{ "cells": [ { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1000 27 sec\n", "2000 57 sec\n", "3000 85 sec\n", "4000 115 sec\n", "5000 144 sec\n", "6000 176 sec\n", "7000 213 sec\n", "8000 232 sec\n", "9000 263 sec\n", "10000 297 sec\n", "read train data 322 sec\n", "read test data 383 sec\n", "steps per minute: 73\n", "process acceleration data to get steps and speed 386 sec\n", "got projected position 388 sec\n", "prepared train coordinates 401 sec\n", "0 5a0546857ecc773753327266 site 403 sec\n", " 0 (1054, 1111) 406 sec\n", " 840 (1054, 1111) 455 sec\n", "1 5c3c44b80379370013e0fd2b site 467 sec\n", " 0 (98, 187) 468 sec\n", "2 5d27075f03f801723c2e360f site 469 sec\n", " 0 (201, 226) 471 sec\n", "3 5d27096c03f801723c31e5e0 site 474 sec\n", " 0 (3274, 696) 476 sec\n", " 1300 (3274, 696) 523 sec\n", " 2600 (3274, 696) 570 sec\n", "4 5d27097f03f801723c320d97 site 595 sec\n", " 0 (805, 580) 597 sec\n", "5 5d27099f03f801723c32511d site 624 sec\n", " 0 (228, 207) 626 sec\n", "6 5d2709a003f801723c3251bf site 627 sec\n", " 0 (1003, 503) 629 sec\n", "7 5d2709b303f801723c327472 site 641 sec\n", " 0 (2625, 829) 644 sec\n", " 660 (2625, 829) 692 sec\n", " 1320 (2625, 829) 740 sec\n", " 1980 (2625, 829) 789 sec\n", "8 5d2709bb03f801723c32852c site 837 sec\n", " 0 (3104, 1121) 841 sec\n", " 440 (3104, 1121) 889 sec\n", " 880 (3104, 1121) 937 sec\n", " 1320 (3104, 1121) 985 sec\n", " 1760 (3104, 1121) 1033 sec\n", " 2200 (3104, 1121) 1081 sec\n", " 2640 (3104, 1121) 1129 sec\n", " 3080 (3104, 1121) 1177 sec\n", "9 5d2709c303f801723c3299ee site 1179 sec\n", " 0 (1330, 2043) 1183 sec\n", " 480 (1330, 2043) 1229 sec\n", " 960 (1330, 2043) 1275 sec\n", "10 5d2709d403f801723c32bd39 site 1311 sec\n", " 0 (4656, 914) 1314 sec\n", " 900 (4656, 914) 1362 sec\n", " 1800 (4656, 914) 1410 sec\n", " 2700 (4656, 914) 1458 sec\n", " 3600 (4656, 914) 1506 sec\n", " 4500 (4656, 914) 1555 sec\n", "11 5d2709e003f801723c32d896 site 1563 sec\n", " 0 (2143, 641) 1566 sec\n", " 1160 (2143, 641) 1613 sec\n", "12 5da138274db8ce0c98bbd3d2 site 1653 sec\n", " 0 (387, 202) 1655 sec\n", "13 5da1382d4db8ce0c98bbe92e site 1656 sec\n", " 0 (1148, 970) 1659 sec\n", " 1020 (1148, 970) 1707 sec\n", "14 5da138314db8ce0c98bbf3a0 site 1713 sec\n", " 0 (963, 651) 1715 sec\n", "15 5da138364db8ce0c98bc00f1 site 1748 sec\n", " 0 (668, 300) 1750 sec\n", "16 5da1383b4db8ce0c98bc11ab site 1753 sec\n", " 0 (1466, 676) 1756 sec\n", " 920 (1466, 676) 1804 sec\n", "17 5da138754db8ce0c98bca82f site 1833 sec\n", " 0 (1950, 590) 1835 sec\n", "18 5da138764db8ce0c98bcaa46 site 1886 sec\n", " 0 (1729, 868) 1889 sec\n", " 1020 (1729, 868) 1937 sec\n", "19 5da1389e4db8ce0c98bd0547 site 1970 sec\n", " 0 (515, 294) 1971 sec\n", "20 5da138b74db8ce0c98bd4774 site 1977 sec\n", " 0 (1241, 1172) 1981 sec\n", " 420 (1241, 1172) 2028 sec\n", " 840 (1241, 1172) 2075 sec\n", "21 5da958dd46f8266d0737457b site 2120 sec\n", " 0 (3307, 1843) 2125 sec\n", " 280 (3307, 1843) 2169 sec\n", " 560 (3307, 1843) 2213 sec\n", " 840 (3307, 1843) 2258 sec\n", " 1120 (3307, 1843) 2302 sec\n", " 1400 (3307, 1843) 2346 sec\n", " 1680 (3307, 1843) 2391 sec\n", " 1960 (3307, 1843) 2435 sec\n", " 2240 (3307, 1843) 2479 sec\n", " 2520 (3307, 1843) 2522 sec\n", " 2800 (3307, 1843) 2566 sec\n", " 3080 (3307, 1843) 2609 sec\n", "22 5dbc1d84c1eb61796cf7c010 site 2645 sec\n", " 0 (2102, 2418) 2651 sec\n", " 200 (2102, 2418) 2694 sec\n", " 400 (2102, 2418) 2738 sec\n", " 600 (2102, 2418) 2781 sec\n", " 800 (2102, 2418) 2825 sec\n", " 1000 (2102, 2418) 2868 sec\n", " 1200 (2102, 2418) 2912 sec\n", " 1400 (2102, 2418) 2956 sec\n", " 1600 (2102, 2418) 2999 sec\n", " 1800 (2102, 2418) 3043 sec\n", " 2000 (2102, 2418) 3087 sec\n", "23 5dc8cea7659e181adb076a3f site 3109 sec\n", " 0 (1681, 1326) 3113 sec\n", " 400 (1681, 1326) 3159 sec\n", " 800 (1681, 1326) 3205 sec\n", " 1200 (1681, 1326) 3252 sec\n", " 1600 (1681, 1326) 3298 sec\n", "finish main fingerprinting loop (39075, 4) 3307 sec\n", "prepared df_xy_pred 3315 sec\n", "put predictions into df_dr - start 3328 sec\n", "put predictions into df_dr - end 3348 sec\n", "Finished 3422 sec\n" ] } ], "source": [ "import os\n", "import glob\n", "import pandas as pd\n", "import numpy as np\n", "from dataclasses import dataclass\n", "from pathlib import Path\n", "import time\n", "from scipy.signal import find_peaks, savgol_filter\n", "from numba import njit\n", "from scipy.spatial.distance import cdist\n", "import gc\n", "import warnings\n", "warnings.filterwarnings('ignore')\n", "\n", "\n", "\n", "\n", "\n", "######################################################################################\n", "# part 1 - read all data #############################################################\n", "######################################################################################\n", "\n", "# init timer\n", "start_time = time.time()\n", "\n", "# data structure\n", "@dataclass\n", "class ReadData:\n", " acce: np.ndarray\n", " ahrs: np.ndarray\n", " wifi: np.ndarray\n", " waypoint: np.ndarray\n", " SiteID: str\n", " FileName: str\n", " FloorNum: int\n", "\n", "# site decode dictionary\n", "site_di = {'5a0546857ecc773753327266':0,'5c3c44b80379370013e0fd2b':1,'5d27075f03f801723c2e360f':2,'5d27096c03f801723c31e5e0':3,\n", " '5d27097f03f801723c320d97':4,'5d27099f03f801723c32511d':5,'5d2709a003f801723c3251bf':6,'5d2709b303f801723c327472':7,\n", " '5d2709bb03f801723c32852c':8,'5d2709c303f801723c3299ee':9,'5d2709d403f801723c32bd39':10,'5d2709e003f801723c32d896':11,\n", " '5da138274db8ce0c98bbd3d2':12,'5da1382d4db8ce0c98bbe92e':13,'5da138314db8ce0c98bbf3a0':14,'5da138364db8ce0c98bc00f1':15,\n", " '5da1383b4db8ce0c98bc11ab':16,'5da138754db8ce0c98bca82f':17,'5da138764db8ce0c98bcaa46':18,'5da1389e4db8ce0c98bd0547':19,\n", " '5da138b74db8ce0c98bd4774':20,'5da958dd46f8266d0737457b':21,'5dbc1d84c1eb61796cf7c010':22,'5dc8cea7659e181adb076a3f':23}\n", "\n", "# all train sites\n", "test_bldg = list(site_di.keys())\n", "\n", "# floor decode dictionary\n", "fl_di = {'F1':0, 'F2':1, 'F3':2, 'F4':3, 'F5':4, 'F6':5, 'F7':6, 'F8':7, '1F':0, '2F':1, '3F':2,\n", " '4F':3, '5F':4, '6F':5, '7F':6, '8F':7, '9F':8, 'B1':-1, 'B2':-2}\n", "\n", "# BSSID decode dictionary - construct it as data is read\n", "BSSID_di = {}\n", "\n", "# this function reads one data file at a time\n", "def read_data_file(data_filename, call_type):# call_type: 0=train, 1=test\n", " acce = []\n", " ahrs = []\n", " wifi = []\n", " waypoint = []\n", " FloorNum = -99\n", " ts = 0\n", " wifi_c = 0\n", "\n", " with open(data_filename, 'r', encoding='utf-8') as file:\n", " lines = file.readlines()\n", "\n", " # assign vals from filename\n", " data_filename = str(data_filename).split('/')\n", " FileName = data_filename[-1].split('.')[0]\n", "\n", " if call_type == 0: # train data, infer from path\n", " SiteID = data_filename[-3]\n", " FloorNum = fl_di[data_filename[-2]]\n", " \n", " for line_data in lines:\n", " line_data = line_data.strip()\n", " if not line_data or line_data[0] == '#':\n", " # read metadata\n", " if 'startTime' in line_data:\n", " ld2 = line_data[10 + line_data.find('startTime'):]\n", " ld2 = ld2.split('\\t')\n", " ld2 = ld2[0].split(':')\n", " startTime = int(ld2[0])\n", " if 'SiteID' in line_data:\n", " ld2 = line_data.split(':')\n", " ld2 = ld2[1].split('\\t')\n", " SiteID = ld2[0]\n", " if 'FloorName' in line_data:\n", " ld2 = line_data[line_data.find('FloorName'):]\n", " ld2 = ld2.split(':')\n", " if FloorNum == -99 and ld2[1] != '':\n", " FloorNum = fl_di[ld2[1]]\n", " continue\n", "\n", " line_data = line_data.split('\\t')\n", "\n", " if len(line_data) < 5: # correct data error\n", " line_data.append(0)\n", "\n", " if call_type > 0 and line_data[1] == 'TYPE_ACCELEROMETER': # only need this for test data. Get tot acce - that is all i need\n", " a = np.sqrt(float(line_data[2])**2 + float(line_data[3])**2 + float(line_data[4])**2)\n", " acce.append([int(line_data[0])-startTime, a])\n", " continue\n", "\n", " if call_type > 0 and line_data[1] == 'TYPE_ROTATION_VECTOR': # only need this for test data\n", " ahrs.append([int(line_data[0])-startTime, float(line_data[2]), float(line_data[3]), float(line_data[4])])\n", " continue\n", "\n", " if line_data[1] == 'TYPE_WIFI':\n", " sys_ts = int(line_data[0])-startTime\n", " bssid_t = line_data[3]\n", " rssi = line_data[4]\n", "\n", " #skip wifis after 20 per timestamp\n", " if sys_ts == ts:\n", " wifi_c += 1\n", " else:\n", " wifi_c = 0\n", " ts = sys_ts\n", " if wifi_c > 20:\n", " continue\n", "\n", " bssid = (BSSID_di.get(bssid_t) or -1)\n", " if bssid == -1: # add each new bssid to the dictionary\n", " BSSID_di[bssid_t] = 1 + len(BSSID_di)\n", " bssid = BSSID_di[bssid_t]\n", " \n", " wifi_data = [int(sys_ts), bssid, int(rssi)]\n", " wifi.append(wifi_data)\n", " continue\n", "\n", " if line_data[1] == 'TYPE_WAYPOINT':\n", " waypoint.append([int(line_data[0])-startTime, float(line_data[2]), float(line_data[3])])\n", "\n", " acce = np.array(acce, dtype=np.float32)\n", " ahrs = np.array(ahrs, dtype=np.float32)\n", " wifi = np.array(wifi, dtype=np.int32)\n", " waypoint = np.array(waypoint, dtype=np.float32)\n", " return ReadData(acce, ahrs, wifi, waypoint, SiteID, FileName, FloorNum)\n", "\n", "\n", "\n", "# read train data - prepare data objects\n", "misc_tr = pd.DataFrame()\n", "waypoint_tr = np.zeros([75278, 5], dtype=np.float32)\n", "wifi_tr = np.zeros([5385467, 6], dtype=np.int32)\n", "train_waypoints = pd.DataFrame()\n", "misc_tr = pd.DataFrame({'waypoint_s':np.zeros(10877, dtype=np.int32)})\n", "misc_tr['wifi_s'] = 0\n", "misc_tr['ahrs_s'] = 0\n", "misc_tr['Floor'] = 0\n", "misc_tr['Site'] = ''\n", "misc_tr['PathName'] = ''\n", "misc_tr['path'] = 0\n", "wifi_s = i = waypoint_s = 0\n", "\n", "# read train data\n", "data_path = Path('../input/indoor-location-navigation/train')\n", "floorplans = []\n", "\n", "# select buildings in test\n", "for f in sorted(glob.glob(f'{data_path}/*/*')):\n", " if f.split('/')[-2] in test_bldg:\n", " floorplans.append(f)\n", "paths = {fp:glob.glob(f'{fp}/*.txt') for fp in floorplans}\n", "\n", "# loop over all sites\n", "for p in paths:\n", " for f in os.listdir(p):\n", " data = read_data_file(os.path.join(p, f), 0)\n", " \n", " if data.waypoint.shape[0] > 0:\n", " df = pd.DataFrame({'x':data.waypoint[:,1], 'y':data.waypoint[:,2], 'site':data.SiteID, 'floor':data.FloorNum, 'path':i, 'pathName':data.FileName})\n", " train_waypoints = train_waypoints.append(df)\n", " \n", " waypoint_tr[waypoint_s:waypoint_s + data.waypoint.shape[0], 0:3] = data.waypoint\n", " waypoint_tr[waypoint_s:waypoint_s + data.waypoint.shape[0], 3] = i\n", " waypoint_tr[waypoint_s:waypoint_s + data.waypoint.shape[0], 4] = data.FloorNum\n", " waypoint_s += data.waypoint.shape[0]\n", "\n", " if data.wifi.shape[0] > 0:\n", " wifi_tr[wifi_s:wifi_s + data.wifi.shape[0], 0] = data.wifi[:,0]\n", " wifi_tr[wifi_s:wifi_s + data.wifi.shape[0], 2] = data.wifi[:,1]\n", " wifi_tr[wifi_s:wifi_s + data.wifi.shape[0], 3] = data.wifi[:,2]\n", " wifi_tr[wifi_s:wifi_s + data.wifi.shape[0], 4] = i\n", " wifi_tr[wifi_s:wifi_s + data.wifi.shape[0], 5] = data.FloorNum\n", " wifi_s += data.wifi.shape[0]\n", "\n", " misc_tr['wifi_s'].iat[i] = wifi_s\n", " misc_tr['waypoint_s'].iat[i] = waypoint_s\n", " misc_tr['Floor'].iat[i] = data.FloorNum\n", " misc_tr['Site'].iat[i] = data.SiteID\n", " misc_tr['PathName'].iat[i] = data.FileName\n", " misc_tr['path'].iat[i] = i\n", "\n", " if i > 0 and i%1000 == 0:\n", " print(i, int(time.time() - start_time), 'sec')\n", " i += 1\n", "print('read train data', int(time.time() - start_time), 'sec')\n", "\n", "\n", "\n", "# read test data - prepare data objects\n", "misc_te = pd.DataFrame()\n", "ahrs = np.zeros([3819802, 9], dtype=np.float32)\n", "acce = np.zeros([3819802, 2], dtype=np.float32)\n", "wifi_te = np.zeros([790894, 6], dtype=np.int32)\n", "misc_te = pd.DataFrame({'waypoint_s':np.zeros(626, dtype=np.int32)})\n", "misc_te['wifi_s'] = 0\n", "misc_te['ahrs_s'] = 0\n", "misc_te['Floor'] = 0\n", "misc_te['Site'] = ''\n", "misc_te['PathName'] = ''\n", "misc_te['path'] = 0\n", "path_di = {}\n", "wifi_s = i = ahrs_s = 0\n", "\n", "# read test data\n", "data_path = Path('../input/indoor-location-navigation/test')\n", "for f in os.listdir(data_path):\n", " data = read_data_file(os.path.join(data_path, f), 1)\n", " path_di[f[:-4]] = i # need this for encoding final submission\n", "\n", " if data.ahrs.shape[0] > 0:\n", " ahrs[ahrs_s:ahrs_s + data.ahrs.shape[0], 8] = site_di[data.SiteID]\n", " ahrs[ahrs_s:ahrs_s + data.ahrs.shape[0], 0:4] = data.ahrs\n", " ahrs[ahrs_s:ahrs_s + data.ahrs.shape[0], 4] = i\n", " acce[ahrs_s:ahrs_s + data.ahrs.shape[0], :] = data.acce\n", " ahrs_s += data.ahrs.shape[0]\n", "\n", " if data.wifi.shape[0] > 0:\n", " wifi_te[wifi_s:wifi_s + data.wifi.shape[0], 0] = data.wifi[:,0]\n", " wifi_te[wifi_s:wifi_s + data.wifi.shape[0], 2] = data.wifi[:,1]\n", " wifi_te[wifi_s:wifi_s + data.wifi.shape[0], 3] = data.wifi[:,2]\n", " wifi_te[wifi_s:wifi_s + data.wifi.shape[0], 4] = i + 100000 # to separate test from train\n", " wifi_s += data.wifi.shape[0]\n", "\n", " misc_te['wifi_s'].iat[i] = wifi_s\n", " misc_te['ahrs_s'].iat[i] = ahrs_s\n", " misc_te['Site'].iat[i] = data.SiteID\n", " misc_te['PathName'].iat[i] = data.FileName\n", " misc_te['path'].iat[i] = i + 100000 # to make path unique\n", " i += 1\n", "print('read test data', int(time.time() - start_time), 'sec')\n", "\n", "\n", "\n", "# read sample submission\n", "sub = pd.read_csv('../input/indoor-location-navigation/sample_submission.csv')\n", "tss = sub['site_path_timestamp'].str.split('_')\n", "sub['path'] = tss.apply(lambda x: x[1]).map(path_di).astype('int32')\n", "sub['ts'] = tss.apply(lambda x: x[2]).astype('int32')\n", "sub = sub.sort_values(by=['path', 'ts']).reset_index(drop=True)\n", "misc_te['waypoint_s'] = sub.groupby('path').size().reset_index()[0].cumsum()\n", "\n", "\n", "\n", "\n", "\n", "######################################################################################\n", "# part 2 - make relative prediction (dead reckoning) for test paths###################\n", "######################################################################################\n", "\n", "# dead reckoning parameters\n", "ang_lim = 19.5\n", "h = 10.59\n", "min_dist = 22\n", "step_length = 0.6717\n", "v_min = 0.02666536\n", "window = 33\n", "v_max = 1.4798\n", "p1 = 0.116315\n", "p2 = 0.03715\n", "\n", "# process acceleration data to get steps and speed\n", "acce[:,1] = savgol_filter(acce[:,1], 15, 1)\n", "peak_times, _ = find_peaks(acce[:,1], height=h, distance=min_dist)\n", "peak_times = np.round(peak_times, 0).astype(np.int32)\n", "print('steps per minute:', int(.5 + 60 * peak_times.shape[0] * 50 / acce.shape[0]))\n", "\n", "# set speed\n", "v = np.zeros(ahrs.shape[0], dtype=np.float32)\n", "i = v0 = 0\n", "for j in range(peak_times.shape[0] - 1):\n", " v[i:peak_times[j]] = v0\n", " i = peak_times[j]\n", " f = acce[peak_times[j]:peak_times[j+1],1]\n", " f = f.std()\n", " v0 = 50 * (p1 * f + step_length - p2 * np.sqrt(peak_times[j+1] - peak_times[j])) / (peak_times[j+1] - peak_times[j])\n", "v[i:] = v0\n", "v = savgol_filter(v, window, 1) # smooth speed\n", "v = np.minimum(v_max, np.maximum(v_min, v)) # cap/floor\n", "print('process acceleration data to get steps and speed', int(time.time() - start_time), 'sec')\n", "\n", "# process ahrs data\n", "cos = np.sqrt(1 - np.minimum(0.9999999, (ahrs[:,1:4] * ahrs[:,1:4]).sum(axis=1)))\n", "x = 2 * (ahrs[:,1] * ahrs[:,2] - ahrs[:,3] * cos)\n", "y = 1 - 2 * (ahrs[:,1] * ahrs[:,1] + ahrs[:,3] * ahrs[:,3])\n", "norm = np.sqrt(x * x + y * y)\n", "x = x / norm\n", "y = y / norm\n", "\n", "# rotate by an angle\n", "ang = np.arctan2(x,y) * 180 / 3.14159 # degrees\n", "\n", "# this is rotation that places most points into +-10 degrees of cardinal directions\n", "ang_rot_site = [ 15, -33, -5, -33, -25, 6, 11, 3, -17, 1, 11, -2, -39, -1, 0, -44, 8, 1, 2, 0, -14, 5, 40, -27]\n", "for i in range(24): # loop over sites\n", " ang_rot = ang_rot_site[i]\n", " idx2 = (i == ahrs[:,8])\n", " # if close to cardinal direction, assume it is equal to that direction\n", " # north\n", " idx = idx2 & (np.abs(ang-ang_rot) < ang_lim)\n", " ang[idx] = 0 + ang_rot\n", " # south\n", " idx = idx2 & (np.abs(np.abs(ang-ang_rot) - 180) < ang_lim)\n", " ang[idx] = 180 + ang_rot\n", " # east\n", " idx = idx2 & (np.abs(ang-ang_rot - 90) < ang_lim)\n", " ang[idx] = 90 + ang_rot\n", " # west\n", " idx = idx2 & (np.abs(ang-ang_rot + 90) < ang_lim)\n", " ang[idx] = -90 + ang_rot\n", "ang_inc_site = [-2.0, 10.0, -1.5, -7.0, -4.5, -7.5, -2.5, -9.0, -10.0, -3.0, -6.5, -9.5, -7.0, -0.5, -5.0, -6.0, -8.0, 0.5, -5.0, -1.5, -1.0, -10.0, 0.0, -0.5]\n", "for i in range(24): # loop over sites\n", " idx = (i == ahrs[:,8])\n", " ang[idx] = ang[idx] + ang_inc_site[i]\n", " \n", "# restate x/y using rotated coords\n", "x = np.sin(ang / 180 * 3.14159)\n", "y = np.cos(ang / 180 * 3.14159)\n", "\n", "# get projected position\n", "ahrs[:,5] = (v * np.append(0, x[:-1] * (ahrs[1:,0] - ahrs[:-1,0]) / 1000)).cumsum()\n", "ahrs[:,6] = (v * np.append(0, y[:-1] * (ahrs[1:,0] - ahrs[:-1,0]) / 1000)).cumsum()\n", "print('got projected position', int(time.time() - start_time), 'sec')\n", "\n", "\n", "\n", "# indices of waypoints - only use them for finding intersecting paths\n", "path1 = np.array(sub['path'].astype('int64'))\n", "ts1 = np.array(sub['ts'].astype('int64'))\n", "i1 = path1 * 10000000 + ts1\n", "path2 = ahrs[:,4].astype(np.int64)\n", "ts2 = ahrs[:,0].astype(np.int64)\n", "i2 = path2 * 10000000 + ts2\n", "indices = []\n", "m0 = 0\n", "for i in range(sub.shape[0]):\n", " m = m0 + (i2[m0:m0 + 20000] >= i1[i]).argmax()\n", " if np.abs(i1[i] - i2[m]) > 100000: # use last point from correct path\n", " m -= 1\n", " indices.append(m)\n", " m0 = m\n", "\n", "# select waypoints only, and get the closest position to each one.\n", "subl = sub.copy()\n", "subl['x'] = ahrs[indices,5] # projected position for waypoints\n", "subl['y'] = ahrs[indices,6]\n", "\n", "# find intersecting paths\n", "misc_te2 = misc_te.copy()\n", "misc_te2['path'] = misc_te['path'] - 100000 # turn it into normal path, for merging\n", "subl = subl.merge(misc_te2[['path','waypoint_s','ahrs_s']], how='left', on='path')\n", "res = []\n", "for i1 in range(subl.shape[0] - 2):\n", " for j1 in range(i1+2, subl.shape[0]):\n", " if subl['path'].iat[i1] != subl['path'].iat[j1]:\n", " break\n", " dt = subl['ts'].iat[j1] - subl['ts'].iat[i1]\n", " if dt > 3700:\n", " dt = max(1, dt) / 1000\n", " d = np.sqrt((subl['x'].iat[i1] - subl['x'].iat[j1])**2 + (subl['y'].iat[i1] - subl['y'].iat[j1])**2)\n", " if d < 6.54 and d / dt < 0.064:\n", " res.append([i1, j1, subl['path'].iat[i1], subl['waypoint_s'].iat[i1], indices[i1], indices[j1], subl['ahrs_s'].iat[i1]])\n", " break # no tripples - move on to next i\n", "res = pd.DataFrame(res)\n", "res.columns = ['i', 'j', 'path', 'waypoint_s', 'i2', 'j2', 'ahrs_s']\n", "\n", "# correct intersecting paths\n", "for k1 in range(res.shape[0]):\n", " i, j, path, waypoint_s , i2, j2, ahrs_s = res.iloc[k1]\n", " ts = np.array(subl['ts'].iloc[i:j+1])\n", " ts = ts - ts[0]\n", " ts = ts / ts[-1]\n", " mult = np.append(ts, np.ones(waypoint_s - 1 - j))\n", " subl['x'].iloc[i:waypoint_s] += (subl['x'].iloc[i] - subl['x'].iloc[j]) * mult\n", " subl['y'].iloc[i:waypoint_s] += (subl['y'].iloc[i] - subl['y'].iloc[j]) * mult\n", "\n", " ts = np.array(ahrs[i2:j2+1, 0])\n", " ts = ts - ts[0]\n", " ts = ts / ts[-1]\n", " mult = np.append(ts, np.ones(ahrs_s - 1 - j2))\n", " ahrs[i2:ahrs_s, 5] += (ahrs[i2, 5] - ahrs[j2, 5]) * mult\n", " ahrs[i2:ahrs_s, 6] += (ahrs[i2, 6] - ahrs[j2, 6]) * mult\n", "\n", "\n", "\n", "\n", "\n", "######################################################################################\n", "# part 3 - fingerprinting ############################################################\n", "######################################################################################\n", "\n", "# assign coordinates to each train wifi point (interpolate between waypoints)\n", "wifi_s = waypoint_s = 0\n", "wifi_xy = np.zeros([wifi_tr.shape[0], 2], dtype=np.float32)\n", "for i in range(misc_tr.shape[0]):\n", " wifi = misc_tr['wifi_s'].iat[i] - wifi_s\n", " waypoint = misc_tr['waypoint_s'].iat[i] - waypoint_s\n", " waypoints = waypoint_tr[waypoint_s:waypoint_s + waypoint, :]\n", " waypoints_t = waypoints[:,0].astype(np.int32)\n", " # here each t is repeated many times - loop over distinct t values\n", " values, counts = np.unique(wifi_tr[wifi_s:wifi_s+wifi,0], return_counts=True)\n", " j = 0\n", " for c in range(values.shape[0]):\n", " t = values[c]\n", " if t <= waypoints_t[0]:\n", " k1 = 0\n", " k2 = k1\n", " w1 = 1\n", " elif t >= waypoints_t[-1]:\n", " k1 = waypoints_t.shape[0] - 1\n", " k2 = k1\n", " w1 = 1\n", " else:\n", " k2 = ((waypoints_t - t) > 0).argmax()\n", " k1 = k2 - 1\n", " w1 = (waypoints_t[k2] - t)/ (waypoints_t[k2] - waypoints_t[k1])\n", " wifi_xy[wifi_s:wifi_s+counts[c], 0] = waypoint_tr[waypoint_s + k1, 1] * w1 + waypoint_tr[waypoint_s + k2, 1] * (1 - w1)\n", " wifi_xy[wifi_s:wifi_s+counts[c], 1] = waypoint_tr[waypoint_s + k1, 2] * w1 + waypoint_tr[waypoint_s + k2, 2] * (1 - w1)\n", " j += counts[c]\n", " wifi_s +=counts[c]\n", " waypoint_s += waypoint\n", "print('prepared train coordinates', int(time.time() - start_time), 'sec')\n", "\n", "# function for data formatting: construct unique index\n", "@njit\n", "def f2id(ts, path):# ts, path: count unique combos of ts/path\n", " j = 0\n", " index = np.zeros(ts.shape[0], dtype=np.int32)\n", " for i in range(1, ts.shape[0]):\n", " if ts[i] != ts[i-1] or path[i] != path[i-1]:\n", " j = j + 1\n", " index[i] = j\n", " return index\n", "\n", "# put id in wifi data\n", "wifi_tr[:,1] = f2id(wifi_tr[:,0], wifi_tr[:,4])\n", "wifi_te[:,1] = 1000000 + f2id(wifi_te[:,0], wifi_te[:,4]) # make test separable from train by adding 1M\n", "\n", "# only keep bssids that are in train data\n", "bssids = set(wifi_tr[:,2])\n", "rows = [i for i in range(wifi_te.shape[0]) if wifi_te[i,2] in bssids]\n", "wifi_te = wifi_te[rows,:]\n", "\n", "# combine train and test data\n", "wifi_xy = np.append(wifi_xy, np.zeros([wifi_tr.shape[0], wifi_xy.shape[1]], dtype=np.float32)).reshape(-1, wifi_xy.shape[1])\n", "wifi_tr = np.append(wifi_tr, wifi_te).reshape(-1, wifi_tr.shape[1]) \n", "misc_tr = misc_tr.append(misc_te)\n", "\n", "# save\n", "wifi_tr0 = wifi_tr.copy()\n", "wifi_xy0 = wifi_xy.copy()\n", "\n", "# loop over all sites******************************************************************\n", "df1_tot = pd.DataFrame()\n", "site_id = 0\n", "for site in misc_tr['Site'].unique():\n", " if site == '':\n", " break\n", " print(site_id, site, 'site', int(time.time() - start_time), 'sec')\n", " site_id += 1\n", " \n", " # select current site only\n", " paths = set(misc_tr['path'].loc[misc_tr['Site'] == site])\n", " rows = [i for i in range(wifi_tr0.shape[0]) if wifi_tr0[i,4] in paths]\n", " wifi_tr = wifi_tr0[rows,:].copy()\n", " wifi_xy = wifi_xy0[rows,:].copy()\n", "\n", " # only keep bssids that are present in both train and val\n", " bssids = set(wifi_tr[wifi_tr[:,1] >= 1000000,2])\n", " bssids2 = set(wifi_tr[wifi_tr[:,1] < 1000000,2])\n", " bssids = bssids.intersection(bssids2)\n", " rows = [i for i in range(wifi_tr.shape[0]) if wifi_tr[i,2] in bssids]\n", " wifi_tr = wifi_tr[rows,:]\n", " wifi_xy = wifi_xy[rows,:]\n", "\n", " # renumber bssids\n", " bssids = pd.DataFrame({'bssid':wifi_tr[:,2]})\n", " wifi_tr[:,2] = np.array(bssids['bssid'].astype('category').cat.codes)\n", "\n", " # format data\n", " df = pd.DataFrame(wifi_tr[:,[0, 1, 2, 3, 4, 5]])\n", " df.columns = ['ts', 'id', 'bssid','rssi','path','f']\n", " df['x'] = wifi_xy[:,0]\n", " df['y'] = wifi_xy[:,1]\n", " x = pd.pivot_table(df, values='rssi', index='id', columns='bssid', aggfunc=np.sum, fill_value=-10000).reset_index()\n", "\n", " # split into train/valid\n", " x_tr = np.array(x.loc[x['id'] < 1000000], dtype=np.int32)\n", " x_val = np.array(x.loc[x['id'] >= 1000000], dtype=np.int32)\n", "\n", " # process all val points in 1 pass\n", " x_val2 = x_val.reshape(-1)\n", " x_val2[x_val2 == -10000] = 10000\n", " x_val = x_val2.reshape(x_val.shape)\n", "\n", " # process in chunks\n", " x_val0 = x_val.copy() # save\n", " chunk_size = int(5.e9 / 3. / 4. / x_tr.shape[0] / x_tr.shape[1]) # back into 5 Gb total\n", " id1 = x_tr[:,0] # id of tr points\n", " x_tr = x_tr[:,1:] # drop id\n", " x_tr = x_tr.reshape(x_tr.shape[0], 1, x_tr.shape[1])\n", " for i in range(1 + x_val.shape[0]//chunk_size): # loop over chunks\n", " if i%20 == 0:\n", " print(' ', i * chunk_size, x_val0.shape, int(time.time() - start_time), 'sec')\n", " \n", " x_val = x_val0[i*chunk_size:(i+1)*chunk_size,:].copy()\n", " \n", " id0 = x_val[:,0] # id of val points\n", " x_val = x_val[:,1:] # drop id\n", " x_val = x_val.reshape(1, x_val.shape[0], x_val.shape[1])\n", "\n", " # find closest match of rec in x_tr\n", " x1 = np.abs(x_tr - x_val)\n", " x1a = x1 < 200\n", " # count of bssid matches\n", " x2 = x1a.sum(axis=-1)\n", " # diff for matched bssids\n", " x3 = (x1a * x1).sum(axis=-1)\n", " \n", " # turn results into dataframe\n", " df1 = pd.DataFrame({'id0':np.tile(id0, id1.shape[0]), 'cc':x2.ravel(), 'id':np.repeat(id1, id0.shape[0]), 'diff':x3.ravel()})\n", " \n", " # select closest matches for each match count\n", " df1['m'] = 28 * df1['cc'] - df1['diff']\n", " df2 = df1.groupby(['id0'])['m'].max().reset_index()\n", " df2.columns = ['id0','m2']\n", " df1 = df1.merge(df2, how='left', on='id0')\n", " df1 = df1.loc[df1['m'] >= df1['m2']].reset_index(drop=True)\n", " df1.drop(['m2', 'm'], axis=1, inplace=True)\n", "\n", " # append to total\n", " df1_tot = df1_tot.append(df1)\n", "print('finish main fingerprinting loop', df1_tot.shape, int(time.time() - start_time), 'sec')\n", "del x3, x2, x1a, x1, x_val, x_tr, x_val0\n", "gc.collect()\n", "\n", "# bring in coordinates\n", "df = pd.DataFrame(wifi_tr0[:,[0, 1, 2, 3, 4, 5]])\n", "df.columns = ['ts', 'id', 'bssid','rssi', 'path','f']\n", "df['x'] = wifi_xy0[:df.shape[0],0]\n", "df['y'] = wifi_xy0[:df.shape[0],1]\n", "df_xy = df.groupby('id')[['x','y','f']].mean().reset_index()\n", "df1_tot = df1_tot.merge(df_xy, how='left', on='id')\n", "\n", "# weight parameters\n", "cc_di = {} # multiple of cc, tabulated\n", "cc_l = [1,1,1,1,1,1,1,1,1,1,1.2,37,60,60,230,260,260,273,440,440,720,720]\n", "for i in range(22):\n", " cc_di[i] = cc_l[i]\n", "diff_mult = 23.9\n", "\n", "# make predicted floor the same for all points on the same path\n", "def f_pred_path(dft):# this replaces f1 with average floor per path\n", " dft1 = pd.DataFrame(wifi_tr0[:,[1, 4]])\n", " dft1.columns = ['id0', 'path']\n", " dft1 = dft1.loc[dft1['path'] >= 100000] # select test from total\n", " dft2 = dft1.groupby('id0').mean().reset_index()\n", " dft3 = dft[['id0','f1']].merge(dft2, how='left', on='id0')\n", " dft4 = dft3.groupby('path')['f1'].mean().reset_index()\n", " dft4['f1'] = np.round(dft4['f1'], 0).astype('int32') # round to nearest. path, f1 - no dups.\n", " dft.drop('f1', axis=1, inplace=True)\n", " dft5 = dft2.merge(dft4, how='inner', on='path') # id0, path, f1\n", " dft = dft.merge(dft5[['id0','f1']], how='left', on='id0')\n", " return dft\n", " \n", "# bring in relative prediction into df_xy_pred: id, x, y\n", "dft = pd.DataFrame(wifi_tr0[:,[0, 1, 4]])\n", "dft.columns = ['ts', 'id', 'path']\n", "df_xy_pred = dft.groupby('id').mean().reset_index()\n", "dtypes = {'ts':'int32', 'x_p':'float32', 'y_p':'float32', 'path':'int32'}\n", "df_xy_pred = df_xy_pred.loc[df_xy_pred['path'] >= 100000].reset_index(drop=True) # select test from total\n", "df_dr = pd.DataFrame(ahrs[:,[0, 5, 6, 4]]) # relative prediction *********************************\n", "paths = np.array(df_xy_pred['path'], dtype=np.int32) - 100000\n", "tss = np.array(df_xy_pred['ts'], dtype=np.int32)\n", "df_xy_pred.drop(['ts', 'path'], axis=1, inplace=True)\n", "y_te = np.zeros([tss.shape[0], 2])\n", "\n", "# now only select data for wifi points (relative prediction was for sensor timestamps)\n", "path0 = -1\n", "df3a_np = np.array(df_dr)\n", "for i in range(y_te.shape[0]):\n", " path = paths[i]\n", " ts = tss[i]\n", " if path != path0:\n", " d = df3a_np[df3a_np[:,3] == path,:]\n", " offset = (df3a_np[:,3] == path).argmax()\n", " path0 = path\n", " if ts <= d[0,0]:\n", " y_te[i,0] = d[0, 1]\n", " y_te[i,1] = d[0, 2]\n", " elif ts >= d[-1,0]:\n", " y_te[i,0] = d[-1, 1]\n", " y_te[i,1] = d[-1, 2]\n", " else:# interpolate between 2 surrounding points\n", " k2 = ((d[:,0] - ts) > 0).argmax()\n", " k1 = k2 - 1\n", " w1 = (d[k2,0] - ts)/ (d[k2,0] - d[k1,0])\n", " y_te[i,0] = d[k1, 1] * w1 + d[k2, 1] * (1 - w1)\n", " y_te[i,1] = d[k1, 2] * w1 + d[k2, 2] * (1 - w1)\n", "print('prepared df_xy_pred', int(time.time() - start_time), 'sec')\n", "del df3a_np\n", "gc.collect()\n", "df_xy_pred['x'] = y_te[:,0]\n", "df_xy_pred['y'] = y_te[:,1]\n", "df_xy_pred.columns = ['id0', 'x', 'y'] # use id0 here for easier merge\n", " \n", "\n", "\n", "# predict in batches based on DR with offset; use x0/y0 as val DR.\n", "# add adjacent points to form a batch. Add offset to them.\n", "\n", "# bring in pred coordinates - need them for offset\n", "df1_tot = df1_tot.merge(df_xy_pred, how='left', on='id0')\n", "df1_tot.columns = ['id0', 'cc', 'id', 'diff', 'x', 'y', 'f', 'x0', 'y0']\n", "\n", "# bring in path for each id\n", "df_p = df.groupby('id')['path'].mean().reset_index()\n", "paths = np.array(df_p['path'])\n", "def in_1(x):\n", " return x in ids2\n", "\n", "df1_tot0 = df1_tot.copy() # save\n", "max_offset = 90 # only add points withing this distance of current\n", "outlier = 18\n", "for shift in range(1, 43): # only add up to 42 points from before/after (up to 85 total)\n", " # next point on the same path\n", " ids = np.array(df_p['id'].iloc[shift:]) # skip first - it is never next\n", " ids2 = set(ids[paths[shift:] == paths[:-shift]]) # this is the list of ids that can be reduced by 1 and still be on the same path\n", " df1_tot_m = df1_tot0.loc[df1_tot0['id0'].map(in_1)].copy()\n", " df1_tot_m['id0'] -= shift # make it the same as base\n", " # get offset for it\n", " df1_tot_m = df1_tot_m.merge(df_xy_pred, how='left', on='id0')\n", " df1_tot_m.columns = ['id0', 'cc', 'id', 'diff', 'x', 'y', 'f', 'x0', 'y0', 'x0a', 'y0a']\n", " # add offset\n", " df1_tot_m['x'] -= df1_tot_m['x0'] - df1_tot_m['x0a']\n", " df1_tot_m['y'] -= df1_tot_m['y0'] - df1_tot_m['y0a']\n", " # only keep if offset < max_offset\n", " idx = ((df1_tot_m['x0'] - df1_tot_m['x0a'])**2 + (df1_tot_m['y0'] - df1_tot_m['y0a'])**2) < max_offset**2\n", " df1_tot_m = df1_tot_m.loc[idx].reset_index(drop=True)\n", " # append next point\n", " df1_tot_m.drop(['x0a', 'y0a'], axis=1, inplace=True)\n", " df1_tot = df1_tot.append(df1_tot_m).reset_index(drop=True)\n", "\n", " # prev point on the same path\n", " ids = np.array(df_p['id'].iloc[:-shift]) # skip last - it is never previous\n", " ids2 = set(ids[paths[shift:] == paths[:-shift]]) # this is the list of ids that can be increased by 1 and still be on the same path\n", " df1_tot_p = df1_tot0.loc[df1_tot0['id0'].map(in_1)].copy()\n", " df1_tot_p['id0'] += shift # make it the same as base\n", " # get offset for it\n", " df1_tot_p = df1_tot_p.merge(df_xy_pred, how='left', on='id0')\n", " df1_tot_p.columns = ['id0', 'cc', 'id', 'diff', 'x', 'y', 'f', 'x0', 'y0', 'x0a', 'y0a']\n", " # add offset\n", " df1_tot_p['x'] -= df1_tot_p['x0'] - df1_tot_p['x0a']\n", " df1_tot_p['y'] -= df1_tot_p['y0'] - df1_tot_p['y0a']\n", " # only keep if offset < max_offset\n", " idx = ((df1_tot_p['x0'] - df1_tot_p['x0a'])**2 + (df1_tot_p['y0'] - df1_tot_p['y0a'])**2) < max_offset**2\n", " df1_tot_p = df1_tot_p.loc[idx].reset_index(drop=True)\n", " # append prev point\n", " df1_tot_p.drop(['x0a', 'y0a'], axis=1, inplace=True)\n", " df1_tot = df1_tot.append(df1_tot_p).reset_index(drop=True)\n", "\n", "\n", "# calc score - raw\n", "# weight of each point\n", "df1_tot['w'] = (np.exp(- df1_tot['diff']/diff_mult) * df1_tot['cc'].map(cc_di)).astype('float32')\n", "df1_tot['x1'] = (df1_tot['w'] * df1_tot['x']).astype('float32')\n", "df1_tot['y1'] = (df1_tot['w'] * df1_tot['y']).astype('float32')\n", "df1_tot['f1'] = (df1_tot['w'] * df1_tot['f']).astype('float32')\n", "df2 = df1_tot.groupby('id0')[['w', 'x1', 'y1', 'f1']].sum().reset_index()\n", "df1_tot.drop(['x1', 'y1', 'f1'], axis=1, inplace=True)\n", "df2['x1'] = df2['x1'] / df2['w']\n", "df2['y1'] = df2['y1'] / df2['w']\n", "df2['f1'] = df2['f1'] / df2['w']\n", "\n", "# calc score - drop outliers\n", "df1_tot = df1_tot.merge(df2[['id0', 'x1', 'y1', 'f1']], how='left', on='id0') # adds x1, y1\n", "dist = np.sqrt((df1_tot['x'] - df1_tot['x1'])**2 + (df1_tot['y'] - df1_tot['y1'])**2)\n", "df1_tot['x1'] = (df1_tot['w'] * df1_tot['x']).astype('float32')\n", "df1_tot['y1'] = (df1_tot['w'] * df1_tot['y']).astype('float32')\n", "df1_tot['f1'] = (df1_tot['w'] * df1_tot['f']).astype('float32')\n", "df2 = df1_tot.loc[dist < outlier].groupby('id0')[['w', 'x1', 'y1', 'f1']].sum().reset_index() # drop outliers here\n", "df1_tot.drop(['w', 'x1', 'y1', 'f1'], axis=1, inplace=True)\n", "df2['x1'] = df2['x1'] / df2['w']\n", "df2['y1'] = df2['y1'] / df2['w']\n", "df2['f1'] = df2['f1'] / df2['w']\n", "df2 = f_pred_path(df2) # make predicted floor the same for all points on the same path\n", "\n", "\n", " \n", "# put predictions into df_dr\n", "print('put predictions into df_dr - start', int(time.time() - start_time), 'sec')\n", "df_tp = df.groupby('id')[['ts','path']].mean().reset_index()\n", "df2 = df2.merge(df_tp, how='left', left_on='id0', right_on='id')\n", "x_p = np.array(df_dr[1])\n", "y_p = np.array(df_dr[2])\n", "df_dr[3] += 100000\n", "for p in df2['path'].unique():\n", " d = df2.loc[df2['path'] == p].reset_index(drop=True)\n", " o1 = (df_dr[3] == p).argmax()\n", " o2 = (df_dr[3] == p).sum() + o1\n", " # start\n", " n1 = (df_dr[0].iloc[o1:o2] < d['ts'].iat[0]).sum()\n", " x_p[o1:o1+n1] += d['x1'].iat[0] - x_p[o1+n1]\n", " y_p[o1:o1+n1] += d['y1'].iat[0] - y_p[o1+n1]\n", " for i in range(1, d.shape[0]): # i is end of the range\n", " n2 = (df_dr[0].iloc[o1:o2] < d['ts'].iat[i]).sum()\n", " t = np.array(df_dr[0].iloc[o1+n1:o1+n2])\n", " t = (t- t[0])/ (t[-1] - t[0]) # 0 to 1\n", " x_p[o1+n1:o1+n2] += (d['x1'].iat[i-1] - x_p[o1+n1]) + t * ((d['x1'].iat[i] - x_p[o1+n2-1]) - (d['x1'].iat[i-1] - x_p[o1+n1]))\n", " y_p[o1+n1:o1+n2] += (d['y1'].iat[i-1] - y_p[o1+n1]) + t * ((d['y1'].iat[i] - y_p[o1+n2-1]) - (d['y1'].iat[i-1] - y_p[o1+n1]))\n", " n1 = n2\n", " # end\n", " x_p[o1+n1:o2] += d['x1'].iat[i] - x_p[o1+n1]\n", " y_p[o1+n1:o2] += d['y1'].iat[i] - y_p[o1+n1]\n", "df_dr[1] = x_p\n", "df_dr[2] = y_p\n", "df_dr.columns = ['ts','x_p','y_p','path']\n", "df2a = df2.groupby('path')['f1'].mean().reset_index()\n", "df_dr = df_dr.merge(df2a[['path','f1']], how='left', on='path')\n", "print('put predictions into df_dr - end', int(time.time() - start_time), 'sec') \n", "\n", "\n", "# now only select data for waypoints\n", "df3a = df_dr[['ts','path','x_p','y_p','f1']]\n", "df3a.columns = ['ts', 'path', 'x_p', 'y_p', 'f_p']\n", "path0 = -1\n", "df3a_np = np.array(df3a[['ts', 'x_p', 'y_p', 'f_p','path']], dtype=np.float32)\n", "for i in range(sub.shape[0]):\n", " path = sub['path'].iat[i]\n", " ts = sub['ts'].iat[i]\n", "\n", " if path != path0:\n", " d = df3a_np[df3a_np[:,4] - 100000 == path,:]\n", " path0 = path\n", " sub['floor'].iat[i] = d[0,3]\n", "\n", " if ts <= d[0,0]:\n", " sub['x'].iat[i] = d[0, 1]\n", " sub['y'].iat[i] = d[0, 2]\n", " elif ts >= d[-1,0]:\n", " sub['x'].iat[i] = d[-1, 1]\n", " sub['y'].iat[i] = d[-1, 2]\n", " else:# interpolate between 2 surrounding wifi points\n", " k2 = ((d[:,0] - ts) > 0).argmax()\n", " k1 = k2 - 1\n", " w1 = (d[k2,0] - ts)/ (d[k2,0] - d[k1,0])\n", " sub['x'].iat[i] = d[k1, 1] * w1 + d[k2, 1] * (1 - w1)\n", " sub['y'].iat[i] = d[k1, 2] * w1 + d[k2, 2] * (1 - w1)\n", "\n", "\n", "\n", "\n", "\n", "######################################################################################\n", "# part 4 - post-processing ###########################################################\n", "######################################################################################\n", "\n", "# post-processing parameters\n", "threshold = 5 # snap to grid if dist to grid point is < x\n", "step_mult = 0.6 # snap next point on the path if dist to grid is < x * dist to current path point\n", "\n", "# save starting prediction\n", "sub['x1'] = sub['x']\n", "sub['y1'] = sub['y']\n", "\n", "# drop duplicate waypoints\n", "train_waypoints = train_waypoints.sort_values(by=['site','floor','x','y'])\n", "train_waypoints = train_waypoints.drop_duplicates(subset=['site','floor','x','y'], ignore_index=True)\n", "\n", "def add_xy(df): # add x/y\n", " df['xy'] = [(x, y) for x,y in zip(df['x'], df['y'])]\n", " return df\n", "\n", "train_waypoints = add_xy(train_waypoints)\n", "\n", "def closest_point(point, points): # find closest point from a list of points\n", " return points[cdist([point], points).argmin()]\n", "\n", "\n", "# snap to grid\n", "sub.drop('path', axis=1, inplace=True)\n", "sub = pd.concat([sub['site_path_timestamp'].str.split('_', expand=True).rename(columns={0:'site',1:'path',2:'timestamp'}), sub], axis=1).copy()\n", "for N in range(20):# loop until converges\n", " ds = []\n", " sub = add_xy(sub)\n", " for (site, myfloor), d in sub.groupby(['site','floor']):\n", " idx = (train_waypoints['floor'] == myfloor) & (train_waypoints['site'] == site)\n", " true_floor_locs = train_waypoints.loc[idx].reset_index(drop=True)\n", " d['matched_point'] = [closest_point(x, list(true_floor_locs['xy'])) for x in d['xy']]\n", " d['x_'] = d['matched_point'].apply(lambda x: x[0])\n", " d['y_'] = d['matched_point'].apply(lambda x: x[1])\n", " ds.append(d)\n", " sub = pd.concat(ds)\n", " sub['dist'] = np.sqrt( (sub.x-sub.x_)**2 + (sub.y-sub.y_)**2 )\n", "\n", " # Snap to grid if within a threshold.\n", " sub['_x_'] = sub['x']\n", " sub['_y_'] = sub['y']\n", " idx = sub['dist'] < threshold\n", " sub.loc[idx, '_x_'] = sub.loc[idx]['x_']\n", " sub.loc[idx, '_y_'] = sub.loc[idx]['y_']\n", " \n", " # shift each path by mean shift, snap again\n", " dft = sub.groupby('path')[['x','_x_','y','_y_']].mean().reset_index()\n", " dft['dx'] = dft['_x_'] - dft['x']\n", " dft['dy'] = dft['_y_'] - dft['y']\n", " sub = sub.merge(dft[['path','dx','dy']], how='left', on='path')\n", " sub['x'] = sub['x'] + sub['dx']\n", " sub['y'] = sub['y'] + sub['dy']\n", " sub = add_xy(sub)\n", " sub.drop(['dx','dy'], axis=1, inplace=True)\n", "\n", "\n", "# proceed 1 step at a time\n", "for N in range(5):# loop until converges\n", " # pass forward\n", " sub['x2'] = sub['_x_'] # init to best prediction\n", " sub['y2'] = sub['_y_']\n", " sub['t'] = 0\n", " for i in range(0, sub.shape[0]):\n", " if i == 0 or sub['path'].iat[i] != sub['path'].iat[i-1]:# process new path\n", " site = sub['site'].iat[i]\n", " myfloor = sub['floor'].iat[i]\n", " idx = (train_waypoints['floor'] == myfloor) & (train_waypoints['site'] == site)\n", " true_floor_locs = train_waypoints.loc[idx].reset_index(drop=True)\n", " points = list(true_floor_locs['xy'])\n", " x = sub['x2'].iat[i]\n", " y = sub['y2'].iat[i]\n", " d0 = np.sqrt((sub['x1'].iat[i] - sub['x1'].iat[i+1])**2 + (sub['y1'].iat[i] - sub['y1'].iat[i+1])**2)\n", " else: # get 1-step predicted current point: last point + dPDR\n", " x = sub['x2'].iat[i-1] + sub['x1'].iat[i] - sub['x1'].iat[i-1]\n", " y = sub['y2'].iat[i-1] + sub['y1'].iat[i] - sub['y1'].iat[i-1]\n", " d0 = np.sqrt((sub['x1'].iat[i] - sub['x1'].iat[i-1])**2 + (sub['y1'].iat[i] - sub['y1'].iat[i-1])**2)\n", " # find closest grid point to it\n", " dists = cdist([(x,y)], points)\n", " ii = dists.argmin()\n", " p = points[ii]\n", " dist = dists.min()\n", " if dist < d0 * step_mult: # if grid point is close, snap to it\n", " sub['x2'].iat[i] = p[0]\n", " sub['y2'].iat[i] = p[1]\n", " sub['t'].iat[i] = 1\n", " sub['_x_'] = sub['x2'] # put this in final sub\n", " sub['_y_'] = sub['y2']\n", "\n", " # pass backward\n", " sub['x3'] = sub['_x_'] # init to best pred\n", " sub['y3'] = sub['_y_']\n", " sub['t'] = 0\n", " for i in range(sub.shape[0] - 1, 0, -1):\n", " if i == sub.shape[0] - 1 or sub['path'].iat[i] != sub['path'].iat[i+1]:# process new path\n", " site = sub['site'].iat[i]\n", " myfloor = sub['floor'].iat[i]\n", " idx = (train_waypoints['floor'] == myfloor) & (train_waypoints['site'] == site)\n", " true_floor_locs = train_waypoints.loc[idx].reset_index(drop=True)\n", " points = list(true_floor_locs['xy'])\n", " x = sub['x3'].iat[i]\n", " y = sub['y3'].iat[i]\n", " d0 = np.sqrt((sub['x1'].iat[i] - sub['x1'].iat[i-1])**2 + (sub['y1'].iat[i] - sub['y1'].iat[i-1])**2)\n", " else: # get 1-step predicted current point: last point + dPDR\n", " x = sub['x3'].iat[i+1] + sub['x1'].iat[i] - sub['x1'].iat[i+1]\n", " y = sub['y3'].iat[i+1] + sub['y1'].iat[i] - sub['y1'].iat[i+1]\n", " d0 = np.sqrt((sub['x1'].iat[i] - sub['x1'].iat[i+1])**2 + (sub['y1'].iat[i] - sub['y1'].iat[i+1])**2)\n", " # find closest grid point to it\n", " dists = cdist([(x,y)], points)\n", " ii = dists.argmin()\n", " p = points[ii]\n", " dist = dists.min()\n", " if dist < d0 * step_mult: # if grid point is close, snap to it\n", " sub['x3'].iat[i] = p[0]\n", " sub['y3'].iat[i] = p[1]\n", " sub['t'].iat[i] = 1\n", " sub['_x_'] = sub['x3'] # put this in final sub\n", " sub['_y_'] = sub['y3']\n", "# blend forward/backward 50/50\n", "sub['_x_'] = (sub['x3'] + sub['x2']) / 2\n", "sub['_y_'] = (sub['y3'] + sub['y2']) / 2\n", "\n", "\n", "\n", "# save submission\n", "sub.drop(['x','y'], axis=1, inplace=True)\n", "sub = sub.rename(columns={'_x_':'x', '_y_':'y'})\n", "sub[['site_path_timestamp','floor','x','y']].to_csv('submission_ym.csv', index=False)\n", "print('Finished', int(time.time() - start_time), 'sec')" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "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": "3.7.7" } }, "nbformat": 4, "nbformat_minor": 4 }