Predictions_Porto_Seguro_dataset.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## COMPETITION: Porto Seguro’s Safe Driver Prediction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "#importing libraries\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "from sklearn.model_selection import train_test_split\n",
    "from xgboost import XGBClassifier\n",
    "from sklearn import metrics\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>id</th>\n",
       "      <th>target</th>\n",
       "      <th>ps_ind_01</th>\n",
       "      <th>ps_ind_02_cat</th>\n",
       "      <th>ps_ind_03</th>\n",
       "      <th>ps_ind_04_cat</th>\n",
       "      <th>ps_ind_05_cat</th>\n",
       "      <th>ps_ind_06_bin</th>\n",
       "      <th>ps_ind_07_bin</th>\n",
       "      <th>ps_ind_08_bin</th>\n",
       "      <th>...</th>\n",
       "      <th>ps_calc_11</th>\n",
       "      <th>ps_calc_12</th>\n",
       "      <th>ps_calc_13</th>\n",
       "      <th>ps_calc_14</th>\n",
       "      <th>ps_calc_15_bin</th>\n",
       "      <th>ps_calc_16_bin</th>\n",
       "      <th>ps_calc_17_bin</th>\n",
       "      <th>ps_calc_18_bin</th>\n",
       "      <th>ps_calc_19_bin</th>\n",
       "      <th>ps_calc_20_bin</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>7</td>\n",
       "      <td>0</td>\n",
       "      <td>2</td>\n",
       "      <td>2</td>\n",
       "      <td>5</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>9</td>\n",
       "      <td>1</td>\n",
       "      <td>5</td>\n",
       "      <td>8</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>9</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>7</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>...</td>\n",
       "      <td>3</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>9</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>13</td>\n",
       "      <td>0</td>\n",
       "      <td>5</td>\n",
       "      <td>4</td>\n",
       "      <td>9</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>...</td>\n",
       "      <td>4</td>\n",
       "      <td>2</td>\n",
       "      <td>7</td>\n",
       "      <td>7</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>16</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>2</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>2</td>\n",
       "      <td>2</td>\n",
       "      <td>4</td>\n",
       "      <td>9</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>17</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>2</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>3</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>3</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>595207</td>\n",
       "      <td>1488013</td>\n",
       "      <td>0</td>\n",
       "      <td>3</td>\n",
       "      <td>1</td>\n",
       "      <td>10</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>4</td>\n",
       "      <td>1</td>\n",
       "      <td>9</td>\n",
       "      <td>6</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>595208</td>\n",
       "      <td>1488016</td>\n",
       "      <td>0</td>\n",
       "      <td>5</td>\n",
       "      <td>1</td>\n",
       "      <td>3</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>4</td>\n",
       "      <td>1</td>\n",
       "      <td>3</td>\n",
       "      <td>8</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>595209</td>\n",
       "      <td>1488017</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>10</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>3</td>\n",
       "      <td>2</td>\n",
       "      <td>2</td>\n",
       "      <td>6</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>595210</td>\n",
       "      <td>1488021</td>\n",
       "      <td>0</td>\n",
       "      <td>5</td>\n",
       "      <td>2</td>\n",
       "      <td>3</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>...</td>\n",
       "      <td>4</td>\n",
       "      <td>1</td>\n",
       "      <td>4</td>\n",
       "      <td>2</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>595211</td>\n",
       "      <td>1488027</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>8</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>4</td>\n",
       "      <td>4</td>\n",
       "      <td>3</td>\n",
       "      <td>8</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>595212 rows × 58 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "             id  target  ps_ind_01  ps_ind_02_cat  ps_ind_03  ps_ind_04_cat  \\\n",
       "0             7       0          2              2          5              1   \n",
       "1             9       0          1              1          7              0   \n",
       "2            13       0          5              4          9              1   \n",
       "3            16       0          0              1          2              0   \n",
       "4            17       0          0              2          0              1   \n",
       "...         ...     ...        ...            ...        ...            ...   \n",
       "595207  1488013       0          3              1         10              0   \n",
       "595208  1488016       0          5              1          3              0   \n",
       "595209  1488017       0          1              1         10              0   \n",
       "595210  1488021       0          5              2          3              1   \n",
       "595211  1488027       0          0              1          8              0   \n",
       "\n",
       "        ps_ind_05_cat  ps_ind_06_bin  ps_ind_07_bin  ps_ind_08_bin  ...  \\\n",
       "0                   0              0              1              0  ...   \n",
       "1                   0              0              0              1  ...   \n",
       "2                   0              0              0              1  ...   \n",
       "3                   0              1              0              0  ...   \n",
       "4                   0              1              0              0  ...   \n",
       "...               ...            ...            ...            ...  ...   \n",
       "595207              0              0              0              0  ...   \n",
       "595208              0              0              0              0  ...   \n",
       "595209              0              1              0              0  ...   \n",
       "595210              0              0              0              1  ...   \n",
       "595211              0              1              0              0  ...   \n",
       "\n",
       "        ps_calc_11  ps_calc_12  ps_calc_13  ps_calc_14  ps_calc_15_bin  \\\n",
       "0                9           1           5           8               0   \n",
       "1                3           1           1           9               0   \n",
       "2                4           2           7           7               0   \n",
       "3                2           2           4           9               0   \n",
       "4                3           1           1           3               0   \n",
       "...            ...         ...         ...         ...             ...   \n",
       "595207           4           1           9           6               0   \n",
       "595208           4           1           3           8               1   \n",
       "595209           3           2           2           6               0   \n",
       "595210           4           1           4           2               0   \n",
       "595211           4           4           3           8               0   \n",
       "\n",
       "        ps_calc_16_bin  ps_calc_17_bin  ps_calc_18_bin  ps_calc_19_bin  \\\n",
       "0                    1               1               0               0   \n",
       "1                    1               1               0               1   \n",
       "2                    1               1               0               1   \n",
       "3                    0               0               0               0   \n",
       "4                    0               0               1               1   \n",
       "...                ...             ...             ...             ...   \n",
       "595207               1               1               0               1   \n",
       "595208               0               1               0               1   \n",
       "595209               0               1               0               0   \n",
       "595210               1               1               1               0   \n",
       "595211               1               0               0               0   \n",
       "\n",
       "        ps_calc_20_bin  \n",
       "0                    1  \n",
       "1                    0  \n",
       "2                    0  \n",
       "3                    0  \n",
       "4                    0  \n",
       "...                ...  \n",
       "595207               1  \n",
       "595208               1  \n",
       "595209               0  \n",
       "595210               0  \n",
       "595211               0  \n",
       "\n",
       "[595212 rows x 58 columns]"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Load cleaned train dataset\n",
    "traindf = pd.read_csv(\"./traindf1.csv\") #this assumes that traindf1 was created running EDA_Porto_Seguro_gh.R script in R\n",
    "traindf"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Prediction Iteration 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Store feature subset and target vector in X, y with id info to use later if needed\n",
    "X = traindf.loc[:,traindf.columns != 'target']\n",
    "y = traindf.iloc[:,0:2]\n",
    "\n",
    "# splitting train and test data with same seed (random_state=1)\n",
    "X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=1) #70%/30% splitting\n",
    "\n",
    "#create classifier object\n",
    "clf1 = XGBClassifier(random_state=1)\n",
    "\n",
    "#train classifier\n",
    "fit_model = clf1.fit(X_train.iloc[:,1:],y_train.iloc[:,1]) #train model with all features of train dataset except \"id\" feature\n",
    "\n",
    "#create predictions vector and prediction probabilities of class 1\n",
    "y_pred = fit_model.predict(X_test.iloc[:,1:]) #prediction vector with classes\n",
    "y_pred_prob = fit_model.predict_proba(X_test.iloc[:,1:]) #prediction probabilities vector [P(0), P(1)]\n",
    "\n",
    "# set up a 2-D array that has predictions and true (test) values side by side\n",
    "pred_vs_true  = np.c_[y_pred,y_test.iloc[:,1]]\n",
    "# convert it to dataframe to be readable in evaluation step\n",
    "pred_vs_true_df = pd.DataFrame(data=pred_vs_true,columns = [\"pred\",\"true\"]) \n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Evaluation (Iteration 1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Create a confusion matrix to represent model's performance (using my own implementation, not sklearn's) - this is the same exercise as lab 6 (evaluation) - question 4:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>0</th>\n",
       "      <th>1</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>172094</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>6470</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "        0  1\n",
       "0  172094  0\n",
       "1    6470  0"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Create a confusion matrix to represent your model's performance (use your own implementation, not sklearn's)\n",
    "# (this is the same exercise as lab 6 (evaluation) - question 4)\n",
    "\n",
    "target_values = pd.unique(pred_vs_true_df[\"true\"])\n",
    "target_values = np.sort(target_values)\n",
    "\n",
    "#create an empty dataframe that will be our confusion matrix to update later\n",
    "confusion_matrix = pd.DataFrame(columns=target_values,index=target_values) \n",
    "\n",
    "#function that counts how many times [predicted values == true values] in a 2D array (predicted, true)\n",
    "def compare_values(array,predicted,true):\n",
    "    count = len(array.loc[(array[\"pred\"]==predicted) & (array[\"true\"]==true)]) \n",
    "    return count\n",
    "\n",
    "#function to set up a matrix using compare_values function defined before\n",
    "def set_matrix(pred_target_df):\n",
    "    for y in range(0,len(target_values)): #for each unique value (level) of target feature \n",
    "        confusion_matrix.iloc[y] = [compare_values(pred_target_df,x,y) for x in target_values] # compare values for the different combination of predicted values (x) and \"true\" values (y)\n",
    "    return confusion_matrix\n",
    "\n",
    "confusion_matrix = set_matrix(pred_vs_true_df)\n",
    "confusion_matrix"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy is 96.3766%\n",
      "Precision for \"0\" class is 96.3766%\n",
      "Recall for \"0\" class is 100.0%\n",
      "Precision for \"1\" class is nan%\n",
      "Recall for \"1\" class is 0.0%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\mtava\\Anaconda3\\lib\\site-packages\\ipykernel_launcher.py:10: RuntimeWarning: invalid value encountered in longlong_scalars\n",
      "  # Remove the CWD from sys.path while we load stuff.\n"
     ]
    }
   ],
   "source": [
    "#Confusion matrix metrics\n",
    "\n",
    "# function to calculate recall per level in a confusion matrix\n",
    "def calc_rec(cm):\n",
    "    recs = [cm.iloc[i,i]/sum(cm.iloc[i,:]) for i in range(len(cm))]\n",
    "    return recs\n",
    "\n",
    "# function to calculate precision per level in a confusion matrix\n",
    "def calc_prec(cm):\n",
    "    precs = [cm.iloc[i,i]/sum(cm.iloc[:,i]) for i in range(len(cm))]\n",
    "    return precs\n",
    "\n",
    "#function to compute confusion matrix metrics\n",
    "def cm_metrics(cm):\n",
    "    acc = round(np.trace(cm)/cm.values.sum()*100,4) # accuracy\n",
    "    print('Accuracy is ' + str(acc) + '%') # accuracy\n",
    "    precs = calc_prec(cm)\n",
    "    recs = calc_rec(cm)\n",
    "    for i in range(len(cm)):\n",
    "        print('Precision for \"' + str(i) + '\" class is ' + str(round(precs[i]*100,4)) + '%')\n",
    "        print('Recall for \"' + str(i) + '\" class is '  + str(round(recs[i]*100,4)) + '%')\n",
    "\n",
    "cm_metrics(confusion_matrix)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Average class accuracy (harmonic mean): 0.0%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\mtava\\Anaconda3\\lib\\site-packages\\ipykernel_launcher.py:12: RuntimeWarning: divide by zero encountered in reciprocal\n",
      "  if sys.path[0] == '':\n"
     ]
    }
   ],
   "source": [
    "# Evaluate average class accuracy using harmonic mean (according to the module book, arithmetic means are more\n",
    "#susceptible to influence of large outliers, which can inflate the apparent performance of a model. So I use harmonic mean)\n",
    "\n",
    "# function to calculate recall per level in a confusion matrix\n",
    "def calc_recall(cm):\n",
    "    recalls = [cm.iloc[i,i]/sum(cm.iloc[i,:]) for i in range(len(cm))]\n",
    "    return recalls\n",
    "\n",
    "# function to calculate average class accuracy using harmonic mean given a confusion matrix as dataframe\n",
    "def calc_avg_class_acc_hm(cm):\n",
    "    recalls = calc_recall(cm)\n",
    "    avg_class_acc_hm = 1/(1/len(recalls)*sum(np.reciprocal(recalls)))\n",
    "    return avg_class_acc_hm\n",
    "\n",
    "print(\"Average class accuracy (harmonic mean): \" + str(round(calc_avg_class_acc_hm(confusion_matrix)*100,4)) + \"%\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From the output above, Xgboost is basically unable to predict positive class label (imbalanced dataset)."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In code below, I also use normalized Gini index evaluation (kaggle competition metric)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Normalized gini coefficient: 0.2722\n"
     ]
    }
   ],
   "source": [
    "#https://www.kaggle.com/cpmpml/extremely-fast-gini-computation\n",
    "\n",
    "def ginic(actual, pred):\n",
    "    actual = np.asarray(actual) #In case, someone passes Series or list\n",
    "    n = len(actual)\n",
    "    a_s = actual[np.argsort(pred)]\n",
    "    a_c = a_s.cumsum()\n",
    "    giniSum = a_c.sum() / a_s.sum() - (n + 1) / 2.0\n",
    "    return giniSum / n\n",
    " \n",
    "def gini_normalizedc(a, p):\n",
    "    p = np.asarray(p)\n",
    "    if p.ndim == 2:#Required for sklearn wrapper\n",
    "        p = p[:,1] #If proba array contains proba for both 0 and 1 classes, just pick class 1\n",
    "    return ginic(a, p) / ginic(a, a)\n",
    "\n",
    "print(\"Normalized gini coefficient: \" + str(round(gini_normalizedc(y_test.iloc[:,1],y_pred_prob),4)))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Prediction Iteration 2 - Oversample rare class\n",
    "\n",
    "I will use a different split of the dataset in an attemp to overcome the class imbalance problem. As suggested in Case Study of the of the module book (page 479), I will divide the training set with an equal proportion of claimers and non-claimers.\n",
    "In this iteration, I will first oversample the test set to match the occurences of the rare class with the frequency of the majority class.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.utils import resample #using sklearn resample\n",
    "\n",
    "# get same instances of the previous train set with labels (and no id)\n",
    "X_train_prev = pd.concat([X_train.iloc[:,1:], y_train.iloc[:,1]], axis=1)\n",
    "\n",
    "claimers = X_train_prev[X_train_prev['target']==1] #get claimers subset\n",
    "nonclaimers = X_train_prev[X_train_prev['target']==0] #get nonclaimers subset\n",
    "\n",
    "# upsample rare class\n",
    "claimers_upsampled = resample(claimers,\n",
    "                          replace=True, # sample with replacement\n",
    "                          n_samples=len(nonclaimers), # match number in majority class\n",
    "                          random_state=1) # reproducible results\n",
    "\n",
    "# get upsampled claimers subset and nonclaimers subset together to make a new balanced train dataset\n",
    "train_upsampled = pd.concat([nonclaimers, claimers_upsampled])\n",
    "\n",
    "# get new X_train and y_train\n",
    "X_train2 = train_upsampled.loc[:,train_upsampled.columns != 'target']\n",
    "y_train2 = train_upsampled.iloc[:,-1]\n",
    "\n",
    "#train classifier (same classifier of first iteration but with different sampling)\n",
    "fit_model2 = clf1.fit(X_train2,y_train2)\n",
    "\n",
    "#create predictions vector and prediction probabilities. I will predict X_test instances (\"unseen\" instances)\n",
    "y_pred2 = fit_model2.predict(X_test.iloc[:,1:]) #prediction vector with classes\n",
    "y_pred_prob2 = fit_model2.predict_proba(X_test.iloc[:,1:]) #prediction probabilities vector [P(0), P(1)]\n",
    "\n",
    "pred_vs_true2  = np.c_[y_pred2,y_test.iloc[:,1]] # set up a 2-D array that has predictions and true (test) values side by side\n",
    "pred_vs_true_df2 = pd.DataFrame(data=pred_vs_true2,columns = [\"pred\",\"true\"]) # convert it to dataframe to be readable by evaluation algorithm\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(y_train)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Evaluation (Iteration 2) - Oversample rare class"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>0</th>\n",
       "      <th>1</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>107865</td>\n",
       "      <td>64229</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>2828</td>\n",
       "      <td>3642</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "        0      1\n",
       "0  107865  64229\n",
       "1    2828   3642"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "confusion_matrix2 = set_matrix(pred_vs_true_df2)\n",
    "confusion_matrix2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy is 62.4465%\n",
      "Precision for \"0\" class is 97.4452%\n",
      "Recall for \"0\" class is 62.678%\n",
      "Precision for \"1\" class is 5.3661%\n",
      "Recall for \"1\" class is 56.2906%\n"
     ]
    }
   ],
   "source": [
    "cm_metrics(confusion_matrix2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Average class accuracy (harmonic mean): 59.3128%\n"
     ]
    }
   ],
   "source": [
    "print(\"Average class accuracy (harmonic mean): \" + str(round(calc_avg_class_acc_hm(confusion_matrix2)*100,4)) + \"%\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Normalized gini coefficient: 0.2739\n"
     ]
    }
   ],
   "source": [
    "print(\"Normalized gini coefficient: \" + str(round(gini_normalizedc(y_test.iloc[:,1], y_pred_prob2),4)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Prediction Iteration 3 - Undersample majority class"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In this iteration I will also divide the training set with an equal proportion of claimers and non-claimers but undersampling the majority class, i.e., randomly removing majority class instances to reach a balanced dataset."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "# majority class undersampling\n",
    "nonclaimers_undersampled = resample(nonclaimers,\n",
    "                          replace=False, # sample without replacement\n",
    "                          n_samples=len(claimers), # match number in majority class\n",
    "                          random_state=1) # reproducible results\n",
    "\n",
    "# get sampled claimers subset and undersampled nonclaimers subset together to make a new balanced train dataset\n",
    "train_undersampled = pd.concat([nonclaimers_undersampled, claimers])\n",
    "\n",
    "# get new X_train and y_train\n",
    "X_train3 = train_undersampled.loc[:,train_undersampled.columns != 'target']\n",
    "y_train3 = train_undersampled.iloc[:,-1]\n",
    "\n",
    "#train classifier (same classifier of first iteration but with different sampling)\n",
    "fit_model3 = clf1.fit(X_train3,y_train3)\n",
    "\n",
    "#create predictions vector and prediction probabilities. I will predict X_test instances (\"unseen\" instances)\n",
    "y_pred3 = fit_model3.predict(X_test.iloc[:,1:]) #prediction vector with classes\n",
    "y_pred_prob3 = fit_model3.predict_proba(X_test.iloc[:,1:]) #prediction probabilities vector [P(0), P(1)]\n",
    "\n",
    "pred_vs_true3 = np.c_[y_pred3,y_test.iloc[:,1]] # set up a 2-D array that has predictions and true (test) values side by side\n",
    "pred_vs_true_df3 = pd.DataFrame(data=pred_vs_true3,columns = [\"pred\",\"true\"]) # convert it to dataframe to be readable by evaluation algorithm\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Evaluation (Iteration 3) - Undersample majority class"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>0</th>\n",
       "      <th>1</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>107525</td>\n",
       "      <td>64569</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>2865</td>\n",
       "      <td>3605</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "        0      1\n",
       "0  107525  64569\n",
       "1    2865   3605"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "confusion_matrix3 = set_matrix(pred_vs_true_df3)\n",
    "confusion_matrix3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy is 62.2354%\n",
      "Precision for \"0\" class is 97.4047%\n",
      "Recall for \"0\" class is 62.4804%\n",
      "Precision for \"1\" class is 5.2879%\n",
      "Recall for \"1\" class is 55.7187%\n"
     ]
    }
   ],
   "source": [
    "cm_metrics(confusion_matrix3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Average class accuracy (harmonic mean): 58.9061%\n"
     ]
    }
   ],
   "source": [
    "print(\"Average class accuracy (harmonic mean): \" + str(round(calc_avg_class_acc_hm(confusion_matrix3)*100,4)) + \"%\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Normalized gini coefficient: 0.2638\n"
     ]
    }
   ],
   "source": [
    "print(\"Normalized gini coefficient: \" + str(round(gini_normalizedc(y_test.iloc[:,1], y_pred_prob3),4)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From the outputs above, either oversampling or undersampling results in significant improvement of the algorithms' performance metrics. Xgboost has proven to outperform other algorithms tested previously (SGD Classifier, Decision Trees, Random Forest). The fact that both precision and normalized gini coefficient are higher for a dataset that has its rare class oversampled suggests that the loss of information in an undersampled dataset is penalising the algorithm."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Optimization"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "\n",
    "A trained XGBoost model automatically calculates feature importance on a predictive modeling problem. I will explore the built-in Plot Importance feature and then compute a feature selection also with built-in feature Importance Scores.\n",
    "I will work with the oversampled dataset that I used previously on Iteration 2 once this resulted in better performance metrics.\n",
    "\n",
    "The information regarding built-in functions for feature selection using xgboost were obtained in https://machinelearningmastery.com/feature-importance-and-feature-selection-with-xgboost-in-python/"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "#import libraries\n",
    "from matplotlib import pyplot as plt\n",
    "from xgboost import plot_importance\n",
    "from sklearn.feature_selection import SelectFromModel\n",
    "from numpy import sort"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAApwAAAHwCAYAAAALlb8ZAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADh0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uMy4xLjEsIGh0dHA6Ly9tYXRwbG90bGliLm9yZy8QZhcZAAAgAElEQVR4nOzde7yVc97/8dcnjWPSjTTIyIxUOtimBv3GdJDEYEhIt1GbZm4yOQwjuQ3CGCEjcd8hKjNjSJMIiYa2Yw5FkpvEaKYcopzaHdDu8/vj+u6sVuu69trHtdZe7+fjsR97r2td13d91+c2j/vb97q+76+5OyIiIiIi9aVJrjsgIiIiIo2bBpwiIiIiUq804BQRERGReqUBp4iIiIjUKw04RURERKReacApIiIiIvVKA04RkUbEzG4zs8ty3Q8RkVSmHE4RETCzpUAroCLl8H7u/mEt2uwF/NXdW9eud4XJzCYDy93997nui4jklmY4RUS+c6y7N0v5qfFgsy6YWdNcfn5tmNlWue6DiOQPDThFRKpgZoeY2Qtm9oWZvR5mLivfO93M3jKz1Wb2TzM7MxzfAXgM2MPMysPPHmY22cz+kHJ9LzNbnvJ6qZldbGYLgTVm1jRcN83MPjWz983s3IS+bmq/sm0zG2Fmn5jZR2Z2vJn93MzeMbPPzOy/U64dZWZ/N7Mp4fu8amYHpLzfwczKQh3eNLNfpH3ueDObaWZrgKHAqcCI8N0fDueNNLP3Qvv/Z2b9U9ooNbPnzGyMmX0evutRKe/vbGaTzOzD8P6DKe8dY2YLQt9eMLMuWf8fWETqnQacIiIJzGxP4FHgD8DOwO+AaWbWMpzyCXAM0Bw4HbjJzH7s7muAo4APazBjOgg4GmgBbAQeBl4H9gT6AOebWb8s2/o+sG249nJgAvBLoCvwM+ByM/thyvnHAVPDd/0b8KCZfc/Mvhf68QSwG3AOcI+ZtUu59j+Ba4AdgT8D9wDXh+9+bDjnvfC5OwFXAn81s91T2jgYWAzsClwP3GVmFt77C7A90DH04SYAM/sxMBE4E9gFuB2YYWbbZFkjEalnGnCKiHznwTBD9kXK7NkvgZnuPtPdN7r7bGAe8HMAd3/U3d/zyNNEA7Kf1bIf49x9mbuvA34CtHT3q9z9G3f/J9Gg8ZQs2/oWuMbdvwXuIxrI3ezuq939TeBNIHU2cL67/z2c/yeiweoh4acZMDr04yngEaLBcaWH3P35UKf1mTrj7lPd/cNwzhRgCXBQyin/cvcJ7l4B3A3sDrQKg9KjgLPc/XN3/zbUG+DXwO3u/pK7V7j73cDXoc8ikgcK9vkgEZF6cLy7/yPt2N7ASWZ2bMqx7wFzAMIt3yuA/Yj+Eb898EYt+7Es7fP3MLMvUo5tBTybZVurwuANYF34vSLl/XVEA8ktPtvdN4bb/XtUvufuG1PO/RfRzGmmfmdkZoOBC4A24VAzokFwpY9TPn9tmNxsRjTj+pm7f56h2b2BIWZ2TsqxrVP6LSI5pgGniEiyZcBf3P3X6W+EW7bTgMFEs3vfhpnRylvAmWJA1hANSit9P8M5qdctA95397Y16XwN7FX5h5k1AVoDlY8C7GVmTVIGnT8A3km5Nv37bvbazPYmmp3tA8x19wozW8B39UqyDNjZzFq4+xcZ3rvG3a/Joh0RyQHdUhcRSfZX4Fgz62dmW5nZtmExTmuiWbRtgE+BDWG284iUa1cAu5jZTinHFgA/Dwtgvg+cX8Xnvwx8FRYSbRf60MnMflJn33BzXc3shLBC/nyiW9MvAi8RDZZHhGc6ewHHEt2mj7MCSH0+dAeiQeinEC24Ajpl0yl3/4hoEdb/mtl/hD70CG9PAM4ys4MtsoOZHW1mO2b5nUWknmnAKSKSwN2XES2k+W+igdIy4CKgibuvBs4F7gc+J1o0MyPl2reBe4F/hudC9yBa+PI6sJToec8pVXx+BdHArgR4H1gJ3Em06KY+PAQMJPo+pwEnhOclvwF+QfQc5Urgf4HB4TvGuQvYv/KZWHf/P+BGYC7RYLQz8Hw1+nYa0TOpbxMt1jofwN3nET3HeWvo97tAaTXaFZF6puB3EREBolgkYF93/2Wu+yIijYtmOEVERESkXmnAKSIiIiL1SrfURURERKReaYZTREREROqVBpwiIiIiUq8U/J7HWrRo4fvuu2+uu5GX1qxZww477JDrbuQt1SeeapNM9Ymn2iRTfeIVS23mz5+/0t1bZnpPA8481qpVK+bNm5frbuSlsrIyevXqletu5C3VJ55qk0z1iafaJFN94hVLbczsX3Hv6Za6iIiIiNQrDThFREREpF5pwCkiIiKSpxYvXkxJScmmn+bNmzN27FimTp1Kx44dadKkSUE8fqdnOEVERETyVLt27ViwYAEAFRUV7LnnnvTv35+1a9fywAMPcOaZZ+a4h9kp6hlOMzvLzAZX85oyM+uW8H5XM3vDzN41s3FmZuH41Wa20MwWmNkTZrZHbfsvIiIixePJJ5/kRz/6EXvvvTcdOnSgXbt2ue5S1op6wOnut7n7n+u42fHAfwFtw8+R4fgN7t7F3UuAR4DL6/hzRUREpBG77777GDRoUK67USMFfUvdzNoAs4CXgAOBd4DBRIO5XwAbgCfc/Xcx148Cyt19jJmVhXZ6Ay2Aoe7+rJltB0wC9gfeArZL6M/uQHN3nxte/xk4HnjM3b9KOXUHoMo9Rdd9W0GbkY9WdVpRurDzBkpVm1iqTzzVJpnqE0+1Sab6xEuqzdLRR2fVxjfffMOMGTO49tpr67JrDaagB5xBO6LB4fNmNhEYDvQH2ru7m1mLarTV1N0PMrOfA1cAhwPDgLXu3sXMugCvJly/J7A85fXycAwAM7uGaED8JdHAdgtm9l9EM6TsumtLLu+8oRrdLx6ttov+ByyZqT7xVJtkqk881SaZ6hMvqTZlZWVZtfHcc8+xzz778NZbb/HWW29tOv7FF18wf/58ysvL66Kr9aYxDDiXufvz4e+/AhcA64E7zexRotvX2Xog/J4PtAl/9wDGAbj7QjNbmHC9ZTi2aSbT3S8FLjWzS4gGxldscbL7HcAdAO3atfNzTj2uGt0vHmVlZZxcBCG6NaX6xFNtkqk+8VSbZKpPvLqozW233cbZZ5+9RYB8ixYt6Nq1K926xS4vyQuN4RnO9FvT3wIHAdOIbmfPqkZbX4ffFWw+GK/y9newHGid8ro18GGG8/4GDKhGv0RERKRIrV27ltmzZ3PCCSdsOjZ9+nRat27N3LlzOfroo+nXr18Oe1i1xjDD+QMz6x6emxwELAB2cveZZvYi8G4t238GOBWYY2adgC5xJ7r7R2a22swOIXoedDBwC4CZtXX3JeHUXwBv17JfIiIiUgS23357Vq1atdmx/v37079//xz1qPoawwznW8CQcKt7Z+BO4JHw+mngt7VsfzzQLLQ3Ani5ivOHhT68C7wHPBaOjzazRaGdI4DzatkvERERaeQU/J4/Nrr7WWnHDsrmQncflfJ3r5S/VxKe4XT3dcAp2XbG3ecBnTIc1y10ERERqZbGEvzeGAacecXMhgPnAz8CWobBK2Z2HHA1sJEorul8d38uZx0VERGRgpIa/F5oCnrA6e5LyTCbmM7MLgVOSjs81d2vqelnm9lLwDZph08DnidaGV+W9t6TwIwQ1dQFuB9oX9PPFxERkeJSyMHv5p7tAuz8UwfB762A24AfhkPD3P0FM3sQ2AvYFrg5RBVhZuXAn4B+wIVJM5RmthToVjnDmfZed2Ciu3dI+n4/+OG+3uTkm5NOKVoXdt7AjW8U9L+X6pXqE0+1Sab6xFNtkqk+8ZJqU53g9z322IM333yTVq1abTreq1cvxowZkxexSGY2390zdqQx/JdRm+D3ccDT7t7fzLYCmoXjZ7j7Z2GXoVfMbJq7ryLaIWiRu9doW0oz6w9cC+wGZPwvTMHv2VHAcDLVJ55qk0z1iafaJFN94in4HXD3gv0hWtjz75TXhxHdzn4duAs4Adg64fpPgW0yHB8V2nidaFegQ8LxDcBWWfZtKbBrzHs9gH9U1cZ+++3nktmcOXNy3YW8pvrEU22SqT7xVJtkqk+8uqjNwIEDfeLEiVsc79mzp7/yyiu1br8uAPM8ZkzTGGKR6jL4HTPrRbSlZXd3PwB4jejWOsB6d6+oVW8Bd38G+JGZ7VrbtkRERKRxU/B7fqhN8PuTRLmZY8Mt9R2AnYDP3X2tmbUHDqmLTprZvsB77u5m9mNga2BVFZeJiIhIkVPwe36oTfD7eUBvM3uDaP/0jkQzok3D9VcDL1anM2Z2rplVbnG50MzuDG8NABaZ2QLgf4CBYfpZREQkUZs2bejcuTMlJSWbFodcdtlldOnShZKSEo444gg+/DDTTsoi+aExzHDWJvh9BXBchreOijm/WabjaeeMI1qMlH78OuC6bPolIiKSbs6cOey663dPYl100UVcffXVAIwbN46rrrqK2267LVfdE0nUGGY484qZdTWzN8zsXTMbZ2YWjl9tZgvNbIGZPWFme+S6ryIiUriaN2++6e81a9YQ/t+NSF4q6AGnuy9196yC38NAL/Xn0pT3qz3Ta2bTM7TZj2jv9f8C2oafI8MlN7h7F3cvIVpJX6NoJRERKT5mxhFHHEHXrl254447Nh2/9NJL2Wuvvbjnnnu46qqrcthDkWQFG/xeB6Hvk4HPwrWvhutuAToTPWowyt0fMrPtgclEuwK9RRTF9BuP9kxPb3N3YI67tw+vBwG93P3MtPMuAX7g7sOSvqOC3+MpYDiZ6hNPtUmm+sSrr9pkE/z94Ycfsscee/DJJ5/Qt29fbrnlFnr06LHp/WuvvZb169dz5ZVX1nn/slVWVkavXr1y9vn5rFhq05iD32sT+g6wH3C4u1eY2R+Bp9z9jHDdy2b2D6JV7J+7excz60S0Cj7OnsDylNfLwzEAzOwaokHxl0DvTA0o+D07ChhOpvrEU22SqT7x6qs22QZ/v/POOwAceOCB3HvvvWzcuHHTe/vssw+XXHIJvXtn/H8tDaK8vDzr71JsVBsKN/id2oe+TwaGpLyeBywiGlAuAP4NdAAeBHqnnPcq0ZaVmdr8CSmB7sDPgIcznHcJcGVV31HB7/EUMJxM9Ymn2iRTfeLlqjbl5eX+1Vdfbfq7e/fu/thjj/k777yz6Zxx48b5gAEDctK/SvpvJ16x1IaE4PdCn+GMC33vA5xCNON5WML1a1L+NmCAuy9OPcGq9xR2ZRxSpdZAppyKvwGPAldUo20RESlCK1as2JS3uGHDBv7zP/+TI488kgEDBrB48WKaNGnC3nvvrRXqktcKfcBZm9D3dI8D55jZOe7uZnagu78GPAecDMwxs/2JnvHMyN0/MrPVZnYI0bOlg4meC8XM2rr7knDqL4C3q/ldRUSkCP3whz/k9ddf3+L4tGnTctAbkZop6FXq1C70Pd3VwPeIwtoXhdcA/wu0DG1eDCwkegYzzrDQj3eB94DHwvHRZrYotHMEUei8iIgEFRUVHHjggRxzzDEADB06lAMOOIAuXbpw4oknUl5enuMeikhNFfoMZ21C30vTXq8Dzsxw6nrgl+6+3sx+RLQd5r8S2p0HbBHV5O4DsumXiEixuvnmm+nQoQNfffUVADfddNOmrMkLLriAW2+9lZEjR+ayiyJSQ4U+w9kQtgeeM7PXgenAMHf/Ju5kMxseQt/dzHbN8P5PzKzCzE6sxz6LiBSU5cuX8+ijj/KrX/1q07HKwaa7s27dOgWbixSwgp3hdPelZJhJTBcC3k9KOzzV3a/J8nNWA1tkSpnZS8A2aYdPA54nWi1fluGarYi2t3w8m88WESkW559/Ptdffz2rV6/e7Pjpp5/OzJkz2X///bnxxhtz1DsRqa2CHXBCtcLfS2KubwXcBvwwHBrm7i+Y2YPAXsC2wM3ufkc4vxz4E9APuNDdn0voW6bD5wDTiOKTqrTu2wrajHw0m1OLzoWdN1Cq2sRSfeKpNsnqoz5VBZs/8sgj7LbbbnTt2nWLrMJJkyZRUVHBOeecw5QpUzj99NPrtG8i0jAKdqch2DTgfB841L8Lf38bGEpK+Lu7fxFz/RRgrruPDbOPzdz9SzPb2d0/M7PtgFeAnu6+yswcGOju92fRt6VEeZ0rw+s9ieKQDiPKCX3E3f+e4brU4Peul4+dUK2aFItW28GKdbnuRf5SfeKpNsnqoz6d99wp8f0JEybwxBNPsNVWW/HNN9+wdu1afvazn3HppZt2IGbBggVMmTKFa6+9tm47Vw3l5eU0a9YsZ5+f71SfeMVSm969e8fuNJTzAPfa/FD78PdPgW0yHB8V2nidaEX6IeH4BmCrLPu2FNg15fXUlHYmAydW1YaC3+MVS4huTak+8VSbZLmuz5w5c/zoo4/2jRs3+pIlS9zdfePGjX7hhRf6hRdemPO+STzVJ16x1IZGHPwOtQ9/34yZ9QIOB7q7+1ozKyO6tQ6w3t0ratjPbsB94Vb7rsDPzWyDuz9Yw/ZERBotd2fIkCF89dVXuDsHHHAA48ePz3W3RKSGGsOAszbh708S5WZW3lLfAdiJaO/0tWbWHjikLjrp7vtU/m1mk4luqWuwKSKSolevXvTq1QuA559/PredEZE60xhikWoT/n4e0NvM3gDmAx2JFiE1DddfDbxYnc6Y2blmVrnF5UIzu7O6X0hERESkMWkMM5y1CX9fARyX4a2jYs6v8olfdx8HjKvinNJs+icikmT9+vX06NGDr7/+mg0bNnDiiSdy5ZVX4u78/ve/Z+rUqWy11VYMGzaMc889N9fdFZEiVlADTjM7C1jr7n+uxjVlwO882gEo0/tdiRbxbAfMBM5zdzezUcCviRYWAfy3u8+see83+8xSorimD+uiPREpTttssw1PPfUUzZo149tvv+XQQw/lqKOO4q233mLZsmW8/fbbNGnShE8++STXXRWRIldQA053vy3t9VKyCH8Hzshwa7sy/H08UQzRi0QDziP5bv/zm9x9TKYGzWw6sE/a4YvdPZtQ91JgEaABp4jUmJltilr59ttv+fbbbzEzxo8fz9/+9jeaNImemtptt91y2U0RkYYdcFYjqP13MdePAsrdfUyYuXwJ6A20AIa6+7MhO3MSsD/R853bARPd/ewM7e0ONA8LjjCzPwPH892AM5a7909pp3IHoTFmdgMwwd1vMbPLgWNDH14g2qt9ANGK9XvMbB3RaviMqXcKfo+n8O5kqk+8QqpNVYHpABUVFXTt2pV3332X3/zmNxx88MG89957TJkyhenTp9OyZUvGjRtH27ZtG6DHIiKZ5WKGsx3R4LAyqH040J+UoPZqtNXU3Q8ys58DVxDFGQ0juu3excy6AK8mXL8nsDzl9fJwrNJwMxsMzCPaWejzmHb+i2i280B332BmO4fjt7r7VQBm9hfgGHf/u5kNJ+Y2f1rwO5d33pBYgGLVarto4CCZqT7xCqk26bvuxBk7dizl5eVcdtlltG/fnrVr1/LBBx8wZswYnnnmGQYMGMC4cYmPlm9SXl6e9ecWG9UmmeoTT7XJzYBzmbtXZl38FbgAWA/caWaPEgW3Z+uB8Hs+UQg8QA/Coh13XxhWm8fJtP9kZa7neKJV6h5+3wicEdPO4cBt7r4hfO5n4XhvMxsBbE+0gv5N4OGkL+TRNpp3ALRr187POTXTmiYpKyvj5BCdIltSfeI15trMnz+fVatWsffeezNixAjatGlDz549ufHGGzdFDVWlrKws63OLjWqTTPWJp9rkJhYpLqh9GtHt7FnVaOvr8LuCzQfP2e7XWRlfVKk14blKd1/h7hXuvhGYQPLKd0v/TDPbFvhfoh2FOoc2ts1wrYhIjXz66ad88UW0c++6dev4xz/+Qfv27Tn++ON56qmnAHj66afZb7/9ctlNEZGczHDWJqg9G88ApwJzzKwT0CXuRHf/yMxWm9khRM+DDgZugej5Tnf/KJzan2iRT5wngLPMrCzllvrG8N5KM2sGnAhU7p2+GtixZl9PRCTy0UcfMWTIECoqKti4cSMnn3wyxxxzDIceeiinnnoqN910E82aNePOOxUHLCK5lYsBZ2VQ++3AEqJ9yx8JM4JGclB7NsYDk8Kt9AXAy1WcP4zvYpEe47sFQ9ebWQnRzOVSogU/ce4E9iMKev+WaNHQrWY2AXgjXP9KyvmTgduqWjQkIpKkS5cuvPbaa1scb9GiBY8+WhgLo0SkOORiwFmboPZRKX/3Svl7JeEZzjB4OyXbzoSFO1tEK7n7adVoYwPRs6gXpB3/PfD7DOdPI3qEQEQkUVy4e2lpKU8//TQ77bQTAJMnT6akpCTHvRURyaygcjgLQUKQ/ElEs7kdgIPiguhFRFLFhbsD3HDDDZx44ok57qGISNUadMCZbVC7mV0KnJR2uDKovUbM7CVgm7TDp7n7G2bWtHKFeRVt9CPK20z1fmomJ/FB8ouAE4Dba/gVRKQIxYW7i4gUkryc4QwDy8TBZXVD5N394LTrJwOnm9mBwKshpP0WoDNRXUa5+0Nmtj3RjGV7oudPvwZ+E5OhGRsk7+5vhWNZ10HB7/EKKbw7F1SfePlWm5qGu48fP55LL72Uq666ij59+jB69Gi22Sb939QiIvnB3LNNEMovYcD5PnBoSoj828BQUkLk3f2LmOsnA7sCx7l7hZn9Efg/d/9rCJ9/mWggOwxo6+5nhlXvC4BDYgac3YDR7n54eP0zou0uj0k5p4zkvd1Tg9+7Xj52QnVLUxRabQcrtNQqluoTL99q03nPnbI+tzLc/dxzz6V58+bsvPPOfPvtt9x4443sscceDBkypNb9KS8v3zSjKptTbZKpPvGKpTa9e/ee7+7dMr2XlzOc1VDbEPmp7l4R/j4C+IWZVW6ruS3wA+BQ4GYAd19UiyD5rCj4PTuNOby7Lqg+8Qq9NpXh7qeffvqmY1tvvTVjxoypk2BpBVTHU22SqT7xVJvcBL/XpdqGyK9J+duAAe5eEn5+EG6DV+dhqdggeRGRmogLd//ooygm2N158MEH6dSpysfjRURyptBnOOsyRP5x4BwzOyfcjj/Q3V8DngNOJgqS35/oGc+MkoLkRURqIi7c/bDDDuPTTz/F3SkpKeG2227LdVdFRGIV+oCzLkPkrwbGEoW3G1FY+zFE21PeHW6lvwYsBL5MaCdjkLyZ9ScafLYEHjWzBe7erxr9E5EiFBfuXrl1pYhIISj0AWdtQuRL016vI/NuQuuBX7r7ejP7EfAk8K+EduOC5KcD07Ppm4hIJQW/i0hjUOgDzoawPdHt9O8RzZoOc/dvctwnESkSCn4XkcagYAecdREib2ZnAWvd/c8Jn7Ma2LTE38zKzOxT4H/IECRPtK3mYOA/3L1ZynWlwA3AB+HQre5+Z1X9F5HipuB3EWkMCnbAma2kEHl3r/FT9ulB8pXMbAfgVqJnStNNcffh2X6Ggt/j5Vt4d75RfeLlW20U/C4ixaBgg9+h+rsNZbh+FFDu7mNCIPtLQG+gBTDU3Z81s+2AScD+RIuU2hCz01Ba2+UZZji7VTXgVPB7dvItvDvfqD7x8q02Cn4vHKpNMtUnXrHUpjEHvwO0IxocVu42NBzoT8puQ9Voq6m7H2RmPweuAA4nWnW+1t27mFkX4NVa9HWAmfUgGhj/1t2XpZ+g4PfsFHp4d31TfeIVem0U/J47qk0y1SeealP4we+w5W5DPfhut6ETgLXVaOuB8Hs+0Uwmob2/Arj7QqJYpJp4GGjj7l2AfwB317AdESkiCn4XkcagMcxwxu021IdoAc9w4LAs2/o6/K5g89rU+rkDd1+V8nICcF1t2xSRxk/B7yLSGDSGAWdd7jaUyTPAqUTRSJ2ALjVpxMx2d/ePwstfED0PKiKSSMHvItIYNIZb6pW7DS0EdgbuJNptaCHwNNXbbSiT8UCz0N4I4OWkk83sejNbDmxvZsvDwiSAc83sTTN7HTgXKK1lv0SkhpYtW0bv3r3p0KEDHTt25Oabbwbg9ddfp3v37nTu3Jljjz2Wr776Ksc9FRFpHBrDDGdtdhsalfJ3r5S/VxKe4Qw7EJ2SbWfcfQTRwDT9+CXAJdm2IyL1p2nTptx44438+Mc/ZvXq1XTt2pW+ffvyq1/9ijFjxtCzZ08mTpzIDTfcwNVXX53r7oqIFLzGMMOZV8ysq5m9YWbvmtm4sC87ZnaDmb1tZgvNbHo1V8+LSB3afffd+fGPfwzAjjvuSIcOHfjggw9YvHgxPXr0AKBv375MmzYtl90UEWk0CnqGsy52GwrvN3X3DdX5bDN7icw7DY0nytF8EZgJHAk8BswGLnH3DWZ2HdFs58VJn6Hg93j5Ft6db4q9PtmEqW86d+lSXnvtNQ4++GA6derEjBkzOO6445g6dSrLlm2RXCYiIjVQsMHvdRD6Phn4LFz7arjuFqAz0UB8lLs/ZGbbA5OB9lQR/G5muwNz3L19eD0I6OXuZ6ad1x840d1PzdCGgt+zkG/h3fmm2OuTFKaeGsC8bt06zjvvPH75y1/So0cP/v3vf3PLLbfw5Zdf8tOf/pQHHniAhx56qKG6nReKJaC6JlSbZKpPvGKpTVLwO+5ekD9EAz8HfhpeTyR6dnIx3w2kWyRcPxl4BNgqvP4j8MvK64gGsDsAvwNuD8c7EQ1ku8W02Q34R8rrnwGPZDjv4crPSvrZb7/9XDKbM2dOrruQ11SfeJW1+eabb/yII47wG2+8MeN5ixcv9p/85CcN2LP8oP924qk2yVSfeMVSG2Cex4xpCv0ZztqGvk9194rw9xHASDNbAJQB2wI/AA4F7gNw90UkB79bhmObTSGH2/sbgHuq6JuI1BN3Z+jQoXTo0IELLrhg0/FPPvkEgI0bN/KHP/yBs85KX48oIiI1UdDPcFL70Pc1KX8bMMDdF6eeULnoJ0vLgdYpr1sDH6a0NQQ4BugT/iUgIjnw/PPP85e//IXOnTtTUlICwB//+EeWLFnC//zP/wBwwgknbLZ9pIiI1FyhDzjrMvT9ceAcMzvH3d3MDnT314DngJOJgt/3J3rGMyN3/8jMVpvZIUTPlg4mei4UMzuSaJFQT3evznabIlLHDj30UOL+zYS9IsAAACAASURBVHfeeec1cG9ERBq/Qr+lXpeh71cD3wMWmtmi8Brgf4GWoc2LiW6pf5nQzrDQj3eB94hWqAPcCuwIzDazBWamfegkL8WFok+dOpWOHTvSpEkT5s3bYs2ciIhIrEKf4axN6Htp2ut1wJkZTl1PtMBnvZn9CHgS+FdCu/PIENXk7vtm0y+RXIsLRe/UqRMPPPAAZ56Z6X8mIiIi8Qp9hrNWzGwfM3vJzJaY2RQz2zoc72Fmr5rZBqJb9c+FLSmnA8Pc/ZtafObx4da8SF6KC0Xv0KED7dq1y3HvRESkEBXsgNPdl7p7VqHv4RZ26s+l4e3rgJvcvS3wOTA0HP830V7nfwPWu3s3dz/A3bu4+2Oh3ZcytBv7fGeK4wENOKUgpIaii4iI1FSDBb/XQVB7K+A24Ifh0DB3f8HMHgT2Iooxutnd7wjnlwN/AvoBF7r7c2ntGfAp8H2Pdv/pThT23i/lnMlEOZp/r+K7jSDaZWgj8Ji7jzSzXxMFuG9N9DznaUAJUfbnl+FngLu/l9aWgt+zUOzB5lVJqk9SKHqq9FD0Sueffz7Dhg0r2NnOYglgrinVJ55qk0z1iVcstcmL4HdqH9Q+BTg//L0V0Wp0gJ3D7+2ARcAu4bUDJye0tyvwbsrrvYBFaedMJtoRKOl7HQW8AGyf1p9dUs75A3BOtm1W/ij4PV6xhOjWVG3rkxSK3rNnT3/llVdq1X4u6b+dZKpPPNUmmeoTr1hqQx4Fv9cmqP0won3KcfcKd69cKX5ueL7yRaJBY9twvAKYltBelSHtWTocmOQh6sjdPwvHO5nZs2b2BnAq0LEGbYs0OI8JRRcREamphh5wxgW1TyN6tnFWdRozs15EA77u7n4A8BrRrXWInr2siLsWWAm0MLPKlfqbhbRXpxtkHqhOBoa7e2fgypR+ieS1ylD0p556ipKSEkpKSpg5cybTp0+ndevWzJ07l6OPPpp+/fpV3ZiIiAgNH4tUm6D2J4kyLsea2VZE+5zvBHzu7mvNrD1wSLYdcXc3sznAiURbVw4BHqrBd3oCuNzM/hb6sXOY5dwR+MjMvkc0w/lBOH91eE8kLyWFovfv37+BeyMiIo1BQ89w1iao/Tygd7hFPZ/oFvUsoGm4/mqi2+rVcTFwgZm9C+wC3AVgZj8xs+XAScDtZvZmXAPuPguYAcwL+7BXLnq6jGiB1Gzg7ZRL7gMuMrPXQq6nSIOLC3f/7LPP6Nu3L23btqVv3758/vnnOe6piIg0Bg09w1mboPYVwHEZ3joq5vwql4O5+z8zfb67v8Lme6JX1c5oYHTasfGEZ07Tjj+PYpEkx+LC3SdPnkyfPn0YOXIko0ePZvTo0Vx33XW57q6IiBS4gs3hrAtmdpaZDa7mNWVmlnnJf/R+VzN7w8zeNbNxIX4JMzvJzN40s41J14s0hLhw94ceeoghQ4YAMGTIEB588MFcdlNERBqJBpvhdPelZNjyMV0IZT8p7fBUd7+mpp9tZtOBfdIOX+zuWe9nHkLd/wLsC9xjZuuAr909PRF7PFGO5ovATOBIov3UFwEnALfX6EuI1JPUcPcVK1aw++67A9Gg9JNPPslx70REpDHIu73Uw8Ayq8FlNcLkM650MLNRQLm7jzGzstBOb6AFMNTdnzWz7YBJRLfB3wK+Bn7j0Z7p6e3tDjQPi6Iwsz8Trb5/zN3fCsey+WoArPu2gjYjH836/GJyYecNlKo2sSYfuUNW55WXlzNgwADGjh1L8+bN67lXIiJSrPJuwFkD7YgGh8+b2URgONAfaB9WoreoRltN3f0gM/s5cAVR5NIwYK27dzGzLsCrCdfvCSxPeb08HMta6k5DLVu25P4sBw7Fpry8POtBVTEqLy+nrKws8ZwNGzZwySWXcPDBB7PzzjtTVlZG8+bNmTZtGrvssgurVq1ixx13rLKdQpNNbYqZ6hNPtUmm+sRTbRrHgDM9TP4CvguTf5RoK8lsPRB+zyfaGQmicPpxAO6+MKyIj1PrMHmPtua8A6Bdu3beq1ev6lxeNMrKylBt4lVVH3dnyJAh/PSnP2Xs2LGbjg8cOJAlS5YwYMAARo8ezSmnnNLo6qz/dpKpPvFUm2SqTzzVpnEsGqrLMPmvw+8KNh+MZztoXM7mq9trGiYvUq/iwt1HjhzJ7Nmzadu2LbNnz2bkyJG57qqIiDQCjWGGszZh8tl4hii4fY6ZdQK6xJ3o7h+Z2WozO4ToedDBwC21/HyROpcU7v7kk082cG9ERKSxawwznLUJk8/GeKBZaG8E8HIV5w8LfXgXeI9ohTpm1j+EyXcHHjWzx2vZL5FEZ5xxBrvtthudOn0XDvH666/TvXt3OnfuzLHHHstXX32Vwx6KiEixaAwznLUJkx+V8nevlL9XEp7hdPd1wCnZdiasXt8i/sndpwPTs21HpLZKS0sZPnw4gwd/FzX7q1/9ijFjxtCzZ08mTpzIDTfcwNVXX53DXoqISDEomBnOegppv8bMlplZedrxm8xsQfh5x8y+SGijl5llXJhkZjOruUpepM706NGDnXfeebNjixcvpkePHgD07duXadOm5aJrIiJSZApmhjNTSHttw+SBh4FbgSVp7W66DW9m5xBlfKa3+RKwDdAMaBn2UT/N3d9IaefnVfVNpCF16tSJGTNmcNxxxzF16lSWLVuW6y6JiEgRaLABZzVC2n8Xc/0oqh/Svh1UHSZfRRj7IKJMzs1U7jBkZr2Aq4BVwBQzewY42903mtlSoBvRoPQx4Dng/wEfAMeF2/WxFPweT8HvsHT00dW+ZuLEiZx77rlcddVV/OIXv2Drrbeuh56JiIhsrqFnOPMppL1KZrY30ZaYT1Vx6kFEg9x/EQ2qTwD+nnZOW2CQu//azO4HBhDlhqZ/poLfs6DgdxJDhCtDhj/++GPWrFmz2bn//d//DcCyZcvYbbfdii6MWAHMyVSfeKpNMtUnnmrT8APOfAppz8YpwN/dvaKK8152938CmNm9wKFsOeB8390XZOjzZhT8nh2F6CarrM/SpUvZYYcdNtXqk08+YbfddmPjxo2UlpZy0UUXFV0d9d9OMtUnnmqTTPWJp9o0/KKhfAppz8YpwL1ZnJf+mZn68HXK3+l9FqlzgwYNonv37ixevJjWrVtz1113ce+997LffvvRvn179thjD04//fRcd1NERIpAQw968iakvSpm1g74D2BuFqcfZGb7EN1SH0iYoRTJpXvvzfxvpfPOO6+BeyIiIsWuoWc48yqk3cyuD2Hs25vZ8rAwqdIg4D6P245lc3OB0cAi4H2UtykNJFO4+4IFCzj77LMpKSmhW7duvPxyVXsViIiI1K+GnuHMt5D2EUQD08TPq6KNMqAs5r024c+VpMQ3ufuYbPsokiRTuPuIESMYMmQIF198MTNnzmTEiBFF/7C6iIjkVsEEv9cHM9vHzF4ysyVmNsXMtg7Hsw5+r8FnHm9m+9dVe1LcMoW7mxlr1qwB4Msvv2SPPfbIRddEREQ2abAZztqGtIcszRpJCWlPdRpwGXCTu99nZrcBQ4HxmYLfzawz8Je0Nr6uzOOshuOJVuP/XzWvE8nK2LFj6dWrF5MmTWLjxo288MILue6SiIgUuXwNft9icGlmrYDbgB+GQ8Pc/QUzexDYC9gWuDnEChG2q/wT0A+40N2fS2vPgMOA/wyH7gZGET0HmmoQcEXYQagk5ruNIBrAbgQec/eRZvZrojzNrYkWQ50Wrv8F0NPMfg8McPf34mqm4Pd4xRL8XpNw9/Hjx3P22WdzxRVXcP/99zN06FD+8Y9/1EPvREREsmPZrYmpgw+KBpzvA4emBL+/TTSruCn43d0z3r42synAXHcfa2ZbAc3c/Usz29ndPwu7DL0C9HT3VWbmwEB3vz+mvV2BF9193/B6L6LBYqeUc/YGXgRax2VxmtlRRDOlh7v72pT+7OLuq8I5fwBWuPstZjYZeMTd03M6K9vbFPy+664tu14+dkJcSYtaq+1gReI+TY1D5z13qvKcjz/+mEsuuYRJkyYBcMwxx3Dvvfey44474u4cc8wxPPpo4x+cZ6u8vJxmzZrluht5S/WJp9okU33iFUttevfuPd/du2V6r5CC3w8jmhElDP6+DMfPNbP+4e+9iHb0WUWUdTktob1M+1mmj76zCX4/HJjk7mtD3z4LxzuFgWYLoq0tH09o47sOpAW/n3PqcdlcVnTKyso4uchDdCulh7vvtddevPfee5x//vk8+eSTtG/fvugDh1MpgDmZ6hNPtUmm+sRTbRp+wBkX/N6HaHA3nGhgmZWwj/nhQPcwu1hGdGsdYH0VA8WVQAsza+ruG4DWwIdp55wC/KaqbpA56H0ycLy7v25mpUCvKtoRqbZBgwZRVlbGypUrad26NVdeeSUTJkzgjDPOYNKkSWy77bbccYdiYUVEJLcKKfj9SaK90itvqe8A7AR8Hgab7YFDsu1IuIU/BzgRuA8YAjxU+X41gt+fAC43s7+l3lIHdgQ+MrPvEYXRfxDOXx3eE6m1uHD3O+64o+j/NS0iIvmjkILfzwN6m9kbRHuRdyRahNQ0XH810fOW1XExcIGZvQvsAtyV8l5Wwe/uPguYAcwzswXA78JblxEtkJpN9KxqpfuAi8zsNTP7UTX7K0UoU7j7wIEDKSkpoaSkhDZt2lBSknE9m4iISF4opOD3FUCmBxqPijm/yqdz3f2fcZ+fbfB7OHc00U5DqcfGs+WKd8IzrMrhlKxlCnefMmXKpr8vvPBCdtqp6sVFIiIiuaLg98zB73ub2ZNmttDMysysdR1+poLfpVoyhbtXcnfuv/9+Bg0a1MC9EhERyV6DDTjdfWlq5FAcM7s0ZZefyp9La/PZZjY9Q5v9gOuIgt/bAp8TRTQBjAH+7O5dgKuAa82sc4Y2XqpBd45HM5xSR5599llatWpF27Ztc90VERGRWMUe/H4PmYPf9+e750nnAA+6e2Vwe6bvpuD3BtZYgt9rEuye6t5779XspoiI5D0Fv2cIfjezvwEvufvNZnYCUZ7nrpVB7mntKPg9BxpL8Hs2we6wZbg7QEVFBSeddBK33347LVu23Oz8YgkZrgnVJpnqE0+1Sab6xCuW2ij4PbOk4PffAbeG/MxniCKNNsS0o+D3HCi24Pf0cHeAWbNm0blzZ0466aQtzlfIcDzVJpnqE0+1Sab6xFNtGn7RUFzw+zSiZxtnVaextOD3A4DXqEHwe3i9Kfjd3T909xPc/UDg0nDsy8zNJAa/D3f3zsCVKf0SqZZBgwbRvXt3Fi9eTOvWrbnrrii967777tPtdBERKQgKfs8Q/B5ut3/m7huBS4CJCU0p+F3qVVy4++TJkxu2IyIiIjWk4PfMwe+9gMVm9g7QigyLmCop+F0qZQporzRmzBjMjJUrV+agZyIiIrml4PcMnx8W9GRc1BPTjoLfJWNAO8CyZcuYPXs2P/jBD3LUMxERkdwq6uD3qpjZ0nB7vTrXbBNC5N8NofJtwvFdzGyOmZWb2a310V/JrbiA9t/+9rdcf/31RElcIiIixafBZjjdfSmQVfA7kL7sdqq7x97WzqLN6cA+aYcvdvesVo6HNjoDf0k7/LW7H5x2bCjRc6X7mtkpROHyA4lW419GVIMq6yCNw4wZM9hzzz054IADct0VERGRnGnoW+pVCgPLjIPLaobHp7bZP1yfGh5/vZmtTgqPT3Mg0YywAwtDEHwmxxEFyEN0W/5WMzN3XwM8Z2b7VlWDSgp+j5eL4PfqhrSvXbuWa665hieeeKKeeiQiIlIY8m7AmYV2wNCU8PjhQH9SwuMTrh0HPO3u/SvD48PxM1LD481sWmrIu5l1JIpH+qm7rzSzzBtbR/YElgG4+wYz+5JoQVJWq0XSgt+5vHNc/Gdxa7VdNOhsSGVlZVWe8/HHH7NmzRrKysr45z//yTvvvEO7du0A+PTTT+nYsSPjx4+P3Ru9rpSXl2fV32Kk2iRTfeKpNslUn3iqTQPuNFQXwgznM+7+g/D6MKLw+L2AecCjRLv4fBNz/adAa3f/Ou34KKJBK0AboJ+7v2hmS4FuRBFO33f3Kvd0N7M3w/XLw+v3gINSdh0qBbq5+/Cq2mrXrp0vXry4qtOKUr6G6C5dupRjjjmGRYsWbfFemzZtmDdvHrvuWq3HgmskX+uTD1SbZKpPPNUmmeoTr1hqY2axOw0V4qKhhgyP33Rahs+Ns5xoAEwIld8J+CzxCmkU4gLaRUREil0h3lLPRXj8k8B0M7sp7NNeGe6eyQyiEPm5RKHyT3khTSNLjcUFtFdaunRpw3REREQkzxTiDGeDh8e7+5tEC5meNrPXgT8lfMZdwC4hTP4CYGTlG+EW/Z+AUjNbbmbK48xDmQLcL7roItq3b0+XLl3o378/X3zxRQ57KCIiUlgKcYazIcPj26T8fTdwdxafsZ4tY522aE/yV6YA9759+3LttdfStGlTLr74Yq699lquu+66HPZSRESkcBTiDGeDqePg91PNbEHKz0YzK6mPfkvtZApwP+KII2jaNPr32SGHHMLy5ctz0TUREZGCVFAznLkMj8/2M4DPyRD87u73APeEazsDD7n7grrqjzSciRMnMnDgwFx3Q0REpGAU1ICzmsHvWwwu04LfAYZlG/xuZoOB37F58Humz3iczMHvqQuHBgHJK0xQ8HuS2gS/VzfAPdU111xD06ZNOfXUU2vchoiISLEpxBzO94FDU4Lf3ybaTnJT8Lu7Z1zRYWZTgLnuXrlKvZm7f1m56rwy+B3oGVajLyXK4WwFPEBK8HvcKnUzWwQcmZbDebC7r0w55z3gOHffIqwxLfi96+VjJ1S/UEWg1XawYl3Nru28505VnvPxxx9zySWXMGnSpE3HZs2axcMPP8yNN97IttumJ2fll/Lycpo1a1b1iUVItUmm+sRTbZKpPvGKpTa9e/eOzeEsqBnOYJm7Px/+/ivRSvD1wJ1m9ijwSMK1hxHNiOLuFcCX4fi5ZlYZ/L4X0BZYlXbd3ysHjQmRSBBldqbbNKo3s4OBtZkGm6HtO4A7IAp+P+fUTGucpKysjJPrMUR36dKl7LDDDpuCemfNmsWMGTN4+umnadmyZb19bl0plpDhmlBtkqk+8VSbZKpPPNWmMBcNFXrw+ylkcTtdcidTgPvw4cNZvXo1ffv2paSkhLPOSg9KEBERkTiFOMNZsMHvZtaEaKFRj+p+aWk4mQLchw4dmoOeiIiINA6FOMNZsMHvRAPN5e7+z6y+qeSEgt9FRETqViEOODe6+1nu3sXdB7j7B+5+UHjdOQS0Z+TuK9z9uHBeibvPdfev3f2ocP1J7t7L3cvC+W1Sntu82907ufsB7l6a8BnrQzv7hn79M+W9MnfPNIMqeaS0tJRZszZ/MqNv374sWrSIhQsXst9++3HttdfmqHciIiKFpxAHnHXGzIaHgHZPDXg3s/ZmNtfMvjaz39XxZ/Yys/9Xl21K3VLwu4iISN0qqGc46yH4/XmiVe1lacc/A84lWoRU289I1wsoB16o4jzJUwp+FxERqZ4GG3BWM7Q90/VZh7a7+zVmdgnRs5b9gAvd/bn0Nt39tdB2+vFPgE/MLDYhPAwsrwnXV4bCn2xm7d39NDM7Fvg9sDVRxNKpwHbAWUCFmf0SOMfdn437DAW/x1Pwu4iISOFosOD3HIS2O9GWkvdn0belQLfUcPZwfBRQ7u5jEq7tSIZQeDP7D+CL8L1+BXRw9wuralPB79lR8HuyYgkZrgnVJpnqE0+1Sab6xCuW2uRT8HtDhrZXEGVz1re4UPjWwBQz251olvP9bBpT8Ht2FPyeTCHD8VSbZKpPPNUmmeoTT7Vp+EVDDRnavj4MTOtbXCj8LcCt7t4ZOJMtw+QlTyn4XUREpG419AxnLkLb61tcKPxOwAfhnCEp568Gmjd0JyV7Cn4XERGpWw09w9ngoe1JzOxcM1tOdPt7oZndGY5/Pxy/APi9mS03s4yDxIRQ+FHAVDN7Fkh9NvRhoL+ZLTCzn1Wnv1J3MoW7T506lY4dO9KkSRPmzZuXw96JiIg0Lg09w7nR3dPvRR6UzYXuvgLI9EDjUTHnV/l0rruPA8ZlOP4x0SA0KyFs/u60Yw8BD2U49x2gS7ZtS/0oLS1l+PDhDB48eNOxTp068cADD3DmmWfmsGciIiKNT1EHv1fFzJamBsJneU0PM3vVzDaY2YkZ3m9uZh+Y2a1111Oprkzh7h06dKBdu3Y56pGIiEjj1WAznPUQ2p41M5sO7JN2+GJ3f7wabexC9Lxmuj7uvirl9b+BUqJczkyuJnp8QERERKQo5N1OQ6mB6ulqGh7v7v3D9anh8deb2eq48PgMH3800YywAwvd/bSY/i8Nn7UxQ/+7Aq3Cd8iYU5VKwe/xqgp+r024u4iIiNStvBtwZqEdMDQlPH440J+U8PiEa8cBT7t7/8rw+HD8jNTweDObljprGcLdLyUl3L26nTazJsCNwGlAn4TzUoPfubzzhup+VFFotV006IxTVlZWZRsff/wxa9as2eLcL774gvnz51NeXl7LXuZOeXl5VjUoRqpNMtUnnmqTTPWJp9oU5oCzIcPjU6/LFO5eHWcDM919WfpWmqkU/J6dugh+Tw93r9SiRQu6du1Kt25VTkLnLYUMx1Ntkqk+8VSbZKpPPNWmMBcNNWR4/KbTMnxudXUHhodtNMcAg81sdC3blBrKFO4+ffp0Wrduzdy5czn66KPp169frrspIiLSKBTiDGcuwuPjwt2z5u6nVv5tZqVEe7ePrE4bUncyhbsD9O/fP+NxERERqblCnOFs8PD4hHD3LZjZT0Jo/EnA7Wb2Zg2+o9SBTOHun332GX379qVt27b07duXzz//PIc9FBERKQ6FOODc6O5nuXsXdx/g7h+4+0HhdecQwp6Ru69w9+PCeSXuPtfdv3b3o8L1J7l7L3cvC+e3SXlu82537+TuB7h7acJnvOLurd19B3ffxd07ZjhnsrsPr30pJElpaSmzZm3+hMXo0aPp06cPS5YsoU+fPoweracaRERE6lshDjgbTA2D37cxsylm9q6ZvRSinCrf62Jmc83sTTN7w8zSnxWVOpQp3P2hhx5iyJBoa/shQ4bw4IMP5qJrIiIiRaWgnuHMZXh8tp8BfE70TOi+ZnYKcB0w0MyaEq2qP83dXw9B8t/WVX8kOytWrGD33XcHYPfdd+eTTz7JcY9EREQav4IacFYz+H2LwWVa8DvAsGyD381sMNHuQanB75k+43FgVHj5d+BWi3KQjgjXvQ6QtjtRRgp+jzf5yB1y3QURERHJUkENOIN8D37fE1gG4O4bzOxLYBdgP8DDgLQlcJ+7X59+sYLfs5NtiG56uHvz5s2ZNm0au+yyC6tWrWLHHXdslGG8ChmOp9okU33iqTbJVJ94qk1hDjjzPfg9U6q7E9X6UOAnwFrgSTOb7+6b7c+u4PfsZBuimx7uPnDgQJYsWcKAAQMYPXo0p5xySqMM41XIcDzVJpnqE0+1Sab6xFNtCnPRUL4Hvy8nGrQSntvcCfgsHH/a3Ve6+1pgJvDj6vRVqidTuPvIkSOZPXs2bdu2Zfbs2YwcqShUERGR+laIM5z5Hvw+AxgCzAVOBJ4Kt/ofB0aY2fbAN0BP4KbqfnnJXly4+5NPPpnxuIiIiNSPQpzhzOvgd+AuYBcze5fodv/I0Mbn4bpXiAbJr7q7VgTVIwW/i4iI5IdCnOHc6O5npR07KJsL3X0FkOmhyKNizm+T8vfdQGyofMp569kyLqnyvb8SPXcqDaC0tJThw4czePDgTccqg99HjhzJ6NGjGT16NNddd10OeykiItL4FeIMZ4OpYfB7DzN71cw2mNmJae8NMbMl4WdI3fZW0in4XUREJD8U1AxngQS/3wOUEmV2pp6/M3AF0I1oAdJ8M5sRbrVLA1Hwu4iISMMrqAFnIQS/p5y/Me1QP2B25WIjM5sNHAlkXtmCgt+TKPhdRESkcBTUgDPI9+D3OJsC4YPl4dhmFPyeHQW/J1PIcDzVJpnqE0+1Sab6xFNtCnPAme/B73HiAuE3P6Dg96wo+D2ZQobjqTbJVJ94qk0y1SeealOYi4byPfg9zqZA+KA18GEt25QECn4XERHJD4U4w5nvwe9xHgf+aGb/EV4fAVxSzTakGhT8LiIikh8KcYYzr4PfzewnZracaAX77Wb2Zmjjs9D+K+HnqhrempcsKfhdREQkPxTigHOju5/l7l3cfYC7f+DuB4XXnUNAe0buvsLdjwvnlbj7XHf/2t2PCtef5O693L0snN8m5bnNu929k7sf4O6lCZ/xiru3dvcd3H0Xd++Y8t5Ed983/Eyqu5JIJqWlpcyatfkTFpXB70uWLKFPnz6MHj06R70TEREpHoU24DwVSFqFvgUzKzOzbgnvdzWzN8zsXTMbZ2YWjh9gZnPDew+bWfOENkrN7NaY916oTn+l7ij4XUREJD8U1IDT3a9x99ZVnWdml5rZAjNbQBS0fk8Ias9kPFEMUdvwc2Q4ficw0t07A9OBixI+4ypgYKbPcPf/l+XXkwag4HcREZGG16CLhqoZ3J7p+lFAubuPMbOy0E5volnPoe7+bMjS7Ew0mH4L+Br4jbvPy9De7kDzsAAJM/sz0Ur3x4jyPp8Jp84mWvRzWeW1Ydeia8J1pURZoD8zs8XA39z9yvBeubs3C6vhRwEriXZLmg/80t1jV78r+D2egt9FREQKRy5WqdcmuD1dU3c/yMx+TrRt5OFEq9DXunsXM+sCvJpw/Z5EcUWVUsPYFxENgh8iWgC0F8kOIhpIriUKj380wyD3QKKFSh8CzwM/BZ5LPUHB79lR8HsyhQzHU22SqT7xVJtkqk881SY3A87aBLeneyD8ng+0CX/3INpRCHdfGFafx0kKYz8DGGdmlwMzgG+q6Mvsyt2JzOwB4FAgfcD5srsvD+csU2+0agAAIABJREFUCH3ebMCp4PfsKPg9mUKG46k2yVSfeKpNMtUnnmqTm2c46zK4/evwu4LNB8/ZhrQvJwpgr7QpjN3d33b3I9y9K9F+5+9V0Vb6Z2bqw9cpf6f3WeqYgt9FRETyQy4GPLUJbs/GM0Sr2eeYWSegS9yJ7v6Rma02s0OIngcdDNwCYGa7ufsnZtYE+D1wWxWf2zfssb6OaOB8Ri2/h9SSgt9FRETyQy5mOGsT3J6N8UCz0N4I4OUqzh8W+vAu0SzmY+H4IDN7B3ibaNazqtzM54C/EA2gp2VapCR17+abb6ZTp0507NiRsWPH5ro7IiIikkEuZjg3uvtZaccOyuZCdx+V8nevlL9XEp7hdPd1wCnZdiYMDDtlOH4zcHOWbUwGJse81yz8LgPKUo4Pz7aPktmiRYuYMGECL7/8MltvvTVHHnkkRx99NG3bts1110RERCRFQeVw1jUz28fMXjKzJWY2xcy2DsdLzezTypxNM/tVHX7m8Wa2f121V8zeeustDjnkELbffnuaNm1Kz549mT59eq67JSIiImkadMDp7kvdfYvZxHSpoeopP3HB7VUys63CwHKzNomey7zJ3dsCnwNDUy6bEra/LHH3O0M7/TL0q7ojnOMBDTjrQKdOnXjmmWdYtWoVa9euZebMmSxbtizX3RIREZE0DXZLvTqh76mh6inXtwqDux+GQ8Pc/QUze5AoI3Nb4OYQK4SZlQN/AvoBF7r7wWntGfApcHQ4dDdRMPv4uO/g7o8TBcCnf7cRwGnARuAxdx9pZr8mytPcmuj50NOAkvBde5rZ74EB7h67+r3Yg9+Xjj468f0OHTpw8cUX07dvX5o1a8YBBxxA06Za+C8iIpJvLGGjm7r9oGjA+T5waEro+9tEs4qbQt/d/YuY66cAc919rJltBTRz9y/NbGd3/yzsMPQK0NPdV5mZAwPd/f6Y9nYFXnT3fcPrvYgGi53CzkHXEg1I3wF+6+4Zp87M7CiiHYgOd/e1Kf3ZJSWX8w/ACne/xcwmA4+4+99j2ksNfu96+dgJCVVt3DrvuVPse+Xl5TRr1myzYxMmTKBly5Ycf/zx9d21vJepPhJRbZKpPvFUm2SqT7xiqU3v3r3nu3u3/8/enYbLVZXp///eEkAgzIPNaEAUEBKDYRSEBBn9YWskHUAEosFuQRpspFUahUTl34goow0KCkgrUxBEEFEDQWaZkhCmJkCUCIoJ4yFhSPL8X6xVUKnU3qfOVHXqnPtzXedK1R5XPa9W1t7rXvX2NXs4qCeh73uQRkSJiMXAy3n7MZLG5s8bk9ZDn0/Kuby65Hploe+/Bi6LiDckfZE0+rlHwXX2BC6KiAW5bS/k7dvkjuYawFDqjIzW4+D3xlRCdJ9//nnWW289/vKXv3D//fdz1113seaaa7a6eS3nkOFirk0516eYa1PO9Snm2jS/w1kU+v4x0szyoynu2C0jr0++J7BzHl2cRnq0DvB67pgWmQesIWlIRCxi6dD3+VXHXQB8t6wZ1A95vxj4VETMyCOmozv7PdZ1BxxwAPPnz2f55Zfnhz/8oTubZmZm/VCzO5w9CX2fSsrMrDxSXwVYHXgxdza3BHZqtCH5Ef4twDjgcuBw0rrpSFo/Ip7Lh/4zKTu0yO+AkyT9ovqROrAq8Jyk5UlB9H/Nx7+a91kvuO2221rdBDMzM+tEs2ORehL6fiwwRtJDpLXTtyZNQhqSz/82cHcX2/M14DhJs4G1gZ/k7cdIeljSDOAYYELRBSLit6S11u/LM9+Pz7u+SZog9XvSu6oVlwP/KelBSe/rYnvNzMzM2k6zRzh7Evr+d6DeC437FRzf6du5EfFUvftHxAnACY20Kx9/KnBqzbbzqDPjPb/D6likLjjjjDO48MILkcTw4cO56KLOFn0yMzOz/mRQB793RtKcPJu9K+fsJukBSYskjavZ91tJL0kqmxxlVf76179y9tlnc9999zFr1iwWL17M5Zdf3upmmZmZWRc0bYQzIuZQZwnJWjng/V9qNl+Vszm7Jed3blqz+Ws5V7PRawwnrZVe7Y3afE/gL6RH8MezrO8BKwP/1uh9DRYtWsTChQtZfvnlWbBgARtssEGrm2RmZmZd0OUOp6Q1gY0jYmYftId6oe9V9x5Gg+HxNdccm89/D2l1oc2A0yS9WhYeX2Nb0ohwADMj4tCC9s/J91pSZ9/UPLO+IYMh+L2zcPcNN9yQ448/nk022YSVVlqJvffem7333ptp06Y1p4FmZmbWYw0Fv+e4oX8mdVCnkwLRb42I4/q0dcu2YxjNDY+fA2wHvAf4JbBLRMyrmole1taLqRPwnjucx0fE/gXnDarg97Jwd4BXX32Vk08+mZNOOomhQ4cyadIkdt99d3beeedBEaLbXYMlZLg7XJtyrk8x16ac61NssNSmN4LfV4+IVyQdQQo5PznPDG+FZobHV583JSLm5XNLO5s94eD3pV111VVsu+22b68e9Oyzz3L33XczdOjQQR+iW8Yhw8Vcm3KuTzHXppzrU8y1aXzS0BBJ6wPjKe/QNUNRePzVwKdIj9wbVhMe/yHgQd4Jj3/7sDr3tSbYZJNNuPvuu1mwYAERwdSpU9lqq61a3SwzMzPrgkY7nN8iLc34ZETcK2kz4Im+a1apTSTtnD8vFR4PfBkYWXJuJTweSctJWo3GwuOnAuMlrZ3PXat3fop1Zscdd2TcuHF8+MMfZvjw4SxZsoR//dd/bXWzzMzMrAsaeqQeEVcBV1V9fwo4oK8a1YlKePyPSJ3eSaTw+HeTRiI7C4//saSJpLXWjySNiH4xvyLwOHXC4yPiYUmnALdKWkwaBZ1Q7waStgeuAdYEPiFpckRsnffdBmwJDJU0F5jYlZnyg9XkyZOZPHlyq5thZmZm3dRQh1PSB0gh5u+JiG0kjQD+OSK+06etq6+Z4fHDqj5fAlzSwD3uJa3LXm/fRxtppy3Nwe9mZmbtrdFH6heQVt55CyBHIh3UV43qL7oZ/L6ipCskzZZ0T55Zj6RhkhZKmp7/zu+LNg80Dn43MzNrf43OUl85Iv4kqXrboj5oT6lWhsc3eg/gRdI7oZtLOgj4LnBg3v9kRJS9Y2p1OPjdzMysvTXa4Zwn6X3kmdp5ycbn+qxVBboY/L5M57Im+B3gyEaD3yUdRlo9qDr4vd49biK9VwowBThXNT31Rjn43cHvZmZmA0Gjwe+bkbIhP0IawXsaOCQi/ty3zVumHcPo58HvkmYB+0bE3Pz9SWBHYCjwMKmT/ArwjYi4rc75Dn6v4uD37hksIcPd4dqUc32KuTblXJ9ig6U2PQp+l/QuYLuI2FPSKsC7IuLV3m5kF/T34Pd6o5lBGhHeJHdkRwHXSto6Il5Z6kAHvy/Fwe/d45DhYq5NOdenmGtTzvUp5to0MGkoIpYAR+fPr7W4swn9P/h9LqnTiqQhpJzPFyLijYiYDxAR9wNPAh/oSlsHIwe/m5mZtb9GZ6n/XtLxkjaWtFblr09bVqy/B79fBxyeP48Dbs6P+tfNj/Erryi8H3iq8587uDn43czMrP01Omno8/nfL1VtC96ZfNNM/Tr4HfgJcKmk2cALvBMftRvwLUmL8r2/2Jdrsg8kDn43MzNrb42uNLRpXzekC/p78PvrLBuXRERcTXrsb13k4HczM7P21uhKQ4fV2x4RP+vd5pgtrRL8/sgjj7DSSisxfvx4Lr/8coYNG9bqppmZmVmDGn2kvn3V53cDHwMeAJra4Wx28HslFqkyO72RewCnk+oyijTT/cCImCNpB/Lsc9Kj/0kRcU1X2jNYOfjdzMysvTX6SP3fq79LWh24tE9a1Atyx7LXVhXqyj0kHUX9lYZmkTqviyStD8yQ9OuIKFyxycHvDn43MzMbCBoKfl/mJGl50mo7Tc2n6eJKQ/XO79JKQ9UjnAUrDdW7x02k0cu7cizS34B1o6rQkjYlTU7asLbD6eD3pTn4vXsGS8hwd7g25VyfYq5NOden2GCpTY+C3wEk/Zp3cijfBXyQ9Pi4FbYAJlatNHQ0MJaqlYZKzj0buDUixlZWGsrbP1+90pCkqyuZmQCStgZOpGqloZJ7bAg8A5BHM18G1iYtD7oj8FPgvcCh9UY3Hfy+NAe/d49Dhou5NuVcn2KuTTnXp5hr0/g7nKdXfV4E/LmydGMLtOtKQ0TEPcDWkrYCLpF0Y57VbgWqg99XWmklpk6dynbb1f3Pk5mZmfVTjQa/fzwibs1/d0TEXEnf7dOWFWvLlYaqD4iIR4HXaGAC1GDn4HczM7P212iHc6862+pmVzZBu640tGnugCLpvaRXA+Z0+muNyZMn89hjjzFr1iwuvfRSVlxxxVY3yczMzLqgtMMp6UhJDwFbSJpZ9fc0MLM5TVxGZaWhmcBawIWklYZmArfS+UpDY/Jvuh/YmjQiOiSf/20KVhoizUi/VdIM4Acl9/gJsHZeaeg44Ot5+66kmenTgWuAo+rFLdmyzjjjDLbeemu22WYbDj74YF5/3W8hmJmZtZPO3uH8BXAj8N+803ECeLWFyzK260pDl9KPo6T6Kwe/m5mZtb/SEc6IeDki5kTEwRHxZ2Ah6V3GoZI2aUoL24ykoyXNlhSS1qnavqWkuyS9IalubJPVVwl+X7RokYPfzczM2lBD73BK+oSkJ4CnSY+t55BGPpsqd34bWmlI0vSavxN7sy2SvlFwjztIk5D+XHPKC8AxLD3j3zpRHfy+/vrrs/rqq7P33nu3ullmZmbWBQ0Fv+f3FvcA/hAR20oaAxwcES2dLtyCIPgO0vub+wBfiYjbS9o2hzrLYkqaBHRERN2Op4Pfl+bg9+4ZLCHD3eHalHN9irk25VyfYoOlNj0Ofgfeioj5kt4l6V0RcUsLY5FqNTMIfhVgVkSc1Fc/xsHvS3Pwe/c4ZLiYa1PO9Snm2pRzfYq5No3HIr0kaShwG/BzSWeRRg/7g9og+N14Jwj+08CCknP3AM6DFAQfEdVB8DNIM9YrQfAAi0l5n9Yk1cHvEcHUqVPZaqumrqhqZmZmPdRoh/OTpI7bl0mPsJ8EPtFXjeqiZgbBv55XKLImcfC7mZlZ+2vokXpEvJbDyt8fEZdIWhlYrm+b1rBNJO0cEXdREwQv6W5gdsm5lSD4M/Mj9VVoLAjemmjy5MlMnjy51c0wMzOzbmp0lvoXgCnAj/KmDYFr+6pRXdT0IPgyko6RNBfYCJgp6cK8/Z/y9uOAb0iam1c6sk44+N3MzKy9NTpp6Eukx9T3AETEE5LW67NWdU0zg+A7nWIWEWeTJiPVbv8bqRNqXeDgdzMzs/bX6Ducb0TEm5UveU3wzvOU2pykOdXh7Q2es6KkK3L4+z05uglJe0m6X9JD+d89+qLNA5GD383MzNpbox3OWyX9F7CSpL2Aq4Bf912zGtPKIHhJ19S55j7ARNI7oJsDZwCV+Kh5wCciYjhwOF7msiEOfjczM2t/jQa/v4vUkdobEHATcGE0cnIvakHQ+xxyeLukw4DjSSO7MyPi0IJ73ARMioi78kjw34B1q2slSaQO6AYR8UbN+Q5+r+Lg9+4ZLCHD3eHalHN9irk25VyfYoOlNt0Ofpe0SUT8JSKWABfkv1ZrZtA7AJK2Bk4Edsmdz7VK7rEh8AxARCyS9DKwNqmDWXEA8GBtZzOf4+D3Kg5+7x6HDBdzbcq5PsVcm3KuTzHXpvNH6m/PRJfUXwLPmxn0Xn3elMoylRHxQsk9VGdb9ejm1qTH7P9Wcg3LHPxuZmbW/jrrcFZ3njYrPKq5mhn0/vZhde5bZC6p01qZXLU68EL+vhFwDXBYRDzZlXYOVg5+NzMza3+ddTij4HMrbSJp5/x5qaB30kpII0vOrQS9I2m5nIPZSND7VGC8pLXzuWWP1K8jTQoCGAfcXPWo/wbghKoRWmvA5MmTeeyxx5g1axaXXnopK664YqubZGZmZl3QWYfzQ5JekfQqMCJ/fkXSq5JeaUYD62h60HtEPAycQpqtPwP4Qck9fgKsLWk2KeT963n70cDmwDerZrX3lyzTlnr88ccZOXLk23+rrbYaZ555ZqubZWZmZr2kdNJQRPSX5SurNTPofVjV50uASxq4x+vAv9TZ/h3gO420c7DZYostmD59OgCLFy9mww03ZOzYsS1ulZmZmfWWRnM4ByRJm+Zw9idyWPsKVfvGS3pE0sOSftGL9/yUpA/21vUGmqlTp/K+972P9773va1uipmZmfWStupw9kHQ+3eBMyLi/cCLpKxRJL0fOIEUg7Q16d3Q7t6j1qcAdzgLXH755Rx88MGtboaZmZn1okbXUu+xZoa2R8Qpkk4gvWu5D/CViLi95noixR19Jm+6BJhEik36AvDDiHgRICKer21PRJxCeq8TSV8FDiVNLFo1Ir4u6QukAPcVgNl5/8j8W3eX9A3ggLLZ6gvfWsywr99QtLttzDn1/zV03Jtvvsl1113Hf//3f/dxi8zMzKyZGlppqFdulDqcTwO7VoW2P0YaVXw7tD0iXio4/wrgrog4sxLaHhEvS1qrOrQd2D0i5ksK4MCIuLLgeusAd+clKJG0MXBjRGyTO7H/B+wCLEdaOahu3JKk/YBvAnvmme6V9qxdCY+X9B3g7xFxjqSLgesjYkrB9d5eaWjdddcddeWVdZs/IN1+++386le/4nvf+16nxw6WVRu6y/Up5tqUc32KuTblXJ9ig6U23V5pqA/UhrYfxzuh7TcA15ecuwdpRJSIWAxUh7ZXZphUQtvnA4tJ2ZxFygLah+TrjAY2Am6TtE1BZ3hP4KKIWJDbVgmF3yZ3NNcgrWh0U0lb3mlAzUpDg2llgvPPP5+jjjqqodUYvGpDOdenmGtTzvUp5tqUc32KuTbNf4ezmaHtr+eOaZF5wBo5nB1Sx/LZ/Hku8KuIeCsingYeZ9nVh95uBvUzSi8Gjo6I4cBklg2TtyoLFizg97//PZ/+9Kdb3RQzMzPrZc3ucLYitL2uSO8S3EIKZ4cU1v6r/PlaYEy+1zrAB4CnCi71O+DzklbOx1dC4VcFnpO0PHBI1fGv5n1WZeWVV2b+/PmsvvrqrW6KmZmZ9bJmdzibHtreia8Bx+WQ9rVJoe2QHn/Pl/QIqVP6n5X3MWvldzuvA+6TNB2oTHr6JmmC1O9J76pWXA78p6QHJb2vi+0dkBz8bmZmNrA1+x3OZoa2d/p2bkQ8Ve/+efTzuPzXSNtOBU6t2XYeacZ77bF34FikpTj43czMbGBrqxzOdiDpaEmzJUV+HF/ZPlrSy1WZnSe1sp39lYPfzczMBp6mjXBGxBygodB2ll0a8qqce9ktkq4BNq3Z/LWIaGjmeL7GcODSms1vRMSONdvuIM22n1bnMrdFxP6N3nMwcvC7mZnZwNPsR+qdqg5U70wXwuTrPp/tYpj8Q5I2pyRMPrf/wXzthn5vGQe/m5mZ2UDQtOD3vtDfwuRrrj0H2C4i5uXvo0nxT3NJ8UvHR8TDdc57O/h9nXXWHXXSmRd0pST90vANG5t57uD33uP6FHNtyrk+xVybcq5PscFSm7Lg94HQ4fxjRGySv+9BmuizMXAfcANpVZ83C87/B7BRRLxRs30SUBkVHQbsExF3S1oErNhJvmflGnNYusO5GmnSVIekj5NGTouyPYEU/P744493dqsB46CDDmKfffbhc5/7XKfHOkS3nOtTzLUp5/oUc23KuT7FBkttJBV2OAfCpKH+FCZf3MiIVyKiI3/+DbB89aSiwc7B72ZmZgPXQOhw9psw+TKS/kn5xU5JO5BqXzfbczBy8LuZmdnANRA6nP0qTF7SMZLmkpbKnCnpwrxrHDBL0gzgbOCgaOf3GXqRg9/NzMwGtn43S70b+luY/NmkDmXt9nOBcxtp12Dj4HczM7OBra1GOCV9UdJhXTxnmqS6L7Dm/aMkPZTD2s+uPPbO+/5d0uOSHpZ0Wsk1Jkiq25mUdGdX2jvYOfjdzMxs4GmrEc6IOL/m+xwaCJMHPl/1aLuiEiZ/HimG6G7gN8C+wI2SxpBGP0dExBuS1qs+uSZMfi1gJUm/rg2Tj4iPNPTjDHDwu5mZ2UDU1A5nF4Lajy84fxLQERGnS5qWrzMGWAOYGBG35ezMi0jrlT8KrAT8NCKOqnO99YHVIuKu/P1npJntN5ImE51aiUyKiOerz60Ok5c0gRSj9B+SzgZ+ERGT876OiBiaZ79PAuaROsn3A58te4/Twe9mZmY2ELRihHMLUuewEtR+NKmz9nZQexeuNSQidsi5lieT4oyOBBZExAhJI4AHSs7fkBTEXjE3bwP4APBRSacAr5OC2u8tudYOpI7kAuBeSTdExH01x2xLmpj0LGkJzF2ApVYrqgl+56Thi0pu2R6mTZvW0HG33347m266KY8++iiPPvpo6bEdHR0NX3cwcn2KuTblXJ9irk0516eYa9OaDuczEXFH/vy/pKD214ELJd1AWoe8Ub/M/95PCmgH2I08aSciZubZ5kXqrT9ZGXEcAqxJikXaHrhS0mYlI5K/j4j5AJJ+CexKCp+v9qeImJuPmZ7bvFSHMyJ+DPwYUvD7vx9Sb07TwHT++edz1FFHNRSOO1hCdLvL9Snm2pRzfYq5NuVcn2KuTWsmDfVmUHtlhaDFLN15bjRuqBJfVLERafSxsu+XkfwJWAKUBbXX3rNeG6pXNKpt86Dm4HczM7OBqxUdzp4EtTfij8AhAJK2AUYUHRgRzwGvStopz04/DPhV3n0tsEe+zgeAFUjvXxbZS9Ja+R3ST5EemVuDHPxuZmY2cLWiw9mToPZGnAcMzdf7KvCnTo4/MrdhNvAkacIQwE+BzSTNAi4HDu8kqP124FJSB/rqOu9vWgEHv5uZmQ1srXik25Og9klVn0dXfZ5HfoczIhYCBzXamNwxXCZaKSLeBD7b4DUuBi4u2Dc0/zsNmFa1/ehG2zjQOfjdzMxsYGub4Pc+Cn0/RdIzkjrq7Bsv6ZEc+v6LkmuMllR3opOk33Rx1v2g5+B3MzOzgaepI5yNBrVLOhH4l5rNV9UGv3eFpHuAFWs2f480uvpEzbHvB04AdomIFyuh75L2Ab5bc41XgZfr3TMiPt7d9g5WDn43MzMbeJrW4exK6HteAeiUmvMnSTq+G6HvAETEjiVtq930BeCHEfFiPvf5/O9NwE01544GvpVXHtqCNGnpqIhYImkOsB0wlPRu6O3AR4C/Ap/Mj/8LOfjdzMzMBoJmv8PZn0Lfy3wAQNIdwHLApIgoi2vagdTJ/TOpU/1pYErNMe8HDo6IL0i6EjiAlEO6FAe/O/i9N7g+xVybcq5PMdemnOtTzLVpfoezP4W+lxlC6iCOJmVz3iZpm4h4qeD4P0XEUwCSLiOFvtd2OJ+OiOl12rwUB787+L03uD7FXJtyrk8x16ac61PMtWn+pKH+FPpeZi7wq4h4KyKeBh4ndUCLOPS9hxz8bmZmNnA1u8PZb0LfO3Et6f1QJK1DesT+VMnxO0jaVNK7gAOpWa7SOufgdzMzs4Gr2R3OfhX6Luk0SXOBlSXNlTQp77oJmC/pEeAW4D8r66QXuAs4FZgFPA1c07OfMTC99NJLjBs3ji233JKtttqKu+66q9VNMjMzsyZo9mPd/hb6/lVSx7R2e5DeLz2ugWtMoyrQvWbfsPxxHlVxUBFxeqNtHEiOPfZY9t13X6ZMmcKbb77JggULWt0kMzMza4K2CX5vF5KOljRbUuTH8ZXth0iamf/ulPShVraz2V555RX++Mc/MnHiRABWWGEF1ljDmfhmZmaDQdNGOHsh9P2Uesc3oiD0/dCIeKgL1xhOWiu92ht18j3vIM22n1az/Wlg9xwkvx9pJnphNuhA89RTT7Huuuvyuc99jhkzZjBq1CjOOussVllllVY3zczMzPpYv5spXS/0vUgXwuTrduwkvQc4H9gsbzoyIu6UdC2wMfBu4KyI+HFEPCRpc+AHwD7AVyJimclBEfFgvnbt9jurvt5Nilsq1U7B752Fuy9atIgHHniAc845hx133JFjjz2WU089lW9/+9tNaqGZmZm1itLriu0pdzifBnatCpN/DJhIVZh8UX6mpCuAuyLiTEnLAUMj4mVJa0XEC3nlontJI5PzJQVwYERc2UDb5gDb5XdMa/cdn9t3RJ191cHvo04684JGStFywzcsn13+wgsvcNRRR3H55ZcDMHPmTH7xi19w6qmndut+HR0dDB06tFvnDgauTzHXppzrU8y1Kef6FBsstRkzZsz9EbFdvX39boSzG3oSJr8HaUSUiFjMO2uiHyNpbP68MSmDcz4pP/PqnjRW0hhSh3jXevsHcvD7GWecwfrrr88WW2zBtGnT+OhHP9rtIFyH6JZzfYq5NuVcn2KuTTnXp5hrMzA6nEVh8h8jzVg/mtSxbEheG31PYOeIWJDXbX933v167ph2S15u80Jgv05ilgakc845h0MOOYQ333yTzTbbjIsuuqjVTTIzM7MmGAgdzk0k7RwRd1ETJi/pbmB2yblTSeuvVx6prwKsDryYO5tbAjv1RiMlbUJajvPQiPi/3rhmuxk5ciT33Xdfq5thZmZmTTYQYpF6EiZ/LDBG0kOk9c23Jk1CGpLP/zZpgk/DJB2Tw+Q3AmZKujDvOglYG/gfSdMlDbqel4PfzczMBqeBMMLZkzD5vwP1XpLcr+D4Tt/4jYizgbPrbD8CWGaS0GDi4HczM7PBaSCMcPYZSXOqw9sbPGdFSVfk8Pd78kz66v2bSOrIM9UHDQe/m5mZDV5t3eGMiDkR0VCYfH6MXf13Yk/uLemaOtfchzQD/cWI2Bw4A/huzalnADf25N7tqDr4fdttt+WII47gtddea3WzzMzMrAna6pF6F4Lelxo9rITJ1wS9j5d0S1HQe517HwYcT5oVPzMixtYek4+7CZiUv04BzpWknAn6KeApoKGeloPfzczMbCBoq+D3FgS9zwHN9801AAAgAElEQVS2A95DmmG+S0TMqxxfcI9ZwL4RMTd/f5K0hOVC4A/AXqSOa0dEnF7nfAe/N2CwhOh2l+tTzLUp5/oUc23KuT7FBkttBlrwezOD3qvPm1JZNaios5mpzrYAJgNnRERH7bKXSx3o4PeGOES3nOtTzLUp5/oUc23KuT7FXJv27HA2M+j97cPq3LfIXFKnda6kIaRczxdIo5zjJJ0GrAEskfR6RJzbaFvbnYPfzczMBqd27HC2Iuh9KnCNpDPyo/bCR+rAdcDhwF3AOODmSO8tfLRygKRJpEfqg6azCQ5+NzMzG6zacZZ604PeI+Jh4BTgVkkzgB+U3OMnwNqSZpMe93+9i79vwHLwu5mZ2eDUjiOczQx6H1b1+RLgkgbu8TrwL50cM6mz6wxEDn43MzMbnNpxhLNpuhn8vpukByQtkjSuavt7Jd2f8zofllTbaR7QHPxuZmY2eLXVCGdEzAEaCnpn2VHGq3IeZ68ougfwc2ACKfqo2nPARyLiDUlDgVmSrouIZ3urTf1ZdfD7jBkzGDVqFGeddRarrLJKq5tmZmZmfaytOpxdDH5fpnNZE/wOcGQPgt8PrXePquOXVH+PiDervq5IA6PLDn43MzOzgaCtOpzZFsDEquD3o4GxVAW/l5x7NnBrRIytBL/n7Z+vDn6XdHVEvJ3DKWlr4ESqgt+703BJGwM3AJsD/1lvdLMm+J2Thi/qzq2abtq0aaX7X3jhBdZZZx0WLlzItGnTeN/73scvfvELPvaxj3Xrfh0dHZ3eczBzfYq5NuVcn2KuTTnXp5hr054dzv4e/F4oIp4BRkjaALhW0pQ8kan6GAe/N8AhuuVcn2KuTTnXp5hrU871KebatGeHs78Hv3cqIp6V9DApm3NKb123v3Pwu5mZ2eDUjh3O/h78XpekjYD5EbFQ0prALpTneQ44Dn43MzMbnNoxFqlfB79L2l7SXNIM9h/lkUyArYB78vm3AqdHxENd+N1tz8HvZmZmg1M7jnD29+D3e4GN6mz/PTCikXYOVA5+NzMzG5zabYTzEKBLaeGSpknarmT/KEkPSZot6WxJytuvyCHt03MA/PSSa0yQVHdddEl3dqW9A5WD383MzAavtupwRsQpEbHM6GEtSSdWOovAdsDPc1B7PeeRYojen//2zfc6MCJGRsRI4GrglyX3+BZwYL17RMRHuvATB6zq4Pdtt92WI444gtdee63VzTIzM7MmaOoj9S4Gt9c7fxLQERGn59nk9wBjSKOeEyPitpylOZzUmX4UeAP4UkQsM1tF0vrAankCEpJ+BnwKuLHqGAHjqZn5nlctOiUfM4GUBfpRSY8Dv4iIyXlfR0QMzbPhJwHzSKsl3Q98NiIKZ787+N3MzMwGgla8w9mT4PZaQyJiB0kfB04mxRsdCSyIiBGSRgAPlJy/ITC36vvcvK3aR4G/R8QTnbRlB1JHcgEpPP6GOp3cbUkTlZ4F7iDNVL+9+gAHvzfGIbrlXJ9irk0516eYa1PO9Snm2rSmw9mT4PZalcfc9wPD8ufdSCsKEREz8+zzIqqzrXbE8WDgsgba8vvK6kSSfgnsCtR2OP8UEXPzMdNzm5fqcDr4vTEO0S3n+hRzbcq5PsVcm3KuTzHXpjUdzt4Mbn8j/7uYpX9LoyHtc1l6RvlGpNFHACQNAT4NjGrgWrX3rNeGN6o+17Z5wHPwu5mZ2eDUig5PT4LbG/FH0mz2WyRtQ0kUUUQ8J+lVSTuR3gc9DDin6pA9gccqo5Kd2Cuvsb6Q9B7o57v7AwYqB7+bmZkNTq3ocFaC238EPEGaSHO9pHeTHnGXBbc34jzgovwofTrwp06OPxK4GFiJNFnoxqp9B9HY43RIj8YvBTYnTRoalD2rYcOGseqqq7LccssxZMgQdzDNzMysJR3OngS3T6r6PLrq8zzyO5wRsZDUUWxI7hhuU7BvQoPXuJjUaa23b2j+dxowrWr70Y22sd3ccsstrLPOOq1uhpmZmfUTbZXDKemLkg7r4jndDX7/nqTHJM2UdE0XZ8931qYJkjboreuZmZmZ9WdN7XBGxJyIqDuaWK06VL3q78SIOD8iftade0u6p841h1MQ/A78HtgmIkaQ8kJPkLRPnWtc043mTAAGZIdTEnvvvTejRo3ixz/+caubY2ZmZv1Afw5+P6XO+ZNoLPj9IuCDpPdFVwKIiB3rXK8w+D0ifld16N3AuIi4CbipznWWA74L7EOanX5BRJwj6STgE7kNdwL/BhzAO6sfLQR2zq8BLKO/Bb93Fu4OcMcdd7DBBhvw/PPPs9dee7Hllluy2267NaF1ZmZm1l85+L3z4HdIM86vKLnOvwKbAttGxKI8Wx3g3Ij4FoCkS4H9I2KKpKOB4wtWP+q3we+Nhtb+3//9HwDbbrstl112GUuWLOn1tjhEt5zrU8y1Kef6FHNtyrk+xVwbB793Gvye10dfBPy85Dp7AudHxKJ83xfy9jGSvgqsDKwFPAz8uuQ6bR38/tprr7FkyRJWXXVVXnvtNf7rv/6Lk046qU/Cbh2iW871KebalHN9irk25VyfYq6Ng987C34/HNgf+FjZmuekjmttR/XdwP8A20XEM/l1gHc32K629Pe//52xY8cCae30z3zmM+y7776dnGVmZmYDnYPfC4LfJe0LfA3YPSIWdHLP3wFflDSt6pF65TnyPElDgXHAlLztVWDV7v28/muzzTZjxowZrW6GmZmZ9TOtiEWqBL/PJD1mvpAU/D4TuJXeCX4fmq/3VRoLfr+Q1NF9kneC388ldQp/n2ejn19yjQuBvwAzJc0APhMRLwEXAA8B1wL3Vh1/MXB+vu5KXflx/cHixYvZdttt2X///VvdFDMzM2sDDn4vCH6PiM27cI1FpHdRj6vZ/g3gG3WOvxq4utHr9zdnnXUWW221Fa+88kqrm2JmZmZtoK2C35tN0hxJXVoyR9KKkq7IQfL35CgoJK0g6aIcMj9D0ug+aHKfmzt3LjfccANHHHFEq5tiZmZmbaKpI5wRMYeCZSSr5Znh/1Kz+aqIWCabs1GS7gFWrNl8aEQ81IVr7EPK26z2dESMrfo+EXgxIjaXdFA+/kDgCwARMVzSesCNkraPiN7PDOpDX/7ylznttNN49dVXW90UMzMzaxOteKTeqdyxrBf8PozGg+Nrr7mjpPcA5wOb5c2r5uteC2xMmkV+Vo4mqr33YcDxpNnoMyPi0ILmfxKYlD9PAc7Ny2V+EJia2/K8pJdIAfCF75i2Ivi9LNz9+uuvZ7311mPUqFGDPk/MzMzMGqfytJ/+JXc4nwZ2rQqOf4w0qvh2cHyesFPv/CuAuyLizLw60NCIeFnSWhHxQp7Acy9pZvp8SXNIncL3kDI/d4mIeZXjC+4xC9g3Iubm708COwKfBvYizczfGHiQFIB/dc351cHvo04684JuVqt7hm+4euG+Cy64gN/97ncst9xyvPnmmyxYsICPfvSjnHjiiU1sYdLR0cHQoUObft924foUc23KuT7FXJtyrk+xwVKbMWPG3B8R29Xb1y9HODvRk+D4PUgjokTEYuDlvP0YSZXH4huT1lSfX3PelDw5iaLOZlYUJv9TYCvgPuDPpKUul1lGqD8Hv1eH1k6bNo3TTz+d66/vSk5/73GIbjnXp5hrU871KebalHN9irk27TlpqCg4/mrSOui/7crF8uSdPUlrmn+INPJYG9C+TLB7ibmkTiuShgCrAy9ExKKI+I+IGBkRnySt//5EV9pqZmZm1o7ascO5iaSd8+elguOBLwMjS86dSsrdRNJyklYjdQhfjIgFkrYEdio4b7yktfO5a9U5puI64PD8eRxwc37Uv7KkVfL5ewGLIuKRBn5vvzR69OiWjW6amZlZe2nHDmdPguOPJa1v/hBp/fWtSSOiQ/L53wburj0pIh4mTWK6NQe7/6DkHj8B1pY0m/S4/+t5+3rAA5IeJa1gVDTpqN9z8LuZmZl1RTu+w9mT4Pi/k2aR19qv4PhhVZ8vAS5p4B6vs2ykUyUSaotG2tnfOfjdzMzMuqIdRzibppvB77tJekDSIknjqraPlHSXpIclzZR0YO+3uO85+N3MzMy6qq1GOFsZHN/oPYCfAxNImZ3VFgCHRcQTkjYA7pd0U1GEU3/l4HczMzPrqrbqcHYx+L1ecHxt8PuREXFnD4LfCzuwkpZaQSgi/q/q87OSngfWBQo7nA5+NzMzs4HAwe+9HPxeda+LgesjYkqdfTuQ3gfdunZpSwe/N2awhOh2l+tTzLUp5/oUc23KuT7FBkttyoLf27HD+ceI2CR/34M0E3xjUqD6DaRO3psF5/8D2Cgi3qjZPgmoBL8PA/aJiLurOpwHA/8UEQ33rIo6nJLWB6YBh0fEMjPiq22xxRbx+OOPN3rLpnLwe//m+hRzbcq5PsVcm3KuT7HBUhtJhR3Odpw01N+D38vutRqpU/yNzjqbZmZmZgNFO3Y4+3vwe12SVgCuAX4WEVd19fz+xsHvZmZm1qh27HD26+B3SdtLmkuawf4jSQ/nXeOB3YAJkqbnv7LOcUu8/vrr7LDDDnzoQx9i66235uSTT251k8zMzKzNtdUs9ay/B7/fC2xUZ/v/Av/bSDtbacUVV+Tmm29m6NChvPXWW+y6667st99+7LRTvYFfMzMzs8612wjnIcAaXTlB0jRJdV9gzftHSXpI0mxJZ0tSzf7jJUVXA+A7adOXJa3cW9frTZLenkn31ltv8dZbb1FTEjMzM7MuaasOZ0ScEhHLjB7WknRi5bE1aZb5z3NQez3nkWKI3p//9q26zsbAXsBfyu5R9dfoLPYvA/2ywwlprfSRI0ey3nrrsddee7Hjjju2uklmZmbWxpr6SL2Lwe31zp8EdETE6ZKm5euMIY16ToyI23KW5nBSZ/pR4A3gSxFxX53rrQ+sFhF35e8/I810vzEfcgbwVeBXtefmVYtOyecNBc4BDspLVk6OiKslnQdsD6wETImIkyUdA2wA3CJpXkSMKapXXwS/lwW7Vyy33HJMnz6dl156ibFjxzJr1iy22abTBZ7MzMzM6mrFO5xbkDqHleD2o0kZmG8Ht3fhWkMiYgdJHwdOJsUbHQksiIgRkkYAD5ScvyEwt+r73LwNSf8M/DUiZjTwSPmbwMsRMTyfu2befmIOlF8OmCppREScLek4YExEzKu9UE3wOycNX9TZvbukqysEDRs2jB/+8IcceGD/Wvq9o6PDqx2VcH2KuTblXJ9irk0516eYa9OaDuczEXFH/vy/pOD214ELJd0AdCVr55f53/tJge2QZoKfDRARM/Ps8yL1epKR3688Edi7wXbsCRz09gUiXswfx+cO5BBgfeCDQFl7yMtq/hhS8Pu/H1JvjlPf+cc//sHyyy/PGmuswcKFC/nmN7/J1772tX4XWDtYQnS7y/Up5tqUc32KuTblXJ9irk1rOpxFwe0fI3Xajgb2aPBalRWDFrP0b2k0pH0uS88o3wh4FngfsClQGd3cCHhA0g4R8bc611kmGF7SpqS117ePiBfzykO1gfL9znPPPcfhhx/O4sWLWbJkCePHj2f//fdvdbPMzMysjbWiw7mJpJ3ze5NLBbdLuhuY3cPr/5E0m/0WSdsAI4oOjIjnJL0qaSfS+6CHAedExEPAepXjKktc1nsEnv2O1FH+cj5+TWA14DXgZUnvIUUvTcvHvwqsChRdr2VGjBjBgw8+2OpmmJmZ2QDSilnqPQlub8R5wNB8va8Cf+rk+CNzG2YDT/LOhKGu+A6wpqRZORh+TETMIC2T+TDwU+COquN/DNwo6ZZu3MvMzMysrbRihLMnwe2Tqj6Prvo8j/wOZ0QspOp9ygaueR9QOgW7OgC+YH8HcHid7RMKjj+HNKu933n99dfZbbfdeOONN1i0aBHjxo1j8uTJrW6WmZmZtbG2yuFsNklzuhr4LmlFSVfkIPl7chQUkpaXdEkOmX9U0gl90eaeqqw0NGPGDKZPn85vf/tb7r57mdU+zczMzBrW1BHOiJhDJ6OJkELVSWuRV7sqZ192i6R7gBVrNh+a39ds9BqfI63HXu2OiPhS1feJwIsRsbmkg4DvAgeSfs+KETE8z4J/RNJluSb9hlcaMjMzs97WL9dSrw5Vr9aT4PiI2DFP3jkf2CxvXjVf91pgY9Is8rNyNFHtvQ8jvV8awMyIOLSg+Z8EJuXPU4Bz83KZAawiaQgpCP5N4JWSMrQs+H3x4sWMGjWK2bNn86UvfckrDZmZmVmPKKLRBKHWyx3Op4Fdq4LjHyONKr4dHB8RLxWcfwVwV0ScmcPYh0bEy5LWygHtKwH3ArtHxPzK7HTgPaTMz10iYl7l+IJ7zAL2jYi5+fuTwI7Ay8ClpPinlYH/KOjYVge/jzrpzAu6U6pCwzdcveFjOzo6+OY3v8kxxxzDpptu2qvt6KmOjo63R2JtWa5PMdemnOtTzLUp5/oUGyy1GTNmzP0RsV29ff1yhLMTPQmO34M0IkpELCZ1AgGOkTQ2f96YtKb6/JrzplRikYo6m1ndMHnSxKjFpGUt1wRuk/SHiHhqqQNbHPxe6/7772f+/Pl87nOfa2k7ajlEt5zrU8y1Kef6FHNtyrk+xVyb9pw0VBQcfzVpHfTfduVikkaTVgraOSI+RIoyqg1oXybYvcRcUqeV/Ph8deAF4DPAbyPirYh4nhSTVPd/Aa30j3/8g5deSgPECxcu5A9/+ANbbrlli1tlZmZm7awdO5ybSNo5f14qOJ4UvD6y5NyppNxNJC0naTVSh/DFiFggaUtgp4LzxktaO5+7Vsk9ruOdiKRxwM2R3lv4C7CHklXyfR7r/Oc213PPPceYMWMYMWIE22+/PXvttZdXGjIzM7MeacdH6pXg+B8BT5Am6Fwv6d2kkciy4PhjgR9Lmkh6vH0kaUT0izko/nFgmQygiHhY0inArZIWk0ZBJxTc4yfApZJmk0Y2K5mgPwQuAmbldl4UEaXrqreCVxoyMzOz3taOHc6eBMf/nTSLvNZ+BccPq/p8CXBJA/d4nWUjnSrh8Mts728c/G5mZma9rR07nNaHKsHvQ4cO5a233mLXXXdlv/32Y6ed6r1pYGZmZta5tnqHMyLmRERDwfGSptf8nVhw7KZ5RaAn8gpBK9TsHycpJG1Xs73he9S55wRJGzRybLM5+N3MzMx624Ac4awXHJ9zN+v5LnBGRFwu6XxSpud5+ZxVgWNIQfOd3qMLJpDe5Xy27CAHv5uZmdlA0NTg956sFJTPr10p6MiIuLNopSBJHcAPgH2Ar0TE7TXXE/AP4J8iYlGe/T4pIvbJ+88E/gAcDxwfEfcVtGs5Usd1H1J80gURcY6kk4BPkFYWuhP4N+AA4GLgr8BCUhzTwqprOfi9AYMlRLe7XJ9irk0516eYa1PO9Sk2WGpTFvxORDTtDxhG6pDtkr//FPgqaXZ4pfO7Rsn5VwBfzp+XI8UhAayV/12JNHK4dv4ewPiS660DzK76vjEwK3/eFrg6f54GbFdynSNJOaBDatqzVtUxlwKfaOR6lb8PfOAD0WqTJk2K733ve61uxjJuueWWVjehX3N9irk25VyfYq5NOden2GCpDXBfFPRpWvEOZ+1KQbvxzkpBnwYWlJy7B/lxd0QsjojqlYJmkCKNKisFQYo+urrkenVXBZL0LuAM4CsN/B5IwfHnR8Si3LbKSkRj8vuhD+W2b93g9VrGwe9mZmbW21rxDmfRSkEfI2VWHk3qnDWkZqWgBZKm8c5KQa9HWsKyyDxgDUlDcmdxI9J7lasC2wDT8oSZfwKuk/TPUf+x+jIrEeVc0P8hjWQ+I2kSy65g1O8899xzHH744SxevJglS5Ywfvx4B7+bmZlZj7Siw7mJpJ0j4i5qVgqSdDcwu+TcykpBZ+b3JlehsZWC6oqIkHQLaUWgy0krBP0qj5yuUzkud2IL3+EEfkcKj58W6V3QtYAled88SUPzPabkba+SOrX9joPfzczMrLe14pF6ZaWgmcBawIWklYJmArfS+UpBY/Ij6vtJj6h/CwzJ53+bOisFdeJrwHF5ZaC1SSsFddWFpKUrZ+ZH+5+JiJeAC4CHgGuBe6uOvxg4P0cprdSN+/XIM888w5gxY9hqq63YeuutOeuss5rdBDMzMxtEWjHC2cyVgjqdEhYRT3V2/4gY3cn+RcBx+a96+zeAb9Q5/mrK3y3tU0OGDOH73/8+H/7wh3n11VcZNWoUe+21Fx/84Adb1SQzMzMbwNoq+L23SfqipMO6eM602hD4mv2nSHomRzLV2183SL6Z1l9/fT784Q8DsOqqq7LVVlvx17/+tVXNMTMzswGuqSOcETGHNBmnVF6xp3bd8asiha13i6RrgNowya9FxE1duMY+wHbAzyVVsjOfjoixVYf9GjgXeKLO+YVB8vV0N/i9kXD3t4+dM4cHH3zQ4e5mZmbWZ5oa/N7beiFIfhLQERGn54lB9wBjgDWAiRFxW37H8iLgg6T3T4cBXyqZQFS5dkftI/1GguR7I/i90XD3hQsXcuyxx/LZz36W3Xbbrcv3aaXBEqLbXa5PMdemnOtTzLUp5/oUGyy1KQt+HwhLW25B6hzeIemnpFilscCWeRb6Gl241pCI2EHSx4GTSXFLRwILImKEpBHAA91ppKRtgY0j4npJdTvAAJFWSfoxwBZbbBH/fki9V1Z77q233mL//ffni1/8Iscdd1znJ/Qz06ZNY/To0a1uRr/l+hRzbcq5PsVcm3KuTzHXZmC8w9mTIPlav8z/3k8aySRf738BImImMLOrDexGkHyfiggmTpzIVltt1ZadTTMzM2svA6HDWRQkfzXwKdIj90a9kf9dzNKjvz1976A6SH4OKSv0ulZNHLrjjju49NJLufnmmxk5ciQjR47kN7/5TSuaYmZmZoPAQHik3pMg+Ub8ETgEuEXSNsCIrl6gG0HyfWrXXXelnd/dNTMzs/YyEEY4exIk34jzgKH5el8F/lR2sKTTJM0FVpY0N09M6lcc/G5mZmbNNBBGOHsSJD+p6vPoqs/zyO9wRsRC0hrvDYmIr5I6pmXHjC7b39cc/G5mZmbN1DYjnM0MaZd0nKRHJM2UNFXSe0uuMVrS9QX7ftPFWfJN4eB3MzMza6a2GeGMiPPrbJtDD4LkKQ5pfxDYLiIWSDoSOA04sOaa9wArAkOBdSVNBw6NiIeq2vfxztrWag5+NzMzs77WtOD3dgtpr9q3LXBuROxSsH808C1gPikT9I/AURGxJM9I347UKb0RuB34CPBX4JP5cX3t9Rz83oDBEqLbXa5PMdemnOtTzLUp5/oUGyy16U/B720R0l5jIqmzWGYHUif3z6RO9aeBKTXHvB84OCK+IOlK4AByvmc1B783xiG65VyfYq5NOdenmGtTzvUp5to0/x3Ofh/SXk3SZ0kjlN/r5NA/RcRTEbEYuAzYtc4xT0fE9DptbjoHv5uZmVkzNbvD2Q4h7QBI2hM4EfjniHijk8Nr71mvDdXXqG1zUzn43czMzJqp2Z2efh/SDm+/t/kjYN+IeL6BU3aQtCnpkfqB5Efi/ZWD383MzKyZmj3C2S4h7d8jTfS5StJ0Sdd1ct+7gFOBWcDTwDU9+A19zsHvZmZm1kzNHuFsi5D2iNizC9eYBkwr2Dcsf5xHVXxTRJze6PX7goPfzczMrJnaJvgd+iz8fZSkhyTNlnS2JOXtIyXdnUc475NU2DGWNEnSMnFOkjaQVDtbveUc/G5mZmbN1LQRzp6GtEfEKfXC3xt0saRFNdsqIe3nkXIv7wZ+A+xLikE6DZgcETfm6KUfSlq+5hpvRERhYnpEPAuM62abm8LB72ZmZtbXmvpIvQvh7yMLzp9E18PfVwIm1At/l7Q+sFqexISkn5Fmy99Immm+Wj50deCJiPhMyc/7kKSbgY2B0yLigvx7r4+IbSRNyL9xZeB9wDX5kX6hhW8tZtjXbyg7pK45p/6/ho7r6OjggAMO4Mwzz2S11Vbr/AQzMzOzbmhFNE9/Cn/fEJhb9X1u3gbwZeAmSaeTXj34SCdtGQHsBKwCPCipXk9xJKmj/QbwuKRzIuKZ6gOqVxpad911uXLfVTq57bKmTZvW6TGLFi3ihBNOYMcdd2SttdZq6Jz+pKOjo+3a3EyuTzHXppzrU8y1Kef6FHNtWtPhrA1/P453wt9vAK7vwrWKwt/PhhT+nmesF1GdbZW8oCOB/4iIqyWNB35C6tAW+VWetLRQ0i2kyVDTa46ZGhEvA0h6BHgvsFSHs3alob5YmSAiOPzww9lll10488wze/36zeBVG8q5PsVcm3KuTzHXppzrU8y1ac2kof4U/j4X2Kjq+0bAs/nz4bzTob2KzmfTt034u4PfzczMrJla0eHpN+HvEfGcpFcl7UR6H/Qw4Jy8+1lgd1Lk0R7AE53c95OS/pv0SH008HVghe7/jL7j4HczMzNrplaMcPar8HfSo/MLSR3dJ0kThgC+AHxf0gzg/yO/V1niT8ANpNnu384z1PslB7+bmZlZM7VihLO/hb/fR524poi4HRjV1XbVbJ9TuXZEXAxcXLVv/0bb2Nsc/G5mZmbN1FbB72UkzZG0ThfP2U3SA5IWSRpXs29xDn3vdGnLont3J6i+GRz8bmZmZs3U1BHO3gh/7+69Jd0DrFiz+WvABGCZVYKAhbV5oJI+Bxxbc9wdFOhBUH3TOPjdzMzM+lqfdTi7EPK+TGcvIk6RdCFwPrBZ3nxLvu61pHD1dwNn5Rih2nsfRupEBjAzIg4tWxFI0pJGflNEXEQKla89//8B/ylpTN70mYiY3UhQfdn9HPxuZmZmA4H6arZy7nA+DexaFfL+GDCRqpD3iHip4PwrgLsi4kxJywFDI+JlSWtFxAt5RaF7gd0jYr6kOcB2wHtIcUa7RMS8yvGdtPVi0opAU6q2LSLNoF8EnBoR15acPwe4IHeUDwPGR9HhXB4AABGlSURBVMT+dTqc90fEV3JQ/XERsUyuZ03w+6grr7yyrOndVgl+33777Rk/fnyf3KMvdXR0MHTo0FY3o99yfYq5NuVcn2KuTTnXp9hgqc2YMWPuj4jt6u3r60fqPQl534M0IkpELAZeztuPkTQ2f94YeD8wv+a8KXkiEZ11NktsEhHPStoMuFnSQxHxZMnxl1X9e0bBMfWC6pfi4PfGOES3nOtTzLUp5/oUc23KuT7FXJu+nzTUmyHvSBpNWu1n54j4EPAg6dH6UofVuW+XVWKNIuIpUhbntp2dUvC5WlFQfVM5+N3MzMyaqa87PT0JeZ9KysisPFJfBVgdeDEiFkjakrR2eb3zrpF0Rn7U3ukj9VqS1iStx/5Gnn2+C3BaJ6cdCJya/72rK/drNge/m5mZWTP19QhnT0LejwXGSHqI9Ah6a9KI6JB8/rdJIetLiYiHgVOAW3No+w+KbiBpe0lzSTPifyTp4bxrK+C+fP4tpHc4H+nkt66YZ8If28nvaorPf/7zrLfeemyzTaehAGZmZmZ9qq9HOHsS8v534JN1du1XcPywqs+XAJc0cI97WXot9cr2O4HhjbSz5t6Ta7ZPqvo8uurz20H1fWXChAkcffTRHHZYv4sBNTMzs0FmwAS/d5ekTSXdI+kJSVdIWiFv30TSLZIelDQzzyzvrXtOkLRBb12vnt1224211lqrL29hZmZm1pA+63BGxJyIaCjkvWpFn8rfib3Zlk7u8V3gjIh4P/AiKbYJ4BvAlRGxLWmpzP+RdE2d6+zTjSZNAPq0w2lmZmbWXzR1pnRJGPyqLB0GX3dFIUnvYekw+CMj4s6iMHhJHaR3OPcBjs7ro1dfT6QYpc/kTZcAk4DzSDPNK2noqwPPRsRY6siTmr6b7xOkTM5zJJ0EfAJYCbgT+DfgAFJe6M8lLSTNuF9Y77plwe+NhrubmZmZtVqfBb/XvVnzw+ADODAi6qan5xnod0fE5vn7xsCNEbGNpPWB3wFrkmbI7xkR9xdc50hSXNOBEbGoqj1vz5CXdClpxPTXOQT++Ii4r8613g5+X2eddUeddOYFdWs5fMPV626v9re//Y0TTjiBiy5aZnGktjdYQnS7y/Up5tqUc32KuTblXJ9ig6U2rQx+r6eZYfCLSZmfRVRnW6UHfjBwcUR8X9LOwKWStomIestg7gmcHxGLctsqMUxjJH0VWJk0S/9h4Ncl7Vkm+P3fD6k3b6oxc+bMYZVVVhmQYbMO0S3n+hRzbcq5PsVcm3KuTzHXpjWThpoZBv967pgWmQesIanS8d4IeDZ/nghcCZBzRN8NrFPUDGp+l6R3A/8DjIuI4cAFLBtS32cOPvhgdt55Zx5//HE22mgjfvKTnzTr1mZmZmZLacUIZyvC4OvKj/BvAcYBlwOHA7/Ku/8CfAy4WNJWpM7iPwou9Tvgi5KmVR6pA5WR0HmShuZ7VNZqf5X03mqfueyyyzo/yMzMzKwJWjHC2fQw+E58DThO0mxgbaAyFPgV4As5/P0yYEIUv/B6IamDOjMf/5n8HuoFwEPAtaR3SysuBs7Ps9xX6mJ7G+LgdzMzM+svWjHC2cww+E7f0M1rpS9z/7yy0C4NtmsR6V3U42q2f4MUr1R7/NWUv1vaYw5+NzMzs/5iwAS/S5qTZ5135ZzdJD0gaZGkcTX7TpP0sKRHJZ2dI5SKrtNRsP1bkvbsSpt6i4PfzczMrL9o6ghnRMwBGgqDJ61vXu2qonzORki6Bti0ZvMPSCHsx9cc+xHS6OaIvOl2YHdJK5LyNqs9XXTPiDipu+01MzMzGyj6rMNZEvJ+EkuHvB9fe25EnCLpQpYOeb8lX7duyHvNvQ8jdSIDmBkRhxaFtufja6OOIl9/BdIM9OWBv0fEo8BNdc7vkPR9YAxptaKDIuIfki4Gro+IKZLmkILlP5Gv9y8R8VhRm8DB72ZmZjYw9PUI5xbAxKqQ96OBsVSFvJecezZwa0SMrYS85+2frw55l3R1RMyvnCRpa+BEYJeImJdnjHdJRNyVZ68/R+pwnps7m0VWAR6IiK/k1YVOzr+11ryI+LCko0gd4iNqD6gJfuek4Yvq3nDatGmd/o6//e1vvPbaaw0d2246OjoG5O/qLa5PMdemnOtTzLUp5/oUc236vsPZzJD36vOmRMS8fO4LdJGkzYGtSLmcAL+XtFtE/LHglCXAFfnz/wK/LDiusv1+4NP1DnDwe2McolvO9Snm2pRzfYq5NuVcn2KuTd9PGmpmyPvbh9W5b1eNJS152RERHcCNdCHfs+T+b+R/F9PHnX0Hv5uZmVl/0dcdzk3yspBQE/IOfBkYWXJuJeQdSctJWo3GQt6nAuMlrZ3P7c5U7b+QJgkNkbQ8sDspP7TIu0jB7gCfIU0yaqnLLruM5557jrfeeou5c+cyceLEVjfJzMzMBqm+7nA2PeQ9Ih4GTgFuzSHsPyi6gaTtJc0lzYj/kaSH864pwJOk0PYZwIyIKFsD/TVga0n3kx7pf6vk2KZw8LuZmZn1F339DmczQ96HVX2+hDQjvLN73Ms772lWb18M/Fsj7czHVyY0fbNm+4SC9t0HjG70+t3h4HczMzPrLwZM8Ht3SNpU0j2SnpB0haQV8vbjJD0iaaakqZLe24v3/JSkD/bW9Yo4+N3MzMz6iz7rcEbEnIhoKOQ9ryle/Xdib7ZF0jcK7vFd4IyIeD8pP7PyouODwHYRMYL0eP20fJ176lxneBeb8ymgzzucZmZmZv1F01YaKgmCX5UUsF4YBJ/Pfw9LB8EfGRF3FgXB5+UmfwDsA3wlIr5Tcz0B/yBN8oH0CH4ScF5E3FJ16N3AZwEiYseCtn0VOJQUj3RjRHxd0hdIeZorALPz/pGk0PvdJX0DOCAiniyqmYPfzczMbCBQRE8ThBq8UepwPg3sWhUE/xhpVPHtIPiIeKng/CuAuyLizEoQfES8LGmt6iB4YPeImC8pgAMj4sqC661Dij7aPH/fmNRZ3KbmuHOBv9V2WKv270d6d3PPPHu+0p61K4H0kr7z/7d377FylHUYx7+PLSi0Sm1BQyiXYipiBMpFAgFrWwqpQqBETUSIjTHBGDQQFaEmIpiQYEgUYwyRW7lEilC5NP4BJYCpIYBAWy6KjVyqIJQDUS4FLIKPf8y7sD3s7DkQd2d79vkkzZl5dzr77pOd019n3nmH6klFv2h/+lDN/tonfj/wrAsu7pjnPrvs0LG93caNG1m6dCnLli0bc9utzaZNm5g6derYGw6p5FMv2XSXfOolm+6ST71hyWb+/Pn32z6o02t9fZY6/Z0I/k2q+T7rqEPbFtW3pJOAg6imRaqzEFhm+9XSt9ZE858qheY0qqckveORmJ1k4vfxySS63SWfesmmu+RTL9l0l3zqJZv+3zTUz4ng/10K0zrPA9MktYrumcDTbfteSPWIzGNtb+7w99/alM4TvV8OfMv2PsA5vHOC+p7KxO8RERExKPp9hnM3SYfavotRE8FLuptqrGOd1kTwrUvqUxjfRPAdlUv4d1BN2H4NsAS4CUDS/sCvgEW2R8bY1SrgLElXt19Spxqb+kyZOP5E4B9l+5fLaz21fPnyXr9FRERExLj0+wxn3yeCH8MZwHckPQrMAFqnAc+nugx+XbkTfWXdDmzfDKwE7pO0Dmjd9PRDqhukbqUaq9pyDXC6pLWSPvYu+xsRERGx1en3Gc5+TgQ/5uhc2493en/bC8fTp7btzwPOG9V2IXBhh23vJNMiRURExBAZ6onfIyIiIqL3+naG0/YGYFwTwVM927zddbbPfa/vLekGYNao5jNsj+vO8bKPfYCrRjVvrpubMyIiIiIq/b6kPqZSWL7n4rJmn8ePvdWY+3iIauL2iIiIiHgX+jbxe7x7kl4G1jfdjwG1I9XUVtFZ8qmXbLpLPvWSTXfJp96wZLO77Z06vTBwZzhjC+vrZuwfdpLuSzb1kk+9ZNNd8qmXbLpLPvWSTW4aioiIiIgeS8EZERERET2VgnOwXdR0BwZYsuku+dRLNt0ln3rJprvkU2/os8lNQxERERHRUznDGRERERE9lYJzAElaJGm9pEclndl0f5om6TJJI5IebmubLulWSX8tPz/cZB+bImlXSXdIekTSnySdWtqTDyDpA5L+KOmBks85pX2WpHtKPr+RtG3TfW2KpEmS1kr6XVlPNoWkDZIekrRO0n2lLccWIGmapBWS/lJ+/xyabCqS9irfmdaflySdNuz5pOAcMJImAb+kekb8J4ETJA37s9cvBxaNajsTuM32bOC2sj6M3gC+a3tv4BDglPJ9ST6VzcAC2/tRPbhhkaRDgJ8APyv5/Av4eoN9bNqpwCNt68lmS/Ntz2mb0ibHVuXnwM22PwHsR/UdSjaA7fXlOzMHOBB4FbiBIc8nBefgORh41Pbjtl8HrgGOa7hPjbK9GvjnqObjgCvK8hXA4r52akDYfsb2mrL8MtUv/V1IPgC4sqmsblP+GFgArCjtQ5uPpJnA0cAlZV0km7EM/bEl6UPAXOBSANuv236BZNPJEcBjtv/GkOeTgnPw7AI82bb+VGmLLX3U9jNQFV3ARxruT+Mk7QHsD9xD8nlLuWS8DhgBbgUeA16w/UbZZJiPsQuA7wP/LeszSDbtDKySdL+kk0tbji3YE3gOWFaGY1wiaQrJppMvA8vL8lDnk4Jz8KhDW6YSiK4kTQV+C5xm+6Wm+zNIbL9ZLm3NpLqCsHenzfrbq+ZJOgYYsX1/e3OHTYcumzaH2T6AaojTKZLmNt2hATEZOAC40Pb+wCsM2eXh8Sjjn48Frmu6L4MgBefgeQrYtW19JvB0Q30ZZM9K2hmg/BxpuD+NkbQNVbH5a9vXl+bkM0q55Pd7qrGu0yS1Hu07rMfYYcCxkjZQDd1ZQHXGM9kUtp8uP0eoxuAdTI4tqP6desr2PWV9BVUBmmy29Dlgje1ny/pQ55OCc/DcC8wud4puS3U6fmXDfRpEK4ElZXkJcFODfWlMGXN3KfCI7Z+2vZR8AEk7SZpWlrcDFlKNc70D+GLZbCjzsb3U9kzbe1D9nrnd9okkGwAkTZH0wdYycBTwMDm2sL0ReFLSXqXpCODPJJvRTuDty+kw5Plk4vcBJOnzVGcaJgGX2T634S41StJyYB6wI/As8CPgRuBaYDfg78CXbI++sWjCk3Q48AfgId4eh/cDqnGcyUfal2pw/iSq/2Bfa/vHkvakOqs3HVgLnGR7c3M9bZakecD3bB+TbColhxvK6mTgatvnSppBji0kzaG62Wxb4HHga5RjjCHPBkDS9lT3Y+xp+8XSNtTfnRScEREREdFTuaQeERERET2VgjMiIiIieioFZ0RERET0VArOiIiIiOipFJwRERER0VOTx94kIiIGhaQ3qabBallse0ND3YmIGJdMixQRsRWRtMn21D6+3+S2Z6tHRLwnuaQeETGBSNpZ0mpJ6yQ9LOkzpX2RpDWSHpB0W2mbLulGSQ9KurtMlI+ksyVdJGkVcKWkSZLOl3Rv2fYbDX7EiNgK5ZJ6RMTWZTtJ68ryE7aPH/X6V4BbylNxJgHbS9oJuBiYa/sJSdPLtucAa20vlrQAuBKYU147EDjc9muSTgZetP1pSe8H7pS0yvYTvfygETFxpOCMiNi6vGZ7TpfX7wUuk7QNcKPtdeXRlatbBWLb4/QOB75Q2m6XNEPSDuW1lbZfK8tHAftKaj1jfQdgNpCCMyLGJQVnRMQEYnu1pLnA0cBVks4HXgA6DdhXp12Un6+M2u7btm/5v3Y2IoZGxnBGREwgknYHRmxfDFwKHADcBXxW0qyyTeuS+mrgxNI2D3je9ksddnsL8M1y1hRJH5c0pacfJCImlJzhjIiYWOYBp0v6D7AJ+Krt58o4zOslvQ8YAY4EzgaWSXoQeBVYUrPPS4A9gDWSBDwHLO7lh4iIiSXTIkVERERET+WSekRERET0VArOiIiIiOipFJwRERER0VMpOCMiIiKip1JwRkRERERPpeCMiIiIiJ5KwRkRERERPZWCMyIiIiJ66n8z6R2hL1Nk9QAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<Figure size 720x576 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "#plotting features importance by F score\n",
    "fig, ax = plt.subplots(figsize=(10,8))\n",
    "plot_importance(fit_model2,ax=ax)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From plot above we can see that there are features with a big difference concerning its importance, which suggests that feature selection might be useful to improve performance metrics.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In code block below, I set up a function to predict and evaluate an xgboost classifier given a threshold of feature importance."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [],
   "source": [
    "#create a vector of thresholds with feature importances\n",
    "thresholds = sort(fit_model2.feature_importances_)\n",
    "\n",
    "#create a function to train and evaluate using a subset of features given a threshold of importance\n",
    "def evaluate_feature_reduction(thresh):\n",
    "    \n",
    "    ### Feature reduction logic ###\n",
    "    \n",
    "    # create selection object to select features above \"thresh\" importance level\n",
    "    selection = SelectFromModel(fit_model2, threshold=thresh, prefit=True) #fit_model2 once it is more sensitive to rare class\n",
    "    \n",
    "    # select train dataset only with features above \"thresh\" importance level\n",
    "    X_train_subset = selection.transform(X_train2) #X_train2 (oversampled training set)\n",
    "    \n",
    "    selection_clf = XGBClassifier(random_state=1) #create classifier object\n",
    "    selection_fit_model = selection_clf.fit(X_train_subset,y_train2) #train classifier with feature reduction\n",
    "    \n",
    "    # select test dataset only with features above \"thresh\" importance level\n",
    "    X_test_subset = selection.transform(X_test.iloc[:,1:])\n",
    "\n",
    "    #create predictions vector and prediction probabilities. I will predict X_test instances (\"unseen\" instances)\n",
    "    y_pred_subset = selection_fit_model.predict(X_test_subset) #prediction vector with classes\n",
    "    y_pred_prob_subset = selection_fit_model.predict_proba(X_test_subset) #prediction probabilities vector [P(0), P(1)]\n",
    "\n",
    "    pred_vs_true_subset = np.c_[y_pred_subset,y_test.iloc[:,1]] # set up a 2-D array that has predictions and true (test) values side by side\n",
    "    pred_vs_true_df_subset = pd.DataFrame(data=pred_vs_true_subset,columns = [\"pred\",\"true\"]) # convert it to dataframe to be readable by evaluation algorithm\n",
    "    \n",
    "    ### Evaluation logic ###\n",
    "    \n",
    "    confusion_matrix_subset = set_matrix(pred_vs_true_df_subset)\n",
    "    n_features = X_train_subset.shape[1]\n",
    "    avg_class_accuracy_hm = round(calc_avg_class_acc_hm(confusion_matrix_subset)*100,4)\n",
    "    normalized_gini = round(gini_normalizedc(y_test.iloc[:,1], y_pred_prob_subset),4)\n",
    "    threshold = thresh\n",
    "    return n_features, avg_class_accuracy_hm, normalized_gini, thresh\n",
    "    \n",
    "fr_results = [evaluate_feature_reduction(thresh) for thresh in thresholds]\n",
    "\n",
    "fr_n_features = [x[0] for x in fr_results] #get number of features used in prediction\n",
    "fr_avg_class_accuracy_hm = [x[1] for x in fr_results]  #get average class accuracy \n",
    "fr_normalized_gini = [x[2] for x in fr_results] #get normalized gini index\n",
    "fr_threshold = [x[3] for x in fr_results] #get threshold to use later\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In plots below, I compute average class accuracy and normalized gini coefficient for each subset of features. Given that the subset of features was chosen based on importance threshold and there are seven features with importance equal to zero, this resulted in the same 56 features (all features) being chosen for the first 8 iterations and consequently outputing the same performance metrics - hence the gap observed from n=55 to n=47 in plots below."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Feature reduction results - average class accuracy:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.bar(fr_n_features,fr_avg_class_accuracy_hm)\n",
    "plt.title('Average Class Accuracy')\n",
    "plt.xlabel('Number of features')\n",
    "plt.ylabel('Average Class Accuracy (%)')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Feature reduction results - normalized gini coefficient:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# plot feature reduction results - normalized gini coefficient\n",
    "plt.bar(fr_n_features,fr_normalized_gini)\n",
    "plt.title('Normalized Gini Coefficient')\n",
    "plt.xlabel('Number of features')\n",
    "plt.ylabel('Normalized Gini Coefficient')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From the outputs below I conclude that feature reduction has little impact in both average class accuracy and normalized gini coefficient."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "However, there is an optimal subset of features that produce the highest gini coefficient that I can retrieve to make better predictions based on this criteria."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum normalized gini coefficent: 0.2762 for 47 features (59.34%).\n"
     ]
    }
   ],
   "source": [
    "#obtain index of maximum gini coefficient\n",
    "max_index = fr_normalized_gini.index(max(fr_normalized_gini))\n",
    "max_threshold = fr_threshold[max_index]\n",
    "max_ginic = fr_normalized_gini[max_index]\n",
    "max_avg_class_accuracy_hm = fr_avg_class_accuracy_hm[max_index]\n",
    "print(\"Maximum normalized gini coefficent: \"+ str(max_ginic) +\" for \" + str(fr_n_features[max_index]) + \" features (\"+ str(round(max_avg_class_accuracy_hm,2))+\"%).\" )"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Build predictive model based on maximum gini coefficient obtained after feature reduction."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [],
   "source": [
    "# create selection object to select features above \"max_threshold\" importance level\n",
    "selection = SelectFromModel(fit_model2, threshold=max_threshold, prefit=True) #fit_model2 once it is more sensitive to rare class\n",
    "    \n",
    "# select train dataset only with features above \"max_thresh\" importance level\n",
    "X_train_subset = selection.transform(X_train2) #X_train2 (oversampled training set)\n",
    "    \n",
    "selection_clf = XGBClassifier(random_state=1) #create classifier object\n",
    "selection_fit_kaggle_model = selection_clf.fit(X_train_subset,y_train2) #train classifier with feature reduction"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Kaggle Prediction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "#load kaggle test dataset\n",
    "testdf = pd.read_csv(\"./testdf1.csv\")\n",
    "\n",
    "# select kaggle test dataset only with features above \"thresh\" importance level\n",
    "kaggle_test_subset = selection.transform(testdf.iloc[:,1:])\n",
    "\n",
    "#create prediction probabilities 2D vector. I will predict kaggle_test_subset\n",
    "kaggle_pred_prob_subset = selection_fit_kaggle_model.predict_proba(kaggle_test_subset) #prediction probabilities vector [P(0), P(1)]\n",
    "\n",
    "preds_index = testdf.iloc[:,0] #get indexes id from testdf\n",
    "preds_test = np.around(kaggle_pred_prob_subset[:,1],4) # round probability predictions vector to 4 decimal places\n",
    "\n",
    "#create dataframe with indexes and probabilities\n",
    "output = pd.DataFrame({'Id': preds_index, 'target': preds_test})\n",
    "\n",
    "# save to txt\n",
    "np.savetxt(\"submission.csv\", output.values, delimiter=\",\", fmt=['%d','%1.4f'], header=\"id,target\",comments='')\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}