cos711-assignment-2-u21497402.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "First five values    Length (major axis)  Width (minor axis)  Thickness (depth)     Area  \\\n",
      "0                  NaN          227.940628         127.759132  22619.0   \n",
      "1                  NaN          234.188126         128.199509  23038.0   \n",
      "2                  NaN          229.418610         125.796547  22386.5   \n",
      "3                  NaN          232.763153         125.918808  22578.5   \n",
      "4                  NaN          230.150742         107.253448  19068.0   \n",
      "\n",
      "    Perimeter  Roundness  Solidity  Compactness  Aspect Ratio  Eccentricity  \\\n",
      "0  643.813269        NaN  0.973384     1.458265           NaN           NaN   \n",
      "1  680.984841        NaN  0.957304     1.601844           NaN           NaN   \n",
      "2  646.943212        NaN  0.967270     1.487772           NaN           NaN   \n",
      "3  661.227483        NaN  0.965512     1.540979           NaN           NaN   \n",
      "4  624.842706        NaN  0.951450     1.629395           NaN           NaN   \n",
      "\n",
      "     Extent  Convex hull(convex area)   Type  \n",
      "0  0.681193                   23237.5  MAMRA  \n",
      "1  0.656353                   24065.5  MAMRA  \n",
      "2  0.683620                   23144.0  MAMRA  \n",
      "3  0.685360                   23385.0  MAMRA  \n",
      "4  0.714800                   20041.0  MAMRA  \n"
     ]
    }
   ],
   "source": [
    "import pandas as pd\n",
    "import numpy as np\n",
    "almonds_data = pd.read_csv('Almond.csv')\n",
    "almonds_data.drop(columns=[\"Unnamed: 0\"],inplace=True)\n",
    "print(\"First five values\", almonds_data.head())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<class 'pandas.core.frame.DataFrame'>\n",
      "RangeIndex: 2803 entries, 0 to 2802\n",
      "Data columns (total 13 columns):\n",
      " #   Column                    Non-Null Count  Dtype  \n",
      "---  ------                    --------------  -----  \n",
      " 0   Length (major axis)       1946 non-null   float64\n",
      " 1   Width (minor axis)        1861 non-null   float64\n",
      " 2   Thickness (depth)         1799 non-null   float64\n",
      " 3   Area                      2803 non-null   float64\n",
      " 4   Perimeter                 2803 non-null   float64\n",
      " 5   Roundness                 1946 non-null   float64\n",
      " 6   Solidity                  2803 non-null   float64\n",
      " 7   Compactness               2803 non-null   float64\n",
      " 8   Aspect Ratio              1004 non-null   float64\n",
      " 9   Eccentricity              1004 non-null   float64\n",
      " 10  Extent                    2803 non-null   float64\n",
      " 11  Convex hull(convex area)  2803 non-null   float64\n",
      " 12  Type                      2803 non-null   object \n",
      "dtypes: float64(12), object(1)\n",
      "memory usage: 284.8+ KB\n",
      "Columns with their Non-null counts and their types None\n"
     ]
    }
   ],
   "source": [
    "print(\"Columns with their Non-null counts and their types\", almonds_data.info())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Stats for numeric attributes        Length (major axis)  Width (minor axis)  Thickness (depth)  \\\n",
      "count          1946.000000         1861.000000        1799.000000   \n",
      "mean            290.609274          171.025915         109.705378   \n",
      "std              62.719433           29.916529          18.940597   \n",
      "min             151.335266           88.050529          59.494278   \n",
      "25%             245.966293          149.453659          97.091682   \n",
      "50%             279.879883          170.168365         110.280136   \n",
      "75%             330.508575          190.640427         121.392773   \n",
      "max             515.352478          258.569794         181.845200   \n",
      "\n",
      "               Area    Perimeter    Roundness     Solidity  Compactness  \\\n",
      "count   2803.000000  2803.000000  1946.000000  2803.000000  2803.000000   \n",
      "mean   26511.117374   743.863770     0.470466     0.955828     1.825233   \n",
      "std    13782.561344   230.632076     0.118673     0.039596     0.794058   \n",
      "min     6037.000000   311.563489     0.173748     0.718772     1.164469   \n",
      "25%    16211.500000   571.730009     0.384810     0.944579     1.357398   \n",
      "50%    23440.500000   707.487369     0.472718     0.970422     1.576412   \n",
      "75%    33451.000000   878.896530     0.577553     0.981484     1.965953   \n",
      "max    89282.000000  1864.947387     0.697293     0.992889     9.660057   \n",
      "\n",
      "       Aspect Ratio  Eccentricity       Extent  Convex hull(convex area)  \n",
      "count   1004.000000   1004.000000  2803.000000               2803.000000  \n",
      "mean       1.753216      0.813114     0.724587              27696.218159  \n",
      "std        0.206616      0.041312     0.047474              14237.347610  \n",
      "min        1.400082      0.699897     0.454538               6355.000000  \n",
      "25%        1.612490      0.784476     0.701673              17088.500000  \n",
      "50%        1.705716      0.810120     0.733720              24589.000000  \n",
      "75%        1.833339      0.838141     0.757551              34863.250000  \n",
      "max        2.731251      0.930563     0.845813              90642.500000  \n"
     ]
    }
   ],
   "source": [
    "print(\"Stats for numeric attributes\", almonds_data.describe())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Length (major axis)          69.425615\n",
      "Width (minor axis)           66.393150\n",
      "Thickness (depth)            64.181234\n",
      "Area                        100.000000\n",
      "Perimeter                   100.000000\n",
      "Roundness                    69.425615\n",
      "Solidity                    100.000000\n",
      "Compactness                 100.000000\n",
      "Aspect Ratio                 35.818766\n",
      "Eccentricity                 35.818766\n",
      "Extent                      100.000000\n",
      "Convex hull(convex area)    100.000000\n",
      "Type                        100.000000\n",
      "dtype: float64\n",
      "      Length (major axis)  Width (minor axis)  Thickness (depth)     Area  \\\n",
      "0                     NaN          227.940628         127.759132  22619.0   \n",
      "1                     NaN          234.188126         128.199509  23038.0   \n",
      "2                     NaN          229.418610         125.796547  22386.5   \n",
      "3                     NaN          232.763153         125.918808  22578.5   \n",
      "4                     NaN          230.150742         107.253448  19068.0   \n",
      "...                   ...                 ...                ...      ...   \n",
      "2798                  NaN          192.709366         122.356506  18471.5   \n",
      "2799                  NaN          186.254745         118.708961  17213.5   \n",
      "2800                  NaN          186.196182         119.147224  17510.5   \n",
      "2801                  NaN          188.660828         120.634438  17941.0   \n",
      "2802           269.356903          176.023636                NaN  36683.5   \n",
      "\n",
      "       Perimeter  Roundness  Solidity  Compactness    Extent  \\\n",
      "0     643.813269        NaN  0.973384     1.458265  0.681193   \n",
      "1     680.984841        NaN  0.957304     1.601844  0.656353   \n",
      "2     646.943212        NaN  0.967270     1.487772  0.683620   \n",
      "3     661.227483        NaN  0.965512     1.540979  0.685360   \n",
      "4     624.842706        NaN  0.951450     1.629395  0.714800   \n",
      "...          ...        ...       ...          ...       ...   \n",
      "2798  653.345233        NaN  0.931000     1.838965  0.725739   \n",
      "2799  581.688379        NaN  0.952706     1.564234  0.714016   \n",
      "2800  608.315795        NaN  0.948821     1.681705  0.718999   \n",
      "2801  630.759446        NaN  0.944810     1.764701  0.738191   \n",
      "2802  887.310743   0.643761  0.947380     1.707933  0.722429   \n",
      "\n",
      "      Convex hull(convex area)    Type  \n",
      "0                      23237.5   MAMRA  \n",
      "1                      24065.5   MAMRA  \n",
      "2                      23144.0   MAMRA  \n",
      "3                      23385.0   MAMRA  \n",
      "4                      20041.0   MAMRA  \n",
      "...                        ...     ...  \n",
      "2798                   19840.5  SANORA  \n",
      "2799                   18068.0  SANORA  \n",
      "2800                   18455.0  SANORA  \n",
      "2801                   18989.0  SANORA  \n",
      "2802                   38721.0  SANORA  \n",
      "\n",
      "[2803 rows x 11 columns]\n"
     ]
    }
   ],
   "source": [
    "percent_not_null = almonds_data.notnull().mean() * 100\n",
    "\n",
    "print(percent_not_null)\n",
    "\n",
    "almonds_data.drop(columns=['Aspect Ratio', 'Eccentricity'], inplace=True)\n",
    "print(almonds_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "      Length (major axis)  Width (minor axis)  Thickness (depth)     Area  \\\n",
      "0                     NaN          227.940628         127.759132  22619.0   \n",
      "1                     NaN          234.188126         128.199509  23038.0   \n",
      "2                     NaN          229.418610         125.796547  22386.5   \n",
      "3                     NaN          232.763153         125.918808  22578.5   \n",
      "4                     NaN          230.150742         107.253448  19068.0   \n",
      "...                   ...                 ...                ...      ...   \n",
      "2798                  NaN          192.709366         122.356506  18471.5   \n",
      "2799                  NaN          186.254745         118.708961  17213.5   \n",
      "2800                  NaN          186.196182         119.147224  17510.5   \n",
      "2801                  NaN          188.660828         120.634438  17941.0   \n",
      "2802           269.356903          176.023636                NaN  36683.5   \n",
      "\n",
      "       Perimeter  Roundness  Solidity  Compactness    Extent  \\\n",
      "0     643.813269        NaN  0.973384     1.458265  0.681193   \n",
      "1     680.984841        NaN  0.957304     1.601844  0.656353   \n",
      "2     646.943212        NaN  0.967270     1.487772  0.683620   \n",
      "3     661.227483        NaN  0.965512     1.540979  0.685360   \n",
      "4     624.842706        NaN  0.951450     1.629395  0.714800   \n",
      "...          ...        ...       ...          ...       ...   \n",
      "2798  653.345233        NaN  0.931000     1.838965  0.725739   \n",
      "2799  581.688379        NaN  0.952706     1.564234  0.714016   \n",
      "2800  608.315795        NaN  0.948821     1.681705  0.718999   \n",
      "2801  630.759446        NaN  0.944810     1.764701  0.738191   \n",
      "2802  887.310743   0.643761  0.947380     1.707933  0.722429   \n",
      "\n",
      "      Convex hull(convex area)  \n",
      "0                      23237.5  \n",
      "1                      24065.5  \n",
      "2                      23144.0  \n",
      "3                      23385.0  \n",
      "4                      20041.0  \n",
      "...                        ...  \n",
      "2798                   19840.5  \n",
      "2799                   18068.0  \n",
      "2800                   18455.0  \n",
      "2801                   18989.0  \n",
      "2802                   38721.0  \n",
      "\n",
      "[2803 rows x 10 columns]\n"
     ]
    }
   ],
   "source": [
    "from sklearn.preprocessing import LabelEncoder\n",
    "\n",
    "one_hot_encode = pd.get_dummies(almonds_data[\"Type\"])\n",
    "#almonds_data[\"Type\"] = [LabelEncoder().fit_transform(almonds_data[\"Type\"])] #check if maybe one hot encoding to prevent form assuming ordianol relationship\n",
    "almonds_data = almonds_data.drop(\"Type\", axis=1)\n",
    "print(almonds_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Index(['Length (major axis)', 'Width (minor axis)', 'Thickness (depth)',\n",
      "       'Area', 'Perimeter', 'Roundness', 'Solidity', 'Compactness', 'Extent',\n",
      "       'Convex hull(convex area)'],\n",
      "      dtype='object')\n",
      "      Length (major axis)  Width (minor axis)  Thickness (depth)     Area  \\\n",
      "1481           256.791199                 NaN          70.634346  13893.0   \n",
      "247            285.089111          182.272736                NaN  38908.5   \n",
      "2025                  NaN          154.184448         102.262245  12135.0   \n",
      "2171           230.673859          147.944733                NaN  26349.0   \n",
      "753            239.546631                 NaN         104.702888  17409.5   \n",
      "...                   ...                 ...                ...      ...   \n",
      "1032           419.353516                 NaN         114.451721  32626.5   \n",
      "2121           343.750488          190.821793                NaN  50067.5   \n",
      "1424           257.446320          133.669800                NaN  25247.5   \n",
      "1725           262.819336                 NaN         119.488258  24278.5   \n",
      "2254           241.192123          140.644928                NaN  25475.0   \n",
      "\n",
      "        Perimeter  Roundness  Solidity  Compactness    Extent  \\\n",
      "1481   589.705625   0.268254  0.970317     1.991887  0.761385   \n",
      "247   1069.903669   0.609528  0.914225     2.341181  0.657238   \n",
      "2025   441.220343        NaN  0.976071     1.276619  0.785284   \n",
      "2171   669.528998   0.630489  0.974770     1.353832  0.743986   \n",
      "753    944.340181   0.386293  0.800014     4.076250  0.590874   \n",
      "...           ...        ...       ...          ...       ...   \n",
      "1032  1169.394003   0.236222  0.823319     3.335350  0.586818   \n",
      "2121  1048.933108   0.539485  0.951845     1.748758  0.733515   \n",
      "1424   660.315797   0.485015  0.988470     1.374280  0.782723   \n",
      "1725   713.830515   0.447525  0.965982     1.670162  0.761272   \n",
      "2254   825.837654   0.557568  0.923191     2.130420  0.718375   \n",
      "\n",
      "      Convex hull(convex area)  \n",
      "1481                   14318.0  \n",
      "247                    42559.0  \n",
      "2025                   12432.5  \n",
      "2171                   27031.0  \n",
      "753                    21761.5  \n",
      "...                        ...  \n",
      "1032                   39628.0  \n",
      "2121                   52600.5  \n",
      "1424                   25542.0  \n",
      "1725                   25133.5  \n",
      "2254                   27594.5  \n",
      "\n",
      "[1962 rows x 10 columns]\n",
      "      Length (major axis)  Width (minor axis)  Thickness (depth)     Area  \\\n",
      "962            221.438278                 NaN          95.086189  16561.5   \n",
      "170            346.594452          177.486313                NaN  48539.0   \n",
      "2603                  NaN          219.762085         126.875633  21603.0   \n",
      "12             372.898560                 NaN         100.815842  28773.5   \n",
      "360            209.845016          127.058708                NaN  20869.0   \n",
      "...                   ...                 ...                ...      ...   \n",
      "2203           232.145523                 NaN         101.575600  17658.0   \n",
      "578            302.212494          194.422806                NaN  45657.5   \n",
      "602            287.414246          172.944992                NaN  38595.0   \n",
      "2044           314.061310          170.138077                NaN  41364.5   \n",
      "1897           300.154755          180.785843                NaN  41661.0   \n",
      "\n",
      "        Perimeter  Roundness  Solidity  Compactness    Extent  \\\n",
      "962    569.345233   0.430036  0.982295     1.557549  0.707756   \n",
      "170   1272.690466   0.514467  0.944954     2.655491  0.721598   \n",
      "2603   603.587874        NaN  0.977998     1.342014  0.762279   \n",
      "12     876.173661   0.263464  0.958510     2.123136  0.777116   \n",
      "360    581.386863   0.603413  0.983366     1.288899  0.742246   \n",
      "...           ...        ...       ...          ...       ...   \n",
      "2203   727.043717   0.417187  0.914448     2.382153  0.686414   \n",
      "578    860.364568   0.636499  0.984942     1.290158  0.743607   \n",
      "602    848.381811   0.594873  0.971200     1.484027  0.748662   \n",
      "2044   825.854899   0.533960  0.988482     1.312109  0.775778   \n",
      "1897   812.724958   0.588775  0.988657     1.261675  0.760362   \n",
      "\n",
      "      Convex hull(convex area)  \n",
      "962                    16860.0  \n",
      "170                    51366.5  \n",
      "2603                   22089.0  \n",
      "12                     30019.0  \n",
      "360                    21222.0  \n",
      "...                        ...  \n",
      "2203                   19310.0  \n",
      "578                    46355.5  \n",
      "602                    39739.5  \n",
      "2044                   41846.5  \n",
      "1897                   42139.0  \n",
      "\n",
      "[841 rows x 10 columns]\n",
      "      MAMRA  REGULAR  SANORA\n",
      "1481    1.0      0.0     0.0\n",
      "247     1.0      0.0     0.0\n",
      "2025    0.0      1.0     0.0\n",
      "2171    0.0      1.0     0.0\n",
      "753     0.0      0.0     1.0\n",
      "...     ...      ...     ...\n",
      "1032    1.0      0.0     0.0\n",
      "2121    0.0      1.0     0.0\n",
      "1424    1.0      0.0     0.0\n",
      "1725    0.0      1.0     0.0\n",
      "2254    0.0      1.0     0.0\n",
      "\n",
      "[1962 rows x 3 columns]\n",
      "      MAMRA  REGULAR  SANORA\n",
      "962     0.0      0.0     1.0\n",
      "170     1.0      0.0     0.0\n",
      "2603    0.0      0.0     1.0\n",
      "12      1.0      0.0     0.0\n",
      "360     0.0      0.0     1.0\n",
      "...     ...      ...     ...\n",
      "2203    0.0      1.0     0.0\n",
      "578     0.0      0.0     1.0\n",
      "602     0.0      0.0     1.0\n",
      "2044    0.0      1.0     0.0\n",
      "1897    0.0      1.0     0.0\n",
      "\n",
      "[841 rows x 3 columns]\n"
     ]
    }
   ],
   "source": [
    "from sklearn.model_selection import train_test_split\n",
    "import torch\n",
    "\n",
    "num_features = almonds_data.select_dtypes(include=['int64', 'float64']).columns\n",
    "print(num_features)\n",
    "\n",
    "X = almonds_data[num_features]\n",
    "# y = almonds_data[\"MAMRA\",  \"REGULAR\", \"SANORA\"]\n",
    "X_train, X_test, y_train, y_test = train_test_split(X, one_hot_encode.astype(\"float64\"), test_size=0.3, random_state=5)\n",
    "\n",
    "\n",
    "print(X_train)\n",
    "print(X_test)\n",
    "print(y_train)\n",
    "print(y_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "      Length (major axis)  Width (minor axis)  Thickness (depth)     Area  \\\n",
      "1481           256.791199          171.269798          70.634346  13893.0   \n",
      "247            285.089111          182.272736         109.939933  38908.5   \n",
      "2025           292.690803          154.184448         102.262245  12135.0   \n",
      "2171           230.673859          147.944733         109.939933  26349.0   \n",
      "753            239.546631          171.269798         104.702888  17409.5   \n",
      "...                   ...                 ...                ...      ...   \n",
      "1032           419.353516          171.269798         114.451721  32626.5   \n",
      "2121           343.750488          190.821793         109.939933  50067.5   \n",
      "1424           257.446320          133.669800         109.939933  25247.5   \n",
      "1725           262.819336          171.269798         119.488258  24278.5   \n",
      "2254           241.192123          140.644928         109.939933  25475.0   \n",
      "\n",
      "        Perimeter  Roundness  Solidity  Compactness    Extent  \\\n",
      "1481   589.705625   0.268254  0.970317     1.991887  0.761385   \n",
      "247   1069.903669   0.609528  0.914225     2.341181  0.657238   \n",
      "2025   441.220343   0.469517  0.976071     1.276619  0.785284   \n",
      "2171   669.528998   0.630489  0.974770     1.353832  0.743986   \n",
      "753    944.340181   0.386293  0.800014     4.076250  0.590874   \n",
      "...           ...        ...       ...          ...       ...   \n",
      "1032  1169.394003   0.236222  0.823319     3.335350  0.586818   \n",
      "2121  1048.933108   0.539485  0.951845     1.748758  0.733515   \n",
      "1424   660.315797   0.485015  0.988470     1.374280  0.782723   \n",
      "1725   713.830515   0.447525  0.965982     1.670162  0.761272   \n",
      "2254   825.837654   0.557568  0.923191     2.130420  0.718375   \n",
      "\n",
      "      Convex hull(convex area)  \n",
      "1481                   14318.0  \n",
      "247                    42559.0  \n",
      "2025                   12432.5  \n",
      "2171                   27031.0  \n",
      "753                    21761.5  \n",
      "...                        ...  \n",
      "1032                   39628.0  \n",
      "2121                   52600.5  \n",
      "1424                   25542.0  \n",
      "1725                   25133.5  \n",
      "2254                   27594.5  \n",
      "\n",
      "[1962 rows x 10 columns]\n",
      "      Length (major axis)  Width (minor axis)  Thickness (depth)     Area  \\\n",
      "962            221.438278          170.441575          95.086189  16561.5   \n",
      "170            346.594452          177.486313         109.161401  48539.0   \n",
      "2603           285.848481          219.762085         126.875633  21603.0   \n",
      "12             372.898560          170.441575         100.815842  28773.5   \n",
      "360            209.845016          127.058708         109.161401  20869.0   \n",
      "...                   ...                 ...                ...      ...   \n",
      "2203           232.145523          170.441575         101.575600  17658.0   \n",
      "578            302.212494          194.422806         109.161401  45657.5   \n",
      "602            287.414246          172.944992         109.161401  38595.0   \n",
      "2044           314.061310          170.138077         109.161401  41364.5   \n",
      "1897           300.154755          180.785843         109.161401  41661.0   \n",
      "\n",
      "        Perimeter  Roundness  Solidity  Compactness    Extent  \\\n",
      "962    569.345233   0.430036  0.982295     1.557549  0.707756   \n",
      "170   1272.690466   0.514467  0.944954     2.655491  0.721598   \n",
      "2603   603.587874   0.472637  0.977998     1.342014  0.762279   \n",
      "12     876.173661   0.263464  0.958510     2.123136  0.777116   \n",
      "360    581.386863   0.603413  0.983366     1.288899  0.742246   \n",
      "...           ...        ...       ...          ...       ...   \n",
      "2203   727.043717   0.417187  0.914448     2.382153  0.686414   \n",
      "578    860.364568   0.636499  0.984942     1.290158  0.743607   \n",
      "602    848.381811   0.594873  0.971200     1.484027  0.748662   \n",
      "2044   825.854899   0.533960  0.988482     1.312109  0.775778   \n",
      "1897   812.724958   0.588775  0.988657     1.261675  0.760362   \n",
      "\n",
      "      Convex hull(convex area)  \n",
      "962                    16860.0  \n",
      "170                    51366.5  \n",
      "2603                   22089.0  \n",
      "12                     30019.0  \n",
      "360                    21222.0  \n",
      "...                        ...  \n",
      "2203                   19310.0  \n",
      "578                    46355.5  \n",
      "602                    39739.5  \n",
      "2044                   41846.5  \n",
      "1897                   42139.0  \n",
      "\n",
      "[841 rows x 10 columns]\n"
     ]
    }
   ],
   "source": [
    "from sklearn.impute import SimpleImputer\n",
    "\n",
    "\n",
    "\n",
    "for feature in [\"Length (major axis)\", \"Width (minor axis)\", \"Thickness (depth)\", \"Roundness\"]:\n",
    "    X_train[feature] = X_train[feature].fillna(X_train[feature].mean())\n",
    "    X_test[feature] = X_test[feature].fillna(X_test[feature].mean())\n",
    "\n",
    "print(X_train)\n",
    "print(X_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "      Length (major axis)  Width (minor axis)  Thickness (depth)      Area  \\\n",
      "1481            -0.958643            0.000000          -4.019193 -0.550244   \n",
      "247             -0.202991            0.569608           0.000000  0.867256   \n",
      "2025             0.000000           -0.884487          -0.785082 -0.649861   \n",
      "2171            -1.656067           -1.207509           0.000000  0.155573   \n",
      "753             -1.419133            0.000000          -0.535514 -0.350982   \n",
      "...                   ...                 ...                ...       ...   \n",
      "1032             3.382332            0.000000           0.461353  0.511287   \n",
      "2121             1.363470            1.012182           0.000000  1.499579   \n",
      "1424            -0.941149           -1.946504           0.000000  0.093157   \n",
      "1725            -0.797671            0.000000           0.976364  0.038249   \n",
      "2254            -1.375193           -1.585411           0.000000  0.106048   \n",
      "\n",
      "      Perimeter  Roundness  Solidity  Compactness    Extent  \\\n",
      "1481  -0.397968  -1.710060 -0.007192     0.664824  0.494451   \n",
      "247    1.150416   1.189624 -1.532872     1.227779 -1.384445   \n",
      "2025  -0.876755   0.000000  0.149306    -0.487968  0.925607   \n",
      "2171  -0.140580   1.367726  0.113917    -0.363524  0.180548   \n",
      "753    0.745540  -0.707126 -4.639364     4.024177 -2.581712   \n",
      "...         ...        ...       ...          ...       ...   \n",
      "1032   1.471219  -1.982224 -4.005463     2.830074 -2.654876   \n",
      "2121   1.082797   0.594497 -0.509636     0.272976 -0.008361   \n",
      "1424  -0.170288   0.131684  0.486558    -0.330569  0.879390   \n",
      "1725   0.002269  -0.186855 -0.125114     0.146302  0.492412   \n",
      "2254   0.363432   0.748144 -1.288994     0.888098 -0.281497   \n",
      "\n",
      "      Convex hull(convex area)  \n",
      "1481                 -0.573465  \n",
      "247                   0.977951  \n",
      "2025                 -0.677045  \n",
      "2171                  0.124922  \n",
      "753                  -0.164557  \n",
      "...                        ...  \n",
      "1032                  0.816936  \n",
      "2121                  1.529579  \n",
      "1424                  0.043124  \n",
      "1725                  0.020683  \n",
      "2254                  0.155878  \n",
      "\n",
      "[1962 rows x 10 columns]\n",
      "      Length (major axis)  Width (minor axis)  Thickness (depth)      Area  \\\n",
      "962             -1.742519            0.000000          -1.393269 -0.405585   \n",
      "170              1.643389            0.337252           0.000000  1.583313   \n",
      "2603             0.000000            2.361118           1.753486 -0.092020   \n",
      "12               2.355007            0.000000          -0.826105  0.353962   \n",
      "360             -2.056157           -2.076866           0.000000 -0.137673   \n",
      "...                   ...                 ...                ...       ...   \n",
      "2203            -1.452851            0.000000          -0.750899 -0.337386   \n",
      "578              0.442703            1.148052           0.000000  1.404093   \n",
      "602              0.042359            0.119846           0.000000  0.964828   \n",
      "2044             0.763255           -0.014529           0.000000  1.137082   \n",
      "1897             0.387034            0.495211           0.000000  1.155523   \n",
      "\n",
      "      Perimeter  Roundness  Solidity  Compactness    Extent  \\\n",
      "962   -0.438186  -0.390575  0.334055    -0.016681 -0.431224   \n",
      "170    2.003779   0.383512 -0.679027     1.865136 -0.190043   \n",
      "2603  -0.319298   0.000000  0.217467    -0.386097  0.518817   \n",
      "12     0.627101  -1.917737 -0.311266     0.952706  0.777339   \n",
      "360   -0.396378   1.198983  0.363110    -0.477134  0.169750   \n",
      "...         ...        ...       ...          ...       ...   \n",
      "2203   0.109332  -0.508371 -1.506670     1.396648 -0.803107   \n",
      "578    0.572213   1.502321  0.405872    -0.474975  0.193466   \n",
      "602    0.530609   1.120693  0.033030    -0.142694  0.281531   \n",
      "2044   0.452398   0.562229  0.501894    -0.437353  0.754029   \n",
      "1897   0.406811   1.064781  0.506638    -0.523794  0.485408   \n",
      "\n",
      "      Convex hull(convex area)  \n",
      "962                  -0.444652  \n",
      "170                   1.602361  \n",
      "2603                 -0.134455  \n",
      "12                    0.335973  \n",
      "360                  -0.185887  \n",
      "...                        ...  \n",
      "2203                 -0.299312  \n",
      "578                   1.305096  \n",
      "602                   0.912618  \n",
      "2044                  1.037610  \n",
      "1897                  1.054962  \n",
      "\n",
      "[841 rows x 10 columns]\n"
     ]
    }
   ],
   "source": [
    "from sklearn.preprocessing import RobustScaler #check diff scalers\n",
    "\n",
    "\n",
    "\n",
    "\n",
    "X_train[num_features] = RobustScaler().fit_transform(X_train[num_features])\n",
    "print(X_train)\n",
    "X_test[num_features] = RobustScaler().fit_transform(X_test[num_features])\n",
    "print(X_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[-0.9586,  0.0000, -4.0192,  ...,  0.6648,  0.4945, -0.5735],\n",
      "        [-0.2030,  0.5696,  0.0000,  ...,  1.2278, -1.3844,  0.9780],\n",
      "        [ 0.0000, -0.8845, -0.7851,  ..., -0.4880,  0.9256, -0.6770],\n",
      "        ...,\n",
      "        [-0.9411, -1.9465,  0.0000,  ..., -0.3306,  0.8794,  0.0431],\n",
      "        [-0.7977,  0.0000,  0.9764,  ...,  0.1463,  0.4924,  0.0207],\n",
      "        [-1.3752, -1.5854,  0.0000,  ...,  0.8881, -0.2815,  0.1559]])\n",
      "tensor([[-1.7425,  0.0000, -1.3933,  ..., -0.0167, -0.4312, -0.4447],\n",
      "        [ 1.6434,  0.3373,  0.0000,  ...,  1.8651, -0.1900,  1.6024],\n",
      "        [ 0.0000,  2.3611,  1.7535,  ..., -0.3861,  0.5188, -0.1345],\n",
      "        ...,\n",
      "        [ 0.0424,  0.1198,  0.0000,  ..., -0.1427,  0.2815,  0.9126],\n",
      "        [ 0.7633, -0.0145,  0.0000,  ..., -0.4374,  0.7540,  1.0376],\n",
      "        [ 0.3870,  0.4952,  0.0000,  ..., -0.5238,  0.4854,  1.0550]])\n",
      "tensor([[1., 0., 0.],\n",
      "        [1., 0., 0.],\n",
      "        [0., 1., 0.],\n",
      "        ...,\n",
      "        [1., 0., 0.],\n",
      "        [0., 1., 0.],\n",
      "        [0., 1., 0.]], dtype=torch.float64)\n",
      "tensor([[0., 0., 1.],\n",
      "        [1., 0., 0.],\n",
      "        [0., 0., 1.],\n",
      "        ...,\n",
      "        [0., 0., 1.],\n",
      "        [0., 1., 0.],\n",
      "        [0., 1., 0.]], dtype=torch.float64)\n"
     ]
    }
   ],
   "source": [
    "X_train_torch = torch.from_numpy(X_train.to_numpy()).float()\n",
    "y_train_torch = torch.tensor(y_train.to_numpy(), dtype=torch.float64) #torch.from_numpy(y_train.to_numpy()).float().view(-1, 1)  \n",
    "X_test_torch = torch.from_numpy(X_test.to_numpy()).float()\n",
    "y_test_torch = torch.tensor(y_test.to_numpy(), dtype=torch.float64) #torch.from_numpy(y_test.to_numpy()).float().view(-1, 1)\n",
    "\n",
    "print(X_train_torch)\n",
    "print(X_test_torch)\n",
    "print(y_train_torch)\n",
    "print(y_test_torch)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "from torch import nn\n",
    "import torch\n",
    "\n",
    "\n",
    "class RpropModel(nn.Module):\n",
    "    def __init__(self, activation_func=nn.ReLU()):\n",
    "        super(RpropModel, self).__init__()\n",
    "        self.activation_func = activation_func\n",
    "#         self.fc1 = nn.Linear(10, 30)\n",
    "#         self.fc2 = nn.Linear(30, 3)\n",
    "        self.fc1 = nn.Linear(10, 35)\n",
    "        self.fc2 = nn.Linear(35, 45)    \n",
    "        self.fc3 = nn.Linear(45, 3)\n",
    "\n",
    "    def forward(self, x):\n",
    "#         x = torch.relu(self.fc1(x))\n",
    "#         x = torch.softmax(self.fc2(x), dim=1)\n",
    "\n",
    "        x = self.activation_func(self.fc1(x))  #torch.relu  \n",
    "        x = self.activation_func(self.fc2(x)) #nn.ReLU()(self.fc2(x))\n",
    "        x = torch.softmax(self.fc3(x), dim=1)\n",
    "        return x\n",
    "\n",
    "model = RpropModel()\n",
    "loss_function = nn.CrossEntropyLoss()\n",
    "optimizer = torch.optim.Rprop(model.parameters())\n",
    "\n",
    "def train(model, optimizer, loss_function):\n",
    "    train_losses = []\n",
    "    test_losses = []\n",
    "    for epoch in range(64):\n",
    "        optimizer.zero_grad()\n",
    "        y_pred = model(X_train_torch)\n",
    "        loss = loss_function(y_pred, y_train_torch)\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "        train_losses.append(loss.item())\n",
    "\n",
    "        model.eval()\n",
    "        with torch.no_grad():\n",
    "            y_pred_test = model(X_test_torch)\n",
    "            test_loss = loss_function(y_pred_test, y_test_torch).item()\n",
    "            test_losses.append(test_loss)\n",
    "        model.train()\n",
    "\n",
    "    return train_losses, test_losses\n",
    "\n",
    "def evaluate(model):\n",
    "    with torch.no_grad():\n",
    "        y_pred = model(X_test_torch)\n",
    "        test_loss = loss_function(y_pred, y_test_torch).item()\n",
    "        #print(y_pred)\n",
    "        #print(y_test_torch)\n",
    "        accuracy = (np.argmax(y_pred.numpy(), axis=1)==np.argmax(y_test_torch.numpy(), axis=1)).sum() / y_test_torch.shape[0]\n",
    "        #print(f'Accuracy: {accuracy*100:.2f}%')\n",
    "    return test_loss, accuracy\n",
    "    #     y_pred_cls = y_pred.round()\n",
    "    #     acc = y_pred_cls.eq(y_test_torch).sum().item() / y_test_torch.shape[0]\n",
    "    #     print(f'Accuracy: {acc*100:.2f}%')\n",
    "\n",
    "        \n",
    "#train(model, optimizer, loss_function) evaluate(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "%%capture\n",
    "!pip install skorch\n",
    "\n",
    "#note that I added %%capture to the start of the cell because the verbose option wasn't working properly in my jupyter notebook instance, so my notebook was producing too much output and craching without %%capture\n",
    "from torch import nn\n",
    "from sklearn.model_selection import KFold\n",
    "from sklearn.model_selection import GridSearchCV\n",
    "import skorch\n",
    "from skorch import NeuralNetClassifier\n",
    "import torch.optim as optim\n",
    "\n",
    "net = NeuralNetClassifier(\n",
    "    train_split=None,\n",
    "    module=RpropModel,\n",
    "    criterion=nn.CrossEntropyLoss,\n",
    "    optimizer=optim.Rprop,\n",
    "    #optimizer__lr=0.001,\n",
    "#     batch_size=32,\n",
    "    max_epochs=64,\n",
    ")\n",
    "\n",
    "hyperparameter_grid = {\n",
    "    'module__activation_func': [nn.ReLU(), nn.LeakyReLU(0.01), nn.Tanh(), nn.ELU(alpha=1.0)],\n",
    "    'optimizer__lr': [0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1]\n",
    "}\n",
    "\n",
    "grid_search = GridSearchCV(net, hyperparameter_grid, cv=5, scoring='accuracy', verbose=0)\n",
    "grid_search.fit(X_train.to_numpy(dtype='float32'), np.argmax(y_train.to_numpy(dtype='float32'), axis=1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Best parameters found:  {'module__activation_func': ReLU(), 'optimizer__lr': 0.005}\n",
      "Highest accuracy found:  0.6778885600041543\n"
     ]
    }
   ],
   "source": [
    "print(\"Best parameters found: \", grid_search.best_params_)\n",
    "print(\"Highest accuracy found: \", grid_search.best_score_)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "import pandas as pd\n",
    "\n",
    "\n",
    "rprop_opt_results = grid_search.cv_results_\n",
    "#print(rprop_opt_results)\n",
    "data = pd.DataFrame({\n",
    "    'Activation Function': rprop_opt_results['param_module__activation_func'],\n",
    "    'Learning Rate': rprop_opt_results['param_optimizer__lr'],\n",
    "    'mean_score': rprop_opt_results['mean_test_score'],\n",
    "    'std_score': rprop_opt_results['std_test_score']\n",
    "})\n",
    "\n",
    "heatmap_data = data.pivot(index='Activation Function', columns='Learning Rate', values='mean_score')\n",
    "heatmap_data_std = data.pivot(index='Activation Function', columns='Learning Rate', values='std_score')\n",
    "\n",
    "plt.figure(figsize=(10, 6))\n",
    "sns.heatmap(heatmap_data, annot=True)\n",
    "plt.title('Heatmap showing the Mean Test Scores of the Grid Search Results for Hyperparameter Optimization (Learning Rate and Activation function)')\n",
    "plt.xlabel('Learning Rate')\n",
    "plt.ylabel('Activation Function')\n",
    "plt.show()\n",
    "\n",
    "plt.figure(figsize=(10, 6))\n",
    "sns.heatmap(heatmap_data_std, annot=True, fmt=\".3f\", cmap=\"coolwarm\")\n",
    "plt.title('Heatmap showing the Standard Deviation of Test Scores of the Grid Search Results for Hyperparameter Optimization (Learning Rate and Activation function)')\n",
    "plt.xlabel('Learning Rate')\n",
    "plt.ylabel('Activation Function')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "T-statistic: 0.22241616790905228, P-value: 0.8348850125956644\n"
     ]
    }
   ],
   "source": [
    "from scipy.stats import ttest_rel\n",
    "import numpy as np\n",
    "\n",
    "mask_relu = np.array([isinstance(act_func, nn.ReLU) for act_func in rprop_opt_results['param_module__activation_func']])\n",
    "scores_relu = np.array([rprop_opt_results[f'split{i}_test_score'][mask_relu] for i in range(5)])\n",
    "\n",
    "# print(mask_relu)\n",
    "# print(scores_relu)\n",
    "# print(rprop_opt_results['param_module__activation_func'])\n",
    "# print(type(rprop_opt_results['param_module__activation_func'][0]) == type(nn.ReLU()))\n",
    "\n",
    "mask_lrelu = np.array([isinstance(act_func, nn.LeakyReLU) for act_func in rprop_opt_results['param_module__activation_func']])\n",
    "scores_lrelu = np.array([rprop_opt_results[f'split{i}_test_score'][mask_lrelu] for i in range(5)])\n",
    "\n",
    "avg_relu_score = scores_relu.mean(axis=1)\n",
    "avg_lrelu_score = scores_lrelu.mean(axis=1)\n",
    "\n",
    "t_statistic, p_value = ttest_rel(avg_relu_score, avg_lrelu_score)\n",
    "print(f\"T-statistic: {t_statistic}, P-value: {p_value}\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x1cd82755a90>"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAkAAAAHHCAYAAABXx+fLAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjcuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8pXeV/AAAACXBIWXMAAA9hAAAPYQGoP6dpAACM9UlEQVR4nOzdd3xN9x/H8de92TuyhZCIETNiRYyiTRujatWq1u5QdKgqHagO2qJatBQ/VEuNolqjRu0VK2YQEgmyBEkkkXnP748rV28TRBK5GZ/n43Eeufes+7knV+7bOd/z/aoURVEQQgghhKhA1IYuQAghhBCipEkAEkIIIUSFIwFICCGEEBWOBCAhhBBCVDgSgIQQQghR4UgAEkIIIUSFIwFICCGEEBWOBCAhhBBCVDgSgIQQQghR4UgAEnoGDx6Mp6dnobadPHkyKpWqeAsqZa5cuYJKpWLJkiWGLuWRlixZgkql4sqVK4YuRYhio1KpGDVq1BN9jSNHjtCqVSusrKxQqVSEhIQ80dcThiEBqIxQqVQFmnbt2mXoUis8T0/PAv2uiitEffnll6xfv75Y9lVccsNwQkKCoUspVy5fvsxLL72Ei4sLFhYW1KpVi48++uiB62dlZVGvXj1UKhXTp08v0Gs87DP7xhtvFNdbKbWysrLo3bs3t27d4ttvv2XZsmVUr179ib3erl279I6xkZERLi4uvPjii4SGhj6x1xVgbOgCRMEsW7ZM7/nPP//Mtm3b8syvW7dukV5nwYIFaDSaQm378ccfM378+CK9fnkwa9YsUlJSdM83bdrEihUr+Pbbb3FyctLNb9WqVbG83pdffsmLL75I9+7d9ea/8sor9OvXDzMzs2J5HWFYISEhtG/fnipVqvDee+/h6OhIVFQUV69efeA2s2fPJioq6rFf69lnn2XgwIF55teuXfux91XWXL58mcjISBYsWMDw4cNL7HXfeustmjdvTlZWFqdOnWLevHns2rWLM2fO4ObmVmJ1VCQSgMqIl19+We/5oUOH2LZtW575/5WWloalpWWBX8fExKRQ9QEYGxtjbCwfqf8GkdjYWFasWEH37t0LfXmxMIyMjDAyMiqx1xNFk5qaipWVVb7LNBoNr7zyCj4+PuzcuRMLC4tH7i8+Pp4pU6bwwQcfMHHixMeqpXbt2o/821JexcfHA2Bvb19s+3zY7zZX27ZtefHFF3XP69Spw4gRI/j5558ZN25csdSRnp6OqakparVc/AG5BFautG/fngYNGnDs2DGeeuopLC0t+fDDDwH4448/6NKlC+7u7piZmeHt7c1nn31GTk6O3j7+2wYot83L9OnT+emnn/D29sbMzIzmzZtz5MgRvW3zawOUe71+/fr1NGjQADMzM+rXr8+WLVvy1L9r1y6aNWuGubk53t7ezJ8/v8Dtivbu3Uvv3r2pVq0aZmZmeHh48O6773L37t0878/a2prr16/TvXt3rK2tcXZ2ZuzYsXmORWJiIoMHD8bOzg57e3sGDRpEYmLiI2spqF9++YWmTZtiYWGBg4MD/fr1y/O/+bCwMHr16oWbmxvm5uZUrVqVfv36kZSUBGiPb2pqKkuXLtWdQh88eDCQfxsgT09Pnn/+efbt20eLFi0wNzenRo0a/Pzzz3nqO3XqFO3atcPCwoKqVavy+eefs3jx4mJtV/TPP//Qtm1brKyssLe3p1u3bnlO+9+5c4d33nkHT09PzMzMcHFx4dlnn+X48eMFPk4Ps3r1at3vwcnJiZdffpnr16/rlk+fPh2VSkVkZGSebSdMmICpqSm3b9/WzTt8+DAdO3bEzs4OS0tL2rVrx/79+/W2y/1cnzt3jpdeeolKlSrRpk2bB9a4detWzpw5w6RJk7CwsCAtLS3P5/W/xo8fT506dZ5YkPn335tWrVphYWGBl5cX8+bNy7NufHw8w4YNw9XVFXNzc3x9fVm6dGme9TQaDd999x0NGzbE3NwcZ2dnOnbsyNGjR/Os+6i/KQX53PzX4MGDadeuHQC9e/dGpVLRvn173fKCfF4f93f7IG3btgW0Z6T+raDHPffS2m+//cbHH39MlSpVsLS0JDk5GXj05z73eFhbWxMeHk5QUBBWVla4u7szZcoUFEV57PdU2sh/18uZmzdv0qlTJ/r168fLL7+Mq6sroP0ytLa2ZsyYMVhbW/PPP/8wceJEkpOT+eabbx653+XLl3Pnzh1ef/11VCoVX3/9NT179iQ8PPyRZ4327dvH2rVrefPNN7GxseH777+nV69eREVF4ejoCMCJEyfo2LEjlStX5tNPPyUnJ4cpU6bg7OxcoPe9evVq0tLSGDFiBI6OjgQHBzN79myuXbvG6tWr9dbNyckhKCgIf39/pk+fzvbt25kxYwbe3t6MGDECAEVR6NatG/v27eONN96gbt26rFu3jkGDBhWonkf54osv+OSTT+jTpw/Dhw/nxo0bzJ49m6eeeooTJ05gb29PZmYmQUFBZGRkMHr0aNzc3Lh+/Tp//fUXiYmJ2NnZsWzZMoYPH06LFi147bXXAPD29n7oa1+6dIkXX3yRYcOGMWjQIP73v/8xePBgmjZtSv369QG4fv06HTp0QKVSMWHCBKysrFi4cGGxXk7bvn07nTp1okaNGkyePJm7d+8ye/ZsWrduzfHjx3VB/I033mDNmjWMGjWKevXqcfPmTfbt20doaChNmjQp0HF6kCVLljBkyBCaN2/O1KlTiYuL47vvvmP//v2630OfPn0YN24cq1at4v3339fbftWqVTz33HNUqlQJ0H5BdurUiaZNmzJp0iTUajWLFy/m6aefZu/evbRo0UJv+969e1OrVi2+/PLLh36hbN++HQAzMzOaNWvGsWPHMDU1pUePHvzwww84ODjorR8cHMzSpUvZt29foW5MSE9Pz7f9lq2tLaamprrnt2/fpnPnzvTp04f+/fuzatUqRowYgampKUOHDgXg7t27tG/fnkuXLjFq1Ci8vLxYvXo1gwcPJjExkbffflu3v2HDhrFkyRI6derE8OHDyc7OZu/evRw6dIhmzZrp1ivI35RHfW7y8/rrr1OlShW+/PJL3SWp3L+hBf285iro7/ZBcv+TkfvZ+reCHPdcn332GaampowdO5aMjAxMTU0L9LnPlZOTQ8eOHWnZsiVff/01W7ZsYdKkSWRnZzNlypTHfl+liiLKpJEjRyr//fW1a9dOAZR58+blWT8tLS3PvNdff12xtLRU0tPTdfMGDRqkVK9eXfc8IiJCARRHR0fl1q1buvl//PGHAih//vmnbt6kSZPy1AQopqamyqVLl3TzTp48qQDK7NmzdfO6du2qWFpaKtevX9fNCwsLU4yNjfPsMz/5vb+pU6cqKpVKiYyM1Ht/gDJlyhS9df38/JSmTZvqnq9fv14BlK+//lo3Lzs7W2nbtq0CKIsXL35kTbm++eYbBVAiIiIURVGUK1euKEZGRsoXX3yht97p06cVY2Nj3fwTJ04ogLJ69eqH7t/KykoZNGhQnvmLFy/We11FUZTq1asrgLJnzx7dvPj4eMXMzEx57733dPNGjx6tqFQq5cSJE7p5N2/eVBwcHPLsMz+5n4UbN248cJ3GjRsrLi4uys2bN3XzTp48qajVamXgwIG6eXZ2dsrIkSMfuJ+CHqf/yszMVFxcXJQGDRood+/e1c3/66+/FECZOHGibl5AQIDe50NRFCU4OFgBlJ9//llRFEXRaDRKrVq1lKCgIEWj0ejWS0tLU7y8vJRnn31WNy/3+PTv379Atb7wwgu6f4cDBgxQ1qxZo3zyySeKsbGx0qpVK73X02g0SosWLXT7zv03/M033xTotYAHTitWrNCtl/v3ZsaMGbp5GRkZut9rZmamoiiKMmvWLAVQfvnlF916mZmZSkBAgGJtba0kJycriqIo//zzjwIob731Vp6a/v3+Cvo35VGfmwfZuXNnvp+ngn5eH/d3m/t6//vf/5QbN24o0dHRypYtW5SaNWsqKpVKCQ4O1lu/oMc9d781atTQ+/v4OJ/73L+Xo0eP1s3TaDRKly5dFFNT04f++y4L5BJYOWNmZsaQIUPyzP93m4E7d+6QkJBA27ZtSUtL4/z584/cb9++ffX+J5J7ejY8PPyR2wYGBuqdlWjUqBG2tra6bXNycti+fTvdu3fH3d1dt17NmjXp1KnTI/cP+u8vNTWVhIQEWrVqhaIonDhxIs/6/72bpW3btnrvZdOmTRgbG+vOCIG2Tc3o0aMLVM/DrF27Fo1GQ58+fUhISNBNbm5u1KpVi507dwLozlz8/fffpKWlFfl1c9WrV0/3+wNwdnamTp06eu9/y5YtBAQE0LhxY908BwcHBgwYUCw1xMTEEBISwuDBg/XOXjRq1Ihnn32WTZs26ebZ29tz+PBhoqOj891XYY/T0aNHiY+P580338Tc3Fw3v0uXLvj4+LBx40bdvL59+3Ls2DG9yxErV67EzMyMbt26AdpGymFhYbz00kvcvHlT93tNTU3lmWeeYc+ePXluMCjoXVW5jeqbN2/OL7/8Qq9evZgyZQqfffYZBw4cYMeOHbp1lyxZwunTp/nqq68KfCz+q1u3bmzbti3P1KFDB731jI2Nef3113XPTU1Nef3114mPj+fYsWOA9t+Sm5sb/fv3161nYmLCW2+9RUpKCrt37wbg999/R6VSMWnSpDz1/Pcs1qP+psCjPzeP43E+r7ke9465oUOH4uzsjLu7Ox07diQpKYlly5bRvHnzPOsW5LjnGjRokN7fx8f53Of6d7cDuc0aMjMzdWcmyyoJQOVMlSpV9E5R5zp79iw9evTAzs4OW1tbnJ2ddW0DCtJOolq1anrPc8PQv9s+FHTb3O1zt42Pj+fu3bvUrFkzz3r5zctPVFSU7o9Tbrue3Gv5/31/uW0LHlQPQGRkJJUrV8ba2lpvvTp16hSonocJCwtDURRq1aqFs7Oz3hQaGqprhOnl5cWYMWNYuHAhTk5OBAUFMXfu3AL9vh7mUb8P0L7/ovw+HiW3PU1+x7Nu3bq64ADw9ddfc+bMGTw8PGjRogWTJ0/W+6Ir7HF6WA0+Pj56bX569+6NWq1m5cqVgPYS6erVq+nUqRO2traA9vcK2i+c//5eFy5cSEZGRp6avLy8Hn6g7sn9Avt3iAB46aWXADhw4AAAycnJTJgwgffffx8PD48C7Ts/VatWJTAwMM+Uezkol7u7e57Gvbl3iuVewomMjKRWrVp5Gt7m3rGae5wvX76Mu7t7nst5+SnIZ/hRn5vH8Tif11wF/d3mmjhxItu2bWPdunUMHDiQpKSkBzZWLshxf1Adj/O5B1Cr1dSoUaNAr1XWSBugcia/u0MSExNp164dtra2TJkyBW9vb8zNzTl+/DgffPBBgW57f9DdREoBrm0XZduCyMnJ4dlnn+XWrVt88MEH+Pj4YGVlxfXr1xk8eHCe92foO6M0Gg0qlYrNmzfnW8u/Q9eMGTMYPHgwf/zxB1u3buWtt95i6tSpHDp0iKpVqxbq9Z/076O49enTh7Zt27Ju3Tq2bt3KN998w1dffcXatWt1ZwifxHH6N3d3d9q2bcuqVav48MMPOXToEFFRUXpnWXI/Z998843embN/+2+gLsjdXLmvD+QJIC4uLsD9/4hMnz6dzMxM+vbtq/tyunbtmm6dK1eu4O7unu9/ksqSgnyGC/K5eZIK+rvN1bBhQwIDAwHtnaRpaWm8+uqrtGnTpkhh9nHrqEjkDFAFsGvXLm7evMmSJUt4++23ef755wkMDMy3cZ0huLi4YG5uzqVLl/Isy2/ef50+fZqLFy8yY8YMPvjgA7p160ZgYKDe5bTHVb16dWJiYvT68wG4cOFCofeZy9vbG0VR8PLyyvd/2S1bttRbv2HDhnz88cfs2bOHvXv3cv36db07Pp5E79vVq1cv9O+joPuH/I/n+fPncXJy0vsfbuXKlXnzzTdZv349ERERODo68sUXX+ht96jj9Dg1XLhwIU/nd3379uXkyZNcuHCBlStXYmlpSdeuXXXLcy/J2Nra5vt7DQwMLHQ3E02bNgXIc5dO7uWd3DOaUVFR3L59m/r16+Pl5YWXl5fucueXX36Jl5cX586dK1QN+YmOjs5z5uPixYsAukbB1atXJywsLM9/RHIvveceZ29vb6Kjo7l161ax1VeQz01BPO7ntThMmzaN9PT0fOstyHF/kMf93Gs0mjxnzgr6WqWdBKAKIPd/S//+31FmZiY//PCDoUrSY2RkRGBgIOvXr9e7Xn/p0iU2b95coO1B//0pisJ3331X6Jo6d+5MdnY2P/74o25eTk4Os2fPLvQ+c/Xs2RMjIyM+/fTTPGddFEXh5s2bgPZyRnZ2tt7yhg0bolarycjI0M2zsrIq1tvzAYKCgjh48KDeEAC3bt3i119/LZb9V65cmcaNG7N06VK92s+cOcPWrVvp3LkzoD3m/71s5OLigru7u+4YFPQ4/VezZs1wcXFh3rx5eutt3ryZ0NBQunTpord+r169MDIyYsWKFaxevZrnn39e70uvadOmeHt7M3369DzBGeDGjRuPOCoP1q1bN8zMzFi8eLFekFi4cCGg7bgQtJ3prVu3Tm+aP38+oL2led26dY99aeZhsrOzdfsH7d+V+fPn4+zsrAttnTt3JjY2Vnf5MHe72bNnY21trbtU3atXLxRF4dNPP83zOo97drIgn5vHUdDPa3Hy9vamV69eLFmyhNjYWL1lBTnuD/K4n3uAOXPm6B4risKcOXMwMTHhmWeeKezbKxXkElgF0KpVKypVqsSgQYN46623UKlULFu2rFRd8pg8eTJbt26ldevWjBgxgpycHObMmUODBg0eOQ6Pj48P3t7ejB07luvXr2Nra8vvv/9eoPZJD9K1a1dat27N+PHjuXLlCvXq1WPt2rVFbn8D2j9sn3/+ORMmTODKlSt0794dGxsbIiIiWLduHa+99hpjx47ln3/+YdSoUfTu3ZvatWuTnZ3NsmXLMDIyolevXrr9NW3alO3btzNz5kzc3d3x8vLC39+/SDWOGzeOX375hWeffZbRo0frboOvVq0at27dKvBZp5kzZ+bpiFOtVvPhhx/yzTff0KlTJwICAhg2bJjutmI7OzsmT54MaBvsV61alRdffBFfX1+sra3Zvn07R44cYcaMGQAFPk7/ZWJiwldffcWQIUNo164d/fv3190O7Onpybvvvqu3vouLCx06dGDmzJncuXOHvn375nlfCxcupFOnTtSvX58hQ4ZQpUoVrl+/zs6dO7G1teXPP/8s0HH7Lzc3Nz766CMmTpxIx44d6d69OydPnmTBggX0799f11C2SZMmeW7xzr0UVr9+/TyddD7IxYsX+eWXX/LMd3V11YUt0F6a++qrr7hy5Qq1a9dm5cqVhISE8NNPP+nOdr322mvMnz+fwYMHc+zYMTw9PVmzZg379+9n1qxZ2NjYANChQwdeeeUVvv/+e8LCwujYsSMajYa9e/fSoUOHxxr/qyCfm8dVkM9rcXv//fdZtWoVs2bNYtq0abr5BTnuD/K4n3tzc3O2bNnCoEGD8Pf3Z/PmzWzcuJEPP/ywwN2UlFolft+ZKBYPug2+fv36+a6/f/9+pWXLloqFhYXi7u6ujBs3Tvn7778VQNm5c6duvQfdBp/fLbSAMmnSJN3zB90Gn9+tqNWrV89z6/aOHTsUPz8/xdTUVPH29lYWLlyovPfee4q5ufkDjsJ9586dUwIDAxVra2vFyclJefXVV3W3xv77lvVBgwYpVlZWebbPr/abN28qr7zyimJra6vY2dkpr7zyiu6W66LcBp/r999/V9q0aaNYWVkpVlZWio+PjzJy5EjlwoULiqIoSnh4uDJ06FDF29tbMTc3VxwcHJQOHToo27dv19vP+fPnlaeeekqxsLBQAN1xfdBt8F26dMlTY7t27ZR27drpzTtx4oTStm1bxczMTKlataoydepU5fvvv1cAJTY29qHvOfd45jcZGRnp1tu+fbvSunVrxcLCQrG1tVW6du2qnDt3Trc8IyNDef/99xVfX1/FxsZGsbKyUnx9fZUffvhBt05Bj9ODrFy5UvHz81PMzMwUBwcHZcCAAcq1a9fyXXfBggUKoNjY2OjdQvzf49azZ0/F0dFRMTMzU6pXr6706dNH2bFjR57j8zi3EWs0GmX27NlK7dq1FRMTE8XDw0P5+OOPdbc9P0hx3gb/789I7t+bo0ePKgEBAYq5ublSvXp1Zc6cOXn2GRcXpwwZMkRxcnJSTE1NlYYNG+b7byg7O1v55ptvFB8fH8XU1FRxdnZWOnXqpBw7dkyvvkf9TSnI5+ZBHnQbvKI8+vOqKI//u33Y6ymKorRv316xtbVVEhMTFUUp+HF/1H4L8rnP/Xt5+fJl5bnnnlMsLS0VV1dXZdKkSUpOTk6B3l9pplKUUnQaQIj/6N69O2fPntXdYSMM65133mH+/PmkpKQYvDG5MKz27duTkJDAmTNnDF1KhVKSx33w4MGsWbMm30u65YG0ARKlxn+HrQgLC2PTpk16XdGLkvPf38fNmzdZtmwZbdq0kfAjhCjzpA2QKDVq1KjB4MGDqVGjBpGRkfz444+YmpoW20CA4vEEBATQvn176tatS1xcHIsWLSI5OZlPPvnE0KUJIUSRSQASpUbHjh1ZsWIFsbGxmJmZERAQwJdffkmtWrUMXVqF1LlzZ9asWcNPP/2ESqWiSZMmLFq0iKeeesrQpQkhRJFJGyAhhBBCVDjSBkgIIYQQFY4EICGEEEJUONIGKB8ajYbo6GhsbGyeyDADQgghhCh+iqJw584d3N3dHziYbC4JQPmIjo4u0uBzQgghhDCcq1evPnIgZAlA+cjtmv3q1avY2toauBohhBBCFERycjIeHh667/GHkQCUj9zLXra2thKAhBBCiDKmIM1XpBG0EEIIISocCUBCCCGEqHAkAAkhhBCiwpE2QEIIUcZoNBoyMzMNXYYQJc7ExKTYBmOWACSEEGVIZmYmERERaDQaQ5cihEHY29vj5uZW5H76JAAJIUQZoSgKMTExGBkZ4eHh8ciO3oQoTxRFIS0tjfj4eAAqV65cpP1JABJCiDIiOzubtLQ03N3dsbS0NHQ5QpQ4CwsLAOLj43FxcSnS5TD574MQQpQROTk5AJiamhq4EiEMJzf8Z2VlFWk/EoCEEKKMkTEKRUVWXJ9/CUBCCCGEqHAkAAkhhChzPD09mTVrVoHX37VrFyqVisTExCdWkyhbJAAJIYR4YlQq1UOnyZMnF2q/R44c4bXXXivw+q1atSImJgY7O7tCvV5BlbWgtWTJEt3vQq1WU7lyZfr27UtUVJShS3vi5C6wErb74g38vRwwNymejpyEEKI0i4mJ0T1euXIlEydO5MKFC7p51tbWuseKopCTk4Ox8aO/mpydnR+rDlNTU9zc3B5rm4rC1taWCxcuoCgKERERvPnmm/Tu3ZvDhw8Xab9ZWVmYmJgUU5XFT84AlaD1J64zeHEwr/58lLuZOYYuRwghnjg3NzfdZGdnh0ql0j0/f/48NjY2bN68maZNm2JmZsa+ffu4fPky3bp1w9XVFWtra5o3b8727dv19vvfS2AqlYqFCxfSo0cPLC0tqVWrFhs2bNAt/++ZmSVLlmBvb8/ff/9N3bp1sba2pmPHjnqBLTs7m7feegt7e3scHR354IMPGDRoEN27dy/08bh9+zYDBw6kUqVKWFpa0qlTJ8LCwnTLIyMj6dq1K5UqVcLKyor69euzadMm3bYDBgzA2dkZCwsLatWqxeLFiwtdS67c30nlypVp1aoVw4YNIzg4mOTkZN06np6efPbZZ/Tv3x8rKyuqVKnC3Llz8+znxx9/5IUXXsDKyoovvvgCgB9//BFvb29MTU2pU6cOy5Yty3e7Tp06YWFhQY0aNVizZk2R39ejGDQA7dmzh65du+Lu7o5KpWL9+vUPXT8mJoaXXnqJ2rVro1areeedd/Jdb/Xq1fj4+GBubk7Dhg11Hx5Dc7U1x8LEiL1hCQxdcoS0zGxDlySEKMMURSEtM9sgk6IoxfY+xo8fz7Rp0wgNDaVRo0akpKTQuXNnduzYwYkTJ+jYsSNdu3Z95GWZTz/9lD59+nDq1Ck6d+7MgAEDuHXr1gPXT0tLY/r06Sxbtow9e/YQFRXF2LFjdcu/+uorfv31VxYvXsz+/ftJTk5+5PfUowwePJijR4+yYcMGDh48iKIodO7cWXdL98iRI8nIyGDPnj2cPn2ar776SneW7JNPPuHcuXNs3ryZ0NBQfvzxR5ycnIpUz3/Fx8ezbt06jIyM8vSx88033+Dr68uJEycYP348b7/9Ntu2bdNbZ/LkyfTo0YPTp08zdOhQ1q1bx9tvv817773HmTNneP311xkyZAg7d+7U2+6TTz6hV69enDx5kgEDBtCvXz9CQ0OL9b39l0EvgaWmpuLr68vQoUPp2bPnI9fPyMjA2dmZjz/+mG+//TbfdQ4cOED//v2ZOnUqzz//PMuXL6d79+4cP36cBg0aFPdbeCwB3o78PLQFgxcf4WD4TQb9L5j/DW6OjXnpPUUohCi97mblUG/i3wZ57XNTgrA0LZ6vkClTpvDss8/qnjs4OODr66t7/tlnn7Fu3To2bNjAqFGjHrifwYMH079/fwC+/PJLvv/+e4KDg+nYsWO+62dlZTFv3jy8vb0BGDVqFFOmTNEtnz17NhMmTKBHjx4AzJkzp0j/oQ4LC2PDhg3s37+fVq1aAfDrr7/i4eHB+vXr6d27N1FRUfTq1YuGDRsCUKNGDd32UVFR+Pn50axZM0B7VqY4JCUlYW1tretpGeCtt97CyspKb73WrVszfvx4AGrXrs3+/fv59ttv9X53L730EkOGDNE979+/P4MHD+bNN98EYMyYMRw6dIjp06fToUMH3Xq9e/dm+PDhgPb3vW3bNmbPns0PP/xQLO8xPwY9A9SpUyc+//xz3YfrUTw9Pfnuu+8YOHDgAxuyfffdd3Ts2JH333+funXr8tlnn9GkSRPmzJlTnKUXWjNPB5YNa4GNuTFHrtzmlUXBJN0tWmdOQghRluV+oedKSUlh7Nix1K1bF3t7e6ytrQkNDX3kGaBGjRrpHltZWWFra6sbNiE/lpaWuvAD2qEVctdPSkoiLi6OFi1a6JYbGRnRtGnTx3pv/xYaGoqxsTH+/v66eY6OjtSpU0d3tuOtt97i888/p3Xr1kyaNIlTp07p1h0xYgS//fYbjRs3Zty4cRw4cOCBr/Xrr79ibW2tm/bu3fvAdW1sbAgJCeHo0aPMmDGDJk2a6C5f/VtAQECe5/89S/Pf32VoaCitW7fWm9e6des82xVk38Wt3DWCPnjwIGPGjNGbFxQU9NDTlhkZGWRkZOie//u655PgV60SK15tycuLDhNyNZGXFx5m2bAW2FtK765CiIKzMDHi3JQgg712cfnvmYaxY8eybds2pk+fTs2aNbGwsODFF18kMzPzofv5b4NblUr10EFj81u/OC/tFcbw4cMJCgpi48aNbN26lalTpzJjxgxGjx5Np06diIyMZNOmTWzbto1nnnmGkSNHMn369Dz7eeGFF/SCVpUqVR74mmq1mpo1awJQt25dLl++zIgRI/K01SmI//4uS7Ny1wg6NjYWV1dXvXmurq7ExsY+cJupU6diZ2enmzw8PJ5McZoc2DYREi7RoIodK15tiaOVKaevJ9F/wWFupmQ8eh9CCHGPSqXC0tTYINOT7I16//79DB48mB49etCwYUPc3Ny4cuXKE3u9/NjZ2eHq6sqRI0d083Jycjh+/Hih91m3bl2ys7P17q66efMmFy5coF69erp5Hh4evPHGG6xdu5b33nuPBQsW6JY5OzszaNAgfvnlF2bNmsVPP/2U72vZ2NhQs2ZN3ZQ7hlZBjB8/npUrV+Z5r4cOHcrzvG7duo98z/v379ebt3//fr33W9h9F1W5OwNUGBMmTNA7a5ScnPxkQtCuqbD/OzizFob+Td3KVfjttZa8tPAwoTHJ9F9wiF+G++NiY178ry2EEGVErVq1WLt2LV27dkWlUvHJJ5889EzOkzJ69GimTp1KzZo18fHxYfbs2dy+fbtA4e/06dPY2NjonqtUKnx9fenWrRuvvvoq8+fPx8bGhvHjx1OlShW6desGwDvvvEOnTp2oXbs2t2/fZufOnbogMHHiRJo2bUr9+vXJyMjgr7/+eiIhwcPDgx49ejBx4kT++usv3fz9+/fz9ddf0717d7Zt28bq1avZuHHjQ/f1/vvv06dPH/z8/AgMDOTPP/9k7dq1ee7qW716Nc2aNaNNmzb8+uuvBAcHs2jRomJ/b/9W7s4Aubm5ERcXpzcvLi7uof0/mJmZYWtrqzc9ES1eB8eakHQVlnWH1ARqudqw8rWWuNmaczEuhQELDpOUJm2ChBAV18yZM6lUqRKtWrWia9euBAUF0aRJkxKv44MPPqB///4MHDiQgIAArK2tCQoKwtz80f9Jfeqpp/Dz89NNuW2HFi9eTNOmTXn++ecJCAhAURQ2bdqkuxyXk5PDyJEjqVu3Lh07dqR27dq6hsCmpqZMmDCBRo0a8dRTT2FkZMRvv/32RN77u+++y8aNGwkODtbNe++99zh69Ch+fn58/vnnzJw5k6Cgh1+C7d69O9999x3Tp0+nfv36zJ8/n8WLF9O+fXu99T799FN+++03GjVqxM8//8yKFSvynCUqbirF0Bc871GpVKxbt67A/Su0b9+exo0b5+kKvW/fvqSlpfHnn3/q5rVq1YpGjRoxb968Au07OTkZOzs7kpKSij8MJV6F/3WE5GtQuTEM+hPMbYm6mUaf+QeJTU7H38uBn4e1wMxYOksUQtyXnp5OREQEXl5eBfoSFsVLo9FQt25d+vTpw2effWbockqUp6cn77zzzgO7nymKx/3+f9i/g8f5/jboGaCUlBRCQkIICQkBICIigpCQEF1L/wkTJjBw4EC9bXLXT0lJ4caNG4SEhHDu3Dnd8rfffpstW7YwY8YMzp8/z+TJkzl69OhDb50sUfYeMHA9WDpCTAis6A9Zd6nmaMmSoc2xMTPmcMQtxq4+hUZTKrKpEEJUSJGRkSxYsICLFy9y+vRpRowYQUREBC+99JKhSxPFwKABKPdUmp+fH6DtH8DPz4+JEycC2o4P/3vbY+76x44dY/ny5fj5+dG5c2fd8latWrF8+XJ++uknfH19WbNmDevXrzd4H0B6nGrBy2vB1AYi98HqIZCThY+bLfNeaYqJkYo/T0bz1d/nDV2pEEJUWGq1miVLltC8eXNat27N6dOn2b59+xNvnCtKRqm5BFaaPNFLYP92ZR/80guy06FRX+g+D9Rq1h6/xphVJwGY0q0+AwM8n1wNQogyQy6BCVFOLoFVeJ5toPdSUBnBqZWwZTwoCj2bVOW9Z2sDMHnDWbadi3vEjoQQQgjxOCQAGVqdjtBjHqCC4Pmw+ysARj1dk37NPdAoMHrFcUKuJhq0TCGEEKI8kQBUGjTqA52/0T7eNRXObUClUvF59wa0q+1MepaGYUuOEHkz1bB1CiGEEOWEBKDSosWr0HKk9vH6EXDjIsZGauYOaEJ9d1tupmYyZPER7qRLH0FCCCFEUUkAKk2e/RSqt4HMFFg5ANKTsTYzZvHg5rjbmROekMon688YfKwaIYQQoqyTAFSaGJlA78Vg4w4JF+GPN0FRcLE15/v+fhipVawPieb349cNXakQQghRpkkAKm2sXaDvMjAyhdA/Yf8sAJp5OvDOM7UAmPjHGcJvpBiwSCGEKJ0mT55M48aNDV2GKAMkAJVGVZtBJ+3dYOyYApf/AeDNDjUJqOFIWmYOo1ecICM7x4BFCiHEo6lUqodOkydPLtK+169frzdv7Nix7Nixo2hFF0BZC1qDBw/WHXMTExO8vLwYN24c6enphi7NYCQAlVZNh4Dfy6BoYM0wuB2JkVrFrH6NcbAy5Wx0MtM2S0/RQojSLSYmRjfNmjULW1tbvXljx44t1teztrbG0dGxWPdZXnTs2JGYmBjCw8P59ttvmT9/PpMmTSrSPhVFITs7u5gqLFkSgEorlQo6zwB3P7h7C1a9All3cbU1Z3rvRgAs3n+FHaHSSaIQovRyc3PTTXZ2dqhUKr15v/32G3Xr1sXc3BwfHx/dyOcAmZmZjBo1isqVK2Nubk716tWZOnUqoB2cE6BHjx6oVCrd8/+emRk8eDDdu3dn+vTpVK5cGUdHR0aOHElW1v07amNiYujSpQsWFhZ4eXmxfPlyPD098wy2/ThOnz7N008/jYWFBY6Ojrz22mukpNxvurBr1y5atGiBlZUV9vb2tG7dmsjISABOnjxJhw4dsLGxwdbWlqZNm3L06NFC15LLzMwMNzc3PDw86N69O4GBgWzbtk23/MqVK6hUKn777TdatWqFubk5DRo0YPfu3Xp1q1QqNm/eTNOmTTEzM2Pfvn1kZGTw1ltv4eLigrm5OW3atOHIkSN5ttu4cSONGjXC3Nycli1bcubMmSK/r8KSAFSamZhDn2X3Bk49CRvfA+BpH1eGtvYCYOzqk8QmVdxTmEJUaIoCmamGmYrhbtRff/2ViRMn8sUXXxAaGsqXX37JJ598wtKlSwH4/vvv2bBhA6tWreLChQv8+uuvuqCT++W6ePFiYmJi9L5s/2vnzp1cvnyZnTt3snTpUpYsWcKSJUt0ywcOHEh0dDS7du3i999/56effiI+Pr7Q7ys1NZWgoCAqVarEkSNHWL16Ndu3b9cNyp2dnU337t1p164dp06d4uDBg7z22muoVCoABgwYQNWqVTly5AjHjh1j/PjxmJiYFLqe/Jw5c4YDBw5gamqaZ9n777/Pe++9x4kTJwgICKBr167cvHlTb53x48czbdo0QkNDadSoEePGjeP3339n6dKlHD9+nJo1axIUFMStW7fy7HvGjBkcOXIEZ2dnunbtqhdGS5KxQV5VFJy9B7z4P1jWA0J+hXrdofZzfNCpDocjbnI2Opl3Vp7g1+EtMVKrDF2tEKIkZaXBl+6Gee0Po8HUqki7mDRpEjNmzKBnz54AeHl5ce7cOebPn8+gQYOIioqiVq1atGnTBpVKRfXq1XXbOjs7A2Bvb4+bm9tDX6dSpUrMmTMHIyMjfHx86NKlCzt27ODVV1/l/PnzbN++nSNHjtCsWTMAFi5cSK1atQr9vpYvX056ejo///wzVlbaYzRnzhy6du3KV199hYmJCUlJSTz//PN4e3sD6A2wGhUVxfvvv4+Pjw9AkWr5t7/++gtra2uys7PJyMhArVYzZ86cPOuNGjWKXr16AfDjjz+yZcsWFi1axLhx43TrTJkyhWeffRbQBr4ff/yRJUuW0KlTJwAWLFjAtm3bWLRoEe+//75uu0mTJum2W7p0KVWrVmXdunX06dOnWN7j45AzQGVBjfYQcK+TxE1jITMNM2MjZvf3w9LUiEPht5i785JBSxRCiMeRmprK5cuXGTZsGNbW1rrp888/5/Lly4D28lVISAh16tThrbfeYuvWrYV6rfr162NkZKR7XrlyZd0ZngsXLmBsbEyTJk10y2vWrEmlSpUK/d5CQ0Px9fXVhR+A1q1bo9FouHDhAg4ODgwePJigoCC6du3Kd999R0xMjG7dMWPGMHz4cAIDA5k2bZrueOTnjTfe0Dt+D9OhQwdCQkI4fPgwgwYNYsiQIbqg828BAQG6x8bGxjRr1ozQ0FC9dXLDIsDly5fJysqidevWunkmJia0aNEiz3b/3reDgwN16tTJs05JkTNAZUW78XBmHSRGwt4Z8Mwn1HC25rNuDXhv9Ulmbb/I0z4uNKhiZ+hKhRAlxcRSeybGUK9dBLntYRYsWIC/v7/estyw0qRJEyIiIti8eTPbt2+nT58+BAYGsmbNmscr9T+Xj1QqFRqNpgjVF93ixYt566232LJlCytXruTjjz9m27ZttGzZksmTJ/PSSy+xceNGNm/ezKRJk/jtt9/o0aNHnv1MmTKlwA3JraysqFmzJgD/+9//8PX1ZdGiRQwbNuyx6/93uCur5AxQWWFmDZ2maR/v/w5uXASgV9OqdGlUGY0CU/48J71EC1GRqFTay1CGmFRFu+Tu6uqKu7s74eHh1KxZU2/y8vLSrWdra0vfvn1ZsGABK1eu5Pfff9e1KzExMSEnp2jdgdSpU4fs7GxOnDihm3fp0iVu375d6H3WrVuXkydPkpp6f/zG/fv3o1arqVOnjm6en58fEyZM4MCBAzRo0IDly5frltWuXZt3332XrVu30rNnTxYvXpzva7m4uOgdu4JSq9V8+OGHfPzxx9y9e1dv2aFDh3SPs7OzOXbsmN4luv/y9vbG1NSU/fv36+ZlZWVx5MgR6tWr98B93759m4sXLz5030+SBKCyxOd5qBUEmizYOEbXCPGjznUxN1ETfOUWm07HGrhIIYQomE8//ZSpU6fy/fffc/HiRU6fPs3ixYuZOXMmADNnzmTFihWcP3+eixcvsnr1atzc3LC3twe0d4Lt2LGD2NjYQgcWHx8fAgMDee211wgODubEiRO89tprWFhY6BolP8jdu3cJCQnRmy5fvsyAAQMwNzdn0KBBnDlzhp07dzJ69GheeeUVXF1diYiIYMKECRw8eJDIyEi2bt1KWFgYdevW5e7du4waNYpdu3YRGRnJ/v37OXLkyBMJCb1798bIyIi5c+fqzZ87dy7r1q3j/PnzjBw5ktu3bzN06NAH7sfKyooRI0bw/vvvs2XLFs6dO8err75KWlpanrNLU6ZMYceOHZw5c4bBgwfj5ORE9+7di/29FYQEoLJEpYLOX4OxBVzZC6dWAeBub8HrT2kb0n25KZT0LOkgUQhR+g0fPpyFCxeyePFiGjZsSLt27ViyZInuDJCNjQ1ff/01zZo1o3nz5ly5coVNmzahVmu/umbMmMG2bdvw8PDAz8+v0HX8/PPPuLq68tRTT9GjRw9effVVbGxsMDc3f+h2Fy9exM/PT296/fXXsbS05O+//+bWrVs0b96cF198kWeeeUbX4NjS0pLz58/Tq1cvateuzWuvvcbIkSN5/fXXMTIy4ubNmwwcOJDatWvTp08fOnXqxKefflro9/cgxsbGjBo1iq+//lrvbNW0adOYNm0avr6+7Nu3jw0bNuDk5PTQfU2bNo1evXrxyiuv0KRJEy5dusTff/+dpy3VtGnTePvtt2natCmxsbH8+eef+d6JVhJUilwzySM5ORk7OzuSkpKwtbU1dDl57Z2h7SHa0glGHwWLStzNzOHpGbuISUpn7HO1GfV08dw1IIQoPdLT04mIiMDLy+uRX86i8K5du4aHhwfbt2/nmWeeMXQ5JebKlSt4eXlx4sSJYu/leteuXXTo0IHbt2/rzuAV1sP+HTzO97ecASqLAkaDUx1IS9AGIcDC1IjxnbS3TM7deVn6BhJCiAL6559/2LBhAxERERw4cIB+/frh6enJU089ZejSxBMkAagsMjaFLjO0j48uhmvaHkJf8HWnafVK3M3K4estMkyGEEIURFZWFh9++CH169enR48eODs7s2vXrmLvfFCULnIJLB+l/hJYrnVvwMkV4NYQXt0FRsacvJpIt7nalvjr3myFX7XC92UhhChd5BKYEHIJTAA8+xmY20PsaTiyAABfD3t6NakKwJS/5LZ4IYQQIj8SgMoya2cIvDeS7z9fQIq2Z9MPOtbBytSIE1GJ/BFioE7ShBBPjPzHRlRkxfX5lwBU1jUZrB0xPvMOHPoRABdbc97soO0Qa9rm86RlZhuwQCFEccntITkzM9PAlQhhOGlpaUDeHr4flwyFUdap1dB2LKwcAEcWQZt3wdyWYW28+O1IFFdv3WXersuMea7Oo/clhCjVjI2NsbS05MaNG5iYmOj6wxGiIlAUhbS0NOLj47G3t9cb360wJACVB3U6g1NtSLgIx5ZA67cwNzHio851eeOX48zfE07fFtWoYm9h6EqFEEWgUqmoXLkyERERREZGGrocIQzC3t4eNze3Iu9HAlB5oFZDq7dgwyg49AP4vw7GZgTVd8Pfy4HDEbdYsCecyS/UN3SlQogiMjU1pVatWnIZTFRIJiYmRT7zk0sCUHnRqA/s/ALuxGiHyGjyCiqVipEdanI4IpjVR68y5rna2JpLvxZClHVqtVpugxeiiOQCcnlhbAYt39Q+3v8daDQAtK3lRG1Xa1Izc1gZfNWABQohhBClhwSg8qTpYDCzg5thcGEToG0zMLS1dmDBJQeukJ2jMWCBQgghROkgAag8MbeF5sO0j/fPgnt9JXT3q4KDlSnXE+/y99k4w9UnhBBClBISgMqbliPAyAyuHYHIAwCYmxjxsn81ABbtCzdkdUIIIUSpIAGovLF2gcYvaR/vn6Wb/XJAdUyN1ByPSuRE1G3D1CaEEEKUEhKAyqNWo0GlhrCtEHcWABcbc7r6ugOwaF+EIasTQgghDE4CUHnk6A11X9A+3v+dbvawNtrG0JvPxHI98a4hKhNCCCFKBQlA5VWbd7Q/T6+BxCgA6rnbElDDkRyNws8HrhisNCGEEMLQDBqA9uzZQ9euXXF3d0elUrF+/fpHbrNr1y6aNGmCmZkZNWvWZMmSJXrLJ0+ejEql0pt8fHyezBsozdz9wKsdKDlwcK5udu5ZoOXBUaRmyCCpQgghKiaDBqDU1FR8fX2ZO3fuo1cGIiIi6NKlCx06dCAkJIR33nmH4cOH8/fff+utV79+fWJiYnTTvn37nkT5pV/uWaDjP0PaLQCe9nHBy8mKO+nZrDl2zXC1CSGEEAZk0KEwOnXqRKdOnQq8/rx58/Dy8mLGjBkA1K1bl3379vHtt98SFBSkW8/Y2LhYBkor82p0gMq+EHNSO0hq2zGo1SqGtvbkkz/Osnh/BK+0rI5arTJ0pUIIIUSJKlNtgA4ePEhgYKDevKCgIA4ePKg3LywsDHd3d2rUqMGAAQOIiop66H4zMjJITk7Wm8oFlQqav6p9fPxn3fAYvZpWxc7ChCs309hxPt6ABQohhBCGUaYCUGxsLK6urnrzXF1dSU5O5u5d7V1N/v7+LFmyhC1btvDjjz8SERFB27ZtuXPnzgP3O3XqVOzs7HSTh4fHE30fJapBTzC1gdsRcGUvAJamxvRvIR0jCiGEqLjKVAAqiE6dOtG7d28aNWpEUFAQmzZtIjExkVWrVj1wmwkTJpCUlKSbrl4tR4OGmlpBo97ax8eX6mYPalUdY7WKQ+G3OBudZKDihBBCCMMoUwHIzc2NuDj9sazi4uKwtbXFwsIi323s7e2pXbs2ly5deuB+zczMsLW11ZvKlaaDtT9D/4TUmwBUtrOgc8PKACw7GGmgwoQQQgjDKFMBKCAggB07dujN27ZtGwEBAQ/cJiUlhcuXL1O5cuUnXV7pVdkXKjeGnEw4uUI3O/cy2MZTMdzNzDFQcUIIIUTJM2gASklJISQkhJCQEEB7m3tISIiu0fKECRMYOHCgbv033niD8PBwxo0bx/nz5/nhhx9YtWoV7777rm6dsWPHsnv3bq5cucKBAwfo0aMHRkZG9O/fv0TfW6nTdJD257ElulHi/b0cqFrJgjsZ2Ww9F2u42oQQQogSZtAAdPToUfz8/PDz8wNgzJgx+Pn5MXHiRABiYmL07uDy8vJi48aNbNu2DV9fX2bMmMHChQv1boG/du0a/fv3p06dOvTp0wdHR0cOHTqEs7Nzyb650qbBi2BiBTfDIEp715xaraJXk6oA0ieQEEKICkWlKPdOBwid5ORk7OzsSEpKKl/tgf4YBSeWQaN+0HM+AFE303jqm52oVLD/g6dxt8+/LZUQQghR2j3O93eZagMkiii3MfS59XD3NgDVHC3x93JAUWDdiesGK00IIYQoSRKAKpIqTcG1AWSnw6n73QK82PT+ZTA5ISiEEKIikABUkahU0CRvY+jODStjaWpEREIqx6NuG64+IYQQooRIAKpoGvUBY3OIPwfXjgJgZWZMpwbabgKkMbQQQoiKQAJQRWNhD/V7aB8fW6KbnXsZ7K+TMaRnSZ9AQgghyjcJQBVR7mWws2shXTvw67/7BPr7rPQJJIQQonyTAFQRVWsJTnUgKw1OrwakTyAhhBAViwSgikil0u8Z+p7cALTvUgIxSXcNUJgQQghRMiQAVVS+/cHIFGJPQfQJQL9PoLXHpU8gIYQQ5ZcEoIrK0gHqvqB9nE9j6N+lTyAhhBDlmASgiiy3Z+jTayAjBbjfJ1B4QirHoxINVpoQQgjxJEkAqsg824CDN2SmwJnfAekTSAghRMUgAagiU6nunwXKt0+gaOkTSAghRLkkAaiia/wSqE0g+jjEnASkTyAhhBDlnwSgis7KCep21T4+thTQ9gnUU/oEEkIIUY5JABL3L4OdWgWZqQD09KsCwP5LCcTfSTdQYUIIIcSTIQFIgGdbcKgBmXfgzFrtLCcrGnvYo1G044MJIYQQ5YkEIAFq9f3xwf7VGLp7Y3cA/giRThGFEEKULxKAhFbjAdrG0NePQuxpAJ73dcdIreLktSTCb6QYuEAhhBCi+EgAElrWzuDTRfv4XmNoJ2sz2tZyAmB9SLShKhNCCCGKnQQgcZ9eY+g0ALo31jaG/iPkugyNIYQQotyQACTu82oHlTwhIwnOrQfg2XquWJgYEXkzjRNXEw1ZnRBCCFFsJACJ+/JpDG1lZkxQfVcA/jghjaGFEEKUDxKAhL7GA0BtDFcPQ9w5ALrd6xPor1MxZOVoDFmdEEIIUSwkAAl9Nq5Qp7P28XFtY+i2NZ1wtDLlZmom+8ISDFicEEIIUTwkAIm8chtDn1wBWXcxNlLT1VfbJ9B66RNICCFEOSABSORVowPYV4P0JDj3BwDd7nWKuPVsHKkZ2YasTgghhCgyCUAiL7UamgzUPj66GIDGHvZUd7TkblYOW8/JCPFCCCHKNglAIn+NXwaVEVw9BPHnUalUuj6B1p+QThGFEEKUbRKARP5sK0PtjtrHx38GoPu9u8H2XUrgxp0MQ1UmhBBCFJkEIPFgusbQyyErHS8nK3yr2pGjUfjrlJwFEkIIUXZJABIPVvMZsK0Kd2/D+b+A+2eBZGwwIYQQZZkEIPFgaiNo8or28b2eoZ9vdG+E+KuJRCSkGq42IYQQoggkAImH83sZVGq4shcSLuFsY0brmvdGiJehMYQQQpRREoDEw9lVhZrPah/f6xm6h5+2TyAZIV4IIURZZdAAtGfPHrp27Yq7uzsqlYr169c/cptdu3bRpEkTzMzMqFmzJkuWLMmzzty5c/H09MTc3Bx/f3+Cg4OLv/iKJLcxdMhyyM7kuXpumJuouXIzjVPXkgxamhBCCFEYBg1Aqamp+Pr6Mnfu3AKtHxERQZcuXejQoQMhISG88847DB8+nL///lu3zsqVKxkzZgyTJk3i+PHj+Pr6EhQURHx8/JN6G+VfrefApjKkJcCFjViZGfNsPTdAhsYQQghRNqmUUnINQ6VSsW7dOrp37/7AdT744AM2btzImTNndPP69etHYmIiW7ZsAcDf35/mzZszZ84cADQaDR4eHowePZrx48cXqJbk5GTs7OxISkrC1ta28G+qPPnnc9jzDdRoDwP/YEdoHMOWHsXJ2oxDE57G2EiupgohhDCsx/n+LlPfWgcPHiQwMFBvXlBQEAcPHgQgMzOTY8eO6a2jVqsJDAzUrSMKye8VQAXhu+BWBG1rOWNvaUJCSgYHw28aujohhBDisZSpABQbG4urq6vePFdXV5KTk7l79y4JCQnk5OTku05s7IPHr8rIyCA5OVlvEv9RqTp4P619fPxnTI3VdGlYGZChMYQQQpQ9ZSoAPSlTp07Fzs5ON3l4eBi6pNKp6SDtz5BfISeLbvfGBvv7bCzpWTkGLEwIIYR4PGUqALm5uREXF6c3Ly4uDltbWywsLHBycsLIyCjfddzc3B643wkTJpCUlKSbrl69+kTqL/PqdAYrF0iJg4tbaFa9ElXsLUjJyGZHqDQyF0IIUXaUqQAUEBDAjh079OZt27aNgIAAAExNTWnatKneOhqNhh07dujWyY+ZmRm2trZ6k8iHkQk0fkn7+NgS1GoVLzS+3yeQEEIIUVYYNAClpKQQEhJCSEgIoL3NPSQkhKioKEB7ZmbgwIG69d944w3Cw8MZN24c58+f54cffmDVqlW8++67unXGjBnDggULWLp0KaGhoYwYMYLU1FSGDBlSou+t3Gpy7/dxaQckRtHtXgDadeEGSWlZBixMCCGEKDiDBqCjR4/i5+eHn58foA0vfn5+TJw4EYCYmBhdGALw8vJi48aNbNu2DV9fX2bMmMHChQsJCgrSrdO3b1+mT5/OxIkTady4MSEhIWzZsiVPw2hRSI7e4NUOUOD4MnzcbPFxsyEzR8PmMzGGrk4IIYQokFLTD1BpIv0APcKZ32HNULB2g3dO88O+KL7ecoGWNRz47bUHX2oUQgghnqRy2w+QKCV8umrDT0osnF3LC77ay2CHI24Rk3TXwMUJIYQQjyYBSDw+Y1No8ar28cG5VLW3oIWnA4oCf56UPoGEEEKUfhKAROE0GwrGFhB7CiL36+4Gk04RhRBClAUSgEThWDqAbz/t44M/0KVhZYzVKs7FJBMWd8ewtQkhhBCPIAFIFF7LEdqfFzZRKf0q7Wo7A/BHiJwFEkIIUbpJABKF51wHagYCChyeTzc/7dAYf5y8jtxcKIQQojSTACSKpuWb2p8nfuFZL3MsTY24eusux6MSDVqWEEII8TASgETReD8NznUhKxWL078QVF875toGGRpDCCFEKSYBSBSNSnW/LVDwT3Rv5ALAX6diyMzWGLAwIYQQ4sEkAImia9QHLB0h6Sptsg7iYmPGzdRMtofGGboyIYQQIl8SgETRmVhAs2EAGAXPo08zDwCWH4562FZCCCGEwUgAEsWj+XAwMoVrwbzicQOVCvZdSiDyZqqhKxNCCCHykAAkioeNKzR4EQDXs4t4qpa2T6AVwVcNWZUQQgiRLwlAovgE3Lsl/twfDGlgDMCaY1elMbQQQohSRwKQKD5uDcGzLSg5tL29FhcbMxJSMtl2ThpDCyGEKF0kAIniFTASAKNjS3i1gfbjtSJYGkMLIYQoXSQAieJVKwiqtoDMOwyM/hRTVbY0hhZCCFHqSAASxUuthhf/B+b2mMWFMNtpHSCNoYUQQpQuEoBE8bP3gB7zAAi6s44g9RFpDC2EEKJUkQAknow6naDVaACmm87HPPWaNIYWQghRakgAEk/OM5OganNsSGOOyfesOnzZ0BUJIYQQgAQg8SQZmcCLi8kxs6exOpx2kbO5kiCNoYUQQhieBCDxZNl7YNRT2x5oqPEWjv+9zMAFCSGEEBKAREmo04mI2kMBeDbsUzJvRBi4ICGEEBWdBCBRIqq+OI3TqtrYkEba8lcgJ8vQJQkhhKjAJACJEmFiasbBJtNJVKywv30aDs41dElCCCEqMAlAosR0btOcz7NfBkCzcyrcjjRwRUIIISoqCUCixFStZMnden05pKmLOicdZdP7oCiGLksIIUQFJAFIlKjxneoyWTOcTMUIVdjfEPqnoUsSQghRAUkAEiXKw8GSwLZtmZfTFQBl8zjIuGPgqoQQQlQ0EoBEiRvR3pvV5n24onFFdScG/vnC0CUJIYSoYCQAiRJnZWbMO518+SR7CABK8HyIDjFsUUIIISoUCUDCIHr4VSG5ylNsyAlApWjgr3dAk2PosoQQQlQQEoCEQajVKiZ1rcdnWa+QrFhC9Ak4ssjQZQkhhKggJAAJg2lSrRKtG9fj6+y+ACg7pkByjIGrEkIIUREYG7oAUbF90MmHZ84+Ry/NXvwyL8HG96DpYLh7O++kaCBwEthXM3TZQgghyrhScQZo7ty5eHp6Ym5ujr+/P8HBwQ9cNysriylTpuDt7Y25uTm+vr5s2bJFb53JkyejUqn0Jh8fnyf9NkQhVLaz4PV2tfgwaxjZqOHCRljeG9a9Bls+gN3TIHg+nF4FZ9bApnGGLlkIIUQ5YPAzQCtXrmTMmDHMmzcPf39/Zs2aRVBQEBcuXMDFxSXP+h9//DG//PILCxYswMfHh7///psePXpw4MAB/Pz8dOvVr1+f7du3654bGxv8rYoHeO2pGqw8UodpKf0ZZbMH+0oOYFHpX5MDmFjAzi/h4maIOgTVWhq6bCGEEGWYSlEMOxaBv78/zZs3Z86cOQBoNBo8PDwYPXo048ePz7O+u7s7H330ESNHjtTN69WrFxYWFvzyyy+A9gzQ+vXrCQkJKVRNycnJ2NnZkZSUhK2tbaH2IR7PnyejGb3iBOYmara8/RSeTlZ5V9rwFhxfCtUCYMhmUKlKvlAhhBCl1uN8fxv0ElhmZibHjh0jMDBQN0+tVhMYGMjBgwfz3SYjIwNzc3O9eRYWFuzbt09vXlhYGO7u7tSoUYMBAwYQFRVV/G9AFJvnG1WmhZcD6VkaXvnfYeKS0/Ou1H48GJtD1EEI21byRQohhCg3DBqAEhISyMnJwdXVVW++q6srsbGx+W4TFBTEzJkzCQsLQ6PRsG3bNtauXUtMzP27h/z9/VmyZAlbtmzhxx9/JCIigrZt23LnTv5DLmRkZJCcnKw3iZKlUqmY85If1R0tuXrrLgMXBZOYlqm/kq07+L+ufbzjU9BoSr5QIYQQ5UKpaAT9OL777jtq1aqFj48PpqamjBo1iiFDhqBW338rnTp1onfv3jRq1IigoCA2bdpEYmIiq1atynefU6dOxc7OTjd5eHiU1NsR/+JiY84vw/xxsTHjQtwdhi45Qlpmtv5Krd8BMzuIO6NtFC2EEEIUgkEDkJOTE0ZGRsTFxenNj4uLw83NLd9tnJ2dWb9+PampqURGRnL+/Hmsra2pUaPGA1/H3t6e2rVrc+nSpXyXT5gwgaSkJN109erVwr8pUSQeDpYsG+aPnYUJx6MSeX3ZMTKz/3Wmx9IB2rytffzP55Cdmf+OhBBCiIcwaAAyNTWladOm7NixQzdPo9GwY8cOAgICHrqtubk5VapUITs7m99//51u3bo9cN2UlBQuX75M5cqV811uZmaGra2t3iQMp46bDf8b3BwLEyP2hiXw7qoQcjT/aqvvPwKs3SAxEo4tMVidQgghyi6DXwIbM2YMCxYsYOnSpYSGhjJixAhSU1MZMkQ7UObAgQOZMGGCbv3Dhw+zdu1awsPD2bt3Lx07dkSj0TBu3P3+YcaOHcvu3bu5cuUKBw4coEePHhgZGdG/f/8Sf3+icJpWr8T8V5piYqRi46kYPvnjDLobFk0tod293/eeryEjxXCFCiGEKJMMHoD69u3L9OnTmThxIo0bNyYkJIQtW7boGkZHRUXpNXBOT0/n448/pl69evTo0YMqVaqwb98+7O3tdetcu3aN/v37U6dOHfr06YOjoyOHDh3C2dm5pN+eKIKnajvzbd/GqFSw/HAU07deuL+wyUBwqAGpN+DQj4YrUgghRJlk8H6ASiPpB6h0WX44ig/XnQZgQicfXm/nrV1w5ndYMxTMbOGtELByNFyRQgghDK7M9AMkREG85F+NcR3rADB183mWHbyiXVCvB7g1goxk2DfTcAUKIYQocyQAiTLhzfY1GdlBe+bnkz/OsubYNVCrIXCydoXgBZAod+8JIYQoGAlAoswY+1wdBrfyBGDcmpP8dSoavJ8Gz7aQkwEbRstt8UIIIQpEApAoM1QqFZO61qNfcw80CrzzWwjbQ+Oh41QwsYTwnbB+hPQQLYQQ4pEkAIkyRaVS8UWPhnRr7E62RuHNX4+z705l6LsM1Mba3qG3jAdp2y+EEOIhChWArl69yrVr13TPg4ODeeedd/jpp5+KrTAhHsRIrWJ6b1+eq+dKZo6GV38+yhHjJtBjvnaF4Pmwd7phixRCCFGqFSoAvfTSS+zcuROA2NhYnn32WYKDg/noo4+YMmVKsRYoRH5MjNTMfsmPdrWduZuVw5DFRzhd6Vno+JV2hX8+h6OLDVukEEKIUqtQAejMmTO0aNECgFWrVtGgQQMOHDjAr7/+ypIlS4qzPiEeyMzYiHkvN8Xfy4GUjGxG/HqMJN9h8NT72hU2joFzfxi2SCGEEKVSoQJQVlYWZmZmAGzfvp0XXngBAB8fH71em4V40ixMjVgwqBkeDhZcu32XD9eeRmn/ITQdAooGfh8OEXsMXaYQQohSplABqH79+sybN4+9e/eybds2OnbsCEB0dDSOjtIbryhZtuYmzO7fBGO1io2nY/jt6DXoMgPqvgA5mbDiJYgOMXSZQgghSpFCBaCvvvqK+fPn0759e/r374+vry8AGzZs0F0aE6IkNfaw5/0gbW/Rn/55los30qDXQvB6CjLvwKqBkJlq4CqFEEKUFoUeCywnJ4fk5GQqVaqkm3flyhUsLS1xcXEptgINQcYCK5s0GoXBS46w5+INartas2FUG8xzUuHHVpB0FVq9Bc99ZugyhRBCPCFPfCywu3fvkpGRoQs/kZGRzJo1iwsXLpT58CPKLrVaxYzevjhZm3ExLoXP/joH5rbay2EAB+dCzEnDFimEEKJUKFQA6tatGz///DMAiYmJ+Pv7M2PGDLp3786PP/5YrAUK8Ticbcz4tq/2kuyvh6PYfDoGagdB/Z6g5MCGtyAn28BVCiGEMLRCBaDjx4/Ttm1bANasWYOrqyuRkZH8/PPPfP/998VaoBCPq20tZ95opx049YPfT3Htdhp0nAbmdhATou0oUQghRIVWqACUlpaGjY0NAFu3bqVnz56o1WpatmxJZGRksRYoRGG891xtGnvYk5yezVsrTpBl6QzP3mv/88/ncFs+p0IIUZEVKgDVrFmT9evXc/XqVf7++2+ee+45AOLj46XRsCgVTIzUzO7vh42ZMcejEpm57SI0GQjV20BWmraTRBkvTAghKqxCBaCJEycyduxYPD09adGiBQEBAYD2bJCfn1+xFihEYXk4WDK1V0MAftx1mS1nY6HrLDAyhUvb4czvhi1QCCGEwRT6NvjY2FhiYmLw9fVFrdbmqODgYGxtbfHx8SnWIkua3AZfvkz58xz/2x+BlakR60e2ptb5H2HnF2DlDCODwdLB0CUKIYQoBk/8NngANzc3/Pz8iI6O1o0M36JFizIffkT5M6GzDy1rOJCamcNry46R3GwkOPtA6g3Y9omhyxNCCGEAhQpAGo2GKVOmYGdnR/Xq1alevTr29vZ89tlnaDSa4q5RiCIxMVIz96UmuNuZE5GQyrurz6F5/jvtwhO/yFhhQghRARUqAH300UfMmTOHadOmceLECU6cOMGXX37J7Nmz+eQT+R+1KH0crc2Y/0ozzIzV7Dgfz6yLDtBsmHbhhrcg445hCxRCCFGiCtUGyN3dnXnz5ulGgc/1xx9/8Oabb3L9+vViK9AQpA1Q+bX2+DXGrNL2Br2oX22e+ac7JF8D3/7QY55hixNCCFEkT7wN0K1bt/Jt6+Pj48OtW7cKs0shSkTPJlUZ3MoTgLfXhXPt6e9ApYaTK+DkSsMWJ4QQosQUKgD5+voyZ86cPPPnzJlDo0aNilyUEE/SR13q4u/lQEpGNgN3GJPe+n3tgo1j4Fa4YYsTQghRIgp1CWz37t106dKFatWq6foAOnjwIFevXmXTpk26YTLKKrkEVv4lpGTQdfY+YpLSCarrxDzNZFSRB8DdD4ZuBWNTQ5cohBDiMT3xS2Dt2rXj4sWL9OjRg8TERBITE+nZsydnz55l2bJlhSpaiJLkZG3G/FeaYmqk5u/QBHbX/wLM7SH6BPzzmaHLE0II8YQVuiPE/Jw8eZImTZqQk5NTXLs0CDkDVHHM3HqB7/+5RGU7c3Y+n4L57wO1C17+HWoGGrY4IYQQj6VEOkIUojx4s0NNqjlYEpOUzsyrtaH5cO2CdW9ASrxhixNCCPHESAASFZq5iRGfdqsPwKJ9EVzwHQcu9bW9RK97A6RjTyGEKJckAIkKr0MdFzrWdyNHo/DRn5fR9FoExhZweQcczHu3oxBCiLLP+HFW7tmz50OXJyYmFqUWIQxmYtd67Am7wdHI26y56kGfjlPhr3dgx6dQtRlUb2XoEoUQQhSjxzoDZGdn99CpevXqDBw48EnVKsQT425vwbuBtQGYuimU2z4vQYNeoMmGVQMh6ZqBKxRCCFGcivUusPJC7gKrmLJyNDz//T4uxN2hfwsPpj7vDYuCIO40VG4MQ7eAiYWhyxRCCPEAcheYEIVgYqTm8x4NAFgRfJVjMZnQ71ewcICYEPjzbZD/LwghRLkgAUiIf2nu6UDvplUB+Hj9GbJtPaDPUlAZwamVcOgHA1cohBCiOJSKADR37lw8PT0xNzfH39+f4ODgB66blZXFlClT8Pb2xtzcHF9fX7Zs2VKkfQrxbxM618Xe0oTQmGSWHowEr6cg6Evtwq0fw+Wdhi1QCCFEkRk8AK1cuZIxY8YwadIkjh8/jq+vL0FBQcTH598J3ccff8z8+fOZPXs2586d44033qBHjx6cOHGi0PsU4t8crEz5oKMPoO0pOjYpHfxfh8YDQNHAmiFwK8LAVQohhCgKgzeC9vf3p3nz5rrR5TUaDR4eHowePZrx48fnWd/d3Z2PPvqIkSNH6ub16tULCwsLfvnll0Lt87+kEbTQaBRenHeA41GJdG7oxg8DmkJWOizpDNePaTtLHLYVzKwNXaoQQoh7ykwj6MzMTI4dO0Zg4P0xl9RqNYGBgRw8eDDfbTIyMjA3N9ebZ2Fhwb59+4q0z+TkZL1JVGxqtYrPuzfESK1i0+lYdp6PBxNz6PsLWLtC/FlYP0IaRQshRBll0ACUkJBATk4Orq6uevNdXV2JjY3Nd5ugoCBmzpxJWFgYGo2Gbdu2sXbtWmJiYgq9z6lTp+r1Z+Th4VEM706UdfXcbRna2hOAiRvOcDczB2zdoc8yUJtA6AbY841hixRCCFEoBm8D9Li+++47atWqhY+PD6ampowaNYohQ4agVhf+rUyYMIGkpCTddPXq1WKsWJRl7wTWxt3OnKu37jJnZ5h2ZjV/eH6m9vHOL+DcH4YrUAghRKEYNAA5OTlhZGREXFyc3vy4uDjc3Nzy3cbZ2Zn169eTmppKZGQk58+fx9ramho1ahR6n2ZmZtja2upNQgBYmRkz+QXtYKk/7QknLO6OdkGTgdDyTe3jdW9AzEkDVSiEEKIwDBqATE1Nadq0KTt27NDN02g07Nixg4CAgIdua25uTpUqVcjOzub333+nW7duRd6nEPl5rr4bgXVdycpR+Gj9GXT3DTz7GXg/A1lpsOIlSJG7DIUQoqww+CWwMWPGsGDBApYuXUpoaCgjRowgNTWVIUOGADBw4EAmTJigW//w4cOsXbuW8PBw9u7dS8eOHdFoNIwbN67A+xTicU1+oR4WJkYER9zi9+PXtTONjOHF/4FjLUi+Br8NgOwMwxYqhBCiQB5rNPgnoW/fvty4cYOJEycSGxtL48aN2bJli64Rc1RUlF77nvT0dD7++GPCw8Oxtramc+fOLFu2DHt7+wLvU4jHVbWSJe8E1mLq5vN8uSmUZ3xcqGRlChb20P83WPg0XAvWDpfR/UdQqQxdshBCiIcweD9ApZH0AyTy8+/BUvs192Bar0b3F17+B355EZQc7aWx1m8ZrlAhhKigykw/QEKUJSZGar64N1jqb0eucvTKrfsLvZ+GjtO0j7dNhIt/G6BCIYQQBSUBSIjH0MzTgX7Ntf1EfbTuDBnZOfcXtngVmg4BFFgzDC5tN0yRQgghHkkCkBCP6YOOPjhYmXIh7g5vrwghO0ejXaBSQedvwLMtZN6BX3rB2tch7dbDdyiEEKLESQAS4jFVsjLl+35+mBqp2XI2lvfXnEKjudeUzshE2yja/w1ABad+gznN4fQaGTZDCCFKEQlAQhRCm1pOzB3QBGO1inUnruv3D2RmDZ2+gmHbwLkupCXA78NgeR9IlF7GhRCiNJAAJEQhPVvPlW/7NkatghXBUXz2Vyh6N1V6NIfX90CHj8DIFMK2wg8t4fBPoMl58I6FEEI8cRKAhCiCrr7ufHXvdvj/7Y9gxtaL+isYm0K7cfD6XvDwh8wU2Pw+fN8YDs6F9KSSL1oIIYQEICGKqnczDz7rph0vbM7OS8zdeSnvSi4+MGQLdJ4OFpUgMQr+/hBm1oPNH8DNyyVctRBCVGzSEWI+pCNEURjzd19m6ubzAEx8vh5D23jlv2JmGpxaCYd+hIQL92aqoE4n8H8d7Dwg7Wb+k7EFWLuCtcu9n87an1Yu2rNNQghRgT3O97cEoHxIABKF9e22i3y3IwyAz7s34OWW1R+8sqJA+E5tEArbWvQXd/aB5sPBtx+Y2RR9f0IIUcZIACoiCUCisBRFYdrm88zfEw7AZ93q80qA56M3TAiDw/Pg1Crtc0sHsHQEi3s/LR3BspJ2sNWUOO3I8//+qcm+vy8zW2g8QNsxo6N38b9JIYQopSQAFZEEIFEUiqIwdfN5froXgj59oT6DWnk+uRfUaLSXx86ug+Cf4GbY/WU1n9X2SeT9NKilyZ8QonyTAFREEoBEUSmKwrQt55m/WxuCJnWtx5DWD2gTVJw0Ggj/R3urfdhW4N4/70qeUKcL1A6C6q20HTYKIUQ5IwGoiCQAieKgKApf/32BH3dp7/D6uEtdhretUXIF3LwMRxbCiV8gI/n+fDNbqPkM1O6oPUNk5VhyNQkhxBMkAaiIJACJ4qIoCjO2XmTOvVvjSzwEAWSkwOUd2hHqL/6t7Zk6l0oNVZpBNX+o2lz72K5KydYnhBDFRAJQEUkAEsVJURS+3XaR7//RhqAPO/vw2lMGapys0cD1Y3BxizYMxZ3Ou46NO1Rtqg1EVVuARwtQG5V8rUII8ZgkABWRBCDxJPz7Fvl3Amvx9jO1UKlUhi0q8SpE7IHrR+HaEYg7B8p/humwcYfGL2knuatMCFGKSQAqIglA4kn5bnsY327XDpfRq0lVpvZsiKlxKbo7KzMVokO0Yej6UbiyD+7evr+8ehvwGwD1uoGplcHKFEKI/EgAKiIJQOJJWn44ik/+OEOORiGghiPzXmmKnUUpvSsrOwMubNY2pL68AxSNdr6pDdTvrh3fzL6adrKrKneXCSEMSgJQEUkAEk/argvxjPz1OKmZOdR0sWbx4OZ4OFgauqyHS7oOJ1dow9DtiLzLVWrt5bLcQFS1Gfh0AVv3kq9VCFEhSQAqIglAoiSci05m6JIjxCan42RtysJBzWnsYW/osh5NUSDygLbjxVvh2oFdE6MgJyP/9as01QYhn+fBuU7J1iqEqFAkABWRBCBRUmKT0hm65AjnYpIxN1HzXT8/guq7Gbqsx6fRQOqNe2EoEm5FwKVtcDUYXWeMAI61tGHItz+4+BisXCFE+SQBqIgkAImSlJKRzajlx9l14QYqFXzQ0YfXn6ph+DvEisOdOLiwCc5vhPBdoMnSzlcbQ4ePoPXbcou9EKLYSAAqIglAoqRl52iYtOEsvx6OAuC5eq5809u39DaOLoz0ZO1ZoZAV2p+gvausxzyw9zBsbUKIckECUBFJABKGoCgKvx6OYsqf58jM0VDNwZIfBjShQRU7Q5dWvBQFQpbD5nGQmQJmdtD1W2jQy9CVCSHKuMf5/i5FHZAIUbGpVCpeblmdNSMCqFrJgqhbafT88QC/BUdRrv6folJp+xJ6fY926I2MJFgzFNa9oT1LJIQQJUACkBClTKOq9vw1ug3P+LiQma1h/NrTjF19iruZOY/euCxx9IahW+Cpcdpb6E+ugHltIOqwoSsTQlQAEoCEKIXsLU1ZMLAZ4zrWQa2C349fo/vc/VyKTzF0acXLyASe/giGbNb2HZQYCYs7wYHZ2ktlQgjxhEgAEqKUUqtVvNm+Jr8Ob4mTtRkX4u7Q+bu9zNx6gfSscnY2qFpLeGMfNOytHYts68ewaqBcEhNCPDESgIQo5QK8Hdn0Vhuequ1MZo6G7/+5RODM3Ww/F2fo0oqXuR30XACdp4PaBEI3wIKnIT7U0JUJIcohuQssH3IXmCiNFEVhy5lYpvx1jpikdACe9nFhUtd6VHcsZwOTXj0CqwdB8nUwsYIXvoeGLxq6KiFEKSe3wReRBCBRmqVmZDNn5yUW7g0nK0fB1FjNiHbejGjvjblJOepUMDVBe3dYxG7tc/834NnPwNj08fel0YBaTngLUd5JACoiCUCiLLgUn8KkDWfYf+kmAO525rzezpu+zT3KTxDS5MDOL2DvDO1zl3rau8eMTLWXyYxMtI+NTLR3kqUnwd1EuHsb0hO1j9MTIScLmg6GoC/A2Mxgb0cI8WRJACoiCUCirFAUhY2nY/j8r1Bik7WXxZysTRnaxouXW1bH1ryc9CR9fpO2n6CMpKLtx70J9PlZep4WopySAFREEoBEWZOelcPqY9eYt+sy1xPvAmBjbszgVp4Mae2Fg1UhLhuVNknX4fI/kJ0OmmzIydSe2cl9rGjAzBYsKoGFPZjb3/954/y9jhYTwcIBei2AmoEGfTtCiOInAaiIJACJsiorR8OfJ6P5YddlXZ9BFiZG9GvhwdDWXng4WBq4QgO6Ham9tT4mBFBB+wnw1PvSNkiIcqTMDYUxd+5cPD09MTc3x9/fn+Dg4IeuP2vWLOrUqYOFhQUeHh68++67pKen65ZPnjwZlUqlN/n4+DzptyGEwZkYqenZpCpb33mKeS83oWEVO+5m5bB4/xXafbOTkcuPcyLqtqHLNIxK1WHo39q2QCiw60tY3gfSbhm6MiGEARg8AK1cuZIxY8YwadIkjh8/jq+vL0FBQcTHx+e7/vLlyxk/fjyTJk0iNDSURYsWsXLlSj788EO99erXr09MTIxu2rdvX0m8HSFKBbVaRccGldkwqjU/D21B21pOaBTYeCqGHj8c4MUfD7DlTCw5mgp2AtjEHLp+B91+AGNz7aj089vBlf2GrkwIUcIMfgnM39+f5s2bM2fOHAA0Gg0eHh6MHj2a8ePH51l/1KhRhIaGsmPHDt289957j8OHD+tCzuTJk1m/fj0hISGFqkkugYnyKDQmmYV7I9hw8jpZOdp/9tUdLRnSypMeflWxsywnDaYLKuaU9pLY7Qjt8xodoP14ba/UQogyqcxcAsvMzOTYsWMEBt5vjKhWqwkMDOTgwYP5btOqVSuOHTumu0wWHh7Opk2b6Ny5s956YWFhuLu7U6NGDQYMGEBUVNQD68jIyCA5OVlvEqK8qVvZlhl9fNn3wdO82d4bOwsTIm+mMfnPczT/YjsjfjnG1rOxZGZrDF1qyajcCF7bpb0kpjaG8J3wvyD4uTtEHTJwcUKIJ82gZ4Cio6OpUqUKBw4cICAgQDd/3Lhx7N69m8OH8x8V+vvvv2fs2LEoikJ2djZvvPEGP/74o2755s2bSUlJoU6dOsTExPDpp59y/fp1zpw5g42NTZ79TZ48mU8//TTPfDkDJMqztMxsVh+9xorgKM7H3tHNr2RpQldfd3o2qYpvVTtUKpUBqywhtyO1fQ2F/Kq9qwzkjJAQZVCZuQusMAFo165d9OvXj88//xx/f38uXbrE22+/zauvvsonn3yS7+skJiZSvXp1Zs6cybBhw/Isz8jIICMjQ/c8OTkZDw8PCUCiwjgXncza49f442Q0N+7c/7dQw8mKvs09eLFpVRytK0AHgvkFIdcGULU5VG0GVZqBU225c0yIUqrMBKDMzEwsLS1Zs2YN3bt3180fNGgQiYmJ/PHHH3m2adu2LS1btuSbb77Rzfvll1947bXXSElJQf2AP0zNmzcnMDCQqVOnPrIuaQMkKqrsHA37L99k7fFr/H02lvQs7eUwUyM1HRu4McC/Gi28HMr/WaH8glAuM1tw99MGokqe2t6qlRztT02Odn0lR9sfUb1u2p9CiBLxON/fxiVUU75MTU1p2rQpO3bs0AUgjUbDjh07GDVqVL7bpKWl5Qk5Rkbabv8flOVSUlK4fPkyr7zySvEVL0Q5ZGykpl1tZ9rVdiYlI5u/TkazPDiKU9eS2HAymg0no6npYs0A/2r0LM8NpytV1w7A2uEjiDoI14/CtWMQfQIykrXjk+WOUfYwm8dDg17QfBhUafLk6xZCFJjB7wJbuXIlgwYNYv78+bRo0YJZs2axatUqzp8/j6urKwMHDqRKlSq6MzeTJ09m5syZ/PTTT7pLYCNGjKBp06asXLkSgLFjx9K1a1eqV69OdHQ0kyZNIiQkhHPnzuHs7PzImuQMkBD6Tl9LYnlwJH+ERJOWmQOAmbGaHn5VGNrGi9quedvWlUs52RB/7n4gSo3XNqBWG2l/qozuP48Ogfiz97d194Nmw7SByLSQHVKmJmjPQBVmQFghKoAycwks15w5c/jmm2+IjY2lcePGfP/99/j7+wPQvn17PD09WbJkCQDZ2dl88cUXLFu2jOvXr+Ps7EzXrl354osvsLe3B6Bfv37s2bOHmzdv4uzsTJs2bfjiiy/w9vYuUD0SgITI3530LNaHRPProUi9htNP1XZmWBsvnqrlVP4vjxWUosDVw3BkEZxbrx2uA8DcDnxfAp8u4OH/6DCTdgvO/A4nV8D1Y9r+i9ybQDV/qBagbZ9k6ZB3u9SbcDMMEi5CQpj20py1C1i7an9a3Xts5aQNbEKUA2UuAJU2EoCEeDhFUTgaeZtFeyPYei6W3P4Ua7taM7S1F939qpSfEemLQ2oCnPgFjv4PEiPvzze1gRrtoOYz2rHJ7Ktp52dnajtpPLkCLmwBTdbD9+/sAx4ttOOhJVzShp67BezhWqXWjo9m5QSWTmDleO/nveeVPLX1SUgSZYAEoCKSACREwV29lcbi/VdYeSSK1HuXxxysTBncypPBrT3Lz4j0xUGjgcs74NQq7cCuaQn6y53qaPsnuvwPpN28P9+tEfj2h4Yvwt1EuHoIog5rf9689ODXs/MAx5raO9eMzSD1BqTEQcq9n6k3gAJ8BVRpCl1mgnvjQrxpIUqOBKAikgAkxONLTs9i1ZGrLN5/RTcivZ2FCcPbeDG4tSc2EoT0aTTagVkv7YBL2+FasPYMTi5rV2jURxt8XOs/eD+pCdpLbdeOakOOUy1wrAWO3mBq9fAacrK1QSv1hjaMpd6b0v718/IuyLyjPVPU4jVtw3Bz+bsoSicJQEUkAUiIwsvO0bDpTCzf7wjTjUhvZ2HCq229GNzaC2szg958WnrdvQ3huyD2NFRrBTXag1EpOFbJMfD3h3B2rfa5tRt0nAr1e4C09xKljASgIpIAJETR5WgU/joVzfc7wrh8IxUAe0sTXm1bg0GtPCUIlTWX/4GN78GtcO1z72egy3RwqGHYuoT4FwlARSQBSIjikxuEvtsRRvi9IORgZcqb7b15uWV1aSxdlmSlw/5ZsHcm5GSA2kTbANvB695U4/5k467dJiMZ0pPu/0xPhow72g4iHb21Db+N5PKoKB4SgIpIApAQxS9Ho/DnSW0QikjQBiE3W3NGP1OTPs08MDGS4SXKjJuXYdNY7VmhB1Eb5+1FOz8qI23Hkw41wMFbG4qqNNVOcolNPCYJQEUkAUiIJyc7R8Pa49f5bkeYrrF0NQdL3n22Fi/4VsFILV96ZYKiaC+H3bys/fnvKTFSP/wYm2s7cDS30zagNrXWNrK+FQ7Zd/Pfv311baeRDXppG4FLGBIFIAGoiCQACfHkZWTnsOJwFHN2XiIhRdtJYG1Xa8Y8W5ug+m7SoWJZlpMNd2K0wcfcVnt3Wn40GkiJvRei7gWphDAI3w1ZqffXc6qj7QKgQS/tGSIhHkACUBFJABKi5KRlZrPkwBXm7bpMcrr2rEF9d1veDazNM3VdJAhVRJlpcHGLtgfssK33e9EGcKmnvTzm7qcdX82lvgwNInQkABWRBCAhSl7S3SwW7g3nf/sidB0qNqpqx7vP1qZ9bWcJQhVVehKE/qUNQ+G7QMnRX25kCq4NtIHIowXU7ggW9oaotGzIzoCUeO04dinx9zrGjNdeYrSpDDZu935W1jZUL2P/7iQAFZEEICEM51ZqJgv2hrNk/xXuZmm/7Pyq2fNuYG3aylhjFVtqAkQdgujjEH0Crh+H9ET9dYxModZz2ktmtTuCicWD95dxR7uPuDPaNksqNaDS/lSptV/+KrV2OJDKjcH60YNplwqKAknXtAP3xp3VTvGhkHw97/F6GCMzsHHVNlD3flo7XItLvVIdiiQAFZEEICEMLyElg5/2hPPzwSukZ2l7SG5WvRIjn64pZ4SElqLA7SvaQHT9uLZX7Ruh95eb2kDd56HBi+D1lLaN0bUj96aj2nX/3fv2o9hW0Qahyr7aYUEqN9aeJcm+q+0iICsNstMh6672Jyows9GfHveW/7u374e96BNw47z2DjtjczCxBJPcnxbabgluhWvDTkbSg/epNrk/KK61C1g5Awrcib03xegPxfJvNpXvjV33rLazzlJ2tk0CUBFJABKi9Ii/k868XeH8cjiSzGztl1W9yraMaO9N54aV5a4xoS/uLJxeDad/h6So+/NV6vzDjl01bZgxtdYuVzSAcv9xThbcuHBvzLVi+Lo0Nr8fhiwqgbm9NkTk/rSopA12MSHa0HM7onCvozbWjgHnUg9c62nbSlXy1Aaeglzays7QXh5Ljrk3ZMt2iNirf9eeykh76dG5zr1+oO51Y+BQQ/v+DEACUBFJABKi9IlLTmfBnnCWB0eRdq+NUHVHS15/ypteTatgZiwdKop/0Wi046udXqMdxiPtpvZMiXsTqNoMqjbX/rRxK9j+Mu5AzCltGIg5CdEhkHCR+6FIpT0LY2IBxhbaMzOKBjJStNs+6Hb/gqjkdb/Rt1tDbZjLunt/yr57/6yTXTVt4HGsVfyNw7PSIeqA9kxb2DZIuPDgda1ctKEo90yTlTNYOWnnWznfP/tUzEFJAlARSQASovS6nZrJ0oNXWHLgColpWQC42JgxvK0XAwM8pWdpkVdOFiRd1YaD4hxfLTNNe4eaiYW27dHDzqrkZGsHlc24N6Unwd1EbZucu4naS125j3Mywa2BNqy5+4GlQ/HVXJwSo7SXE2/m9gF1WdulQVpCwbav0wX6Ly/WkiQAFZEEICFKv7TMbFYEX2XBnnBik9MB8HCwYMoLDejg42Lg6oSowNKTtIHo9hVIuQGpN7R3naUmaB+n3HvcoCe88H2xvrQEoCKSACRE2ZGZrWH9ievM3HZRF4SC6rsysWt9qtg/5A4gIYRhaXJAXbxnbCUAFZEEICHKnpSMbL7fEcaifRHkaBQsTIx465laDGvjhamxjDMmREUgAaiIJAAJUXZdiL3DJ+vPEHzlFgA1Xaz5rFsDArwdDVyZEOJJe5zvb/lvkRCiXKnjZsPK11syo7cvjlamXIpPof+CQ3yw5hR30rMMXZ4QopSQACSEKHdUKhW9mlbln/fa83LLaqhUsPLoVTrO2suh8Ad08CaEqFAkAAkhyi07SxM+796Q315tSdVKFlxPvEv/BYf4/K9zpGflPHoHQohySwKQEKLc86/hyJZ3nqJfcw8UBRbui+D52fs4fe0hwwUIIco1CUBCiArB2syYab0a8b/BzXCyNuNSfAo9ftjPd9vDyMp5jPGghBDlggQgIUSF8rSPK1vffYrODd3I1ih8u/0iQbP2sO7ENbIlCAlRYcht8PmQ2+CFKP8URWHDyWgmbTirG1LDy8mKUR1q0q2xO8ZG8v9DIcoa6QeoiCQACVFx3EnP4ueDkSzYG64LQp6OlozsUJMeflUkCAlRhkgAKiIJQEJUPCkZ2fx88AoL9oRz+14QquZgyZvtvenuV0UGWRWiDJAAVEQSgISouFIzsll2KJIFe8K5mZoJgJO1GUNaezLAvxr2lqYGrlAI8SASgIpIApAQIi0zm+WHo1i0L4KYJO0gq5amRvRp5sGwNl54OFgauEIhxH9JACoiCUBCiFxZORr+OhXNT3siCI1JBkCtgs4NK/Nq2xr4etgbtkAhhI4EoCKSACSE+C9FUdh3KYGf9oSzNyxBN7+xhz2DW3nSqaEbZsbSTkgIQ5IAVEQSgIQQD3MuOpmFe8P581Q0WTnaP6FO1qb0b1GNl/yrUdnOwsAVClExSQAqIglAQoiCuHEng5VHovjlUBSxydp2QkZqFR3ru/FKQHX8vRxQqVQGrlKIikMCUBFJABJCPI6sHA3bzsWx9MAVDkfc0s2v6WJN/xbV6NWkitw9JkQJeJzv71LRw9fcuXPx9PTE3Nwcf39/goODH7r+rFmzqFOnDhYWFnh4ePDuu++Snp5epH0KIURhmRip6dywMitfD2DLO23p36IalqZGXIpP4bO/ztHiyx2MWRnCkSu3kP9zClE6GPwM0MqVKxk4cCDz5s3D39+fWbNmsXr1ai5cuICLi0ue9ZcvX87QoUP53//+R6tWrbh48SKDBw+mX79+zJw5s1D7/C85AySEKKo76VmsD4lm+eEo3d1jALVdtWeFejapip2FiQErFKL8KVOXwPz9/WnevDlz5swBQKPR4OHhwejRoxk/fnye9UeNGkVoaCg7duzQzXvvvfc4fPgw+/btK9Q+/0sCkBCiuCiKQsjVRFYER7HhZDTpWdoBVy1MjOjuV4WBAdWpW1n+zghRHMrMJbDMzEyOHTtGYGCgbp5arSYwMJCDBw/mu02rVq04duyY7pJWeHg4mzZtonPnzoXepxBCPCkqlQq/apX4+kVfgj8KZEq3+tR2teZuVg4rgqPo9N1eXvzxAH+EXCczW0ajF6KkGBvyxRMSEsjJycHV1VVvvqurK+fPn893m5deeomEhATatGmDoihkZ2fzxhtv8OGHHxZ6nxkZGWRkZOieJycn57ueEEIUha25CQMDPHmlZXUOR9xi2cFI/j4by9HI2xyNvM1n1qH0b+HBKwHVcbExN3S5QpRrpaIR9OPYtWsXX375JT/88APHjx9n7dq1bNy4kc8++6zQ+5w6dSp2dna6ycPDoxgrFkIIfSqVipY1HJk7oAn7xz/NO4G1cLExIyElg9n/XKLtVzuZ9McZrifeNXSpQpRbBg1ATk5OGBkZERcXpzc/Li4ONze3fLf55JNPeOWVVxg+fDgNGzakR48efPnll0ydOhWNRlOofU6YMIGkpCTddPXq1eJ5g0II8Qiutua8E1ib/eOfZu5LTfCrZk9GtoalByNp9/VOPlhziisJqYYuU4hyx6AByNTUlKZNm+o1aNZoNOzYsYOAgIB8t0lLS0Ot1i/byEjb/byiKIXap5mZGba2tnqTEEKUJBMjNV0aVWbtiFYsH+5PQA1HsjUKK49e5ekZu3j7txNcjLtj6DKFKDcM2gYIYMyYMQwaNIhmzZrRokULZs2aRWpqKkOGDAFg4MCBVKlShalTpwLQtWtXZs6ciZ+fH/7+/ly6dIlPPvmErl276oLQo/YphBCllUqlolVNJ1rVdOJY5C3m/HOJnRdu8EdINH+ERBNY14XhbWtIL9NCFJHBA1Dfvn25ceMGEydOJDY2lsaNG7NlyxZdI+aoqCi9Mz4ff/wxKpWKjz/+mOvXr+Ps7EzXrl354osvCrxPIYQoC5pWd2DxkBacuZ7E3J2X2Hwmlu2h8WwPjadRVTuGt61B5wZuGBuVueacQhicwfsBKo2kHyAhRGl0KT6FRfsiWHv8Ghn3bpmvYm/BkNae9G3ugY25dKwoKrYy1RFiaSQBSAhRmt1MyWDZoUiWHYzkZmomADZmxrzYrCp9m3vg4yZ/t0TFJAGoiCQACSHKgvSsHNaduM7CveFcvnH/TjHfqnb0bubBC43dsZWzQqICkQBURBKAhBBliUajsDvsBquOXGV7aBxZOdo/62bG2kFa+zTzwN/LAbVaGk2L8k0CUBFJABJClFU3UzJYd+I6q45e5WJcim5+1UoWdGvsTrfGVajtamPACoV4ciQAFZEEICFEWacoCievJbHyyFX+PBlNSka2bpmPmw3dGlfhhcbuVLG3MGCVQhQvCUBFJAFICFGe3M3MYcf5OP4IiWbXhXjdJTKAFp4OdPNz5wVfd7mLTJR5EoCKSAKQEKK8SkrLYvOZGP4IieZQxE1yvwEsTIx4wded/v7V8K1qJ50sijJJAlARSQASQlQEsUnp/HkympVHr3Ip/n57oXqVbenvX41ucheZKGMkABWRBCAhREWiKApHI2+z4nAUf52OIfNeJ4sWJkZ09a3MoFae1He3M3CVQjyaBKAikgAkhKioEtMyWXv8OsuDo/TOCrWs4cDQ1l48U9cVI7mdXpRSEoCKSAKQEKKiyz0r9PPBSDadjiFHo/2qqOZgyaBWnvRpVlUaTYtSRwJQEUkAEkKI+6IT77LsUCTLD0eRdDcLAGszY3o3q8rgVp5Ud7QycIVCaEkAKiIJQEIIkdfdzBzWnrjG4v1XdJfHVCoIrOvK0NZetKzhIHePCYOSAFREEoCEEOLBFEVhb1gCi/ZFsPviDd38upVtGdrak66+7pibGBmwQlFRSQAqIglAQghRMJfiU1hyIILfj13nblYOAE7Wpgzwr04Pvyp4OsnlMVFyJAAVkQQgIYR4PIlpmfx25CpLD1whJildN9/LyYr2dZzpUMeFFl4OcmZIPFESgIpIApAQQhROVo6Gv8/GsiI4isPht8jW3P+KsTAxopW3I+19XAis60JlOxmHTBQvCUBFJAFICCGK7k56Fvsv3WTXhXh2XognLjlDb3ljD3s6NXCjU4PKVHO0NFCVojyRAFREEoCEEKJ4KYpCaMwddl6I55/z8RyPus2/v33qu9vSqYEbHRtUpqaLteEKFWWaBKAikgAkhBBPVnxyOn+fjWXzmVgOhd/kX1fKqOliTWBdVwLruuBXrZL0PC0KTAJQEUkAEkKIknMzJYNt5+LYfCaWA5cTyMq5/7XkYGXK0/faDLWt5YyVmbEBKxWlnQSgIpIAJIQQhpF0N4vdF2+wIzSOnefjSU7P1i0zNVLT0tuR9rWdaV/HGS8nK+l4UeiRAFREEoCEEMLwsnI0HLlyix2h8WwPjSPyZprecg8HC9rXdqFdbWcCvB3l7JCQAFRUEoCEEKJ0URSFyzdS2Hn+Brsv3iA44haZORrdclMjNc29KtGhjgtP+7hQw1kaUldEEoCKSAKQEEKUbqkZ2RwKv8nuizfYdeEGUbf0zw55OlrS/l4Y8q/hgJmxdMBYEUgAKiIJQEIIUXYoikJEQiq7Ltxg54V4DoXf1GtIbWlqROuaTjxVy4k2tZzxdLSUtkPllASgIpIAJIQQZVdKRjb7LyWw83z+HTBWsbegbS0n2tRyorW3E5WsTA1UqShuEoCKSAKQEEKUD4qicDY6md0Xb7A37AbHIm/rnR1SqaCBux1tajnRtqYTTapXkvHKyjAJQEUkAUgIIcqntMxsDkfcYl9YAvvCErgQd0dvubmJmuaeDrSpqT1DVNfNFrV0xFhmSAAqIglAQghRMcQnp7PvkjYM7buUQPwd/ctljlam+FWzp767HfXdbWlQxY7KdubShqiUkgBURBKAhBCi4lEUhbD4FF0YOhR+k7TMnDzrOViZUt/dlvrudjSoYotvVXuqVrKQUFQKSAAqIglAQgghMrM1nLqWyOnrSZy5nszZ6CTC4lPI0eT92rS3NKFhFTsaVbWjUVV7GlW1w81WzhSVNAlARSQBSAghRH7Ss3K4GHeHM9eTOROdxJnrSYTGJOs1rM7lZG1Ggyq2NKxiR4MqdjSUy2dPnASgIpIAJIQQoqAysnO4GJvCqeuJnL6WxKlrSVyIu5PvmSJHK1PqV7GjgbstddxsqONmQw0na0yN1QaovPyRAFREEoCEEEIURXpWDqExyZy5nsTp60mcvp5MWNwdsvMJRcZqFTWcrajtakMdVxtqu9lQ29WGag6WGMkdaI9FAlARSQASQghR3NKzcjgfe4fT9y6bXYy9w4W4O9z514j3/2ZqrMbb2ZrartbUcrGmlqs2GFV3sJRb8x+gzAWguXPn8s033xAbG4uvry+zZ8+mRYsW+a7bvn17du/enWd+586d2bhxIwCDBw9m6dKlesuDgoLYsmVLgeqRACSEEKIkKIpCbHI6F2LvaKe4O1yMu8Ol+BTSszT5bmNpaoSPmw313e2o525LfXdbarvaSAeOPN73t3EJ1fRAK1euZMyYMcybNw9/f39mzZpFUFAQFy5cwMXFJc/6a9euJTMzU/f85s2b+Pr60rt3b731OnbsyOLFi3XPzczMntybEEIIIQpBpVJR2c6CynYWtK9z/ztPo1G4dvsuF+PuEBafQljcHS7Ga4NRWmYOx6MSOR6VqFvfSK3C29kKH7d7bYtcte2LqthbyNmiBzD4GSB/f3+aN2/OnDlzANBoNHh4eDB69GjGjx//yO1nzZrFxIkTiYmJwcrKCtCeAUpMTGT9+vWFqknOAAkhhCiNcjQKEQkpnI1O5lx0Mudikjkbncyt1Mx817c0NaKWqw11XK2pfe8SWm1XG1xtzcrl3Whl5gxQZmYmx44dY8KECbp5arWawMBADh48WKB9LFq0iH79+unCT65du3bh4uJCpUqVePrpp/n8889xdHTMdx8ZGRlkZNzv/TM5ObkQ70YIIYR4sozUKmq62FDTxYZujasA2stocckZnI3W3n2mbVuUwuV7Z4tOXk3k5NVEvf3YmhtT29XmXrsia2o4W+NRyQJ3e4sKcynNoAEoISGBnJwcXF1d9ea7urpy/vz5R24fHBzMmTNnWLRokd78jh070rNnT7y8vLh8+TIffvghnTp14uDBgxgZ5f3FTp06lU8//bRob0YIIYQwAJVKhZudOW525jxT9/73aVaOhsibqVyITeFCbDIX41K4GH+HyJtpJKdnczTyNkcjb+fZn6utGVUrWeJRyYKqlSyp5mCJl7MVNZyscLAyLTdnjgx6CSw6OpoqVapw4MABAgICdPPHjRvH7t27OXz48EO3f/311zl48CCnTp166Hrh4eF4e3uzfft2nnnmmTzL8zsD5OHhIZfAhBBClDsZ2TmE30jl4r0G1xfjUoi8mcrVW3e5m5V36I9/szU3poazNTWcrKjhbIWXkzWeTpZ4OlphZWbwZsVl5xKYk5MTRkZGxMXF6c2Pi4vDzc3todumpqby22+/MWXKlEe+To0aNXBycuLSpUv5BiAzMzNpJC2EEKJCMDM2om5lW+pW1g8IiqJwKzWTa7fvcvV2mvbnrTSibqURfiOV64l3SU7PJuRqIiH/uaQG4GJjhqeTFZ6Olng6ac8Y1XSxprqjFSZGpa+jR4MGIFNTU5o2bcqOHTvo3r07oG0EvWPHDkaNGvXQbVevXk1GRgYvv/zyI1/n2rVr3Lx5k8qVKxdH2UIIIUS5o1KpcLQ2w9HaDF8P+zzL72bmcOVmKhEJqYTfSCE8Qfv4SkIqt9OyiL+TQfydDIIjbultZ2KkooaTNbXuNcTO7dPI09ESYwMGI4OfrxozZgyDBg2iWbNmtGjRglmzZpGamsqQIUMAGDhwIFWqVGHq1Kl62y1atIju3bvnadickpLCp59+Sq9evXBzc+Py5cuMGzeOmjVrEhQUVGLvSwghhChPLEzzP3MEkJSWxZWbqbqAdCUhlfCEVN1t+xfitH0cQYxum7a1nFg2zL8E34E+gwegvn37cuPGDSZOnEhsbCyNGzdmy5YtuobRUVFRqNX6CfHChQvs27ePrVu35tmfkZERp06dYunSpSQmJuLu7s5zzz3HZ599Jpe5hBBCiCfAztIEX0v7PGeONBqF64l3CYu/Q1hcChfjUnSPvZ2tDVPsPQbvB6g0kn6AhBBCiCdHo1HIyNZgYVq8t9w/zvd36WuVJIQQQohyTa1WFXv4eewaDPrqQgghhBAGIAFICCGEEBWOBCAhhBBCVDgSgIQQQghR4UgAEkIIIUSFIwFICCGEEBWOBCAhhBBCVDgSgIQQQghR4UgAEkIIIUSFIwFICCGEEBWOBCAhhBBCVDgSgIQQQghR4UgAEkIIIUSFY2zoAkojRVEASE5ONnAlQgghhCio3O/t3O/xh5EAlI87d+4A4OHhYeBKhBBCCPG47ty5g52d3UPXUSkFiUkVjEajITo6GhsbG1Qq1WNvn5ycjIeHB1evXsXW1vYJVFh2ybF5ODk+DybH5uHk+DyYHJuHK0/HR1EU7ty5g7u7O2r1w1v5yBmgfKjVaqpWrVrk/dja2pb5D9OTIsfm4eT4PJgcm4eT4/Ngcmwerrwcn0ed+ckljaCFEEIIUeFIABJCCCFEhSMB6AkwMzNj0qRJmJmZGbqUUkeOzcPJ8XkwOTYPJ8fnweTYPFxFPT7SCFoIIYQQFY6cARJCCCFEhSMBSAghhBAVjgQgIYQQQlQ4EoCEEEIIUeFIACpmc+fOxdPTE3Nzc/z9/QkODjZ0SQaxZ88eunbtiru7OyqVivXr1+stVxSFiRMnUrlyZSwsLAgMDCQsLMwwxZawqVOn0rx5c2xsbHBxcaF79+5cuHBBb5309HRGjhyJo6Mj1tbW9OrVi7i4OANVXLJ+/PFHGjVqpOuULSAggM2bN+uWV+Rj81/Tpk1DpVLxzjvv6OZV5OMzefJkVCqV3uTj46NbXpGPDcD169d5+eWXcXR0xMLCgoYNG3L06FHd8or2d1kCUDFauXIlY8aMYdKkSRw/fhxfX1+CgoKIj483dGklLjU1FV9fX+bOnZvv8q+//prvv/+eefPmcfjwYaysrAgKCiI9Pb2EKy15u3fvZuTIkRw6dIht27aRlZXFc889R2pqqm6dd999lz///JPVq1eze/duoqOj6dmzpwGrLjlVq1Zl2rRpHDt2jKNHj/L000/TrVs3zp49C1TsY/NvR44cYf78+TRq1EhvfkU/PvXr1ycmJkY37du3T7esIh+b27dv07p1a0xMTNi8eTPnzp1jxowZVKpUSbdOhfu7rIhi06JFC2XkyJG65zk5OYq7u7sydepUA1ZleICybt063XONRqO4ubkp33zzjW5eYmKiYmZmpqxYscIAFRpWfHy8Aii7d+9WFEV7LExMTJTVq1fr1gkNDVUA5eDBg4Yq06AqVaqkLFy4UI7NPXfu3FFq1aqlbNu2TWnXrp3y9ttvK4oin51JkyYpvr6++S6r6Mfmgw8+UNq0afPA5RXx77KcASommZmZHDt2jMDAQN08tVpNYGAgBw8eNGBlpU9ERASxsbF6x8rOzg5/f/8KeaySkpIAcHBwAODYsWNkZWXpHR8fHx+qVatW4Y5PTk4Ov/32G6mpqQQEBMixuWfkyJF06dJF7ziAfHYAwsLCcHd3p0aNGgwYMICoqChAjs2GDRto1qwZvXv3xsXFBT8/PxYsWKBbXhH/LksAKiYJCQnk5OTg6vr/9u4tJKqujQP4f3Sa0bFMyRinotFOpkVRVmJDUBmkdlFhZWExFWRalkEFQUUllQVmpRdTvqBGRVLRdKSjmhdCRzWVzKzsAGUS0cGOF/N8Fy9tvp0dv9dm+737/4MFs9dae3zmYdg8rL22Y1X1W61WtLS0aBRV5/QlH8wV4PF4sGLFCjgcDgwdOhTA3/kxmUwICgpSzdVTfurq6tC1a1eYzWakpaXB7XYjKiqKuQFQUlKCqqoqZGdntxvTe35iYmJQXFyMc+fOweVyobm5GePGjcPbt291n5sHDx7A5XJh4MCBOH/+PNLT07F8+XLs27cPgD6vy/w1eCINLV26FPX19ap9CgRERESgpqYGr1+/xtGjR+F0OlFRUaF1WJp78uQJMjMzcfHiRfj5+WkdTqeTkJCgvB42bBhiYmJgt9tx+PBh+Pv7axiZ9jweD0aNGoWtW7cCAEaMGIH6+nrs2bMHTqdT4+i0wRWgDhISEgJfX992TxQ8f/4coaGhGkXVOX3Jh95zlZGRgdOnT6O8vBx9+vRR+kNDQ/H582e8evVKNV9P+TGZTBgwYACio6ORnZ2N4cOHY/fu3brPzc2bN9Ha2oqRI0fCaDTCaDSioqICeXl5MBqNsFqtus7P14KCgjBo0CDcu3dP998dm82GqKgoVV9kZKRyi1CP12UWQB3EZDIhOjoapaWlSp/H40FpaSliY2M1jKzzCQ8PR2hoqCpXb968wdWrV3WRKxFBRkYG3G43ysrKEB4erhqPjo5Gly5dVPlpbGzE48ePdZGfb/F4PPj06ZPucxMXF4e6ujrU1NQobdSoUUhJSVFe6zk/X2tra8P9+/dhs9l0/91xOBzt/t3G3bt3YbfbAej0uqz1Lux/k5KSEjGbzVJcXCy3b9+W1NRUCQoKkpaWFq1D87q3b99KdXW1VFdXCwDJzc2V6upqefTokYiIbNu2TYKCguTEiRNSW1srU6dOlfDwcPnw4YPGkf956enp0r17d7l8+bI8e/ZMae/fv1fmpKWlSd++faWsrExu3LghsbGxEhsbq2HU3rNmzRqpqKiQ5uZmqa2tlTVr1ojBYJALFy6IiL5z8y3//RSYiL7zs3LlSrl8+bI0NzdLZWWlTJo0SUJCQqS1tVVE9J2ba9euidFolC1btkhTU5McPHhQLBaLHDhwQJmjt+syC6AOlp+fL3379hWTySRjxoyRK1euaB2SJsrLywVAu+Z0OkXk70cu169fL1arVcxms8TFxUljY6O2QXvJt/ICQIqKipQ5Hz58kCVLlkhwcLBYLBaZPn26PHv2TLugvWjhwoVit9vFZDJJz549JS4uTil+RPSdm2/5ugDSc36Sk5PFZrOJyWSS3r17S3Jysty7d08Z13NuREROnTolQ4cOFbPZLIMHD5aCggLVuN6uywYREW3WnoiIiIi0wT1AREREpDssgIiIiEh3WAARERGR7rAAIiIiIt1hAURERES6wwKIiIiIdIcFEBEREekOCyAi0h2DwYDjx49rHQYRaYgFEBF5zfz582EwGNq1+Ph4rUP7qfHjx8NgMKCkpETVv2vXLoSFhWkTFBH9z4xaB0BE+hIfH4+ioiJVn9ls1iia3+Pn54d169YhKSkJXbp00TocIvoHuAJERF5lNpsRGhqqasHBwcq4wWCAy+VCQkIC/P390a9fPxw9elT1HnV1dZg4cSL8/f3Ro0cPpKamoq2tTTWnsLAQQ4YMgdlshs1mQ0ZGhmr8xYsXmD59OiwWCwYOHIiTJ0/+NPY5c+bg1atX+Ouvv344z+VyoX///jCZTIiIiMD+/ft/+t5E5F0sgIio01m/fj2SkpJw69YtpKSkYPbs2WhoaAAAvHv3DpMnT0ZwcDCuX7+OI0eO4NKlS6oCx+VyYenSpUhNTUVdXR1OnjyJAQMGqP7Gpk2bMGvWLNTW1iIxMREpKSl4+fLlD+MKDAzE2rVrkZWVhXfv3n1zjtvtRmZmJlauXIn6+nosXrwYCxYsQHl5+T/MChF1KK1/jZWI9MPpdIqvr68EBASo2pYtW5Q5ACQtLU11XkxMjKSnp4uISEFBgQQHB0tbW5syfubMGfHx8ZGWlhYREenVq5esXbv2u3EAkHXr1inHbW1tAkDOnj373XO+/Or6x48fxW63S1ZWloiI7Ny5U+x2uzJv7NixsmjRItW5M2fOlMTExO++NxF5H1eAiMirJkyYgJqaGlVLS0tTzYmNjW13/GUFqKGhAcOHD0dAQIAy7nA44PF40NjYiNbWVjx9+hRxcXE/jGPYsGHK64CAAAQGBqK1tfWn8ZvNZmRlZSEnJwcvXrxoN97Q0ACHw6HqczgcSvxE1DlwEzQReVVAQEC721Edyd/f/5fmfb2J2WAwwOPx/NK5c+fORU5ODjZv3swnwIj+T3EFiIg6nStXrrQ7joyMBABERkbi1q1bqj04lZWV8PHxQUREBLp164awsDCUlpb+sfh8fHyQnZ0Nl8uFhw8fqsYiIyNRWVmp6qusrERUVNQfi4eIfh8LICLyqk+fPqGlpUXVvr6VdOTIERQWFuLu3bvYsGEDrl27pmxyTklJgZ+fH5xOJ+rr61FeXo5ly5Zh3rx5sFqtAICNGzdix44dyMvLQ1NTE6qqqpCfn9+hn2PKlCmIiYnB3r17Vf2rV69GcXExXC4XmpqakJubi2PHjmHVqlXKnMGDB8PtdndoPET0e3gLjIi86ty5c7DZbKq+iIgI3LlzRznetGkTSkpKsGTJEthsNhw6dEhZQbFYLDh//jwyMzMxevRoWCwWJCUlITc3Vznf6XTi48eP2LlzJ1atWoWQkBDMmDGjwz/L9u3bMXbsWFXftGnTsHv3buTk5CAzMxPh4eEoKirC+PHjlTmNjY14/fp1h8dDRL/OICKidRBERF8YDAa43W5MmzZN61CI6F+Mt8CIiIhId1gAERERke5wDxARdSq8K09E3sAVICIiItIdFkBERESkOyyAiIiISHdYABEREZHusAAiIiIi3WEBRERERLrDAoiIiIh0hwUQERER6Q4LICIiItKd/wBcPFkC+0AdpQAAAABJRU5ErkJggg==",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "#now i use the best parameters to train my model and then evaluate that model\n",
    "\n",
    "model = RpropModel(nn.ReLU())\n",
    "loss_function = nn.CrossEntropyLoss()\n",
    "optimizer = torch.optim.Rprop(model.parameters(), 0.005)#hard code best learning rate\n",
    "\n",
    "training_losses, testing_losses = train(model, optimizer, loss_function)\n",
    "\n",
    "#plt.figure(figsize=(12, 6))\n",
    "plt.plot(list(range(1,65)), training_losses, label='Training Loss - Rprop')\n",
    "plt.plot(list(range(1,65)), testing_losses, label='Testing Loss - Rprop')\n",
    "plt.title('Training and Testing Loss over 64 Epochs for Rprop')\n",
    "plt.xlabel('Epoch No.')\n",
    "plt.ylabel('Loss')\n",
    "plt.legend()\n",
    "\n",
    "#evaluate(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Run 1\n",
      "Run 2\n",
      "Run 3\n",
      "Run 4\n",
      "Run 5\n",
      "Results after 5 runs:\n",
      "Training Loss (mean): 0.8783\n",
      "Training Loss (standard deviation): 0.0709\n",
      "Test Loss (mean): 0.8264\n",
      "Test Loss (standard deviation): 0.0064\n",
      "Test Accuracy (mean): 71.82%\n",
      "Test Accuracy (standard deviation): 0.91\n"
     ]
    }
   ],
   "source": [
    "accuracy_values = []\n",
    "train_losses_every_run = []\n",
    "test_losses = []\n",
    "\n",
    "\n",
    "for run in range(5):\n",
    "    print(f\"Run {run+1}\")\n",
    "    model = RpropModel(nn.ReLU())\n",
    "    loss_function = nn.CrossEntropyLoss()\n",
    "    optimizer = torch.optim.Rprop(model.parameters(), lr=grid_search.best_params_[\"optimizer__lr\"])#hard code best learning rate\n",
    "\n",
    "    train_losses = train(model, optimizer, loss_function)\n",
    "    test_loss, accuracy = evaluate(model)\n",
    "    \n",
    "    accuracy_values.append(accuracy)\n",
    "    train_losses_every_run.append(train_losses[-1])\n",
    "    test_losses.append(test_loss)\n",
    "    \n",
    "\n",
    "train_losses_every_run = np.array(train_losses_every_run)\n",
    "\n",
    "accuracy_values = np.array(accuracy_values)\n",
    "\n",
    "test_losses = np.array(test_losses)\n",
    "\n",
    "print(\"Results after 5 runs:\")\n",
    "print(f\"Training Loss (mean): {train_losses_every_run.mean():.4f}\")\n",
    "print(f\"Training Loss (standard deviation): {train_losses_every_run.std():.4f}\")\n",
    "print(f\"Test Loss (mean): {test_losses.mean():.4f}\")\n",
    "print(f\"Test Loss (standard deviation): {test_losses.std():.4f}\")\n",
    "print(f\"Test Accuracy (mean): {accuracy_values.mean()*100:.2f}%\")\n",
    "print(f\"Test Accuracy (standard deviation): {accuracy_values.std()*100:.2f}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from torch import nn\n",
    "import torch\n",
    "\n",
    "\n",
    "class AdamModel(nn.Module):\n",
    "    def __init__(self, activation_func=nn.ReLU()):\n",
    "        super(AdamModel, self).__init__()\n",
    "        self.activation_func = activation_func\n",
    "#         self.fc1 = nn.Linear(10, 30)\n",
    "#         self.fc2 = nn.Linear(30, 3)\n",
    "        self.fc1 = nn.Linear(10, 35)\n",
    "        self.fc2 = nn.Linear(35, 45)    \n",
    "        self.fc3 = nn.Linear(45, 3)\n",
    "\n",
    "    def forward(self, x):\n",
    "#         x = torch.relu(self.fc1(x))\n",
    "#         x = torch.softmax(self.fc2(x), dim=1)\n",
    "\n",
    "        x = self.activation_func(self.fc1(x))\n",
    "        x = self.activation_func(self.fc2(x))\n",
    "        x = torch.softmax(self.fc3(x), dim=1)\n",
    "        return x\n",
    "\n",
    "model = AdamModel()\n",
    "loss_function = nn.CrossEntropyLoss()\n",
    "optimizer = torch.optim.Adam(model.parameters())\n",
    "\n",
    "#the train and evaluate function are the same so we don't have to define them again\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%%capture\n",
    "!pip install skorch\n",
    "\n",
    "#note that I added %%capture to the start of the cell because the verbose option wasn't working properly in my jupyter notebook instance, so my notebook was producing too much output and craching without %%capture\n",
    "from torch import nn\n",
    "from sklearn.model_selection import KFold\n",
    "from sklearn.model_selection import GridSearchCV\n",
    "import skorch\n",
    "from skorch import NeuralNetClassifier\n",
    "import torch.optim as optim\n",
    "\n",
    "net = NeuralNetClassifier(\n",
    "    train_split=None,\n",
    "    module=AdamModel,\n",
    "    criterion=nn.CrossEntropyLoss,\n",
    "    optimizer=optim.Rprop,\n",
    "    #optimizer__lr=0.001,\n",
    "#     batch_size=32,\n",
    "    max_epochs=64,\n",
    ")\n",
    "\n",
    "hyperparameter_grid = {\n",
    "    'module__activation_func': [nn.ReLU(), nn.LeakyReLU(0.01), nn.Tanh(), nn.ELU(alpha=1.0)],\n",
    "    'optimizer__lr': [0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1]\n",
    "}\n",
    "\n",
    "grid_search = GridSearchCV(net, hyperparameter_grid, cv=5, scoring='accuracy', verbose=0)\n",
    "grid_search.fit(X_train.to_numpy(dtype='float32'), np.argmax(y_train.to_numpy(dtype='float32'), axis=1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(\"Best parameters found: \", grid_search.best_params_)\n",
    "print(\"Highest accuracy found: \", grid_search.best_score_)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "import pandas as pd\n",
    "\n",
    "\n",
    "adam_opt_results = grid_search.cv_results_\n",
    "#print(adam_opt_results)\n",
    "data = pd.DataFrame({\n",
    "    'Activation Function': adam_opt_results['param_module__activation_func'],\n",
    "    'Learning Rate': adam_opt_results['param_optimizer__lr'],\n",
    "    'mean_score': adam_opt_results['mean_test_score'],\n",
    "    'std_score': adam_opt_results['std_test_score']\n",
    "})\n",
    "\n",
    "heatmap_data = data.pivot(index='Activation Function', columns='Learning Rate', values='mean_score')\n",
    "heatmap_data_std = data.pivot(index='Activation Function', columns='Learning Rate', values='std_score')\n",
    "\n",
    "plt.figure(figsize=(10, 6))\n",
    "sns.heatmap(heatmap_data, annot=True)\n",
    "plt.title('Heatmap showing the Mean Test Scores of the Grid Search Results for Hyperparameter Optimization (Learning Rate and Activation function)')\n",
    "plt.xlabel('Learning Rate')\n",
    "plt.ylabel('Activation Function')\n",
    "plt.show()\n",
    "\n",
    "plt.figure(figsize=(10, 6))\n",
    "sns.heatmap(heatmap_data_std, annot=True, fmt=\".3f\", cmap=\"coolwarm\")\n",
    "plt.title('Heatmap showing the Standard Deviation of Test Scores of the Grid Search Results for Hyperparameter Optimization (Learning Rate and Activation function)')\n",
    "plt.xlabel('Learning Rate')\n",
    "plt.ylabel('Activation Function')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#now i use the best parameters to train my model and then evaluate that model\n",
    "\n",
    "model = AdamModel(nn.ReLU())\n",
    "loss_function = nn.CrossEntropyLoss()\n",
    "optimizer = torch.optim.Adam(model.parameters(), 0.005)#hard code best learning rate\n",
    "\n",
    "training_losses, testing_losses = train(model, optimizer, loss_function)\n",
    "\n",
    "#plt.figure(figsize=(12, 6))\n",
    "plt.plot(list(range(1,65)), training_losses, label='Training Loss - Adam')\n",
    "plt.plot(list(range(1,65)), testing_losses, label='Testing Loss - Adam')\n",
    "plt.title('Training and Testing Loss over 64 Epochs for Adam')\n",
    "plt.xlabel('Epoch No.')\n",
    "plt.ylabel('Loss')\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "accuracy_values = []\n",
    "train_losses_every_run = []\n",
    "test_losses = []\n",
    "\n",
    "\n",
    "for run in range(5):\n",
    "    print(f\"Run {run+1}\")\n",
    "    model = RpropModel(nn.ReLU())\n",
    "    loss_function = nn.CrossEntropyLoss()\n",
    "    optimizer = torch.optim.Adam(model.parameters(), lr=0.005)#hard code best learning rate\n",
    "\n",
    "    train_losses = train(model, optimizer, loss_function)\n",
    "    test_loss, accuracy = evaluate(model)\n",
    "    \n",
    "    accuracy_values.append(accuracy)\n",
    "    train_losses_every_run.append(train_losses[-1])\n",
    "    test_losses.append(test_loss)\n",
    "    \n",
    "\n",
    "train_losses_every_run = np.array(train_losses_every_run)\n",
    "\n",
    "accuracy_values = np.array(accuracy_values)\n",
    "\n",
    "test_losses = np.array(test_losses)\n",
    "\n",
    "print(\"Results after 5 runs:\")\n",
    "print(f\"Training Loss (mean): {train_losses_every_run.mean():.4f}\")\n",
    "print(f\"Training Loss (standard deviation): {train_losses_every_run.std():.4f}\")\n",
    "print(f\"Test Loss (mean): {test_losses.mean():.4f}\")\n",
    "print(f\"Test Loss (standard deviation): {test_losses.std():.4f}\")\n",
    "print(f\"Test Accuracy (mean): {accuracy_values.mean()*100:.2f}%\")\n",
    "print(f\"Test Accuracy (standard deviation): {accuracy_values.std()*100:.2f}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from torch import nn\n",
    "import torch\n",
    "\n",
    "class HybridOptimizer():\n",
    "    def __init__(self, *optimizers):\n",
    "        self.optimizers = optimizers\n",
    "\n",
    "    def step(self):\n",
    "        for optimizer in self.optimizers:\n",
    "            optimizer.step()\n",
    "    \n",
    "    def zero_grad(self):\n",
    "        for optimizer in self.optimizers:\n",
    "            optimizer.zero_grad()\n",
    "\n",
    "\n",
    "\n",
    "class HybridModel(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(HybridModel, self).__init__()\n",
    "#         self.fc1 = nn.Linear(10, 30)\n",
    "#         self.fc2 = nn.Linear(30, 3)\n",
    "        self.fc1 = nn.Linear(10, 35)\n",
    "        self.fc2 = nn.Linear(35, 45)    \n",
    "        self.fc3 = nn.Linear(45, 3)\n",
    "\n",
    "    def forward(self, x):\n",
    "#         x = torch.relu(self.fc1(x))\n",
    "#         x = torch.softmax(self.fc2(x), dim=1)\n",
    "\n",
    "        x = torch.relu(self.fc1(x))\n",
    "        x = torch.relu(self.fc2(x))\n",
    "        x = torch.softmax(self.fc3(x), dim=1)\n",
    "        return x\n",
    "\n",
    "model = HybridModel()\n",
    "loss_function = nn.CrossEntropyLoss()\n",
    "hybrid_optimizer = HybridOptimizer(torch.optim.Adam(model.parameters()), \n",
    "                                   torch.optim.Rprop(model.parameters()), \n",
    "                                   torch.optim.Adagrad(model.parameters()))\n",
    "\n",
    "\n",
    "\n",
    "# train(model, hybrid_optimizer, loss_function)\n",
    "# evaluate(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "accuracy_values = []\n",
    "train_losses_every_run = []\n",
    "test_losses = []\n",
    "\n",
    "\n",
    "for run in range(5):\n",
    "    print(f\"Run {run+1}\")\n",
    "    model = HybridModel()\n",
    "    loss_function = nn.CrossEntropyLoss()\n",
    "    optimizer = HybridOptimizer(torch.optim.Adam(model.parameters()), \n",
    "                                   torch.optim.Rprop(model.parameters()), \n",
    "                                   torch.optim.Adagrad(model.parameters()))\n",
    "\n",
    "    train_losses = train(model, optimizer, loss_function)\n",
    "    test_loss, accuracy = evaluate(model)\n",
    "    \n",
    "    accuracy_values.append(accuracy)\n",
    "    train_losses_every_run.append(train_losses[-1])\n",
    "    test_losses.append(test_loss)\n",
    "    \n",
    "\n",
    "train_losses_every_run = np.array(train_losses_every_run)\n",
    "\n",
    "accuracy_values = np.array(accuracy_values)\n",
    "\n",
    "test_losses = np.array(test_losses)\n",
    "\n",
    "print(\"Results after 5 runs:\")\n",
    "print(f\"Training Loss (mean): {train_losses_every_run.mean():.4f}\")\n",
    "print(f\"Training Loss (standard deviation): {train_losses_every_run.std():.4f}\")\n",
    "print(f\"Test Loss (mean): {test_losses.mean():.4f}\")\n",
    "print(f\"Test Loss (standard deviation): {test_losses.std():.4f}\")\n",
    "print(f\"Test Accuracy (mean): {accuracy_values.mean()*100:.2f}%\")\n",
    "print(f\"Test Accuracy (standard deviation): {accuracy_values.std()*100:.2f}\")"
   ]
  }
 ],
 "metadata": {
  "kaggle": {
   "accelerator": "none",
   "dataSources": [
    {
     "datasetId": 5344793,
     "sourceId": 8883521,
     "sourceType": "datasetVersion"
    }
   ],
   "dockerImageVersionId": 30761,
   "isGpuEnabled": false,
   "isInternetEnabled": true,
   "language": "python",
   "sourceType": "notebook"
  },
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.11.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}