Predict classes given bounding boxes.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "\n",
    "from fastai.vision import *\n",
    "import stable.predict\n",
    "from tqdm import tqdm_notebook as tqdm"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "path = Path(\"data\")\n",
    "train_fnames = (path/\"train_fnames.txt\").read_text().split(\"\\n\")\n",
    "valid_fnames = (path/\"valid_fnames.txt\").read_text().split(\"\\n\")\n",
    "df_train = pd.read_csv(path/\"train.csv\")\n",
    "unicode_map = {codepoint: char for codepoint, char in pd.read_csv('data/unicode_translation.csv').values}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "bbox_prediction_path = \"\"\n",
    "unlabelled = pd.read_csv(bbox_prediction_path)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>image_id</th>\n",
       "      <th>labels</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>test_00145af3</td>\n",
       "      <td>nan 588 608 78 81 nan 1408 612 91 70 nan 104 6...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>test_001c37e2</td>\n",
       "      <td>nan 1024 756 57 62 nan 1176 772 19 77 nan 1316...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>test_003aa33a</td>\n",
       "      <td>nan 840 748 23 51 nan 1156 756 36 65 nan 520 7...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>test_00665e33</td>\n",
       "      <td>nan 488 548 57 81 nan 1216 548 103 86 nan 636 ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>test_006964dc</td>\n",
       "      <td>nan 1784 580 96 14 nan 464 596 35 89 nan 636 5...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4145</td>\n",
       "      <td>test_ffb1f141</td>\n",
       "      <td>nan 1376 716 17 58 nan 1872 732 24 66 nan 1196...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4146</td>\n",
       "      <td>test_ffe0bb66</td>\n",
       "      <td>nan 520 536 76 67 nan 340 540 103 102 nan 984 ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4147</td>\n",
       "      <td>test_fff039a9</td>\n",
       "      <td>nan 1820 792 42 91 nan 2264 804 48 139 nan 138...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4148</td>\n",
       "      <td>test_fff50dbc</td>\n",
       "      <td>nan 1320 592 37 76 nan 1796 592 71 40 nan 1648...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4149</td>\n",
       "      <td>test_fff8c890</td>\n",
       "      <td>nan 1056 1024 90 99 nan 572 1036 105 79 nan 12...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>4150 rows × 2 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "           image_id                                             labels\n",
       "0     test_00145af3  nan 588 608 78 81 nan 1408 612 91 70 nan 104 6...\n",
       "1     test_001c37e2  nan 1024 756 57 62 nan 1176 772 19 77 nan 1316...\n",
       "2     test_003aa33a  nan 840 748 23 51 nan 1156 756 36 65 nan 520 7...\n",
       "3     test_00665e33  nan 488 548 57 81 nan 1216 548 103 86 nan 636 ...\n",
       "4     test_006964dc  nan 1784 580 96 14 nan 464 596 35 89 nan 636 5...\n",
       "...             ...                                                ...\n",
       "4145  test_ffb1f141  nan 1376 716 17 58 nan 1872 732 24 66 nan 1196...\n",
       "4146  test_ffe0bb66  nan 520 536 76 67 nan 340 540 103 102 nan 984 ...\n",
       "4147  test_fff039a9  nan 1820 792 42 91 nan 2264 804 48 139 nan 138...\n",
       "4148  test_fff50dbc  nan 1320 592 37 76 nan 1796 592 71 40 nan 1648...\n",
       "4149  test_fff8c890  nan 1056 1024 90 99 nan 572 1036 105 79 nan 12...\n",
       "\n",
       "[4150 rows x 2 columns]"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "unlabelled"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "classifier = load_learner(path=path/\"train_images_chars_256\",\n",
    "                          file=\"20190925_TEST\").to_fp32()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "51e162ae4a4a4579a8ac7ee9df2563c6",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=1, bar_style='info', max=1), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "test_01e41977 820 2924 3 6\n",
      "test_021661ed 1416 1208 108 0\n",
      "test_021661ed 308 1912 139 4\n",
      "test_051829d5 248 1392 180 0\n",
      "test_051829d5 732 4796 42 0\n",
      "test_051829d5 376 4812 218 0\n",
      "test_071bf9b6 292 1604 188 1\n",
      "test_0778b88d 1192 2876 78 0\n",
      "test_08e0ba73 2224 3768 211 0\n",
      "test_09f22cd6 1068 4532 175 0\n",
      "test_0a9bab86 840 1348 160 0\n",
      "test_0a9bab86 396 2580 115 0\n",
      "test_0a9bab86 1352 3660 80 0\n",
      "test_0adbe8a5 2268 2508 184 4\n",
      "test_0adbe8a5 352 4724 168 0\n",
      "test_0dc6ea9a 904 1628 117 2\n",
      "test_145a0fc4 1832 776 3 8\n",
      "test_14d8e349 2272 2508 89 0\n",
      "test_1ada762d 1728 1336 0 0\n",
      "test_1ada762d 2528 2896 167 1\n",
      "test_1b2235cf 2228 2812 133 0\n",
      "test_1b612d1f 1184 4284 178 1\n",
      "test_1d9aebbc 2920 1516 150 0\n",
      "test_1e4be7b4 1108 1756 164 0\n",
      "test_1e4be7b4 2656 3504 106 0\n",
      "test_1e552627 692 1940 43 1\n",
      "test_1ed95720 420 2308 153 0\n",
      "test_1ed95720 2680 2492 175 0\n",
      "test_213b4f83 1684 1820 184 0\n",
      "test_213b4f83 2960 2284 149 0\n",
      "test_21cdd6dc 1788 1364 45 4\n",
      "test_228bff59 772 1440 170 0\n",
      "test_228bff59 2668 4844 144 0\n",
      "test_233bc8ae 448 1896 3 13\n",
      "test_24047790 180 1148 4 6\n",
      "test_2580e7d3 3308 2456 0 0\n",
      "test_2580e7d3 260 2640 161 2\n",
      "test_27dc67d6 532 3716 4 143\n",
      "test_2bd58b75 1580 1448 69 0\n",
      "test_2cc9e4c0 2724 3264 197 0\n",
      "test_2d932b85 324 784 2 5\n",
      "test_2eff1ee9 632 2524 107 2\n",
      "test_2f22b9b7 176 2804 2 4\n",
      "test_30acb487 2548 1400 197 0\n",
      "test_30acb487 620 2624 192 0\n",
      "test_3387e39d 1504 2728 86 0\n",
      "test_34126276 736 3352 1 5\n",
      "test_36b8e315 432 2400 101 0\n",
      "test_36b8e315 1584 2460 95 0\n",
      "test_36b8e315 1284 2544 84 0\n",
      "test_39fdc916 792 1416 78 0\n",
      "test_39fdc916 2712 2584 73 1\n",
      "test_39fdc916 1168 3720 212 0\n",
      "test_3b165e69 1472 1516 145 0\n",
      "test_3b165e69 1436 2312 189 2\n",
      "test_3b165e69 696 2356 78 0\n",
      "test_3d6f0a0a 776 3728 127 2\n",
      "test_4184ee59 660 2764 0 21\n",
      "test_4317dba4 556 488 2 3\n",
      "test_460f3567 440 4240 180 0\n",
      "test_47069a0b 1884 1256 177 0\n",
      "test_479558e9 1664 748 86 3\n",
      "test_49996c72 1364 3980 2 2\n",
      "test_4b0b882a 2312 2816 127 0\n",
      "test_511b6834 2344 3320 85 0\n",
      "test_512c1837 684 3360 111 1\n",
      "test_520ee8fb 1144 1372 204 1\n",
      "test_520ee8fb 2716 1884 167 0\n",
      "test_551b9c8f 932 1200 2 3\n",
      "test_552a8ee2 1192 2948 21 0\n",
      "test_552a8ee2 2704 4412 79 1\n",
      "test_55f7f9a3 1552 1516 15 4\n",
      "test_57c5e4db 2340 2632 136 0\n",
      "test_58551df9 1352 1260 4 98\n",
      "test_585fa713 248 2632 4 8\n",
      "test_598c3749 1992 2920 117 3\n",
      "test_5a19ec59 152 2288 1 5\n",
      "test_5a19ec59 148 2376 2 11\n",
      "test_5a19ec59 152 2460 4 6\n",
      "test_5a1b1282 1888 1912 116 0\n",
      "test_5a1b1282 1476 2364 176 2\n",
      "test_5a1b1282 760 3748 198 0\n",
      "test_5a92c70d 912 1208 61 0\n",
      "test_5c986af9 52 940 1 3\n",
      "test_5ca809a5 476 2456 106 2\n",
      "test_5d16fcd1 2320 1760 209 0\n",
      "test_5e2a9a80 584 1692 89 2\n",
      "test_60ae4efe 2096 1172 3 6\n",
      "test_60ae4efe 1600 3240 104 0\n",
      "test_60fa1e93 432 1564 98 0\n",
      "test_63c040a1 132 2612 58 2\n",
      "test_66d35adb 392 844 96 0\n",
      "test_66fff521 736 1456 174 0\n",
      "test_682db5cb 1512 1508 164 0\n",
      "test_682db5cb 856 3420 130 4\n",
      "test_69f12a50 2312 1912 211 0\n",
      "test_6b3ef6bd 356 1204 63 3\n",
      "test_6bb2b057 1592 1780 167 0\n",
      "test_6c0b06b9 108 432 3 4\n",
      "test_6c2e8861 960 2720 100 0\n",
      "test_6c386409 1744 2236 181 0\n",
      "test_6c386409 2944 2620 186 0\n",
      "test_6eaa2571 936 2308 119 2\n",
      "test_6f3997f5 1780 1476 36 4\n",
      "test_72e175f7 2892 2296 48 0\n",
      "test_72e175f7 236 3288 31 0\n",
      "test_72fbe275 2572 1796 176 0\n",
      "test_7315d707 824 2804 2 2\n",
      "test_732c37fa 1808 2228 154 0\n",
      "test_73f1114d 1520 2380 217 0\n",
      "test_75c4a24b 464 804 70 0\n",
      "test_782b38ba 1580 2880 3 7\n",
      "test_7c546bc6 2216 2544 184 0\n",
      "test_8115c9e4 2620 3904 91 3\n",
      "test_8115c9e4 2612 4540 113 1\n",
      "test_8115c9e4 672 4872 99 1\n",
      "test_828e95fd 1928 2652 142 0\n",
      "test_851b8f57 876 3696 107 0\n",
      "test_87c4fec3 220 3664 181 0\n",
      "test_87d03486 500 576 2 4\n",
      "test_88cbdd12 624 2936 190 1\n",
      "test_8aa41681 772 864 67 0\n",
      "test_8aa41681 764 1508 201 0\n",
      "test_8aa41681 2276 2376 224 4\n",
      "test_8b371868 2316 1656 201 0\n",
      "test_8b5af446 1904 2828 4 7\n",
      "test_901750e9 1564 1348 15 0\n",
      "test_901750e9 380 3248 195 0\n",
      "test_90313f9f 2288 2512 223 0\n",
      "test_941f2e91 1360 2236 4 6\n",
      "test_9661fa3b 1940 1712 153 0\n",
      "test_9661fa3b 404 2508 138 3\n",
      "test_96cdc506 248 1512 4 26\n",
      "test_9a0c37d7 600 1012 54 0\n",
      "test_9b89acfb 432 1392 157 0\n",
      "test_9b89acfb 2296 4316 148 0\n",
      "test_a1d2e6c9 1800 356 0 0\n",
      "test_a272d114 1628 1684 141 1\n",
      "test_a2ee43d0 1532 916 172 0\n",
      "test_a2ee43d0 1180 2792 39 0\n",
      "test_a4c67f42 2320 2552 184 0\n",
      "test_a61cb908 392 892 184 3\n",
      "test_a61cb908 2308 2420 208 4\n",
      "test_a61cb908 784 3456 209 0\n",
      "test_a7d173f5 2276 1820 1 5\n",
      "test_aba0e48e 744 1724 184 2\n",
      "test_aba0e48e 368 3544 176 0\n",
      "test_adb2a677 348 1760 114 0\n",
      "test_ae7bd27f 2020 620 3 17\n",
      "test_b01658e0 332 2876 209 0\n",
      "test_b02004e5 732 1252 204 0\n",
      "test_b02004e5 1824 2620 184 0\n",
      "test_b030be20 276 3632 162 0\n",
      "test_b1129cf3 312 3276 216 0\n",
      "test_b1129cf3 684 3520 220 3\n",
      "test_b242e447 1672 3756 92 0\n",
      "test_b768c298 1460 1232 193 2\n",
      "test_b925f680 1148 2820 184 1\n",
      "test_bbd86475 1808 3668 130 0\n",
      "test_bc27916f 1948 1724 181 0\n",
      "test_bffd598c 1584 1880 74 0\n",
      "test_c071c07d 1088 2552 144 0\n",
      "test_c071c07d 1820 3392 187 4\n",
      "test_c0fe3e51 648 1680 2 19\n",
      "test_c25d1939 1204 2424 97 3\n",
      "test_c62bc520 1368 2624 4 27\n",
      "test_c844018f 1784 1420 4 23\n",
      "test_cbd5a1e5 1452 876 164 0\n",
      "test_cbd5a1e5 2964 2748 171 0\n",
      "test_cbd5a1e5 612 4796 174 0\n",
      "test_ccab2e2f 2172 1504 90 0\n",
      "test_cce86c49 2648 2724 183 0\n",
      "test_cce86c49 1164 3460 187 1\n",
      "test_cce86c49 2288 3776 188 4\n",
      "test_cce86c49 1600 4452 132 2\n",
      "test_cd58b520 1280 3732 4 7\n",
      "test_cdd9c1de 2496 2764 1 2\n",
      "test_d15e7da1 720 1700 198 0\n",
      "test_d32caea3 1376 2404 4 107\n",
      "test_d3e0b10a 220 3484 201 0\n",
      "test_d3e0b10a 284 3884 216 0\n",
      "test_d3e0b10a 632 4428 204 0\n",
      "test_d4c84d71 1596 1260 112 0\n",
      "test_d592f2f7 576 824 160 0\n",
      "test_d592f2f7 1716 1336 221 3\n",
      "test_d5fa8f03 316 2896 233 1\n",
      "test_d5fa8f03 1776 3332 177 0\n",
      "test_d60a7247 248 1656 49 0\n",
      "test_d92081d0 1928 4720 146 0\n",
      "test_d92081d0 1192 4736 52 0\n",
      "test_da77ad1b 684 2192 116 4\n",
      "test_dc6e5d08 1956 3584 168 1\n",
      "test_dc6e5d08 1280 4780 123 1\n",
      "test_dec7d28b 216 800 3 15\n",
      "test_df27bdc1 1332 2408 71 0\n",
      "test_df4138ed 1320 1180 97 3\n",
      "test_e02992ef 1224 896 195 3\n",
      "test_e02992ef 848 1608 129 1\n",
      "test_e02992ef 2368 2348 213 0\n",
      "test_e02992ef 2720 2384 171 0\n",
      "test_e1f70b60 2648 876 186 0\n",
      "test_e1f70b60 1500 908 199 3\n",
      "test_e1f70b60 2628 1568 184 0\n",
      "test_e1f70b60 2280 1648 230 0\n",
      "test_e1f70b60 1904 1900 217 4\n",
      "test_e43ac7e8 428 3200 212 0\n",
      "test_e79c3214 1892 1880 2 9\n",
      "test_e7cbff78 1424 1604 140 0\n",
      "test_e7cbff78 668 2248 197 0\n",
      "test_e7cbff78 2216 2336 176 1\n",
      "test_e7cbff78 316 3572 206 0\n",
      "test_e7cbff78 304 3888 166 0\n",
      "test_e7cbff78 1828 3916 18 0\n",
      "test_e7cbff78 1452 4864 115 0\n",
      "test_e7f873f4 1600 3740 159 2\n",
      "test_e9449d0a 2072 3204 4 14\n",
      "test_e9f440d2 2620 1836 191 0\n",
      "test_ecdd2c00 1964 2268 104 2\n",
      "test_edb581b9 372 1516 88 0\n",
      "test_ee9e37d9 2416 3324 203 0\n",
      "test_f28ae506 444 2404 100 1\n",
      "test_f28ae506 1596 2464 85 0\n",
      "test_f28ae506 1296 2552 93 0\n",
      "test_f3886e0b 1184 4776 202 0\n",
      "test_f5bdedd7 2264 4044 3 3\n",
      "test_f61b6b3b 904 1780 74 0\n",
      "test_f61b6b3b 408 1908 93 3\n",
      "test_f6a2943f 292 2356 180 4\n",
      "test_f754da8e 2272 3540 168 0\n",
      "test_f754da8e 1460 3816 114 4\n",
      "test_f7e9d2e1 1152 1352 4 17\n",
      "test_f9a9f9c8 788 3272 151 0\n",
      "test_f9a9f9c8 1936 3904 225 0\n",
      "test_fa0f4dd8 1780 1360 109 0\n",
      "test_fa0f4dd8 2956 2628 58 0\n",
      "test_fa0f4dd8 2932 4792 93 1\n",
      "test_fbc2f322 588 800 3 4\n",
      "test_fc5d020a 1636 584 2 39\n",
      "test_fccb432a 372 1880 140 2\n",
      "test_fccb432a 1932 4460 150 0\n",
      "test_fccb432a 1228 4524 110 0\n",
      "test_fda27377 1260 1360 99 0\n",
      "test_ffb1f141 1580 1872 106 4\n"
     ]
    }
   ],
   "source": [
    "for i, row in tqdm(unlabelled.iterrows()):\n",
    "    fname = row.image_id\n",
    "    im = PIL.Image.open(path/\"test_images\"/f\"{fname}.jpg\")\n",
    "    im_arr = np.array(im)\n",
    "    im_height, im_width, colour_depth = im_arr.shape \n",
    "    \n",
    "    R = 128\n",
    "    \n",
    "    predictions = []    \n",
    "    if type(row.labels) == str:\n",
    "        lbls = np.array(row.labels.split()).reshape(-1,5)\n",
    "        for _, center_x, center_y, width, height in lbls:\n",
    "            center_x, center_y = int(center_x), int(center_y)\n",
    "            width, height = int(width), int(height)\n",
    "            if width < 5 or height < 5:\n",
    "                # Ignore from prediction\n",
    "                print(fname, center_x, center_y, width, height)\n",
    "            else:\n",
    "                x0, y0 = max(0, center_x - R), max(0, center_y - R)\n",
    "                x1, y1 = center_x + R, center_y + R\n",
    "                character_region = im_arr[y0:y1, x0:x1]\n",
    "                prediction = classifier.predict(stable.predict.preprocess_input(character_region))[0].obj\n",
    "                if prediction != \"BACKGROUND\":\n",
    "                    predictions.append([prediction, center_x, center_y])\n",
    "            \n",
    "    if len(predictions) > 0:\n",
    "        predictions_str = \" \".join([\" \".join(map(str, i)) for i in predictions])\n",
    "        unlabelled.iloc[i].labels = predictions_str\n",
    "    else:\n",
    "        unlabelled.iloc[i].labels = np.nan"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "submission_no = 20"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "unlabelled.to_csv(f\"submissions/submission{submission_no}.csv\", index=False)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "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": 2
}