data_dir : ./train_data/ic15_data\n",
      "[2021/12/23 20:28:15] root INFO:         label_file_list : ['./train_data/ic15_data/rec_gt_test.txt']\n",
      "[2021/12/23 20:28:15] root INFO:         name : SimpleDataSet\n",
      "[2021/12/23 20:28:15] root INFO:         transforms : \n",
      "[2021/12/23 20:28:15] root INFO:             DecodeImage : \n",
      "[2021/12/23 20:28:15] root INFO:                 channel_first : False\n",
      "[2021/12/23 20:28:15] root INFO:                 img_mode : BGR\n",
      "[2021/12/23 20:28:15] root INFO:             CTCLabelEncode : None\n",
      "[2021/12/23 20:28:15] root INFO:             RecResizeImg : \n",
      "[2021/12/23 20:28:15] root INFO:                 image_shape : [3, 32, 100]\n",
      "[2021/12/23 20:28:15] root INFO:             KeepKeys : \n",
      "[2021/12/23 20:28:15] root INFO:                 keep_keys : ['image', 'label', 'length']\n",
      "[2021/12/23 20:28:15] root INFO:     loader : \n",
      "[2021/12/23 20:28:15] root INFO:         batch_size_per_card : 256\n",
      "[2021/12/23 20:28:15] root INFO:         drop_last : False\n",
      "[2021/12/23 20:28:15] root INFO:         num_workers : 4\n",
      "[2021/12/23 20:28:15] root INFO:         shuffle : False\n",
      "[2021/12/23 20:28:15] root INFO:         use_shared_memory : False\n",
      "[2021/12/23 20:28:15] root INFO: Global : \n",
      "[2021/12/23 20:28:15] root INFO:     cal_metric_during_train : True\n",
      "[2021/12/23 20:28:15] root INFO:     character_dict_path : ppocr/utils/ic15_dict.txt\n",
      "[2021/12/23 20:28:15] root INFO:     character_type : EN\n",
      "[2021/12/23 20:28:15] root INFO:     checkpoints : None\n",
      "[2021/12/23 20:28:15] root INFO:     debug : False\n",
      "[2021/12/23 20:28:15] root INFO:     distributed : False\n",
      "[2021/12/23 20:28:15] root INFO:     epoch_num : 40\n",
      "[2021/12/23 20:28:15] root INFO:     eval_batch_step : [0, 200]\n",
      "[2021/12/23 20:28:15] root INFO:     infer_img : doc/imgs_words_en/word_19.png\n",
      "[2021/12/23 20:28:15] root INFO:     infer_mode : False\n",
      "[2021/12/23 20:28:15] root INFO:     log_smooth_window : 20\n",
      "[2021/12/23 20:28:15] root INFO:     max_text_length : 25\n",
      "[2021/12/23 20:28:15] root INFO:     pretrained_model : rec_mv3_none_bilstm_ctc_v2.0_train/best_accuracy\n",
      "[2021/12/23 20:28:15] root INFO:     print_batch_step : 10\n",
      "[2021/12/23 20:28:15] root INFO:     save_epoch_step : 3\n",
      "[2021/12/23 20:28:15] root INFO:     save_inference_dir : ./\n",
      "[2021/12/23 20:28:15] root INFO:     save_model_dir : ./output/rec/ic15/\n",
      "[2021/12/23 20:28:15] root INFO:     save_res_path : ./output/rec/predicts_ic15.txt\n",
      "[2021/12/23 20:28:15] root INFO:     use_gpu : True\n",
      "[2021/12/23 20:28:15] root INFO:     use_space_char : False\n",
      "[2021/12/23 20:28:15] root INFO:     use_visualdl : False\n",
      "[2021/12/23 20:28:15] root INFO: Loss : \n",
      "[2021/12/23 20:28:15] root INFO:     name : CTCLoss\n",
      "[2021/12/23 20:28:15] root INFO: Metric : \n",
      "[2021/12/23 20:28:15] root INFO:     main_indicator : acc\n",
      "[2021/12/23 20:28:15] root INFO:     name : RecMetric\n",
      "[2021/12/23 20:28:15] root INFO: Optimizer : \n",
      "[2021/12/23 20:28:15] root INFO:     beta1 : 0.9\n",
      "[2021/12/23 20:28:15] root INFO:     beta2 : 0.999\n",
      "[2021/12/23 20:28:15] root INFO:     lr : \n",
      "[2021/12/23 20:28:15] root INFO:         learning_rate : 0.0005\n",
      "[2021/12/23 20:28:15] root INFO:     name : Adam\n",
      "[2021/12/23 20:28:15] root INFO:     regularizer : \n",
      "[2021/12/23 20:28:15] root INFO:         factor : 0\n",
      "[2021/12/23 20:28:15] root INFO:         name : L2\n",
      "[2021/12/23 20:28:15] root INFO: PostProcess : \n",
      "[2021/12/23 20:28:15] root INFO:     name : CTCLabelDecode\n",
      "[2021/12/23 20:28:15] root INFO: Train : \n",
      "[2021/12/23 20:28:15] root INFO:     dataset : \n",
      "[2021/12/23 20:28:15] root INFO:         data_dir : ./train_data/ic15_data/\n",
      "[2021/12/23 20:28:15] root INFO:         label_file_list : ['./train_data/ic15_data/rec_gt_train.txt']\n",
      "[2021/12/23 20:28:15] root INFO:         name : SimpleDataSet\n",
      "[2021/12/23 20:28:15] root INFO:         transforms : \n",
      "[2021/12/23 20:28:15] root INFO:             DecodeImage : \n",
      "[2021/12/23 20:28:15] root INFO:                 channel_first : False\n",
      "[2021/12/23 20:28:15] root INFO:                 img_mode : BGR\n",
      "[2021/12/23 20:28:15] root INFO:             CTCLabelEncode : None\n",
      "[2021/12/23 20:28:15] root INFO:             RecResizeImg : \n",
      "[2021/12/23 20:28:15] root INFO:                 image_shape : [3, 32, 100]\n",
      "[2021/12/23 20:28:15] root INFO:             KeepKeys : \n",
      "[2021/12/23 20:28:15] root INFO:                 keep_keys : ['image', 'label', 'length']\n",
      "[2021/12/23 20:28:15] root INFO:     loader : \n",
      "[2021/12/23 20:28:15] root INFO:         batch_size_per_card : 256\n",
      "[2021/12/23 20:28:15] root INFO:         drop_last : True\n",
      "[2021/12/23 20:28:15] root INFO:         num_workers : 8\n",
      "[2021/12/23 20:28:15] root INFO:         shuffle : True\n",
      "[2021/12/23 20:28:15] root INFO:         use_shared_memory : False\n",
      "[2021/12/23 20:28:15] root INFO: train with paddle 2.1.2 and device CUDAPlace(0)\n",
      "[2021/12/23 20:28:15] root INFO: Initialize indexs of datasets:['./train_data/ic15_data/rec_gt_train.txt']\n",
      "[2021/12/23 20:28:15] root INFO: Initialize indexs of datasets:['./train_data/ic15_data/rec_gt_test.txt']\n",
      "W1223 20:28:15.851713   306 device_context.cc:404] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1\n",
      "W1223 20:28:15.857080   306 device_context.cc:422] device: 0, cuDNN Version: 7.6.\n",
      "[2021/12/23 20:28:19] root INFO: loaded pretrained_model successful from rec_mv3_none_bilstm_ctc_v2.0_train/best_accuracy.pdparams\n",
      "[2021/12/23 20:28:19] root INFO: train dataloader has 17 iters\n",
      "[2021/12/23 20:28:19] root INFO: valid dataloader has 9 iters\n",
      "[2021/12/23 20:28:19] root INFO: During the training process, after the 0th iteration, an evaluation is run every 200 iterations\n",
      "[2021/12/23 20:28:19] root INFO: Initialize indexs of datasets:['./train_data/ic15_data/rec_gt_train.txt']\n",
      "[2021/12/23 20:28:23] root INFO: epoch: [1/40], iter: 10, lr: 0.000500, loss: 9.336592, acc: 0.203125, norm_edit_dis: 0.674909, reader_cost: 0.27284 s, batch_cost: 0.40185 s, samples: 2816, ips: 700.75290\n",
      "[2021/12/23 20:28:24] root INFO: epoch: [1/40], iter: 16, lr: 0.000500, loss: 6.955496, acc: 0.210938, norm_edit_dis: 0.678930, reader_cost: 0.00008 s, batch_cost: 0.05430 s, samples: 1536, ips: 2828.80514\n",
      "[2021/12/23 20:28:24] root INFO: save model in ./output/rec/ic15/latest\n",
      "[2021/12/23 20:28:24] root INFO: Initialize indexs of datasets:['./train_data/ic15_data/rec_gt_train.txt']\n",
      "[2021/12/23 20:28:28] root INFO: epoch: [2/40], iter: 20, lr: 0.000500, loss: 6.402417, acc: 0.246094, norm_edit_dis: 0.695874, reader_cost: 0.24180 s, batch_cost: 0.34361 s, samples: 1024, ips: 298.00945\n",
      "[2021/12/23 20:28:29] root INFO: epoch: [2/40], iter: 30, lr: 0.000500, loss: 4.007382, acc: 0.412109, norm_edit_dis: 0.743064, reader_cost: 0.00013 s, batch_cost: 0.08982 s, samples: 2560, ips: 2849.98954\n",
      "[2021/12/23 20:28:29] root INFO: epoch: [2/40], iter: 33, lr: 0.000500, loss: 3.906031, acc: 0.458984, norm_edit_dis: 0.770415, reader_cost: 0.00004 s, batch_cost: 0.02684 s, samples: 768, ips: 2861.80304\n",
      "^C\n",
      "main proc 306 exit, kill process group 306\n"
     ]
    }
   ],
   "source": [
    "!python3 tools/train.py -c configs/rec/rec_icdar15_train.yml \\\n",
    "   -o Global.pretrained_model=rec_mv3_none_bilstm_ctc_v2.0_train/best_accuracy \\\n",
    "   Global.character_dict_path=ppocr/utils/ic15_dict.txt \\\n",
    "   Global.eval_batch_step=[0,200] \\\n",
    "   Global.epoch_num=40"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "根据配置文件中设置的的 `save_model_dir` 字段，会有以下几种参数被保存下来：\n",
    "\n",
    "```\n",
    "output/rec/ic15\n",
    "├── best_accuracy.pdopt  \n",
    "├── best_accuracy.pdparams  \n",
    "├── best_accuracy.states  \n",
    "├── config.yml  \n",
    "├── iter_epoch_3.pdopt  \n",
    "├── iter_epoch_3.pdparams  \n",
    "├── iter_epoch_3.states  \n",
    "├── latest.pdopt  \n",
    "├── latest.pdparams  \n",
    "├── latest.states  \n",
    "└── train.log\n",
    "```\n",
    "其中 best_accuracy.* 是评估集上的最优模型；iter_epoch_x.* 是以 `save_epoch_step` 为间隔保存下来的模型；latest.* 是最后一个epoch的模型。\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "**总结：**\n",
    "\n",
    "如果需要训练自己的数据需要修改：\n",
    "\n",
    "1. 训练和评估数据路径（必须）\n",
    "2. 字典路径（必须）\n",
    "3. 预训练模型 （可选）\n",
    "4. 学习率、image shape、网络结构（可选）"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "### 4.2 模型评估\n",
    "\n",
    "\n",
    "评估数据集可以通过 `configs/rec/rec_icdar15_train.yml`  修改Eval中的 `label_file_path` 设置。\n",
    "\n",
    "这里默认使用 icdar2015 的评估集，加载刚刚训练好的模型权重："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[2021/12/23 14:27:51] root INFO: Architecture : \n",
      "[2021/12/23 14:27:51] root INFO:     Backbone : \n",
      "[2021/12/23 14:27:51] root INFO:         model_name : large\n",
      "[2021/12/23 14:27:51] root INFO:         name : MobileNetV3\n",
      "[2021/12/23 14:27:51] root INFO:         scale : 0.5\n",
      "[2021/12/23 14:27:51] root INFO:     Head : \n",
      "[2021/12/23 14:27:51] root INFO:         fc_decay : 0\n",
      "[2021/12/23 14:27:51] root INFO:         name : CTCHead\n",
      "[2021/12/23 14:27:51] root INFO:     Neck : \n",
      "[2021/12/23 14:27:51] root INFO:         encoder_type : rnn\n",
      "[2021/12/23 14:27:51] root INFO:         hidden_size : 96\n",
      "[2021/12/23 14:27:51] root INFO:         name : SequenceEncoder\n",
      "[2021/12/23 14:27:51] root INFO:     Transform : None\n",
      "[2021/12/23 14:27:51] root INFO:     algorithm : CRNN\n",
      "[2021/12/23 14:27:51] root INFO:     model_type : rec\n",
      "[2021/12/23 14:27:51] root INFO: Eval : \n",
      "[2021/12/23 14:27:51] root INFO:     dataset : \n",
      "[2021/12/23 14:27:51] root INFO:         data_dir : ./train_data/ic15_data\n",
      "[2021/12/23 14:27:51] root INFO:         label_file_list : ['./train_data/ic15_data/rec_gt_test.txt']\n",
      "[2021/12/23 14:27:51] root INFO:         name : SimpleDataSet\n",
      "[2021/12/23 14:27:51] root INFO:         transforms : \n",
      "[2021/12/23 14:27:51] root INFO:             DecodeImage : \n",
      "[2021/12/23 14:27:51] root INFO:                 channel_first : False\n",
      "[2021/12/23 14:27:51] root INFO:                 img_mode : BGR\n",
      "[2021/12/23 14:27:51] root INFO:             CTCLabelEncode : None\n",
      "[2021/12/23 14:27:51] root INFO:             RecResizeImg : \n",
      "[2021/12/23 14:27:51] root INFO:                 image_shape : [3, 32, 100]\n",
      "[2021/12/23 14:27:51] root INFO:             KeepKeys : \n",
      "[2021/12/23 14:27:51] root INFO:                 keep_keys : ['image', 'label', 'length']\n",
      "[2021/12/23 14:27:51] root INFO:     loader : \n",
      "[2021/12/23 14:27:51] root INFO:         batch_size_per_card : 256\n",
      "[2021/12/23 14:27:51] root INFO:         drop_last : False\n",
      "[2021/12/23 14:27:51] root INFO:         num_workers : 4\n",
      "[2021/12/23 14:27:51] root INFO:         shuffle : False\n",
      "[2021/12/23 14:27:51] root INFO:         use_shared_memory : False\n",
      "[2021/12/23 14:27:51] root INFO: Global : \n",
      "[2021/12/23 14:27:51] root INFO:     cal_metric_during_train : True\n",
      "[2021/12/23 14:27:51] root INFO:     character_dict_path : ppocr/utils/ic15_dict.txt\n",
      "[2021/12/23 14:27:51] root INFO:     character_type : EN\n",
      "[2021/12/23 14:27:51] root INFO:     checkpoints : output/rec/ic15/best_accuracy\n",
      "[2021/12/23 14:27:51] root INFO:     debug : False\n",
      "[2021/12/23 14:27:51] root INFO:     distributed : False\n",
      "[2021/12/23 14:27:51] root INFO:     epoch_num : 72\n",
      "[2021/12/23 14:27:51] root INFO:     eval_batch_step : [0, 2000]\n",
      "[2021/12/23 14:27:51] root INFO:     infer_img : doc/imgs_words_en/word_10.png\n",
      "[2021/12/23 14:27:51] root INFO:     infer_mode : False\n",
      "[2021/12/23 14:27:51] root INFO:     log_smooth_window : 20\n",
      "[2021/12/23 14:27:51] root INFO:     max_text_length : 25\n",
      "[2021/12/23 14:27:51] root INFO:     pretrained_model : None\n",
      "[2021/12/23 14:27:51] root INFO:     print_batch_step : 10\n",
      "[2021/12/23 14:27:51] root INFO:     save_epoch_step : 3\n",
      "[2021/12/23 14:27:51] root INFO:     save_inference_dir : ./\n",
      "[2021/12/23 14:27:51] root INFO:     save_model_dir : ./output/rec/ic15/\n",
      "[2021/12/23 14:27:51] root INFO:     save_res_path : ./output/rec/predicts_ic15.txt\n",
      "[2021/12/23 14:27:51] root INFO:     use_gpu : True\n",
      "[2021/12/23 14:27:51] root INFO:     use_space_char : False\n",
      "[2021/12/23 14:27:51] root INFO:     use_visualdl : False\n",
      "[2021/12/23 14:27:51] root INFO: Loss : \n",
      "[2021/12/23 14:27:51] root INFO:     name : CTCLoss\n",
      "[2021/12/23 14:27:51] root INFO: Metric : \n",
      "[2021/12/23 14:27:51] root INFO:     main_indicator : acc\n",
      "[2021/12/23 14:27:51] root INFO:     name : RecMetric\n",
      "[2021/12/23 14:27:51] root INFO: Optimizer : \n",
      "[2021/12/23 14:27:51] root INFO:     beta1 : 0.9\n",
      "[2021/12/23 14:27:51] root INFO:     beta2 : 0.999\n",
      "[2021/12/23 14:27:51] root INFO:     lr : \n",
      "[2021/12/23 14:27:51] root INFO:         learning_rate : 0.0005\n",
      "[2021/12/23 14:27:51] root INFO:     name : Adam\n",
      "[2021/12/23 14:27:51] root INFO:     regularizer : \n",
      "[2021/12/23 14:27:51] root INFO:         factor : 0\n",
      "[2021/12/23 14:27:51] root INFO:         name : L2\n",
      "[2021/12/23 14:27:51] root INFO: PostProcess : \n",
      "[2021/12/23 14:27:51] root INFO:     name : CTCLabelDecode\n",
      "[2021/12/23 14:27:51] root INFO: Train : \n",
      "[2021/12/23 14:27:51] root INFO:     dataset : \n",
      "[2021/12/23 14:27:51] root INFO:         data_dir : ./train_data/ic15_data/\n",
      "[2021/12/23 14:27:51] root INFO:         label_file_list : ['./train_data/ic15_data/rec_gt_train.txt']\n",
      "[2021/12/23 14:27:51] root INFO:         name : SimpleDataSet\n",
      "[2021/12/23 14:27:51] root INFO:         transforms : \n",
      "[2021/12/23 14:27:51] root INFO:             DecodeImage : \n",
      "[2021/12/23 14:27:51] root INFO:                 channel_first : False\n",
      "[2021/12/23 14:27:51] root INFO:                 img_mode : BGR\n",
      "[2021/12/23 14:27:51] root INFO:             CTCLabelEncode : None\n",
      "[2021/12/23 14:27:51] root INFO:             RecResizeImg : \n",
      "[2021/12/23 14:27:51] root INFO:                 image_shape : [3, 32, 100]\n",
      "[2021/12/23 14:27:51] root INFO:             KeepKeys : \n",
      "[2021/12/23 14:27:51] root INFO:                 keep_keys : ['image', 'label', 'length']\n",
      "[2021/12/23 14:27:51] root INFO:     loader : \n",
      "[2021/12/23 14:27:51] root INFO:         batch_size_per_card : 256\n",
      "[2021/12/23 14:27:51] root INFO:         drop_last : True\n",
      "[2021/12/23 14:27:51] root INFO:         num_workers : 8\n",
      "[2021/12/23 14:27:51] root INFO:         shuffle : True\n",
      "[2021/12/23 14:27:51] root INFO:         use_shared_memory : False\n",
      "[2021/12/23 14:27:51] root INFO: train with paddle 2.1.2 and device CUDAPlace(0)\n",
      "[2021/12/23 14:27:51] root INFO: Initialize indexs of datasets:['./train_data/ic15_data/rec_gt_test.txt']\n",
      "W1223 14:27:51.861889  5192 device_context.cc:404] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1\n",
      "W1223 14:27:51.865501  5192 device_context.cc:422] device: 0, cuDNN Version: 7.6.\n",
      "[2021/12/23 14:27:56] root INFO: resume from output/rec/ic15/best_accuracy\n",
      "[2021/12/23 14:27:56] root INFO: metric in ckpt ***************\n",
      "[2021/12/23 14:27:56] root INFO: acc:0.48531535869041886\n",
      "[2021/12/23 14:27:56] root INFO: norm_edit_dis:0.7895228681338454\n",
      "[2021/12/23 14:27:56] root INFO: fps:3266.1877400927865\n",
      "[2021/12/23 14:27:56] root INFO: best_epoch:24\n",
      "[2021/12/23 14:27:56] root INFO: start_epoch:25\n",
      "eval model:: 100%|████████████████████████████████| 9/9 [00:02<00:00,  3.32it/s]\n",
      "[2021/12/23 14:27:59] root INFO: metric eval ***************\n",
      "[2021/12/23 14:27:59] root INFO: acc:0.48531535869041886\n",
      "[2021/12/23 14:27:59] root INFO: norm_edit_dis:0.7895228681338454\n",
      "[2021/12/23 14:27:59] root INFO: fps:4491.015930181665\n"
     ]
    }
   ],
   "source": [
    "!python tools/eval.py -c configs/rec/rec_icdar15_train.yml -o Global.checkpoints=output/rec/ic15/best_accuracy \\\n",
    "        Global.character_dict_path=ppocr/utils/ic15_dict.txt\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "评估后，可以看到训练模型在验证集上的精度。\n",
    "\n",
    "PaddleOCR支持训练和评估交替进行, 可在 `configs/rec/rec_icdar15_train.yml` 中修改 `eval_batch_step` 设置评估频率，默认每2000个iter评估一次。评估过程中默认将最佳acc模型，保存为 `output/rec/ic15/best_accuracy` 。\n",
    "\n",
    "如果验证集很大，测试将会比较耗时，建议减少评估次数，或训练完再进行评估。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "### 4.3 预测\n",
    "\n",
    "使用 PaddleOCR 训练好的模型，可以通过以下脚本进行快速预测。\n",
    "\n",
    "预测图片：\n",
    "![](https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.3/doc/imgs_words_en/word_19.png)\n",
    "\n",
    "默认预测图片存储在 `infer_img` 里，通过 `-o Global.checkpoints` 加载训练好的参数文件："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[2021/12/23 14:29:19] root INFO: Architecture : \n",
      "[2021/12/23 14:29:19] root INFO:     Backbone : \n",
      "[2021/12/23 14:29:19] root INFO:         model_name : large\n",
      "[2021/12/23 14:29:19] root INFO:         name : MobileNetV3\n",
      "[2021/12/23 14:29:19] root INFO:         scale : 0.5\n",
      "[2021/12/23 14:29:19] root INFO:     Head : \n",
      "[2021/12/23 14:29:19] root INFO:         fc_decay : 0\n",
      "[2021/12/23 14:29:19] root INFO:         name : CTCHead\n",
      "[2021/12/23 14:29:19] root INFO:     Neck : \n",
      "[2021/12/23 14:29:19] root INFO:         encoder_type : rnn\n",
      "[2021/12/23 14:29:19] root INFO:         hidden_size : 96\n",
      "[2021/12/23 14:29:19] root INFO:         name : SequenceEncoder\n",
      "[2021/12/23 14:29:19] root INFO:     Transform : None\n",
      "[2021/12/23 14:29:19] root INFO:     algorithm : CRNN\n",
      "[2021/12/23 14:29:19] root INFO:     model_type : rec\n",
      "[2021/12/23 14:29:19] root INFO: Eval : \n",
      "[2021/12/23 14:29:19] root INFO:     dataset : \n",
      "[2021/12/23 14:29:19] root INFO:         data_dir : ./train_data/ic15_data\n",
      "[2021/12/23 14:29:19] root INFO:         label_file_list : ['./train_data/ic15_data/rec_gt_test.txt']\n",
      "[2021/12/23 14:29:19] root INFO:         name : SimpleDataSet\n",
      "[2021/12/23 14:29:19] root INFO:         transforms : \n",
      "[2021/12/23 14:29:19] root INFO:             DecodeImage : \n",
      "[2021/12/23 14:29:19] root INFO:                 channel_first : False\n",
      "[2021/12/23 14:29:19] root INFO:                 img_mode : BGR\n",
      "[2021/12/23 14:29:19] root INFO:             CTCLabelEncode : None\n",
      "[2021/12/23 14:29:19] root INFO:             RecResizeImg : \n",
      "[2021/12/23 14:29:19] root INFO:                 image_shape : [3, 32, 100]\n",
      "[2021/12/23 14:29:19] root INFO:             KeepKeys : \n",
      "[2021/12/23 14:29:19] root INFO:                 keep_keys : ['image', 'label', 'length']\n",
      "[2021/12/23 14:29:19] root INFO:     loader : \n",
      "[2021/12/23 14:29:19] root INFO:         batch_size_per_card : 256\n",
      "[2021/12/23 14:29:19] root INFO:         drop_last : False\n",
      "[2021/12/23 14:29:19] root INFO:         num_workers : 4\n",
      "[2021/12/23 14:29:19] root INFO:         shuffle : False\n",
      "[2021/12/23 14:29:19] root INFO:         use_shared_memory : False\n",
      "[2021/12/23 14:29:19] root INFO: Global : \n",
      "[2021/12/23 14:29:19] root INFO:     cal_metric_during_train : True\n",
      "[2021/12/23 14:29:19] root INFO:     character_dict_path : ppocr/utils/ic15_dict.txt\n",
      "[2021/12/23 14:29:19] root INFO:     character_type : EN\n",
      "[2021/12/23 14:29:19] root INFO:     checkpoints : output/rec/ic15/best_accuracy\n",
      "[2021/12/23 14:29:19] root INFO:     debug : False\n",
      "[2021/12/23 14:29:19] root INFO:     distributed : False\n",
      "[2021/12/23 14:29:19] root INFO:     epoch_num : 72\n",
      "[2021/12/23 14:29:19] root INFO:     eval_batch_step : [0, 2000]\n",
      "[2021/12/23 14:29:19] root INFO:     infer_img : doc/imgs_words_en/word_19.png\n",
      "[2021/12/23 14:29:19] root INFO:     infer_mode : False\n",
      "[2021/12/23 14:29:19] root INFO:     log_smooth_window : 20\n",
      "[2021/12/23 14:29:19] root INFO:     max_text_length : 25\n",
      "[2021/12/23 14:29:19] root INFO:     pretrained_model : None\n",
      "[2021/12/23 14:29:19] root INFO:     print_batch_step : 10\n",
      "[2021/12/23 14:29:19] root INFO:     save_epoch_step : 3\n",
      "[2021/12/23 14:29:19] root INFO:     save_inference_dir : ./\n",
      "[2021/12/23 14:29:19] root INFO:     save_model_dir : ./output/rec/ic15/\n",
      "[2021/12/23 14:29:19] root INFO:     save_res_path : ./output/rec/predicts_ic15.txt\n",
      "[2021/12/23 14:29:19] root INFO:     use_gpu : True\n",
      "[2021/12/23 14:29:19] root INFO:     use_space_char : False\n",
      "[2021/12/23 14:29:19] root INFO:     use_visualdl : False\n",
      "[2021/12/23 14:29:19] root INFO: Loss : \n",
      "[2021/12/23 14:29:19] root INFO:     name : CTCLoss\n",
      "[2021/12/23 14:29:19] root INFO: Metric : \n",
      "[2021/12/23 14:29:19] root INFO:     main_indicator : acc\n",
      "[2021/12/23 14:29:19] root INFO:     name : RecMetric\n",
      "[2021/12/23 14:29:19] root INFO: Optimizer : \n",
      "[2021/12/23 14:29:19] root INFO:     beta1 : 0.9\n",
      "[2021/12/23 14:29:19] root INFO:     beta2 : 0.999\n",
      "[2021/12/23 14:29:19] root INFO:     lr : \n",
      "[2021/12/23 14:29:19] root INFO:         learning_rate : 0.0005\n",
      "[2021/12/23 14:29:19] root INFO:     name : Adam\n",
      "[2021/12/23 14:29:19] root INFO:     regularizer : \n",
      "[2021/12/23 14:29:19] root INFO:         factor : 0\n",
      "[2021/12/23 14:29:19] root INFO:         name : L2\n",
      "[2021/12/23 14:29:19] root INFO: PostProcess : \n",
      "[2021/12/23 14:29:19] root INFO:     name : CTCLabelDecode\n",
      "[2021/12/23 14:29:19] root INFO: Train : \n",
      "[2021/12/23 14:29:19] root INFO:     dataset : \n",
      "[2021/12/23 14:29:19] root INFO:         data_dir : ./train_data/ic15_data/\n",
      "[2021/12/23 14:29:19] root INFO:         label_file_list : ['./train_data/ic15_data/rec_gt_train.txt']\n",
      "[2021/12/23 14:29:19] root INFO:         name : SimpleDataSet\n",
      "[2021/12/23 14:29:19] root INFO:         transforms : \n",
      "[2021/12/23 14:29:19] root INFO:             DecodeImage : \n",
      "[2021/12/23 14:29:19] root INFO:                 channel_first : False\n",
      "[2021/12/23 14:29:19] root INFO:                 img_mode : BGR\n",
      "[2021/12/23 14:29:19] root INFO:             CTCLabelEncode : None\n",
      "[2021/12/23 14:29:19] root INFO:             RecResizeImg : \n",
      "[2021/12/23 14:29:19] root INFO:                 image_shape : [3, 32, 100]\n",
      "[2021/12/23 14:29:19] root INFO:             KeepKeys : \n",
      "[2021/12/23 14:29:19] root INFO:                 keep_keys : ['image', 'label', 'length']\n",
      "[2021/12/23 14:29:19] root INFO:     loader : \n",
      "[2021/12/23 14:29:19] root INFO:         batch_size_per_card : 256\n",
      "[2021/12/23 14:29:19] root INFO:         drop_last : True\n",
      "[2021/12/23 14:29:19] root INFO:         num_workers : 8\n",
      "[2021/12/23 14:29:19] root INFO:         shuffle : True\n",
      "[2021/12/23 14:29:19] root INFO:         use_shared_memory : False\n",
      "[2021/12/23 14:29:19] root INFO: train with paddle 2.1.2 and device CUDAPlace(0)\n",
      "W1223 14:29:19.803710  5290 device_context.cc:404] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1\n",
      "W1223 14:29:19.807695  5290 device_context.cc:422] device: 0, cuDNN Version: 7.6.\n",
      "[2021/12/23 14:29:25] root INFO: resume from output/rec/ic15/best_accuracy\n",
      "[2021/12/23 14:29:25] root INFO: infer_img: doc/imgs_words_en/word_19.png\n",
      "pred idx: Tensor(shape=[1, 25], dtype=int64, place=CUDAPlace(0), stop_gradient=True,\n",
      "       [[29, 0 , 0 , 0 , 22, 0 , 0 , 0 , 25, 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 33]])\n",
      "[2021/12/23 14:29:25] root INFO: \t result: slow\t0.8795223\n",
      "[2021/12/23 14:29:25] root INFO: success!\n"
     ]
    }
   ],
   "source": [
    "!python tools/infer_rec.py -c configs/rec/rec_icdar15_train.yml -o Global.checkpoints=output/rec/ic15/best_accuracy Global.character_dict_path=ppocr/utils/ic15_dict.txt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "得到输入图像的预测结果：\n",
    "\n",
    "```\n",
    "infer_img: doc/imgs_words_en/word_19.png\n",
    "        result: slow\t0.8795223\n",
    "```\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "### 作业\n",
    "\n",
    "**【题目1】**\n",
    "\n",
    "可视化出 PaddleOCR 中的实现的[数据增强](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.4/ppocr/data/imaug/rec_img_aug.py)结果：noise、jitter, 并用语言解释效果。\n",
    "\n",
    "可选测试图片：\n",
    "\n",
    "![](https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.4/doc/imgs_words/ch/word_1.jpg)\n",
    "\n",
    "![](https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.4/doc/imgs_words/ch/word_2.jpg)\n",
    "\n",
    "![](https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.4/doc/imgs_words/ch/word_3.jpg)\n",
    "\n",
    "\n",
    "**【题目2】**\n",
    "\n",
    "更换 configs/rec/rec_icdar15_train.yml 配置中的 backbone 为 PaddleOCR 中的 [ResNet34_vd](https://github.com/PaddlePaddle/PaddleOCR/blob/6ee301be36eb54d91dc437842f754593dce13967/ppocr/modeling/backbones/rec_resnet_vd.py#L176)，当输入图片shape为（3，32，100）时，Head 层最终输出的特征尺寸是多少？\n",
    "\n",
    "\n",
    "**【题目3】**\n",
    "\n",
    "下载10W中文数据集[rec_data_lesson_demo](https://paddleocr.bj.bcebos.com/dataset/rec_data_lesson_demo.tar)，修改 configs/rec/rec_icdar15_train.yml  配置文件训练一个识别模型，提供训练log。\n",
    "\n",
    "可加载预训练模型： https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_none_bilstm_ctc_v2.0_train.tar \n",
    "\n",
    "\n",
    "## 总结\n",
    "\n",
    "至此，一个基于CRNN的文本识别任务就全部完成了，更多功能和代码可以参考 [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR)。\n",
    "\n",
    "如果对项目任何问题或者疑问，欢迎在评论区留言提出"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "py35-paddle1.2.0"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}