06-8. RMSprop, relu+sigmoid.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:41:01.694377Z",
     "start_time": "2019-11-27T16:40:54.210691Z"
    }
   },
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "import seaborn as sns\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.preprocessing import LabelEncoder\n",
    "from sklearn.decomposition import PCA"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:41:26.091872Z",
     "start_time": "2019-11-27T16:41:07.516688Z"
    }
   },
   "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>beer_id</th>\n",
       "      <th>username</th>\n",
       "      <th>date</th>\n",
       "      <th>text</th>\n",
       "      <th>look</th>\n",
       "      <th>smell</th>\n",
       "      <th>taste</th>\n",
       "      <th>feel</th>\n",
       "      <th>overall</th>\n",
       "      <th>score</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>271781</td>\n",
       "      <td>bluejacket74</td>\n",
       "      <td>2017-03-17</td>\n",
       "      <td>750 ml bottle, 2016 vintage, bottle #304 of...</td>\n",
       "      <td>4.00</td>\n",
       "      <td>4.00</td>\n",
       "      <td>4.00</td>\n",
       "      <td>4.25</td>\n",
       "      <td>4.00</td>\n",
       "      <td>4.03</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>125646</td>\n",
       "      <td>_dirty_</td>\n",
       "      <td>2017-12-21</td>\n",
       "      <td></td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.50</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>125646</td>\n",
       "      <td>CJDUBYA</td>\n",
       "      <td>2017-12-21</td>\n",
       "      <td></td>\n",
       "      <td>4.75</td>\n",
       "      <td>4.75</td>\n",
       "      <td>4.75</td>\n",
       "      <td>4.75</td>\n",
       "      <td>4.75</td>\n",
       "      <td>4.75</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>125646</td>\n",
       "      <td>GratefulBeerGuy</td>\n",
       "      <td>2017-12-20</td>\n",
       "      <td>0% 16 oz can. Funny story: As I finally wal...</td>\n",
       "      <td>4.75</td>\n",
       "      <td>4.75</td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.58</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>125646</td>\n",
       "      <td>LukeGude</td>\n",
       "      <td>2017-12-20</td>\n",
       "      <td>Classic TH NEIPA. Overflowing head and bouq...</td>\n",
       "      <td>4.25</td>\n",
       "      <td>4.50</td>\n",
       "      <td>4.25</td>\n",
       "      <td>4.25</td>\n",
       "      <td>4.25</td>\n",
       "      <td>4.31</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   beer_id         username        date  \\\n",
       "0   271781     bluejacket74  2017-03-17   \n",
       "1   125646          _dirty_  2017-12-21   \n",
       "2   125646          CJDUBYA  2017-12-21   \n",
       "3   125646  GratefulBeerGuy  2017-12-20   \n",
       "4   125646         LukeGude  2017-12-20   \n",
       "\n",
       "                                                text  look  smell  taste  \\\n",
       "0     750 ml bottle, 2016 vintage, bottle #304 of...  4.00   4.00   4.00   \n",
       "1                                                     4.50   4.50   4.50   \n",
       "2                                                     4.75   4.75   4.75   \n",
       "3     0% 16 oz can. Funny story: As I finally wal...  4.75   4.75   4.50   \n",
       "4     Classic TH NEIPA. Overflowing head and bouq...  4.25   4.50   4.25   \n",
       "\n",
       "   feel  overall  score  \n",
       "0  4.25     4.00   4.03  \n",
       "1  4.50     4.50   4.50  \n",
       "2  4.75     4.75   4.75  \n",
       "3  4.50     4.50   4.58  \n",
       "4  4.25     4.25   4.31  "
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df = pd.read_csv('reviews.csv')\n",
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:42:31.199539Z",
     "start_time": "2019-11-27T16:42:31.197539Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(9073128, 10)\n"
     ]
    }
   ],
   "source": [
    "print(df.shape) #원본 데이터 900만개를 사용"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:42:35.466500Z",
     "start_time": "2019-11-27T16:42:35.277458Z"
    }
   },
   "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>beerid</th>\n",
       "      <th>userid</th>\n",
       "      <th>rating</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>271781</td>\n",
       "      <td>bluejacket74</td>\n",
       "      <td>4.03</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>125646</td>\n",
       "      <td>_dirty_</td>\n",
       "      <td>4.50</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>125646</td>\n",
       "      <td>CJDUBYA</td>\n",
       "      <td>4.75</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>125646</td>\n",
       "      <td>GratefulBeerGuy</td>\n",
       "      <td>4.58</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>125646</td>\n",
       "      <td>LukeGude</td>\n",
       "      <td>4.31</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   beerid           userid  rating\n",
       "0  271781     bluejacket74    4.03\n",
       "1  125646          _dirty_    4.50\n",
       "2  125646          CJDUBYA    4.75\n",
       "3  125646  GratefulBeerGuy    4.58\n",
       "4  125646         LukeGude    4.31"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "#잠깐 전처리\n",
    "\n",
    "cols = ['beer_id','username','score']\n",
    "df = df[cols]\n",
    "df.rename(columns = {'beer_id':'beerid','username':'userid','score':'rating'}, inplace = True)\n",
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:42:43.052209Z",
     "start_time": "2019-11-27T16:42:38.373155Z"
    }
   },
   "outputs": [],
   "source": [
    "#딥러닝에 이용할 컬럼 변경: user, beer, rating\n",
    "\n",
    "user_enc = LabelEncoder()\n",
    "uid = df['userid'].values\n",
    "df['user'] = user_enc.fit_transform(uid.tolist()) #array 인식이 안돼서 list로 빼냄.(userid의 경우만 오류 발생)\n",
    "n_users = df['user'].nunique()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(164935, 309542, 1.0, 5.0)"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "item_enc = LabelEncoder()\n",
    "df['beer'] = item_enc.fit_transform(df['beerid'].values)\n",
    "n_beers = df['beer'].nunique()\n",
    "\n",
    "df['rating'] = df['rating'].values.astype(np.float32)\n",
    "min_rating = min(df['rating'])\n",
    "max_rating = max(df['rating'])\n",
    "\n",
    "n_users, n_beers, min_rating, max_rating #n_users, n_beers: 유니크한 값 개수, rating의 최솟값과 최댓값."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:42:46.866068Z",
     "start_time": "2019-11-27T16:42:46.855069Z"
    }
   },
   "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>user</th>\n",
       "      <th>beer</th>\n",
       "      <th>rating</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>102213</td>\n",
       "      <td>231507</td>\n",
       "      <td>4.03</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>94429</td>\n",
       "      <td>112229</td>\n",
       "      <td>4.50</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>15635</td>\n",
       "      <td>112229</td>\n",
       "      <td>4.75</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>34826</td>\n",
       "      <td>112229</td>\n",
       "      <td>4.58</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>55007</td>\n",
       "      <td>112229</td>\n",
       "      <td>4.31</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "     user    beer  rating\n",
       "0  102213  231507    4.03\n",
       "1   94429  112229    4.50\n",
       "2   15635  112229    4.75\n",
       "3   34826  112229    4.58\n",
       "4   55007  112229    4.31"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df2=df[['user','beer','rating']] \n",
    "df2.head() #완료된 데이터셋. 여기서 'user','beer','rating'만 사용"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:42:50.351854Z",
     "start_time": "2019-11-27T16:42:49.645695Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((7258502, 2), (1814626, 2), (7258502,), (1814626,))"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "#train, test 셋 분할: test size는 0.2, random_state는 임의로 지정\n",
    "X = df[['user', 'beer']].values\n",
    "y = df['rating'].values\n",
    "\n",
    "X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)\n",
    "\n",
    "X_train.shape, X_test.shape, y_train.shape, y_test.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:42:54.302743Z",
     "start_time": "2019-11-27T16:42:54.300743Z"
    }
   },
   "outputs": [],
   "source": [
    "#딥러닝 모델을 만들 때 필요한 또다른 변수는: 잠재요인의 개수를 정하는 것!\n",
    "#잠재요인의 개수는 원하는 대로 설정하면 되지만, 주의할 것은 user와 beer 모두에 대해 같은 사이즈를 줘야 한다는 점.\n",
    "\n",
    "#우리는 50개로 간다!(X에 대해서만 적용)\n",
    "n_factors = 50\n",
    "\n",
    "#X를 array로 만들기\n",
    "X_train_array = [X_train[:, 0], X_train[:, 1]]\n",
    "X_test_array = [X_test[:, 0], X_test[:, 1]]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "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>name</th>\n",
       "      <th>brewery_id</th>\n",
       "      <th>state</th>\n",
       "      <th>country</th>\n",
       "      <th>style</th>\n",
       "      <th>availability</th>\n",
       "      <th>abv</th>\n",
       "      <th>notes</th>\n",
       "      <th>retired</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>202522</td>\n",
       "      <td>Olde Cogitator</td>\n",
       "      <td>2199</td>\n",
       "      <td>CA</td>\n",
       "      <td>US</td>\n",
       "      <td>English Oatmeal Stout</td>\n",
       "      <td>Rotating</td>\n",
       "      <td>7.3</td>\n",
       "      <td>No notes at this time.</td>\n",
       "      <td>f</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>82352</td>\n",
       "      <td>Konrads Stout Russian Imperial Stout</td>\n",
       "      <td>18604</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NO</td>\n",
       "      <td>Russian Imperial Stout</td>\n",
       "      <td>Rotating</td>\n",
       "      <td>10.4</td>\n",
       "      <td>No notes at this time.</td>\n",
       "      <td>f</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>214879</td>\n",
       "      <td>Scottish Right</td>\n",
       "      <td>44306</td>\n",
       "      <td>IN</td>\n",
       "      <td>US</td>\n",
       "      <td>Scottish Ale</td>\n",
       "      <td>Year-round</td>\n",
       "      <td>4.0</td>\n",
       "      <td>No notes at this time.</td>\n",
       "      <td>t</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>320009</td>\n",
       "      <td>MegaMeow Imperial Stout</td>\n",
       "      <td>4378</td>\n",
       "      <td>WA</td>\n",
       "      <td>US</td>\n",
       "      <td>American Imperial Stout</td>\n",
       "      <td>Winter</td>\n",
       "      <td>8.7</td>\n",
       "      <td>Every time this year</td>\n",
       "      <td>f</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>246438</td>\n",
       "      <td>Peaches-N-Cream</td>\n",
       "      <td>44617</td>\n",
       "      <td>PA</td>\n",
       "      <td>US</td>\n",
       "      <td>American Cream Ale</td>\n",
       "      <td>Rotating</td>\n",
       "      <td>5.1</td>\n",
       "      <td>No notes at this time.</td>\n",
       "      <td>f</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "       id                                  name  brewery_id state country  \\\n",
       "0  202522                        Olde Cogitator        2199    CA      US   \n",
       "1   82352  Konrads Stout Russian Imperial Stout       18604   NaN      NO   \n",
       "2  214879                        Scottish Right       44306    IN      US   \n",
       "3  320009               MegaMeow Imperial Stout        4378    WA      US   \n",
       "4  246438                       Peaches-N-Cream       44617    PA      US   \n",
       "\n",
       "                     style availability   abv                   notes retired  \n",
       "0    English Oatmeal Stout     Rotating   7.3  No notes at this time.       f  \n",
       "1   Russian Imperial Stout     Rotating  10.4  No notes at this time.       f  \n",
       "2             Scottish Ale   Year-round   4.0  No notes at this time.       t  \n",
       "3  American Imperial Stout       Winter   8.7    Every time this year       f  \n",
       "4       American Cream Ale     Rotating   5.1  No notes at this time.       f  "
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "beers = pd.read_csv('beers.csv')\n",
    "beers.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "## create a dictionary out of beerid, beer original title\n",
    "beers_dict = beers.set_index('id')['name'].to_dict()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "##시각화(각 epoch마다의 MSE, MAE 값 비교)\n",
    "\n",
    "def plot_history(hist):\n",
    "    \n",
    "    plt.figure(figsize=(10,15))\n",
    "\n",
    "    plt.subplot(2,1,1)\n",
    "    plt.xlabel('Epoch')\n",
    "    plt.ylabel('Mean Abs Error [rating]')\n",
    "    plt.plot(hist['epoch'], hist['mae'], label='Train Error')\n",
    "    plt.plot(hist['epoch'], hist['val_mae'],label = 'Test Error')\n",
    "    plt.legend()\n",
    "\n",
    "    plt.subplot(2,1,2)\n",
    "    plt.xlabel('Epoch')\n",
    "    plt.ylabel('Mean Square Error [$rating^2$]')\n",
    "    plt.plot(hist['epoch'], hist['mse'], label='Train Error')\n",
    "    plt.plot(hist['epoch'], hist['val_mse'],label = 'Test Error')\n",
    "    plt.legend()\n",
    "    plt.show()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Recommdation System using Keras \n",
    "- 각 유저와 각 맥주를 나타내기 위해 'embedding'이라는 것을 활용한다.->(설명부분 캡처)\n",
    "- 'embeddings'는 벡터이다(n_factor 사이즈의). 시작은 랜덤한 숫자로 가고, 각 유저/맥주의 필수적인 특징을 잡으면 fit model을 해보자.\n",
    "- fit model 과정) 유저 벡터와 맥주 벡터를 스칼라곱(dot product)하여 예측 평점을 계산함.\n",
    "- 이를 위해선 유니크한 유저 수와 맥주 수가 필요, -> 각각의 embedding matrix의 사이즈를 정의하기 위함."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2019-11-27T16:43:14.561307Z",
     "start_time": "2019-11-27T16:43:05.998377Z"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "#keras 라이브러리 로드\n",
    "from keras.models import Model\n",
    "from keras.layers import Input, Reshape, Dot\n",
    "from keras.layers import Embedding, Dense, Dropout\n",
    "from keras.optimizers import Adam\n",
    "from keras.regularizers import l2\n",
    "from keras.layers import Add, Activation, Lambda, Concatenate"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 옵티마이저 변경\n",
    "- RMSprop"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "class EmbeddingLayer:\n",
    "    def __init__(self, n_items, n_factors):\n",
    "        self.n_items = n_items\n",
    "        self.n_factors = n_factors\n",
    "    \n",
    "    def __call__(self, x):\n",
    "        x = Embedding(self.n_items, self.n_factors, embeddings_initializer='he_normal')(x) #initializer는 정규분포\n",
    "        x = Reshape((self.n_factors,))(x)\n",
    "        return x\n",
    "\n",
    "def RecommenderV9(n_users, n_beers, n_factors, min_rating, max_rating):\n",
    "    \n",
    "    #input: user, beer\n",
    "    user = Input(shape=(1,))\n",
    "    u = EmbeddingLayer(n_users, n_factors)(user) #u는 user * n_factor\n",
    "    ub = EmbeddingLayer(n_users, 1)(user) #user_bias\n",
    "    \n",
    "    beer = Input(shape=(1,))\n",
    "    b = EmbeddingLayer(n_beers, n_factors)(beer) #b은 beer * n_factor\n",
    "    bb = EmbeddingLayer(n_beers, 1)(beer) #beer_bias\n",
    "    \n",
    "    #input 연산- activation 2가지를 추가. relu와 sigmoid\n",
    "    x = Concatenate()([u, b])\n",
    "    \n",
    "    x = Dense(10, kernel_initializer='he_normal')(x)\n",
    "    x = Activation('relu')(x)\n",
    "    \n",
    "    x = Dense(1, kernel_initializer='he_normal')(x)\n",
    "    x = Activation('sigmoid')(x)\n",
    "    x = Lambda(lambda x: x * (max_rating - min_rating) + min_rating)(x)\n",
    "\n",
    "\n",
    "    #최종 모델 만들기(input은 user, beer 2개를 받고, output은 x를 반환.)\n",
    "    model = Model(inputs=[user, beer], outputs=x)\n",
    "    model.compile(loss=['mse'], metrics = ['mae','mse'] ,optimizer='RMSprop') \n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"model_1\"\n",
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "input_1 (InputLayer)            (None, 1)            0                                            \n",
      "__________________________________________________________________________________________________\n",
      "input_2 (InputLayer)            (None, 1)            0                                            \n",
      "__________________________________________________________________________________________________\n",
      "embedding_1 (Embedding)         (None, 1, 50)        8246750     input_1[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "embedding_3 (Embedding)         (None, 1, 50)        15477100    input_2[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "reshape_1 (Reshape)             (None, 50)           0           embedding_1[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "reshape_3 (Reshape)             (None, 50)           0           embedding_3[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "concatenate_1 (Concatenate)     (None, 100)          0           reshape_1[0][0]                  \n",
      "                                                                 reshape_3[0][0]                  \n",
      "__________________________________________________________________________________________________\n",
      "dense_1 (Dense)                 (None, 10)           1010        concatenate_1[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "activation_1 (Activation)       (None, 10)           0           dense_1[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "dense_2 (Dense)                 (None, 1)            11          activation_1[0][0]               \n",
      "__________________________________________________________________________________________________\n",
      "activation_2 (Activation)       (None, 1)            0           dense_2[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "lambda_1 (Lambda)               (None, 1)            0           activation_2[0][0]               \n",
      "==================================================================================================\n",
      "Total params: 23,724,871\n",
      "Trainable params: 23,724,871\n",
      "Non-trainable params: 0\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "re9model = RecommenderV9(n_users, n_beers, n_factors, min_rating, max_rating)\n",
    "re9model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:From C:\\Users\\suyn3\\.conda\\envs\\tf0\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:422: The name tf.global_variables is deprecated. Please use tf.compat.v1.global_variables instead.\n",
      "\n",
      "Train on 7258502 samples, validate on 1814626 samples\n",
      "Epoch 1/10\n",
      "7258502/7258502 [==============================] - 1421s 196us/step - loss: 0.2087 - mae: 0.3236 - mse: 0.2087 - val_loss: 0.1873 - val_mae: 0.3050 - val_mse: 0.1873\n",
      "Epoch 2/10\n",
      "7258502/7258502 [==============================] - 1427s 197us/step - loss: 0.1799 - mae: 0.2973 - mse: 0.1799 - val_loss: 0.1838 - val_mae: 0.3013 - val_mse: 0.1838\n",
      "Epoch 3/10\n",
      "7258502/7258502 [==============================] - 1408s 194us/step - loss: 0.1744 - mae: 0.2920 - mse: 0.1744 - val_loss: 0.1822 - val_mae: 0.3003 - val_mse: 0.1822\n",
      "Epoch 4/10\n",
      "7258502/7258502 [==============================] - 1406s 194us/step - loss: 0.1708 - mae: 0.2886 - mse: 0.1708 - val_loss: 0.1819 - val_mae: 0.3000 - val_mse: 0.1819\n",
      "Epoch 5/10\n",
      "7258502/7258502 [==============================] - 1415s 195us/step - loss: 0.1681 - mae: 0.2860 - mse: 0.1681 - val_loss: 0.1819 - val_mae: 0.3004 - val_mse: 0.1819\n",
      "Epoch 6/10\n",
      "7258502/7258502 [==============================] - 25917s 4ms/step - loss: 0.1659 - mae: 0.2838 - mse: 0.1659 - val_loss: 0.1816 - val_mae: 0.2994 - val_mse: 0.1816\n",
      "Epoch 7/10\n",
      "7258502/7258502 [==============================] - 1459s 201us/step - loss: 0.1638 - mae: 0.2819 - mse: 0.1638 - val_loss: 0.1826 - val_mae: 0.3010 - val_mse: 0.1826\n",
      "Epoch 8/10\n",
      "7258502/7258502 [==============================] - 1420s 196us/step - loss: 0.1621 - mae: 0.2801 - mse: 0.1621 - val_loss: 0.1820 - val_mae: 0.2996 - val_mse: 0.182001 - mse\n",
      "Epoch 9/10\n",
      "7258502/7258502 [==============================] - 1431s 197us/step - loss: 0.1605 - mae: 0.2787 - mse: 0.1605 - val_loss: 0.1827 - val_mae: 0.2996 - val_mse: 0.1827\n",
      "Epoch 10/10\n",
      "7258502/7258502 [==============================] - 1420s 196us/step - loss: 0.1592 - mae: 0.2774 - mse: 0.1592 - val_loss: 0.1830 - val_mae: 0.3018 - val_mse: 0.1830\n"
     ]
    }
   ],
   "source": [
    "## fit model\n",
    "seed = 7\n",
    "np.random.seed(seed)\n",
    "history_mf = re9model.fit(x=X_train_array, y=y_train, batch_size=1000, epochs=10,\n",
    "                    verbose=1, validation_data=(X_test_array, y_test)) "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "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>val_loss</th>\n",
       "      <th>val_mae</th>\n",
       "      <th>val_mse</th>\n",
       "      <th>loss</th>\n",
       "      <th>mae</th>\n",
       "      <th>mse</th>\n",
       "      <th>epoch</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.187263</td>\n",
       "      <td>0.304957</td>\n",
       "      <td>0.187263</td>\n",
       "      <td>0.208715</td>\n",
       "      <td>0.323569</td>\n",
       "      <td>0.208715</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.183751</td>\n",
       "      <td>0.301345</td>\n",
       "      <td>0.183751</td>\n",
       "      <td>0.179891</td>\n",
       "      <td>0.297287</td>\n",
       "      <td>0.179891</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.182215</td>\n",
       "      <td>0.300274</td>\n",
       "      <td>0.182216</td>\n",
       "      <td>0.174359</td>\n",
       "      <td>0.291985</td>\n",
       "      <td>0.174359</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.181885</td>\n",
       "      <td>0.300009</td>\n",
       "      <td>0.181885</td>\n",
       "      <td>0.170839</td>\n",
       "      <td>0.288638</td>\n",
       "      <td>0.170839</td>\n",
       "      <td>3</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.181937</td>\n",
       "      <td>0.300394</td>\n",
       "      <td>0.181937</td>\n",
       "      <td>0.168124</td>\n",
       "      <td>0.286028</td>\n",
       "      <td>0.168124</td>\n",
       "      <td>4</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>5</td>\n",
       "      <td>0.181562</td>\n",
       "      <td>0.299389</td>\n",
       "      <td>0.181562</td>\n",
       "      <td>0.165865</td>\n",
       "      <td>0.283837</td>\n",
       "      <td>0.165865</td>\n",
       "      <td>5</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>6</td>\n",
       "      <td>0.182567</td>\n",
       "      <td>0.300999</td>\n",
       "      <td>0.182567</td>\n",
       "      <td>0.163841</td>\n",
       "      <td>0.281859</td>\n",
       "      <td>0.163841</td>\n",
       "      <td>6</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>7</td>\n",
       "      <td>0.181956</td>\n",
       "      <td>0.299566</td>\n",
       "      <td>0.181956</td>\n",
       "      <td>0.162059</td>\n",
       "      <td>0.280141</td>\n",
       "      <td>0.162059</td>\n",
       "      <td>7</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>8</td>\n",
       "      <td>0.182664</td>\n",
       "      <td>0.299620</td>\n",
       "      <td>0.182664</td>\n",
       "      <td>0.160548</td>\n",
       "      <td>0.278695</td>\n",
       "      <td>0.160548</td>\n",
       "      <td>8</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>9</td>\n",
       "      <td>0.182969</td>\n",
       "      <td>0.301787</td>\n",
       "      <td>0.182970</td>\n",
       "      <td>0.159157</td>\n",
       "      <td>0.277390</td>\n",
       "      <td>0.159157</td>\n",
       "      <td>9</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   val_loss   val_mae   val_mse      loss       mae       mse  epoch\n",
       "0  0.187263  0.304957  0.187263  0.208715  0.323569  0.208715      0\n",
       "1  0.183751  0.301345  0.183751  0.179891  0.297287  0.179891      1\n",
       "2  0.182215  0.300274  0.182216  0.174359  0.291985  0.174359      2\n",
       "3  0.181885  0.300009  0.181885  0.170839  0.288638  0.170839      3\n",
       "4  0.181937  0.300394  0.181937  0.168124  0.286028  0.168124      4\n",
       "5  0.181562  0.299389  0.181562  0.165865  0.283837  0.165865      5\n",
       "6  0.182567  0.300999  0.182567  0.163841  0.281859  0.163841      6\n",
       "7  0.181956  0.299566  0.181956  0.162059  0.280141  0.162059      7\n",
       "8  0.182664  0.299620  0.182664  0.160548  0.278695  0.160548      8\n",
       "9  0.182969  0.301787  0.182970  0.159157  0.277390  0.159157      9"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "## show loss & mae & mse at each epoch\n",
    "hist9 = pd.DataFrame(history_mf.history)\n",
    "hist9['epoch'] = history_mf.epoch\n",
    "hist9"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1814626/1814626 [==============================] - 6s 3us/step\n",
      "## evaluation loss and_metrics ##\n",
      "[0.18296934611134513, 0.30178743600845337, 0.1829695701599121]\n"
     ]
    }
   ],
   "source": [
    "#모델 평가하기; model.evaluate()\n",
    "loss_and_metrics = re9model.evaluate(X_test_array, y_test, batch_size=1000)\n",
    "print('## evaluation loss and_metrics ##')\n",
    "print(loss_and_metrics)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAm0AAANcCAYAAAANWpgCAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADh0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uMy4xLjEsIGh0dHA6Ly9tYXRwbG90bGliLm9yZy8QZhcZAAAgAElEQVR4nOzdeXyU5b3//9c1k4TsJCQZwhaWAAk7RlABBRKoxbr2aG21tlbbemy11fZ4ztf252mtbc+xu1Y9bW0rXTxqtcupa63IJpusYSfsS9iykEAgZJ3r98c9WYCwJDPJPZO8n4/HPDJzzz0zn7CEN5/rvq7LWGsRERERkfDmcbsAEREREbk4hTYRERGRCKDQJiIiIhIBFNpEREREIoBCm4iIiEgEiHK7gM6Wnp5uhwwZ4nYZIiIiIhe1Zs2aMmttRlvPdfvQNmTIEFavXu12GSIiIiIXZYzZd77nNDwqIiIiEgEU2kREREQigEKbiIiISATo9te0iYiISPvV19dTXFxMTU2N26V0S7GxsQwcOJDo6OhLfo1Cm4iIiJyjuLiYpKQkhgwZgjHG7XK6FWst5eXlFBcXM3To0Et+nYZHRURE5Bw1NTWkpaUpsHUCYwxpaWnt7mIqtImIiEibFNg6T0d+bRXaRERERCKAQpuIiIiEnfLyciZOnMjEiRPJzMxkwIABzY/r6uou6T3uueceioqKLvkzf/Ob35CRkdH8ORMnTmzX6zubJiKIiIhI2ElLS6OwsBCAxx9/nMTERB555JEzzrHWYq3F42m7BzV37tx2f+6nP/1pnnrqqfM+39DQQFRUS3y6WA2tNTY24vV6211TE3XaREREJGLs3LmTsWPHcv/995OXl8fhw4e57777mDRpEmPGjOGJJ55oPvfqq6+msLCQhoYGUlJSePTRR5kwYQJTpkyhpKTkkj9z3rx5zJ49m0996lNcdtllbdbw4osvMm7cOMaOHcs3v/lNgObPfeyxx7jiiitYuXJlUN+7Om0iIiJyQd95YzNbDp0I6XuO7p/Mt28c06HXbtmyhblz5/LLX/4SgCeffJI+ffrQ0NBAfn4+t912G6NHjz7jNcePH2fGjBk8+eSTfP3rX+eFF17g0UcfPee9//d//5eFCxc2P24KWitWrGDLli1kZWWxc+fOM2ooLi7mscceY/Xq1fTu3ZvZs2fz5ptvMmfOHI4fP05eXh7f+973OvS9tqZOm4iIiESU7OxsJk+e3Pz45ZdfJi8vj7y8PLZu3cqWLVvOeU1cXBzXXXcdAJdffjl79+5t870//elPU1hY2HyLiYkBYMqUKWRlZbVZw4cffkhBQQHp6elER0dz5513snjxYgBiYmL4+Mc/HpLvW502ERERuaCOdsQ6S0JCQvP9HTt28PTTT7Ny5UpSUlK466672lz/rCl8AXi9XhoaGjr8mWc/ttae93VxcXEhWzpFnbYglVbV8tP3trOr9KTbpYiIiPQ4J06cICkpieTkZA4fPsy7777b5TVcddVVLFiwgPLychoaGnjllVeYMWNGyD9HnbYg+a3l5+/vIDbaw5dnDne7HBERkR4lLy+P0aNHM3bsWIYNG8a0adOCer+zr2n71a9+ddHXDBw4kCeeeIKZM2direXGG2/k+uuvb3c372LMhVp63cGkSZPs6tWrO/UzbnjmA+Kivbx2/9RO/RwREZGusnXrVkaNGuV2Gd1aW7/Gxpg11tpJbZ2v4dEQKMjxsWZfBZXVl7bYn4iIiEh7KbSFQH6uD7+FRdtL3S5FREREuimFthCYMDCFtIQY5m+79IX6RERERNpDoS0EPB7DzBwfi7aX0ujv3tcIioiIiDsU2kKkINdHZXU96/ZXuF2KiIiIdEMKbSFyzch0ojyG9zVEKiIiIp1AoS1EkmOjmTQklQUKbSIiIkErLy9n4sSJTJw4kczMTAYMGND8uK7u0ldreOGFFzhy5Eibz911110MHTq0+X2vueaaUJXfKbS4bgjNyu3L99/eysHK0wxIiXO7HBERkYiVlpZGYWEhAI8//jiJiYk88sgj7X6fF154gby8PDIzM9t8/mc/+xm33HLLeV/f0NBAVFTUeR9f6utCQaEthPJzfXz/7a0s2FbCXVcNdrscERGRbun3v/89zz33HHV1dUydOpVnn30Wv9/PPffcQ2FhIdZa7rvvPvr27UthYSGf/OQniYuLY+XKlWfsQXo+jz32GKWlpezevZvMzExmzJjBvHnzOHnyJLW1tbz77rs88sgj/POf/8QYw7e//W1uu+025s2bx5NPPkl6ejqbN29m48aNIf2+FdpCKDsjgaw+8cxXaBMRke7knUfhSGgDCJnj4Lon2/2yTZs28be//Y1ly5YRFRXFfffdxyuvvEJ2djZlZWXNQamyspKUlBSeeeYZnn32WSZOnNjm+33ta1/j8ccfB2D8+PH84Q9/AGDdunUsXryY2NhYfvOb37B8+XIKCwtJTU3lT3/6E1u2bGH9+vWUlpYyefJkpk+fDsCKFSvYsmULWVlZHfhFuTCFthAyxlCQ6+OVVfupqW8kNtrrdkkiIiLdyrx581i1ahWTJjk7PZ0+fZpBgwbx0Y9+lKKiIh566CE+9rGPce21117S+51vePTmm28mNja2+fG1115LamoqAEuWLOHOO+/E6/WSmZnJ1VdfzerVq4mJiWHKlCmdEthAoS3kCnJ9/G7ZXpbvKic/1+d2OSIiIsHrQEess1hruffee/nud797znMbNmzgnXfe4ec//zl/+ctfeP755zv8OQkJCed9fKF9289+XShp9miIXTmsD/ExXu2OICIi0glmz57Nq6++SllZGeDMMt2/fz+lpaVYa/nEJz7Bd77zHdauXQtAUlISVVVVIa1h+vTpvPLKKzQ2NnL06FGWLl3a3PnrTOq0hVivKC/Thqczf1sJT1iLMcbtkkRERLqNcePG8e1vf5vZs2fj9/uJjo7ml7/8JV6vl89//vPYwL+9P/jBDwC45557+MIXvnDeiQitr2kDWLNmzUVruO2221ixYgUTJkzAGMNPf/pTfL7OH10zF2rxdQeTJk2yq1ev7tLPfHnlfr7x1428+/B0cjKTuvSzRUREQmHr1q2MGjXK7TK6tbZ+jY0xa6y1bbbtNDzaCfJznLStIVIREREJFYW2TpDZO5Yx/ZO1O4KIiIiEjEJbJynI9bF63zEqqy99qw0REZFw0t0voXJTR35tFdo6SX6uD7+FRdtL3S5FRESk3WJjYykvL1dw6wTWWsrLy89YB+5SaPZoJ5kwMIU+CTEs2FbCzRMHuF2OiIhIuwwcOJDi4mJKS9V86AyxsbEMHDiwXa9RaOskXo9hZk4G87eV0Oi3eD1a+kNERCJHdHQ0Q4cOdbsMacWV4VFjzBxjTJExZqcx5tE2nr/fGLPRGFNojFlijBkdOP4RY8yawHNrjDEFXV/9pSvI9VFZXc+6/RVulyIiIiIRrstDmzHGCzwHXAeMBu5oCmWtvGStHWetnQj8EPhp4HgZcKO1dhxwN/DHLiq7Q64ZkYHXY7T0h4iIiATNjU7bFcBOa+1ua20d8Apwc+sTrLUnWj1MAGzg+Dpr7aHA8c1ArDGmVxfU3CG946KZNDhVoU1ERESC5kZoGwAcaPW4OHDsDMaYB4wxu3A6bV9t431uBdZZa2vbeO19xpjVxpjVbl9AOWuUj21HqjhUedrVOkRERCSyuRHa2roi/5z5xNba56y12cD/Ax474w2MGQP8APjXtj7AWvu8tXaStXZSRkZGCEruuIJcZ3eEBUXqtomIiEjHuRHaioFBrR4PBA6d51xwhk9vaXpgjBkI/A34rLV2V6dUGELZGYkM6hPH/K0KbSIiItJxboS2VcAIY8xQY0wM8Cng9dYnGGNGtHp4PbAjcDwFeAv4hrV2aRfVGxRjDAU5PpbuKqOmvtHtckRERCRCdXlos9Y2AA8C7wJbgVettZuNMU8YY24KnPagMWazMaYQ+DrOTFECrxsO/GdgOZBCY4yvq7+H9ioY1Zeaej/Ld5e7XYqIiIhEKFcW17XWvg28fdaxb7W6/9B5Xvc94HudW13oXTm0D3HRXhZsKyE/J+wzpoiIiIQh7T3aBWKjvUwbns77W0u0h5uIiIh0iEJbFynI9XGw8jQ7Sk66XYqIiIhEIIW2LtK09IcW2hUREZGOUGjrIpm9YxndL1mhTURERDpEoa0LFeT6WLOvguPV9W6XIiIiIhFGoa0L5ef6aPRbFu1wd2stERERiTwKbV1o4qAU+iTEsEBDpCIiItJOCm1dyOsxzByZwcKiEhr9WvpDRERELp1CWxfLz/VRUV1P4YEKt0sRERGRCKLQ1sWmj8zA6zGaRSoiIiLtotDWxXrHRTNpcCrzt2kygoiIiFw6hTYXFOT62Hr4BIePn3a7FBEREYkQCm0u0O4IIiIi0l4KbS4Y7ktkYGqclv4QERGRS6bQ5gJjDLNyfSzdWU5NfaPb5YiIiEgEUGhzSX6uj9P1jazYXe52KSIiIhIBFNpcctWwNOKivbquTURERC6JQptLYqO9TBuexvxtJVir3RFERETkwhTaXJSf66O44jQ7S066XYqIiIiEOYU2F2npDxEREblUCm0u6tc7jlH9knlfoU1EREQuQqHNZQW5GazZV8Hx6nq3SxEREZEwptDmsoJcH41+y+Id2otUREREzk+hzWUTB6WSGh+t3RFERETkghTaXOb1GGbm+FhQVEKjX0t/iIiISNsU2sJAfq6Piup6Cg9Uul2KiIiIhCmFtjAwY0QGXo/REKmIiIicl0JbGOgdH83lg1O1XpuIiIicl0JbmCjI9bHl8AkOHz/tdikiIiIShhTawkTT7ggLtmnpDxERETmXQluYGOFLZEBKnIZIRUREpE0KbWHCGMOsUT6W7iyjpr7R7XJEREQkzCi0hZH8XB+n6xv5cM8xt0sRERGRMKPQFkamDEsjNtrD/K1H3S5FREREwoxCWxiJjfYyLTud+UUlWKvdEURERKSFQluYKRjl48Cx0+wqPel2KSIiIhJGFNrCTH6Os/SHZpGKiIhIawptYaZ/Shy5mUm8v1WhTURERFootIWhglwfq/dVcPx0vduliIiISJhQaAtDs0b5aPRbPtih3RFERETEodAWhiYOSiU1PlrXtYmIiEgzhbYw5PUYZozMYGFRKY1+Lf0hIiIiCm1hKz/Xx7FTdawvrnS7FBEREQkDCm1hasbIDLwewwINkYqIiAgKbWErJT6Gy7NSdV2biIiIAAptYS0/18fmQyc4crzG7VJERETEZQptYawg19kdYUGRum0iIiI9nUJbGBvZN5EBKXEaIhURERGFtnBmjKEg18fSnWXU1De6XY6IiIi4SKEtzBXk+qiua+TDPcfcLkVERERcpNAW5qZkpxEb7dHSHyIiIj2cQluYi432MjU7nfnbSrBWuyOIiIj0VAptEaAg18f+Y9XsKj3ldikiIiLiEoW2CJAfWPpj/rajLlciIiIiblFoiwADUuLIzUzS0h8iIiI9mEJbhMjP9bF6bwUnaurdLkVERERcoNAWIWbl+mjwWz7YXuZ2KSIiIuIChbYIcVlWKinx0byv69pERER6JIW2COH1GGaMzGBRUSl+v5b+EBER6WkU2iJIQa6P8lN1rC+udLsUERER6WIKbRFkxsgMPAbtjiAiItIDKbRFkJT4GC4fnMr7Cm0iIiI9jkJbhMnP9bH50AmOnqhxuxQRERHpQgptEaYgsDuChkhFRER6FldCmzFmjjGmyBiz0xjzaBvP32+M2WiMKTTGLDHGjA4cTzPGLDDGnDTGPNv1lbsvp28SA1LitDuCiIhID9Ploc0Y4wWeA64DRgN3NIWyVl6y1o6z1k4Efgj8NHC8BvhP4JGuqjfcGGPIz81gyc4yahsa3S5HREREuogbnbYrgJ3W2t3W2jrgFeDm1idYa0+0epgA2MDxU9baJTjhrccqyPVRXdfIh7uPuV2KiIiIdBE3QtsA4ECrx8WBY2cwxjxgjNmF02n7ans+wBhznzFmtTFmdWlpaVDFhqMpw9LpFeXREKmIiEgP4kZoM20cO2eJf2vtc9babOD/AY+15wOstc9baydZaydlZGR0sMzwFRfjZdrwdBYUlWCtdkcQERHpCdwIbcXAoFaPBwKHLnD+K8AtnVpRBMrP9bGvvJrdZafcLkVERES6gBuhbRUwwhgz1BgTA3wKeL31CcaYEa0eXg/s6ML6IkLT0h/zt2qIVEREpCfo8tBmrW0AHgTeBbYCr1prNxtjnjDG3BQ47UFjzGZjTCHwdeDuptcbY/bizCb9nDGmuI2Zpz3CgJQ4cvom6bo2ERGRHiLKjQ+11r4NvH3WsW+1uv/QBV47pPMqiyz5uT5+88FuTtTUkxwb7XY5IiIi0om0I0IEmzXKR4PfsmRHmduliIiISCdTaItglw1KoXdcNO/rujYREZFuT6EtgkV5PcwYmcGi7SX4/Vr6Q0REpDtTaItwBbk+yk7WseHgcbdLERERkU6k0BbhZozMwGPQLFIREZFuTqEtwqUmxJCXlcr8bUfdLkVEREQ6kUJbN5Cf62PTwROUnKhxuxQRERHpJO1ep80Y8/VLOO2UtfZXHahHOqAg18eP3i1iQVEJn5yc5XY5IiIi0gk60mn7dyARSLrA7d9CVaBcXG5mEv17x+q6NhERkW6sIzsi/NFa+8SFTjDGJHSwHukAYwz5uT7+tu4gtQ2N9Iryul2SiIiIhFi7O23W2v8IxTkSWgW5PqrrGlm555jbpYiIiEgn6PDeo+e5tu04sMZaW9jxkqQjpman0yvKw/xtJVwzIsPtckRERCTEgpk9Ogm4HxgQuN0HzAR+bYxRp62LxcV4mZqdxvxtJVir3RFERES6m2BCWxqQZ639N2vtv+GEuAxgOvC5ENQm7VSQ62NfeTW7y065XYqIiIiEWDChLQuoa/W4HhhsrT0N1AZVlXRIfq4PgAWaRSoiItLtBBPaXgJWGGO+bYz5NrAUeDkwc3RLSKqTdhmYGs/Ivola+kNERKQb6nBos9Z+F+c6tkqcCQj3W2ufsNaestZ+OlQFSvsU5PZl5Z5jVNXUu12KiIiIhFCw21itA14D/gqUGGO0HL/LCnJ9NPgtH+woc7sUERERCaEOhzZjzFeAo8B7wJvAW4Gv4qK8rBR6x0VriFRERKSb6fA6bcBDQI61tjxUxUjworwepo/MYGFRCX6/xeMxbpckIiIiIRDM8OgBnGvZJMzMyvVRdrKOjQf12yMiItJdBNNp2w0sNMa8RaslPqy1Pw26KgnKjJEZeAy8v62ECYNS3C5HREREQiCYTtt+nOvZYoCkVjdxWWpCDJdlpWq9NhERkW6kw502a+13QlmIhFZBro8fvVtEyYkafMmxbpcjIiIiQWp3p80Y81Tg6xvGmNfPvoW+ROmIgsDuCAuLSl2uREREREKhI522Pwa+/jiUhUho5WYm0a93LO9vO8rtkwe5XY6IiIgEqd2dNmvtmsDdidbaRa1vwMTQlicdZYwhP9fHkh1l1DY0ul2OiIiIBCmYiQh3t3Hsc0G8n4RYQY6PU3WNrNpT4XYpIiIiEqR2D48aY+4A7gSGnnUNWxKghXbDyNThacREeZi/rYSrR6S7XY6IiIgEoSPXtC0DDgPpwE9aHa8CNoSiKAmN+JgopmansaCohG/dONrtckRERCQIHbmmbZ+1dqG1dspZ17SttdY2dEaRYa3uFMy9Hta/Ao31bldzjoJcH3vKTrG79KTbpYiIiEgQgtkw/ipjzCpjzEljTJ0xptEYcyKUxUWEE4fgdAX87V/h55fBil86QS5M5Oc4S39oA3kREZHIFsxEhGeBO4AdQBzwBeCZUBQVUdJHwJeWwp2vQe9B8I//Bz8bCwt/ANXH3K6OQX3iGeFLZEGRQpuIiEgkCya0Ya3dCXittY3W2rlAfmjKijDGwMhr4d534N5/QtZVsPC/nPD2j2/A8WJXyysY5WPlnmNU1YTf8K2IiIhcmmBCW7UxJgYoNMb80BjzNSAhRHVFrqwr4Y6X4csrYPRNsPJ5eHoC/N+XobTIlZIKcnzUN1qW7Chz5fNFREQkeMGEts8EXv8gcAoYBNwaiqK6Bd8o+Pgv4auFMPmLsPlv8NwV8PKdcGBVl5Zy+eBUkmOjdF2biIhIBOtQaDPGeIHvW2trrLUnrLXfsdZ+PTBcKq2lDILrnoSHN8GMR2H/MvjtbGfG6Y73wNpOLyHK62H6yAwWFJXi93f+54mIiEjodSi0WWsbgYzA8KhcioQ0yP8GfG0zzHkSKvbC/94Gv7wGNv4ZGjt3tZRZo3yUnaxl06Hjnfo5IiIi0jmCGR7dCyw1xvynMebrTbcQ1dV9xSTAVV+Cr66DW34B/nr4y+fhmTxY+WuoP90pHztjpA9j4P2tGiIVERGJRMGEtkPAm4H3SGp1k0sRFQMT74QvLYdPvQyJfeHtR5wZp4t/DKcrQ/pxfRJiuGxQipb+EBERiVAd2cYKAGvtd0JZSI/l8UDuxyDnOti/HJb8DOZ/1/k66R646gFI7heSjyrI9fHjf26npKoGX1JsSN5TREREuka7O23GmMdDcY6cxRgYPBU+/Rrcv8QJccv/B54eD39/EMp2BP0RBbl9AVhYVBr0e4mIiEjX6kin7QsX2a7KAJ8CHu9QRQKZ4+DW30DBY7DsWVj3R1j3Ioy6Ea5+GAZc3qG3HdUviX69Y5m/tYTbJw0KcdEiIiLSmTpyTduvOfMatrNviYFzJFipQ+D6HzvLhUx/BPYsgl8XwO9vhF3z271ciDGGmTk+luwso67B3zk1i4iISKdod6dN17K5IDHD6bpNewjW/A6WPwd//Dj0mwDTHobRN4PHe0lvVZDr4+WV+1m19xjThqd3bt0iIiISMkHtPSpdrFcSTP0KPLQebnoW6k7Bn++BZyfB6rlQX3PRt5g2PI2YKI92RxAREYkwCm2RKKoX5H0GHlgJt/8RYlPgzYedSQtLfgY1519ANz4miinD0hTaREREIkyHt7EKbBAvbvJ4nU3pvzgf7n4D+o6BeY87a73Nexyqjrb5soJcH3vKTrGn7FSXlisiIiIdF8w2VjeHuBbpKGNg6HT4zN/gvkUwfDYsfRqeGgdvPAzlu844vSDXB6Bum4iISAQJZnh0qTHmWWPMNcaYvKZbyCqTjuk/ET4xFx5c7ey4UPiSc83ba5+DQ4UADOoTzwhfIgsU2kRERCJGh3dEAKYGvj7R6pgFCoJ4TwmVtGy48SmY+Q348Bew6rew+W+QXQDTHqYgJ4MXlu3lZG0Dib2C+WMgIiIiXcHYdq71FWkmTZpkV69e7XYZ7qs5DqtfgBW/gJNHOZk2nn87XMDH77iPOeMGuF2diIiIAMaYNdbaSW091+HhUWNMb2PMT40xqwO3nxhjene8TOlUsb3h6q/BQxvghqdI8Ffxq5inyHtjDqz9AzTUul2hiIiIXEAw17S9AFQBtwduJ4C5oShKOlF0LEy6B/OVNfy237cor4uC178CT0+AZc9AbZXbFYqIiIQPfyMcWAUL/hvW/8nVUoK5mCnbWntrq8ffMcYUBluQdBGPl9TJn+S6V3OYf4tlWNHz8M/HYPGPYPIX4cr7nZ0YREREepqqI7Dzfdg5z9k2sqYSMDDpHpjwSdfKCia0nTbGXG2tXQJgjJkGnA5NWdIVZozMwBjD6ydH8vDdb8DBNbDkKfjgJ7D8WbjsMzD1QWcPVBERke6qoQ4OrHBC2s734egm53hiX8i93pnEl10A8X1cLbPDExGMMROAPwBN17FVAHdbazeEqLaQ0ESEC/v4/yzF77f8/cGrWw6W7YBlP4fCl8H6Yey/wPhPQp9hkJIF3mj3ChYREQmFY3tg1/tOSNuzGOpOgicKsqbA8FnOmqd9xzproXahC01E6FCnzRjjAXKstROMMckA1toTQdQoLinI8fGT97ZTWlVLRlIv52D6CLjpGWe5kBX/4+xruvE15znjhZRBkDrUCXF9Al9ThzoduZh4174XERGR86qrhr1LAt20eXAssPB8SpbTmBg+G4Ze4+zzHaaC6bQtttZOD3E9IadO24VtPnSc63++hB/eNp7bJw1q+6SaE3B0M1TsgWO7A7fA/ZrKM89N6tcS4voMPTPUxaV0/jckIiICYC2UbmsZ8ty3DBprISoOhlzthLThs511Tbu4m3YhIe+0BbxnjHkE+BPQvImltfZYEO8pXWx0v2Qyk2NZsK3k/KEtNhkGT3FuZ6s+FghzgVtTsNs5D04eOfPcuD5nhrjWnbqEjLD6SyMinczvdzodB9fAwbVweL3zsyZznDMklTne+fng8bpdqUSS05WwZ1FLUDtx0DmekQtXfNEZ9sya6qykEIGCCW33Br4+0OqYBYYF8Z7SxYwx5Odm8Mb6w9Q1+ImJaucqMPF9nNuAy899ru4UVOw9szNXsQcOfAib/uJcL9ckOqGlM3d2oEseoB/cIpHMWucfz4Nr4dBaJ6gdKoTawFU10fFOSDte7PxDaxtbjvcdEwhx45ybbzT0SnTve5Hw4vfD4ULnz82u9+HASufPT69kGDYTZvwHZM9yLuvpBoK5pu0ua+3SENcjLsjP8fHyygOs3nuMqcPTQ/fGMQmBH7hjzn2uoQ4q97cacg18LS2C7e9CY13Lud4YSBncdpcuZTBExYSuZhEJXvWxQDhbFwhoa+HkUec5T5TzM2HcbdA/DwbkQXoOeAP/HNXXOENaRzfBkY1wZBNs+iusaVoG1Dh//zPHQeZY6BsIc8n91a3vKU6WOstwNC3HUV3mHO830VlEfvhsGDipW06a61Bos9b6jTE/BtoYL5NIM214OjFRHt7fVhLa0HYhUTGQPty5nc3fCCcOtXTmmkPdHueahLqTLecaDyQPPPf6uaZQF5PQNd+PSE9VdwoOb2gJZwfXOn9vm6SPhGH5TjgbcLnTNbvQ0FR0LPSf6NyaWAvHD7SEuCMbnO7Klv9rOScuNTC0Oq4l0KXn6D913UFjPRSvaplAcHi9czw+3RnuzJ7lLMfRA9YWDWYiwneADcBfbRhvYKqJCJfmsy+spPhYNfMfmel2KRdmLZwqPXO4tXWn7vRZl1Qm9m17pmufocuHBPwAACAASURBVK6vtyPtYC3Un3YCe91JqD3ra91J8PaC1MFO9zWxL3iC2fBF2tRYDyVbWq5DO7TOedx0qUPyQBhwmRPO+uc5wSu2E3c3bJokdWQjHN3ofC3ZCg01zvOeaOdapqYQ13S9nP7uh7/KA4HlOObB7kXOULrxwqArWpbjyJzQLf+eX2giQjChrQpIABqAGsAA1lqbfAmvnQM8DXiB31hrnzzr+ftxrpVrBE4C91lrtwSe+wbw+cBzX7XWvnuhz1JouzS/W7qHx9/YwsJHZjIkPYK7U6crW02M2H3mJImqQ2eeG9u7JcSlDHKuq4vqBdFxEBUb+NrLmWkUHdvqa2yr5wP3vcFcHtoNWQv11U4XpraqVcA6BXVVrcLWpTwfuN90ndOlaB3gUgc7y9G0vt+ZQaK7aJ4o0Oo6tCMbWwJRXGpgePNyp4vWPw+S+rpbM0Bjg1P3kY0tt6ObWoZnwQmXrUNc5jjn50A3DAARo74G9i1t2YWgrMg5njywJaQNnd4jViHolNAWRDFeYDvwEaAYWAXc0RTKAuckN637Zoy5CfiytXaOMWY08DJwBdAfmAeMtPb8P80V2i7N/vJqpv9oAd+6YTT3Xj3U7XI6R101VO47d2LEsd3OcGzr6+jayxPVRqgLBL2LBcEOB8Xo0F3D0xSymjtYVYEAdbIlVDUHqKpWAaut55tClv/inwvO9xKTADGJzvpIMYnO416JEJMU+HoJz9fXOL+/FXtbvlbsc+7XHD/zM2NT2g5zKUOcAB/VKzS/rpHCWufvQFM4O7g2MFEg8OsWHe9cLzQgD/oHOmmpQyLrGrKTJWeGuCMbnYXEm/75iElse9KD1p7sHNZC+c6WWZ57l0DDaec/XIOntizHkZETWX/OQiCkS34YY+6y1r4YuD+t9WQEY8yD1tpnL/IWVwA7rbW7A695BbgZaA5tZy3Um4AzK5XAea9Ya2uBPcaYnYH3W97e70POlJUWz3BfIguKSrpvaIuJB98o59YWf6PTRaivcX54NNQ6Q3INNYGvtc7xjjx/suTc5+trnDWDOsp4Lhzqolsd8/ZyPqt1qDp7iJFL/A9cVNy5oSk+zQk/Tcean28KWucLXYmhvVg4Y2Tbx09XtAS41mHu6GYoeueswG6ci9rb6tKlDHbWIoz0jkz1MWdos3UX7eyJAmP/paWL1nqiQKRK9AU6NrNajtWfdoZTW0962PgarP6t87zxQJ/sVsOr451Ql5TZ44JESNRWOTsPNF2bVrnfOZ42HC6/2wlpg6cpKF9AR/4Wfh14MXD/GSCv1XP3AhcLbQOAA60eFwNXnn2SMeaBwGfFAAWtXrvirNcOaOO19wH3AWRlZV2kHGlSkOtj7tI9nKxtILFXhP+A7giPNxAuunB42O93wlRTuGsdGutrWh272PNtBMXqYy3PN9Q6s3CbQlV8emAHi7ND1dmh66xOV0xiZP7jHZcaGM6beO5zfr+zpmBF6y5d4P6exbD+Fc4ItN5eTjcuJRDomodhA/fjUrvkW7pkddXOhdtNkwQOrjlzokDaiJaJAv3znIASoWtYtVt0XGCCRKt/xqx1/gwc2dTSlTu4Gjb/teWc+PTAzNVAkMsc60y46IazFYNirfNr2DTLc/9y8Dc4P0eGzoBpDzmTCPp000ZBJ+jIT19znvttPb7Y65uc8198a+1zwHPGmDuBx4C72/Ha54HnwRkevYSaBGfpj+cX72bJjjLmjM10u5yeweMBT5zzj4e4w+NxOmvJ/dteQLqh1rkounLvucHu4JpzdwXp1TvQoWsd5oY491OyOjcQNU8UWNsS0kq2tgwBNk0UyPtsYLJAJ08UiETGtPyejbqh5fjpSqczezQwe/XIJlj565ZuuTcmMOlh/JnXy/WAa7DOUH0ssBxHYN20pg5u33Ew5UGnmzboSs3q7aCOhDZ7nvttPW5LMdB6lbuBwKHznAvwCvCLDr5W2mHSkFSSYqNYsK1EoU2kSVSv8y9PA871cm116UqLYMd7LRfuN0nqd/4uXVK/S19I2u93rsdsfR3akQ3nThTI+Vh4TRSIVHEpMGSac2vS2ADlO86c9LD9H1D4Yss5vbNaQlxKYOTH+p0ulPW3ulmgjWPtPsff6jMudI5t+/2DOaf2hBNmsc6fv+yCluU4kvt1xe9St9fuiQjGmGpgJ07XKztwn8DjYdbaC44tGWOicCYizAIO4kxEuNNau7nVOSOstTsC928Evm2tnWSMGQO8RMtEhPeBEZqIEDoPvLSWVXuOseIbs/B4dM2GSFD8fqfT0DrMtb5/4iBn/F/XE+38w97WzNe4VKfT07we2rqzJgpMCHTPInSiQHdhrfN73rSeXNP1cuU7L31yziUzznV3Z9zMmfc77Zw2XuONdrpow2c7fw61k02HhHrv0fNcxX1prLUNxpgHgXdxlvx4wVq72RjzBLDaWvs68KAxZjZQD1TgDI0SOO9VnEkLDcADFwps0n4FOT7e2nCYzYdOMG6ghk1EguLxOB2G5H6QddW5zzfUOYvGnj1BomKvM3vz7HUH4ayJAoElN7rDRIHuwhhnokJSJoyY3XK8rhpOlbQEnPOGIHMJ5zQFKelpunzJj66mTlv7lJ+sZdL35/HwrJE8NHuE2+WI9Gw1J1o6c9XlzhIUPWmigEgPFOpOm3RjaYm9mDAwhflFJQptIm6LTW5ZM0xEerwIX2xIOsOsXB8biisprQpiDTEREREJqZCENmNMqjFmfCjeS9yXn+vDWlhYVOJ2KSIiIhLQ4dBmjFlojEk2xvQB1gNzjTE/DV1p4pYx/ZPpm9yLBQptIiIiYSOYTlvvwHZT/wLMtdZeDsy+yGskAhhjyM/x8cH2MuobQz1FXURERDoimNAWZYzpB9wOvBmieiRMFOT6qKptYNXeNpYcEBERkS4XTGh7AmettV3W2lXGmGHAjtCUJW6bNjydGK+H+Vs1RCoiIhIOOhzarLWvWWvHW2u/FHi821p7a+hKEzcl9IriymF9mK/r2kRERMJCMBMRhhlj3jDGlBpjSowxfzfGDA1lceKuglwfu0tPsa/8lNuliIiI9HjBDI++BLwK9MPZB/Q1nM3dpZsoyPUBMH+bum0iIiJuCya0GWvtH621DYHbi5yx87FEusFpCWRnJCi0iYiIhIF2hzZjTJ/A2mwLjDGPGmOGGGMGG2P+A3gr9CWKmwpyfXy4+xinahvcLkVERKRH68jeo2twOmom8PhfWz1nge8GW5SEj/xcH7/+YA9Ldpbx0TGZbpcjIiLSY7U7tFlrzzvZwBgTHVw5Em4mD+lDUq8oXl11gBkjM4iN9rpdkoiISI8U9N6jxlFgjPkNUByCmiSMRHs9fHH6MN7fVsLNzy5l86HjbpckIiLSIwWz5MeVxpingX3A68AHQG6oCpPw8dVZI5j7uclUVNdxy3NLeW7BThq0vZWIiEiX6shEhO8bY3YA/wVsBC4DSq21v7fWVoS6QAkP+bk+3n14Oh8dk8mP3i3iE79azp4yrd8mIiLSVTrSabsPOAr8AnjRWluOlvroEVITYnj2zjx+fsdl7C49xXVPL+YPy/fi9+u3X0REpLN1JLRlAt8HbgJ2GmP+CMQZYzoyE1Ui0E0T+vPPr03nyqFpfOvvm7l77koOHz/tdlkiIiLdWrtDm7W20Vr7jrX2s8Bw4O/AMuCgMealUBco4alvciy/u2cy3//4WNbsq+Dany3mb+uKsVZdNxERkc4Q1OxRa22NtfbPgY3iRwDvhqYsiQTGGD595WDeeegacvom8bU/refL/7uW8pO1bpcmIiLS7QS95EcTa+0Ja+3vQ/V+EjkGpyXwp3+dwqPX5fL+1hI++tRi3tty1O2yREREupWQhTbp2bwew/0zsnn9K9PISIrli39YzX/8eT1VNfVulyYiItItKLRJSOVmJvP3B6bxQH42f15TzJynPmD5rnK3yxIREYl4QYU2Y8xUY8ydxpjPNt1CVZhErpgoD//+0Vxeu38qMVEe7vj1Cr775hZq6hvdLk1ERCRiBbMjwh+BHwNXA5MDt0khqku6gcsHp/LWV6/ms1MG89sle7jhmSVsKK50uywREZGIZDq6RIMxZisw2ob5Gg+TJk2yq1evdruMHu+DHaX8+2sbKD1Zy1cKhvNA/nCivRqdFxERac0Ys8Za22YTLJh/NTfhLLQrclHXjMjg3Yenc9OE/jw1bwe3/mIZO0uq3C5LREQkYgQT2tKBLcaYd40xrzfdQlWYdD+946P52Scn8otP53HgWDXX/3wJv12yR9tgiYiIXIJgtp56PFRFSM9y3bh+XD4klW/+dSPffXML7205wo8/MYGBqfFulyYiIhK2OnxNW6TQNW3hy1rLa6uLeeLNLQB868bRfOLygRhjXK5MRETEHZ1yTZsx5ipjzCpjzEljTJ0xptEYc6LjZUpPY4zh9smDeOehaxjTP5n/+PMGvviHNZRWaRssERGRswVzTduzwB3ADiAO+ELgmEi7DOoTz8tfvIrHrh/F4h2lfPSpxfxj02G3yxIREQkrwW4YvxPwWmsbrbVzgZkhqUp6HI/H8IVrhvHWV65mQEoc97+4lq//qZDjp7UNloiICAQX2qqNMTFAoTHmh8aYrwEJIapLeqgRfZP465en8tCsEfx9/SHmPLWYJTvK3C5LRETEdcGEts8EXv8gcAoYBNwaiqKkZ4v2evjaR0by1y9NJT7Gy12//ZBv/30Tp+u0DZaIiPRcQc0eNcbEAVnW2qLQlRRamj0a2WrqG/nhP4p4YekehqUn8JPbJ3BZVqrbZYmIiHSKzpo9eiNQCPwj8HiiFteVUIuN9vKtG0fz0hevpLbBz62/WMaP3y2irsHvdmkiIiJdKpjh0ceBK4BKAGttITAk+JJEzjU1O513Hr6GW/MG8uyCndzy3FKKjmgbLBER6TmCCW0N1trjIatE5CKSY6P50Scm8PxnLqekqoYbn1nCrxbtolHbYImISA8Q1Ibxxpg7Aa8xZoQx5hlgWYjqEjmva8dk8u7D08nPzeC/39nGp55fzv7yarfLEhER6VTBhLavAGOAWuBl4ATwcCiKErmYtMRe/PKuy/nJJyaw7XAVc55ezEsf7qe7b8smIiI9l/YelYh3sPI0//Hn9SzdWc7MnAx+eOt4fMmxbpclIiLSbheaPdru0HaxGaLW2pva9YadTKGtZ/D7LX9Yvpcn/7GN2Ggv37tlLDeM7+92WSIiIu1yodAW1YH3mwIcwBkS/RAwQdQmEhIej+Fz04ZyzcgMvv7qeh58aR3vbj7Kd28eQ0p8jNvliYiIBK0j17RlAt8ExgJPAx8Byqy1i6y1i0JZnEh7ZWck8pf7p/BvHxnJOxsPc+3PFrOwqMTtskRERILW7tAW2Bz+H9bau4GrgJ3AQmPMV0JenUgHRHk9fGXWCP7vgWmkxEfzubmr+ObfNnKqtsHt0kRERDqsQ7NHjTG9jDH/ArwIPAD8HPhrKAsTCdbYAb15/cGruW/6MF5euZ/rnv6A1XuPuV2WiIhIh7Q7tBljfo+zHlse8B1r7WRr7XettQdDXp1IkGKjvXzzY6N45YtXYbF84lfL+e93tlLboM3nRUQksnRk9qgfOBV42PrFBrDW2uQQ1RYSmj0qTU7WNvD9t7bw8soD5GYm8ZPbJzCmf2+3yxIREWkW0g3jrbUea21S4Jbc6pYUboFNpLXEXlH897+M54XPTaL8VB23PLeU5xbspKFRm8+LiEj4C2ZHBJGIVJDbl38+PJ1rR2fyo3eLuP1Xy9lTduriLxQREXGRQpv0SKkJMTx752U8/amJ7Cw5yUefWsy/v7aeLYdOuF2aiIhImzqyuK5It2CM4eaJA7hyaBrPzN/BX9ce5LU1xVw5tA/3TBvKR0b3xevR2tEiIhIetPeoSMDx6nr+tHo/v1+2j4OVpxmYGsfdU4Zw++RB9I6Ldrs8ERHpAUK692ikUWiT9mpo9DNv61FeWLqXlXuOERft5dbLB/C5qUMZ7kt0uzwREenGFNoU2qSDNh08zu+W7eX1wkPUNfqZPjKDe6YNYcaIDDwaOhURkRBTaFNokyCVnazlpQ/388cV+yitqmVYRgKfmzqEW/MGktBLl4aKiEhoKLQptEmI1DX4eXvjYeYu3cP64uMkxUbxyUmDuHvqEAb1iXe7PBERiXAKbQptEmLWWtbur2Tu0j28s+kIfmuZPaov90wbwpRhaRijoVMREWm/C4U2jeuIdIAxhssHp3L54FQOHz/Niyv28dKH+3lvy1FyM5O4Z9oQbp44gNhor9uliohIN6FOm0iI1NQ38vfCg8xdupdtR6pIjY/mziuz+MxVQ8jsHet2eSIiEgE0PKrQJl3IWsvy3eXMXbqXeVuP4jWG68b1455pQ8jLSnW7PBERCWMaHhXpQsYYpmanMzU7nf3l1fx++V5eXXWAN9YfYsKgFO6dNoTrxvYjJkq7yImIyKVz5V8NY8wcY0yRMWanMebRNp7/ujFmizFmgzHmfWPM4FbP/cAYsylw+2TXVi7SPllp8fznDaNZ8c1ZPHHzGKpO1/PQK4Vc/YP5PPP+DspO1rpdooiIRIguHx41xniB7cBHgGJgFXCHtXZLq3PygQ+ttdXGmC8BM621nzTGXA88DFwH9AIWAQXW2vPu8q3hUQknfr9l0Y5S5i7dy+LtpcREebhpQn/umTaEMf17u12eiIi4LNyGR68AdlprdwMYY14BbgaaQ5u1dkGr81cAdwXujwYWWWsbgAZjzHpgDvBqVxQuEiyPx5Cf4yM/x8fOkip+t2wvf1lzkD+vKeaKoX24d9oQPjI6UxvVi4jIOdwYHh0AHGj1uDhw7Hw+D7wTuL8euM4YE2+MSQfygUFnv8AYc58xZrUxZnVpaWmIyhYJreG+JL53yzhWfGMW3/xYLgcrTnP/i2uZ/sMFPL94F8er690uUUREwogbnba2WghtjtEaY+4CJgEzAKy1/zTGTAaWAaXAcqDhnDez9nngeXCGR0NTtkjn6B0fzX3Ts7l32tDmjer/6+1t/Oy9HdqoXkREmrkR2oo5szs2EDh09knGmNnA/wfMsNY2X61trf0+8P3AOS8BOzq1WpEuEuX1MGdsP+aM7cfmQ8f53dK9vLq6mBdX7OeaEencO20oM0Zqo3oRkZ7KjYkIUTgTEWYBB3EmItxprd3c6pzLgD8Dc6y1O1od9wIp1tpyY8x44CVgYuAatzZpIoJEsrKTtbwc2Ki+pKqWYekJ3D11CLdePpBEbVQvItLthN3iusaYjwFPAV7gBWvt940xTwCrrbWvG2PmAeOAw4GX7LfW3mSMiQXWBo6dAO631hZe6LMU2qQ7qGvw886mw7ywdC/rD1SS1CuK2ycP4u4pQ8hK00b1IiLdRdiFtq6k0Cbdzdr9Fcxdupd3Nh6mURvVi4h0K+G25IeIBCEvK5W8rFSOfGyUs1H9ypaN6j83dQi3XKaN6kVEuiN12kQiXE19I68XHuKFpXuaN6q/44osPjNlMP16x7ldnoiItIOGRxXapAew1rJi9zHmLt3De1uP4jGG68Zmcs+0oeRlpWjoVEQkAmh4VKQHMMYwJTuNKdlpHDhWze+X7eVPqw/w5obDTBjYm09MGsR1YzNJS+zldqkiItIB6rSJdGOnahv469pifr98HztLTuL1GKYMS+OG8f346JhMUhNi3C5RRERa0fCoQpv0cNZath2p4s0Nh3hzw2H2lVcT5TFMG57O9eP78dHRmfSOj3a7TBGRHk+hTaFNpJm1ls2HTvDmhsO8ueEQxRWnifYarhmRwQ3j+zF7dF+SYxXgRETcoNCm0CbSJmstG4qP8+aGQ7y14TCHjtcQ4/UwI8cJcLNG9dXOCyIiXUihTaFN5KL8fsu6A5W8teEwb288zJETNfSK8pCf4+OGCf0oyPURH6MAJyLSmRTaFNpE2sXvt6zZX8Gb6w/x9qYjlFbVEhftpWCUjxvG9SM/16cFfEVEOoFCm0KbSIc1+i0r9xzjrY2HeGfjEcpP1REf42X2qL7cML4f00dmKMCJiISIQptCm0hINDT6+XDPMd7ccIh/bDpCRXU9Sb2i+Mjovlw/vh/XjMggJsrjdpkiIhFLoU2hTSTk6hv9LNtVzluBAHeipoHk2CiuHZPJDeP7MW14OtFeBTgRkfZQaFNoE+lUdQ1+lu4s440Nh3hv81GqahtIiY9mzphMrh/fjynD0ohSgBMRuSiFNoU2kS5T29DI4u1lvLXhEO9tOcqpukb6JMQwZ6zTgbtyaBpej/ZBFRFpi0KbQpuIK2rqG1lYVMqbGw7x/tYSTtc3kp7Yi4+Ny+T6cf2YPKQPHgU4EZFmCm0KbSKuO13XyPxtJby54RDzt5VQ2+Cnb3Ivrhvbjxsn9OOyQakKcCLS4ym0KbSJhJVTtQ3M23qUtzYcZuH2Uuoa/PTvHcvHxvXjhgn9mTCwN8YowIlIz6PQptAmEraqauqZt/Uob64/zOIdpdQ3WgakxHHD+H7cML4/YwckK8CJSI+h0KbQJhIRjp+u55+bj/DWxsMs2VFGg98yOC2e68c5AW5UvyQFOBHp1hTaFNpEIk7FqTr+ueUIb244zLJd5TT6LcPSE7hhfD+uH9+fnMwkt0sUEQk5hTaFNpGIVn6yln9sPsJbGw6zYnc5fgsjfIlcHxhCHe5LdLtEEZGQUGhTaBPpNkqqanh30xHe2HCYVXuPYS3kZiYxZ2wmM3N8jB/QW7NQRSRiKbQptIl0S0dP1PD2xsO8ueEwa/dXYC30SYjhmhHpzMzJYPqIDNISe7ldpojIJVNoU2gT6fbKT9ayZGcZC4tKWby9lPJTdRgD4wb0ZubIDGbk+Jg4KEW7MYhIWFNoU2gT6VH8fsumQ8dZWFTKou2lrNtfgd9C77horhmRzoyRGczIycCXFOt2qSIiZ1BoU2gT6dEqq+uau3CLtpdSWlULwOh+yczMyWBmjo/LslKI1qb2IuIyhTaFNhEJsNay5fCJ5gC3Zl8FjX5LUq8orm7VhevXO87tUkWkB1JoU2gTkfM4UVPP0h1lLNpeysKiUo6cqAEgp28SM3OcADdpcB9iotSFE5HOp9Cm0CYil8Bay/ajJ1lYVMKi7aWs2nuM+kZLQoyXqcOdGakzRmYwMDXe7VJFpJtSaFNoE5EOOFnbwLKdLV24g5WnARjuS2TGyAxm5mQweUgfYqO9LlcqIt2FQptCm4gEyVrLrtJTzV24D3cfo67RT1y0lynZac1duMFpCW6XKiIR7EKhLaqrixERiUTGGIb7EhnuS+QL1wyjuq6BFbvLWVRUysLtpczfVgLA0PSE5skMU4alqQsnIiGjTpuISAjsKTvFoqISFm4vZfmucmob/PSK8nDlsDRmBoZSh6YnYIwW9xWR89PwqEKbiHShmvpGPtxzrHkodXfpKQAG9Ylj5kgfM0ZmMHV4GvExGuwQkTMptCm0iYiLDhyrZuH2UhYVlbB0Zzmn6xuJ8XqYPDSVmSN9zMzJYLgvUV04EVFoU2gTkXBR29DI6r0VzV247UdPAjAgJY7pI53JDNOGp5EUG+1ypSLiBoU2hTYRCVMHK0+zqKiURdudLtzJ2gaiPIZJQ1KZEejC5WYmqQsn0kMotCm0iUgEqGvws3Z/RfMWW1sPnwCgb3Ivpo/IYNrwdKZkp9E3WRvdi3RXCm0KbSISgY6eqAl04Ur5YEcpJ2oaAMjOSGBKdhpTs9O5algafRJiXK5UREJFoU2hTUQiXKPfsuXQCZbvLmPZrnJW7TnGqbpGAHIzk5pD3BVD+9A7TtfDiUQqhTaFNhHpZuob/WwoPs6K3eUs21XG6r0V1Db48RgYO6A3U7LTmDIsjclD+pDQS0uLiEQKhTaFNhHp5mobGlm3v5Jlu8pZsaucdQcqqG+0RHkMEwalMDUQ4vIGp2qXBpEwptCm0CYiPUx1XQNr9lWwbFc5y3eVs6G4Er+FmCgPeVkpTM12JjVMGJhCTJTH7XJFJEChTaFNRHq4qpp6Vu09xrKd5SzfXc6WwyewFuKivUwe2ocpw9KYmp3GmP7JRHkV4kTcotCm0CYicoaKU3V8uMfpwi3bVc6OEmeR36ReUVw5rA9XDXMmNuRmJuHxaI04ka5yodCmq1NFRHqg1IQY5oztx5yx/QAoqaphxe5jLN9VzvJdZczbWuKcFx8dCHBpTMlOIztD222JuEWdNhEROcehytNOgNvtdOMOVp4GICOpV/NQ6pTsNLL6xCvEiYSQhkcV2kREOsxay/5j1c1Dqct3l1NaVQs4e6a27sT1T4lzuVqRyKbQptAmIhIy1lp2lZ48I8RVVtcDMCQtnimBmalThqWRkdTL5WpFIotCm0KbiEin8fst245UsWxXGSt2l/Ph7mNU1Tpbbo3wJTZ34a4alkZKvLbcErkQhTaFNhGRLtPQ6GfzoRPNXbhVe45xur4RY2BUZjJTs9OYOtzZrSEpVltuibSm0KbQJiLimroGPxuKnd0alu0qY+3+Suoa/Hg9hnGBLbemZqeRl5WqLbekx1NoU2gTEQkbNfWNrN1XwfLdzjVx6w9U0uC3eAzkZCaTl5VCXlYqeYNTGZKm2anSsyi0KbSJiIStU7UNrNp7jLX7Kli7v5LCA5WcDFwT1ychhssGpZA3OJXLslKYMDBF3Tjp1rS4roiIhK2EXlHMzPExM8cHQKPfsqOkirX7Klm7v4K1+yt4f5uz2K/HQG5mMnmDA924rFQGqxsnPYQ6bSIiEvYqq+tYt78lxBXur+RUXSPgdOPyslK4LEvdOIl86rSJiEhES4mPIT/XR35uSzdu+9Eq1u6vaA5zTVtveT2GnL5J6sZJt6NOm4iIdAsVp+ooPNB2Ny4tIYbLAt24vKxUJgzqTXyM+hYSftRpExGRbi814fzduLX7Kll3VjcuNzMpMEvV6chpH1UJd+q0iYhIj1Fxqo51ByqaJzmsP6BunIQXpd1Y0wAAIABJREFUddpERERwunEFuX0pyO0LqBsnkUWdNhERkVYu1I1LT4xh4qCWEDd+oLpxElrqtImIiFyitrpxRUeqmic4rPv/2bvz+Kjqe//jr08m+8KWBZB9DSAqYkREhUTRajfbq1ar3ra21rb3tte21/Z62/6qte2trbeLVu+t1urVW692sYu1brUFV1BAESUQocgSQQj7GrJ9fn+cSZjEADNkJieTvJ+PxzxmzpkzM58hYt581/U7eXrFZiBojZs8tKhtlurJIweoNU5SJpSWNjM7H7gViAB3u/vNHZ7/MnA10ATUAZ9093XR534AvA/IAP4CXOtH+BJqaRMRkWTbvq+BV6Mh7pV1O3mtdif7D9Mad9LwAeRlR0KuWNJFj2ppM7MIcAdwLlALLDKzR9y9OuayV4EKd99vZp8DfgBcamazgDOAE6PXPQ/MAeZ3V/0iIiKDCrI5Z/JgzpkctMY1NbdQs3kPr6wPxsV11ho3bcQAjj+uP5OH9qN8cJGCnCQsjO7RGcBqd18DYGYPARcCbaHN3efFXL8QuLL1KSAXyAYMyAI2d0PNIiIih5UZyeD44/pz/HH9+ceZo4B3t8b94dWN/HLheiDYjmtsaSFThvZjynH92u5LCnPC/BrSw4UR2oYBG2KOa4HTjnD9p4DHAdx9gZnNAzYRhLbb3X1FxxeY2TXANQAjR45MUtkiIiLx69ga19Li1O44QPWmXVRv3E31pt0sXrudR17b2PaasqKcdiFuytB+jCouIJKhMXISTmjr7L+8TsekmdmVQAVBFyhmNh6YDAyPXvIXM5vt7s+2ezP3u4C7IBjTlqS6RUREjllGhjGyOJ+RxfmcP3Vo2/kd+xpYsSkIcdWbdlO9cTfPr9pKU0vw6ys/O8KkIUVMjglyk4b0U/dqHxRGaKsFRsQcDwc2drzIzOYCXwfmuPvB6OkPAwvdfW/0mseBmcCzHV8vIiKSDgYWZDNrfAmzxpe0nTvY1MyqzXvbQlz1pt08snQjD7x0qHt1TEkBU47r365VrrRI3au9WRihbREwwczGAG8DlwGXx15gZicDdwLnu/uWmKfWA582s+8RtNjNAX7SLVWLiIh0k5zMCFOH9WfqsP5t59yD7tXlG3e3tcy9sm4Hf4rpXi0tymkLcZOHBkFuTIm6V3uLbg9t7t5kZp8HniRY8uMed19uZjcBi939EeAWoBD4TXStm/Xu/kHgt8DZwOsEXapPuPufuvs7iIiIdDczY8SgfEYMyuf8qUPazu/a39iua7V6025eeHZNW/dqblYGk4a0n/AwaUiRFgVOQ9oRQUREpJc52NTM6i17WbFpTzTIBZMfdtc3AWCt3asdZq+WFeWGXLn0qHXaREREJLVyMiNtS5BwSnDO3Xl754G21rjqjbtZumEnjy7b1Pa6ksLsdhMejj+uH2NKCtW92kMotImIiPQBZsbwgfkMH5jPecfHdK8eaGRlh+7Ve55/i8bmQ92r5UPaL0MyaUgRBTmKEN1N3aMiIiLSTkNTC3+v29uuVa560252HWgEot2rxQXtWuWC7tUc7bvaReoeFRERkbhlZ2YweWgwA/Wi6Dl3Z9Ou+nZB7vW3d/Hn1w91r/bPy2Li4EImDC6ifHAREwYXMnFwkXZ6SBKFNhERETkqM+O4AXkcNyCPuVMGt53fXd/Iyk17qN64ize37GXV5j38edkm/u/A+rZrBhVkMzEa4CYMLmJiWfB4YEF2GF8lbSm0iYiIyDHrl5vFjDGDmDFmUNs5d2fLnoO8uXkPb24OglzN5j387pW32Xuwqe260qKcoGWurIiJg4soHxK00vXLzQrjq/R4Cm0iIiKSVGbG4H65DO6Xy1kTStvOt3axBmHuUKD79eIN7G9obrtuSL9cJgwupHxwUbR1LghzhX188kPf/vYiIiLSbWK7WCvLy9rOt7QEy5F0bJn734XrONjU0nbdsAF57bpZywcXMb6ssM/sw6rQJiIiIqHKyDi028M5kw+Nl2tucTZs39+uZe7NzXt4YfU2GpqDMGcGIwbmt4W51pa5caWF5Gb1rjCn0CYiIiI9UiTDGF1SwOiSgnZryzU1t7B2235WtQa5LXt48509zK+pa9u+K8NgdHFB2wzW1pa5MSUFZGdmhPWVukShTURERNJKZiSD8WWFjC8r5IITDp1vaGph7bZ9QavcO4cC3V+qNxPNcmRGg2DrBIjyIUVMHFzIqOICsiI9O8wptImIiEivkJ2Z0dZFyomHztc3NrOmbh+rthzqZl2+cTePv/EOrXsMZEWMsSWFTBwSLEkyYfChMNdTtvFSaBMREZFeLTcrEuzccFy/ducPNDTz97pgnFzN5j2s2ryXV9fv4E+vbWy7Jiczg3GlhUwcXEhleRkfOnlYd5ffRqFNRERE+qS87AhTh/Vn6rD+7c7vO9jEqi1BmGsdN/fSW9spyMlUaBMRERHpKQpyMpk2YgDTRgxod765Jdz92nv2iDsRERGRHiLssW0KbSIiIiJpQKFNREREJA0otImIiIikAYU2ERERkTSg0CYiIiKSBhTaRERERNKAQpuIiIhIGlBoExEREUkDCm0iIiIiaUChTURERCQNKLSJiIiIpAGFNhEREZE0oNAmIiIikgYU2kRERETSgLl72DWklJnVAeu64aNKgK3d8DmSGvr5pT/9DNOffobpTz/Drhvl7qWdPdHrQ1t3MbPF7l4Rdh1ybPTzS3/6GaY//QzTn36GqaXuUREREZE0oNAmIiIikgYU2pLnrrALkC7Rzy/96WeY/vQzTH/6GaaQxrSJiIiIpAG1tImIiIikAYU2ERERkTSg0NZFZna+mdWY2Wozuz7seiQxZjbCzOaZ2QozW25m14ZdkyTOzCJm9qqZPRp2LZI4MxtgZr81s5XRv4unh12TJMbMvhT9f+gbZvagmeWGXVNvpNDWBWYWAe4ALgCmAB81synhViUJagL+1d0nAzOBf9bPMC1dC6wIuwg5ZrcCT7j7JOAk9LNMK2Y2DPgXoMLdpwIR4LJwq+qdFNq6Zgaw2t3XuHsD8BBwYcg1SQLcfZO7vxJ9vIfgl8WwcKuSRJjZcOB9wN1h1yKJM7N+wGzgFwDu3uDuO8OtSo5BJpBnZplAPrAx5Hp6JYW2rhkGbIg5rkW/8NOWmY0GTgZeCrcSSdBPgK8CLWEXIsdkLFAH3Bvt4r7bzArCLkri5+5vA/8JrAc2Abvc/alwq+qdFNq6xjo5pzVU0pCZFQIPA190991h1yPxMbP3A1vcfUnYtcgxywSmA//t7icD+wCND04jZjaQoJdpDHAcUGBmV4ZbVe+k0NY1tcCImOPhqEk47ZhZFkFge8Ddfxd2PZKQM4APmtlaguEJZ5vZL8MtSRJUC9S6e2sL928JQpykj7nAW+5e5+6NwO+AWSHX1CsptHXNImCCmY0xs2yCgZePhFyTJMDMjGAszQp3/1HY9Uhi3P3f3X24u48m+Pv3N3fXv/DTiLu/A2wws/LoqXOA6hBLksStB2aaWX70/6nnoMkkKZEZdgHpzN2bzOzzwJMEs2XucfflIZcliTkD+EfgdTNbGj33NXd/LMSaRPqaLwAPRP/xuwa4KuR6JAHu/pKZ/RZ4hWBG/qtoO6uU0DZWIiIiImlA3aMiIiIiaUChTURERCQNKLSJiIiIpAGFNhEREZE0oNAmIiIikgYU2kSkTzOzZjNbGnNL2mr8ZjbazN5I1vuJSN+mddpEpK874O7Twi5CRORo1NImItIJM1trZt83s5ejt/HR86PM7K9mtix6PzJ6frCZ/d7MXoveWrfxiZjZz81suZk9ZWZ5oX0pEUlrCm0i0tfldegevTTmud3uPgO4HfhJ9NztwP3ufiLwAHBb9PxtwDPufhLB3pmtu6NMAO5w9+OBncBFKf4+ItJLaUcEEenTzGyvuxd2cn4tcLa7rzGzLOAddy82s63AUHdvjJ7f5O4lZlYHDHf3gzHvMRr4i7tPiB7/G5Dl7t9J/TcTkd5GLW0iIofnh3l8uGs6czDmcTMaSywix0ihTUTk8C6NuV8QffwicFn08RXA89HHfwU+B2BmETPr111FikjfoH/xiUhfl2dmS2OOn3D31mU/cszsJYJ/4H40eu5fgHvM7CtAHXBV9Py1wF1m9imCFrXPAZtSXr2I9Bka0yYi0onomLYKd98adi0iIqDuUREREZG0oJY2ERERkTSgljYRERGRNKDQJiIiIpIGFNpERERE0oBCm4iIiEgaUGgTERERSQMKbSIiIiJpQKFNREREJA0otImIiIikAYU2ERERkTSg0CYiIiKSBhTaRERERNKAQpuIiIhIGlBoExEREUkDCm0iIiIiaUChTURERCQNKLSJiIiIpAGFNhEREZE0oNAmIiIikgYU2kRERETSgEKbiIiISBpQaBMRERFJAwptIiIiImlAoU1EREQkDSi0iYiIiKQBhTYRERGRNKDQJiIiIpIGFNpERERE0oBCm4iIiEgaUGgTERERSQMKbSIiIiJpQKFNREREJA0otImIiIikAYU2ERERkTSg0CYiIiKSBhTaRERERNKAQpuIiIhIGlBoExEREUkDCm0iIiIiaUChTURERCQNKLSJiIiIpAGFNhEREZE0kBl2AalWUlLio0ePDrsMERERkaNasmTJVncv7ey5Xh/aRo8ezeLFi8MuQ0REROSozGzd4Z5T96iIiIhIGlBoExEREUkDCm0iIiIiaSCUMW1mdj5wKxAB7nb3mzs8/2XgaqAJqAM+6e7ros89AcwEnnf393dr4SIiIn1EY2MjtbW11NfXh11Kr5Sbm8vw4cPJysqK+zXdHtrMLALcAZwL1AKLzOwRd6+OuexVoMLd95vZ54AfAJdGn7sFyAc+041li4iI9Cm1tbUUFRUxevRozCzscnoVd2fbtm3U1tYyZsyYuF8XRvfoDGC1u69x9wbgIeDC2AvcfZ67748eLgSGxzz3V2BPdxUrIiLSF9XX11NcXKzAlgJmRnFxccKtmGGEtmHAhpjj2ui5w/kU8HgiH2Bm15jZYjNbXFdXdwwlioiIiAJb6hzLn20Yoa2zKr3TC82uBCoIukTj5u53uXuFu1eUlna6Pp2IiIhIWgkjtNUCI2KOhwMbO15kZnOBrwMfdPeD3VSbiIiI9ADbtm1j2rRpTJs2jSFDhjBs2LC244aGhrje46qrrqKmpibuz7z77rspLS1t+5xp06Yl9PpUC2P26CJggpmNAd4GLgMuj73AzE4G7gTOd/ct3V+iiIiIhKm4uJilS5cCcOONN1JYWMh1113X7hp3x93JyOi8Deree+9N+HOvuOIKfvKTnxz2+aamJjIzD8Wno9UQq7m5mUgkknBNrbq9pc3dm4DPA08CK4Bfu/tyM7vJzD4YvewWoBD4jZktNbNHWl9vZs8BvwHOMbNaM3tPN38FERERCcnq1auZOnUqn/3sZ5k+fTqbNm3immuuoaKiguOPP56bbrqp7dozzzyTpUuX0tTUxIABA7j++us56aSTOP3009myJf42oaeffpq5c+dy2WWXcfLJJ3dawy9/+UtOOOEEpk6dyte+9jWAts/9xje+wYwZM3j55Ze79N1DWafN3R8DHutw7psxj+ce4bVnpbA0ERER6eBbf1pO9cbdSX3PKcf144YPHH9Mr62urubee+/lZz/7GQA333wzgwYNoqmpiaqqKi6++GKmTJnS7jW7du1izpw53HzzzXz5y1/mnnvu4frrr3/Xez/wwAPMnz+/7bg1aC1cuJDq6mpGjhzJ6tWr29VQW1vLN77xDRYvXkz//v2ZO3cujz76KOeffz67du1i+vTpfOc73zmm7xpLOyKIiIhIWhk3bhynnnpq2/GDDz7I9OnTmT59OitWrKC6uvpdr8nLy+OCCy4A4JRTTmHt2rWdvvcVV1zB0qVL227Z2dkAnH766YwcObLTGl566SXOPvtsSkpKyMrK4vLLL+fZZ58FIDs7mw9/+MNJ+d6htLSJiIhI+jjWFrFUKSgoaHu8atUqbr31Vl5++WUGDBjAlVde2en6Z63hCyASidDU1HTMn9nx2L3TRTCAICwma+kUtbR10bpt+7j85wt5+a3tYZciIiLS5+zevZuioiL69evHpk2bePLJJ7u9hpkzZzJv3jy2bdtGU1MTDz30EHPmzEn656ilrYuKC3NYtHY7f125mRljBoVdjoiISJ8yffp0pkyZwtSpUxk7dixnnHFGl96v45i2O++886ivGT58ODfddBOVlZW4Ox/4wAd43/vel3Br3tHYkZr0eoOKigpfvHhxSj/j8p8vZNveBp780uyUfo6IiEh3WbFiBZMnTw67jF6tsz9jM1vi7hWdXa/u0SSoLC+lZvMeNu48EHYpIiIi0ksptCVBVXkZAPNrtM+piIiIpIZCWxKMLytk2IA85tVo8wYRERFJDYW2JDAzKstLeXH1Vg42NYddjoiIiPRCCm1JUlVexr6GZhav3RF2KSIiItILKbQlyazxxWRHMpivLlIRERFJAYW2JMnPzuS0sYOYp8kIIiIiXbZt2zamTZvGtGnTGDJkCMOGDWs7bmhoiPt97rnnHt55551On7vyyisZM2ZM2/uedVbP3t5ci+smUWV5Gd9+tJoN2/czYlB+2OWIiIikreLiYpYuXQrAjTfeSGFhIdddd13C73PPPfcwffp0hgwZ0unzP/7xj/nQhz502Nc3NTWRmZl52ON4X5cMCm1JVFleyrcfhflv1vGPM0eFXY6IiEivdN9993HHHXfQ0NDArFmzuP3222lpaeGqq65i6dKluDvXXHMNgwcPZunSpVx66aXk5eXx8ssvt9uD9HC+8Y1vUFdXx5o1axgyZAhz5szh6aefZu/evRw8eJAnn3yS6667jqeeegoz44YbbuDiiy/m6aef5uabb6akpITly5fz+uuvJ/V7K7Ql0diSAkYOymf+yi0KbSIi0ns8fj28k9wAwpAT4IKbE37ZG2+8we9//3tefPFFMjMzueaaa3jooYcYN24cW7dubQtKO3fuZMCAAfz0pz/l9ttvZ9q0aZ2+35e+9CVuvPFGAE488UTuv/9+AF599VWeffZZcnNzufvuu1mwYAFLly5l4MCB/OpXv6K6uprXXnuNuro6Tj31VGbPDnZFWrhwIdXV1YwcOfIY/lCOTKEticyMqvJSfrV4A/WNzeRmRcIuSUREpFd5+umnWbRoERUVwU5PBw4cYMSIEbznPe+hpqaGa6+9lve+972cd955cb3f4bpHL7zwQnJzc9uOzzvvPAYOHAjA888/z+WXX04kEmHIkCGceeaZLF68mOzsbE4//fSUBDZQaEu6yvIy7luwjpff2s7siaVhlyMiItJ1x9Ailiruzic/+Um+/e1vv+u5ZcuW8fjjj3Pbbbfx8MMPc9dddx3z5xQUFBz2+Ej7tnd8XTJp9miSzRxbTE5mhnZHEBERSYG5c+fy61//mq1btwLBLNP169dTV1eHu3PJJZfwrW99i1deeQWAoqIi9uzZk9QaZs+ezUMPPURzczObN2/mhRdeaGv5SyW1tCVZXnaE08cV80xNHXwg7GpERER6lxNOOIEbbriBuXPn0tLSQlZWFj/72c+IRCJ86lOfwt0xM77//e8DcNVVV3H11VcfdiJC7Jg2gCVLlhy1hosvvpiFCxdy0kknYWb86Ec/oqysLKnfszN2pCa+3qCiosIXL17crZ/5Py+8xY1/qmb+dZWMLkldM6mIiEiqrFixgsmTJ4ddRq/W2Z+xmS1x906b7dQ9mgKV5UHa1u4IIiIikiwKbSkwuqSAMSUFzH9TuyOIiIhIcii0pUhleSkL/r6NAw3NYZciIiJyTHr7EKowHcufrUJbilSVl3GwqYWFa7aFXYqIiEjCcnNz2bZtm4JbCrg727Zta7cOXDw0ezRFZowZRF5WhPk1W6ialPoZJSIiIsk0fPhwamtrqavTUJ9UyM3NZfjw4Qm9RqEtRXKzIswaV8y8mjpujE4/FhERSRdZWVmMGTMm7DIkhrpHU6hyUhnrt+/nra37wi5FRERE0pxCWwpVRrexmlejpmURERHpGoW2FBoxKJ/xZYVar01ERES6TKEtxSonlvLSmu3sb2gKuxQRERFJYwptKVY1qYyG5hZeXK2lP0REROTYKbSlWMXogRRkR5inLlIRERHpAoW2FMvJjDBrfAnza+q0QKGIiIgcM4W2blBVXsbbOw+wesvesEsRERGRNKXQ1g0qy4OlP+Zr6Q8RERE5Rgpt3eC4AXmUDy7SuDYRERE5Zgpt3aRyUimL1m5n70Et/SEiIiKJU2jrJpUTy2hsdl5YvTXsUkRERCQNKbR1k4rRAynMydTuCCIiInJMFNq6SVYkg7MmlDBvpZb+EBERkcQptHWjyvJS3tldT83mPWGXIiIiImlGoa0bVZaXATBvpZb+EBERkcQotHWjwf1ymTK0n8a1iYiISMIU2rpZZXkpi9ftYHd9Y9iliIiISBpRaOtmVZPKaG5xnl+lpT9EREQkfgpt3ezkEQPol6ulP0RERCQxCm3dLDOSwVkTS5lfo6U/REREJH4KbSGoKi9jy56DVG/aHXYpIiIikiYU2kIwZ2IpAPNrtPSHiIiIxEehLQSlRTmcMKw/81ZqXJuIiIjER6EtJFXlpbyyfge79mvpDxERETk6hbaQzCkvo8Xh2VXqIhUREZGjU2gLybQRAxiQn8U8Lf0hIiIicVBoC0kkw5g9oZRn36yjpUVLf4iIiMiRKbSFqGpSKVv3NvDGxl1hlyIiIiI9nEJbiGZPKMVMS3+IiIjI0Sm0hai4MIcThw/QuDYRERE5KoW2kFWVl7J0w06272sIuxQRERHpwRTaQlZVXoY7PKelP0REROQIQgltZna+mdWY2Wozu76T579sZtVmtszM/mpmo2Ke+7iZrYrePt69lSffCcP6U1yQrd0RRERE5Ii6PbSZWQS4A7gAmAJ81MymdLjsVaDC3U8Efgv8IPraQcANwGnADOAGMxvYXbWnQkaGMWdiKc+8WUezlv4QERGRwwijpW0GsNrd17h7A/AQcGHsBe4+z933Rw8XAsOjj98D/MXdt7v7DuAvwPndVHfKzCkvZcf+RpbV7gy7FBEREemhwghtw4ANMce10XOH8yng8URea2bXmNliM1tcV9fzx4rNnlBKhsE8Lf0hIiIihxFGaLNOznXaL2hmVwIVwC2JvNbd73L3CnevKC0tPeZCu8vAgmxOHjmQZ7T0h4iIiBxGGKGtFhgRczwc2NjxIjObC3wd+KC7H0zktemocmIpr9Xuom7PwaNfLCIiIn1OGKFtETDBzMaYWTZwGfBI7AVmdjJwJ0Fgi21+ehI4z8wGRicgnBc9l/aqJpUB8Oyb6iIVERGRd+v20ObuTcDnCcLWCuDX7r7czG4ysw9GL7sFKAR+Y2ZLzeyR6Gu3A98mCH6LgJui59LelKH9KC3KYb5Cm4iIiHQiM4wPdffHgMc6nPtmzOO5R3jtPcA9qasuHK1Lf/ylejNNzS1kRrTusYiIiByiZNCDVJWXsetAI0s3aOkPERERaU+hrQc5c0IJkQxjvpb+EBERkQ4U2nqQ/nlZnDJyIPO09IeIiIh0oNDWw1ROKmX5xt1s2V0fdikiIiLSgyi09TCVE4OlPzSLVERERGIptPUwk4cWMbhfDvPVRSoiIiIxFNp6GDOjqryM51ZtpbG5JexyREREpIdQaOuBKstL2VPfxCvrdoRdioiIiPQQCm090BnjS8jMMOZp6Q8RERGJUmjrgYpys6gYPVDj2kRERKSNQlsPVVVexsp39rBp14GwSxEREZEeQKGth6qaFCz98Yy6SEVERASFth5rQlkhx/XP1e4IIiIiAii09VhmRuWkMp5ftZWGJi39ISIi0tcptPVgVeVl7GtoZvG67WGXIiIiIiFTaOvBZo0rJjuSwXyNaxMREenzFNp6sIKcTGaMGcS8lRrXJiIi0tcptPVwleWlrNqyl9od+8MuRUREREKk0NbDVZYHS3+oi1RERKRvU2jr4caVFjBiUJ5Cm4iISB+n0NbDmRmVE8t4YfVWDjY1h12OiIiIhEShLQ1UTSrlQGMzL7+lpT9ERET6KoW2NHD62BKyM7X0h4iISF+m0JYG8rIjzBxbrC2tRERE+jCFtjRRVV7Kmrp9rN+mpT9ERET6IoW2NNG29Mebam0TERHpixTa0sSYkgJGF+drdwQREZE+SqEtjVSWl7FgzTbqG7X0h4iISF+j0JZGKstLqW9sYeGabWGXIiIiIt1MoS2NzBxbTG6Wlv4QERHpixTa0khuVoRZ40qYr6U/RERE+hyFtjRTWV7K2m37eWvrvrBLERERkW6k0JZmKidGl/5Qa5uIiEifotCWZkYW5zO2tIB5GtcmIiLSpyi0paGq8jIWrtnGgQYt/SEiItJXKLSloaryMhqaWliwZmvYpYiIiEg3yYznIjMbFMdlLe6+s4v1SBxOHTOQ/OwI81bWcfakwWGXIyIiIt0grtAGbIze7AjXRICRXa5IjionM1j6Y17NFtwdsyP9WERERKQ3iDe0rXD3k490gZm9moR6JE5Vk0p5esVm/l63j/FlhWGXIyIiIikW75i205N0jSRJZbmW/hAREelL4gpt7l6fjGskeYYNyGPi4EJtaSUiItJHHDW0mdm5ZvZzM5sWPb4m9WVJPCrLy3jprW3sO9gUdikiIiKSYvG0tP0T8BXgSjM7G5iW2pIkXpXlpTQ2Oy+s1tIfIiIivV08oa3O3Xe6+3XAecCpKa5J4lQxahCFOZnMf1NdpCIiIr1dPKHtz60P3P164P7UlSOJyM7M4IzxxcxfGSz9ISIiIr3XUUObu/+x9bGZrQKqzOwGM7vQzEansDaJQ1V5GRt31fPm5r1hlyIiIiIplOg2VncC7wDbgAuAN8zsdTO7ycyykl6dHJWW/hAREekbEg1tV7rgQ4FLAAAgAElEQVT7P7n77e7+WeBM4G/AbuBHSa9OjmpI/1wmDSlinkKbiIhIr5ZoaNtlZie2Hrj7UmCmu/8ncEZSK5O4VU0qY/HaHeypbwy7FBEREUmRREPbZ4B7zewXZvYFM7sdaIk+l53c0iRelRNLaWrR0h8iIiK9WUKhzd1XAjOAJ4AyYDXwfjMrAB5KfnkSj+mjBlKUm8m8lVr6Q0REpLeKd8P4Nu7eDPwmeov1naRUJAnLimQwe0Ip898Mlv4ws7BLEhERkSRLqKXNzM6Odo3+0MyuMrNTzCwnVcVJ/OaUl7J590FWbNoTdikiIiKSAomOafsl8CiwEBgLfBNYnuyiJHGVE0sBNItURESkl0q0e3S1u/8++rhj96iEqKxfLlOH9eOZmjr+uWp82OWIiIhIkiXa0vaMmX3JNGiqR6qcWMaS9TvYdUBLf4iIiPQ2iYa244HPAZvM7M9m9l0zuyQFdckxqJpUSnOL8/wqLf0hIiLS28QV2szshujDHwKnAGOAG4BVwGmpKU0SNW3EQPrnZWlcm4iISC8U75i2p6L31xK0tmUB1cAygkkJ0gNEMozZE0uZX1NHS4uTkaFebBERkd4irpY2d18QffhVdz8eOAG4Cfg7x9DSZmbnm1mNma02s+s7eX62mb1iZk1mdnGH575vZm9Eb5cm+tm9XVV5KVv3HqR60+6wSxEREZEkSnRM2+8A3P2gu7/i7vcBDyfyBmYWAe4ALgCmAB81sykdLlsPfAL4vw6vfR8wHZhGEBa/Ymb9EvwOvdrs1qU/VqqLVEREpDeJd0zbR8zsZqDIzCZHg1eruxL8zBkES4escfcGgu2vLoy9wN3XuvsyDu1r2moK8Iy7N7n7PuA14PwEP79XKynM4aTh/TWuTUREpJeJt6XtBYIxbAOBHwGrot2XjwIHEvzMYcCGmOPa6Ll4vAZcYGb5ZlYCVAEjOl5kZteY2WIzW1xX1/f246wsL2Pphp3s2NcQdikiIiKSJPGOaXvb3e8HLnT3C9x9LDCXYAbp2Ql+Zmej4z3OOp4CHgNeBB4EFgBNnVx3l7tXuHtFaWlpguWlv8ryUlocnl3V9wKriIhIb5XQmDZ3fyHm8XZ3XxLtpkxELe1bx4YDGxOo4bvuPs3dzyUIgKsS/Pxe78ThAxhUkM0zNQptIiIivUVC21iZWTHwEaCeYM/R19090e7RRcAEMxsDvA1cBlwe5+dHgAHuvs3MTgRO5NByJBIVyTBmTyhh/pta+kNERKS3SHT26O+BUuA/gFuAXWa2MpE3cPcm4PPAk8AK4NfuvtzMbjKzDwKY2almVgtcAtxpZq2b0mcBz5lZNcEEiCuj7ycdVE0qY/u+Bpa9vSvsUkRERCQJEt0wvsjdbzKzf3D3OWZ2EZDw7uTu/hjB2LTYc9+MebyIoNu04+vqCWaQylHMnlCKGcyv2cK0EQPCLkdERES6KNGWtvro/UEzy3P3h4H3JrkmSYKBBdlMGzGAeRrXJiIi0iskGtr+08wGAb8C7jGzLxD/ch3SzarKy1hWu5Ntew+GXYqIiIh0UdyhzcwygPLorNEfEXRvjqDDwrjSc1SVl+Fa+kNERKRXiDu0uXsLcE7M8f+6+1fdffkRXiYhOv64fpQUZjNvpUKbiIhIuku0e3Spmd0QbXWTHi4jw5gzsYxnV9XR3BLX+sUiIiLSQyUavkYQrKu20cz+aGbfNrNLUlCXJElleSk79zeydMPOsEsRERGRLkh0R4SPuPtkYBTwLWA1wQbw0kPNnlBKRnTpDxEREUlfx9TN6e4H3f0Vd7/P3b+S7KIkefrnZ3HKqIHM19IfIiIiaS2u0GZmryTjGglHZXkZr7+9iy176o9+sYiIiPRI8ba0TTazZUe4vQ6UpLJQOXaV5aUA2kBeREQkjcW7jdWkOK5p7kohkjpThvajrCiH+W/WcUnFiLDLERERkWMQV2hz93WpLiStNTdCJCvsKg7LzKgsL+WJN96hqbmFzIhWbBEREUk3+u3dVfu2wg8nwWNfgbo3w67msKrKy9hd38SrWvpDREQkLSWyjZWZmfrWOmqqh/FzYcn/wB2nwv0fgpWPQUvP6i0+Y0IJmRnGvJVa+kNERCQdJbKNlQN/SGEt6an/cPiHO+FLy+Hsb0BdDTz0UbhtGrxwK+zfHnaFAPTLDZb+mKfJCCIiImkp0e7RhWZ2akoqSXeFZTD7K/DFZXDJfdB/BPzlm/CjyfDHz8OmZWFXSNWkMlZs2s07u7T0h4iISLpJNLRVAQvM7O+tS32YWfhppCeJZMHxH4KrHoPPvgAnXQav/xbuPAvuOR/eeDiYuBCCtqU/3lQXqYiISLqJd8mPVhekpIreashU+MCtMPdGePUBWPRz+O0noXAIVHwSTvkEFA3utnLKBxcxtH8u81bWcempI7vtc0VERKTrEt17dB0wAPhA9DZAy4HEIW8gzPo8fOEVuPzXQZib/x/w4+Ph4athw8vgnvIygqU/ynh+9VYam1tS/nkiIiKSPAmFNjO7FngAKIvefmlmX0hFYb1SRgQmvgeufBg+vwROvRrefBJ+cS7cVRm0xjWmdrxZZXkpew82sXjtjpR+joiIiCRXomPaPgWc5u7fdPdvAjOBTye/rD6gZDxccDN8uRre98Ng6ZA//lMwceHpG2HnhpR87BnjS8iKGPM1rk1ERCStJBrajPbbVTVHz8mxyikKWtz+aSF87BEYNStYKuTWE+GhK2DNM0ntOi3MyeTU0YOYv1JLf4iIiKSTRCci3Au8ZGa/jx5/CPhFckvqo8xg7JzgtnM9LL4HltwHKx+F0kkw49Nw4mWQU9jlj6oqL+O7j61g484DHDcgLwnFi4iISKoltCMC8BvgKmA7sAO4yt1/kqLa+q4BI4MZp1+uhgv/CzJz4M//GnSdPn49bF3dpbevmhQs/TFfC+2KiIikjbhb2tzdzewP7n4K8EoKa5JWWXlw8hUw7XKoXQQv3wWL7oaX/jvYOmvGNTD+XMhIrJd7XGkhwwbkMa9mC5efpqU/RERE0oF2REgHZjBiBlx0N3zpDaj8GrzzBvzfR+CnJ8OLt8OB+GeDmhlVk0p5cfVWDjb1rD1SRUREpHPaESHdFA2Byn+DL74OF98TLNT71NfhR1PgT9fC5uVxvU1VeRn7Gpq19IeIiEiaiLt7NDqm7bOAFtPtCTKzYepFwW3Ta/Dyz+G1h2DJ/8CoM4OJC5PeD5HOf8SnjysmO5LBvJVbOGN8SffWLiIiIgmLu6XN3R34sbuv63hLYX0Sj6EnwYW3w5dXwLk3wa718JuPB8uGPHsL7H33hIP87ExOGzuI+W9qMoKIiEg60Ji23iR/EJxxLfzLUrjsQSiZCH/7Dvx4CvzuM1C7pN3lleVlrN6ylw3b94dUsIiIiMTrWMa0LdSYth4uIwKT3gsf+wP886JgY/qVj8LdZ8PPzw66UZsOUlXeuvSHdkcQERHp6cwTWG3fzEZ1dr4nd5FWVFT44sWLwy4jfPW7g7D28l2wbRXkl+CnfIJLFk+i/+DR/OITakAVEREJm5ktcfeKzp6Lq6XNzL4KbeFsRofxbJ9JXqmSMrn94LRr4POL4B9/DyNmYM/9kF8d+AyXvPV1Dq5+NqnbZYmIiEhyxds9elnM43/v8Nz5SapFuoMZjDsbPvogXLuUtyddxUzeIOeXH4D/PgMW3wsN+8KuUkRERDqIN7TZYR53dizpYuBoyi76AXOa/4tHRv07WAY8+sVgu6wnvw7b14RdoYiIiETFG9r8MI87O5Y0kpsV4eRxx/HjbTPhs8/BVU/AuHPgpZ/BbdPhgY/AqqehpSXsUkVERPq0eBfXPcnMdhO0quVFHxM9zk1JZdJtqsrLuOGR5azdtp/Ro06HUafD7o3BQr2L74UHLoLcATDkhOA2eCoMmQqlk4LN7EVERCTl4gpt7h5JdSESnsqYpT8+UTImONnvOKj6Gpx1Hax4BNY+B++8HoS4pgPBNRmZUFIeDXNTo4HuBCgoDumbiIiI9F5xb2Mlvdeo4gLGlhQwr6aOT5wxpv2TmdlwwsXBDaClORjr9s6yYNP6d16Ht56BZQ8dek3RcUGIGzz1UOvcoLHB+nEiIiJyTBTaBAh2R3jgpXUcaGgmL/sI4SojAiUTgtvUiw6d37c1CHCbo0HunTfg73+Dlqbg+ax8KJtyqFVu8Akw+HjIKUztFxMREeklFNoECLpI73nhLRau2UbVpLLE36CgBMZVBbdWTQehbuWhFrnNb8Dy38GSe6MXGAwac6hbtbWLtd+wYGkSERERaaPQJgDMGDOIvKwI82u2HFto60xmTrCZ/dCTDp1zh121Ma1yy2DTMqj+46FrYic9tE58KJ0UdNWKiIgkorkRGvdDY31w31QPjQeCW1P0vt1z0Wvbnmu9th5Gng6zPh/aV0kotJmZAVcAY939JjMbCQxx95dTUp10m9ysCGeML2ZeTR03umOpaukygwEjgtuk9x46X78btlRHu1ajga7dpIcsKC2PGSc3VZMeRETSkXsQpN4VijoLUB2fS+Ta6HPefGx1RnIgK+/QLTMvaEAIUaItbf8FtABnAzcBe4CHAW1c2QvMKS/j6RVbWLN1H+NKu3msWW4/GDkzuLVqaYZtf4fNrx8aJ7dmfieTHk6ImfhwYtDlqkkPIiLJ19QA+7fCvrrobWv0Vgf7t0VbqY4WrvaDH+Pan5l5kJUbvY8NVLlQODh4Lis/OI4NW1l5HZ7LP8z7xLxfRrxL2XafREPbae4+3cxeBXD3HWamPqteonJi69Ifdd0f2jqTEYHSicEtdtLD3rpokHvj0MSH1U8f+tdUVn4wyaF1PbkhJwaTIDTpQSQ87rB/O+xaDzs3wK4Nwf3O9cG5XW8Hf+fzBkH+oOj9wOA+b2DMudj7gVorsqtamuHAjpgQVgf7tnU4bg1lW6F+V+fvE8mG/GLILmgfkvIGtA9JbaGpkwB12LAVvUVyemSQ6k6JhrZGM4sQ3QXBzEoJWt6kFxgxKJ/xZYXMr9nCp84cc/QXhKWwFArPDvZQbdVYH0x62PzGoYkP75r0MLb9enJDTgjWo9OkB5Gua2mGPe8cCmMdw9muDUELS6zsQugfHS4x/NSg9WX/9iBE7FwHG1+FA9uD1pnDySo4FOA6BrrOQl7ewGDcbG/95e8OB3cfClrvuo8JYvu3Bq1jnbV6WUYQwgpKg/uhJwWPC0qDiWftHpdATj/9v7QbJBrabgN+D5SZ2XeBi4FvJL0qCU1VeSn3vbiOfQebKMhJo3kqWblw3LTg1so9+EXRNnv19XdPesgbGJ3oUA45RcEvgLZ/2eUH99mdnIu9Tl2x0hc0HQwmEcWGsLb79bD77UNL/LTKLw5CWelEGD83CGetIa3/iODvXzy/6Bv2B+Ft//bg/sCOQ4/374h5bkdQ4/7tUL/z8F1wlhEEt46td3kDD7XudRYAs/O7/ud4LBoPxNcK1vq4uaHz98ntD/nRwFU8LhiO0jF8tR7nDdT/23qguH8rRychPAssAc4h2MLqQ+6+IkW1SQiqysv4+XNvseDv25g7ZXDY5XSNGQwYGdyONunh9d8EvxhaGhP/nNbBqkcLd23n8oP/+Xc813bL63Bdvv7nKal3cE/7ENYxnO3dTPutpi1oqe4/AkbMiAljI6P3w4O/E8mQHf270H94/K9paQmCW1vA29E++MXe794Im5cHjxv3Hf49M3Pf3WrXabdtTMtf7gCIdPhV29wYtHAdrRWs9XHD3sPUk3cocBUOiU7OKuk8hOUXqyu5F4g7tLm7m9kf3P0UYGUKa5IQVYweREF2hHk1W9I/tB1OZ5MeWjU3xgyi3RfzODq4tmFf++PG/R0eHwjCX+P+4H7ftpj3iT5/uH8FH0kkO45w11kIbH2cG7QueEvQAukOePS49VzLYc7R/rjddR3PeZKvawlyQjzXZWRGg3NBcB97y8oPuuKyC6IBoDB6riB43NuXk2kdT7ZzXSctZdFuzPqd7V8TyQ7WTBwwItpKNrJ9S1m/YRDJCuf7xCMjIwhN+YOCVqV4NR08TIteh5a9AzugrubQc0eaoZjbPwhyGZlBGDuw4zA1Zx5qCSsogYFjoo+LY1rEYronkxWKJW0k2v+10MxOdfdFKalGQpedmcEZ40uYX1OHp3Lpj54qkhXccvul7jOam94d9NqFxI4BMfa6Duca9kdbBzqEyGMJhilhQVi0jKDls/Vx23k7dL7duXiuy2h/riX659qwN/iz69hVdyTvCnwxIa8t8OXHhMDOgmEnr41kd884ny6PJ5sRE8hGBveFg3vvuK8jycyBfkODW7xax5G1a9HrpGWvpfEw48KiLWG9eaydJEWioa0K+IyZrQP2EXSRurufmPTKJDRVk8p4qnozq7fsZcLgorDL6X0imRDpl/pg2NQh3MFhAtLhQtIxBKd250IO/E0NQYBr3B+EuIa9wZ9Dw74g/DZ0uMUGvtbr9tXFHLeGwQS60C1yhMDXSYtf63WdhcKMzGDcWHePJ5OjMwta03L7Az14EpekvURD2wUpqUJ6lMryYOmPeTVbFNrSVSQTIkXB5Iq+KjMbMgcBg5L7vk0N0dAXEwbbgmFsCNzXPvDFXrN/WxC2YoNiQq2jBkVDg1axVI8nE5EeI6HQ5u7rzGwgMAHIjXlqXVKrklAN7Z/HpCFFzK+p45rZCYwFEekLMrODW97A5L5vc2NMi18nLYPNjVA0JAho/Yb1/nF4IvIuiW5jdTVwLTAcWArMBBYQ7JAgvcic8lLuef4t9tQ3UpTbgwcbi/QWkaxgIdK8AWFXIiI9VKIjHq8l2LJqnbtXAScDdUmvSkJXVV5GY7PzwuptYZciIiIiJB7a6t29HsDMctx9JVCe/LIkbKeMGkhRTibPvLkl7FJERESExCci1JrZAOAPwF/MbAewMfllSdiyIhmcOaGEeSv76NIfIiIiPUxCLW3u/mF33+nuNwL/D/gFcGEqCpPwVZWX8c7uemo27wm7FBERkT4v0YkI3+zk9DTgpuSUIz3JnNalP1bWMWlICtcUExERkaNKdEzbvphbM8G6baMT/VAzO9/MasxstZld38nzs83sFTNrMrOLOzz3AzNbbmYrzOw2U79dygzul8uUof2YV6NxbSIiImFLdJ22H8Yem9l/Ao8k8h5mFgHuAM4FaoFFZvaIu1fHXLYe+ARwXYfXzgLOAFp3YHgemAPMT6QGiV/VpFJ+9swadtc30k9Lf4iIiISmq5uc5QNjE3zNDGC1u69x9wbgITqMi3P3te6+DGjp8FonWNQ3G8gBsoDNx1K4xKeyvIzmFuf5VVvDLkVERKRPSyi0mdnrZrYselsO1AC3JviZw4ANMce10XNH5e4LgHnApujtSXdf0Umd15jZYjNbXFenZeS64uQRA+iXm8lDizawvyGBDbhFREQkqRJtaXs/8IHo7TzgOHe/PcH36GwMmsf1QrPxwGSCHRmGAWeb2ex3vZn7Xe5e4e4VpaWlCZYnsTIjGfxT1XiefbOOC259jpff2h52SSIiIn1Sokt+rIu5ve3ux9L0UguMiDkeTvxrvX0YWOjue919L/A4wVZakkKfnTOOh66ZiTtcetcCbnxkuVrdREREulmiS358+UjPu/uP4nibRcAEMxsDvA1cBlweZwnrgU+b2fcIWuzmAD+J87XSBTPHFvPEF8/iB0/U8D8vrmVezRZuufgkZowZFHZpIiIifUKi3aMVwOcIuiaHAZ8FpgBF0dtRRVvnPg88CawAfu3uy83sJjP7IICZnWpmtcAlwJ3R8XMAvwX+DrwOvAa85u5/SvA7yDHKz87kxg8e367V7Vt/Ws6BhuawSxMREen1zD2u4WTBxWZPARe5+57ocRHwG3c/P0X1dVlFRYUvXrw47DJ6nf0NTXz/8ZXct2Ado4rz1eomIiKSBGa2xN0rOnsu0Za2kUBDzHEDx7C4rqS//OxMvnXhVB789Exa3Ln0rgXc9KdqtbqJiIikSKKh7X+Bl83sRjO7EXgZuD/pVUnaOH1cMU9cO5uPzRzFPS+8xQW3PsuitZphKiIikmwJdY8CmNl04CyCZTqec/dXU1FYsqh7tPss+Ps2vvrwa9TuOMBVs8bwlfeUk5cdCbssERGRtNHl7tHoxIAhAO7+CrATmAtcZWYayCTAoVa3f4y2ur33tufU6iYiIpIk8XaP3kl0LFt0MdvvAfcBu4C7UlOapKOCnExuunAq//fp02hqaeEjd2qsm4iISDLEG9oi7t7aZHIpcJe7P+zu/w8Yn5rSJJ3NGlfyrla3xWp1ExEROWZxhzYza12I9xzgbzHPJbRAr/Qdsa1ujc0tXHLnAr79qFrdREREjkW8oe1B4Bkz+yNwAHgO2vYC3ZWi2qSXmDWuhCe/OJsrTxvFL55Xq5uIiMixiCu0uft3gX8F/gc40w9NOc0AvpCa0qQ3KcjJ5Nsfat/q9h21uomIiMQt4SU/0o2W/Oh59h1s4ubHV/K/C9cxpqSAWy4+kYrRmoQsIiKSzB0RRLqsrdXt6vatbvWNanUTERE5HIU2Cc2s8cFYtytOG8ndz7/Fe299jiXrNNZNRESkMwptEqqCnEy+86ET+L+rT+NgUwsX/0ytbiIiIp1JaLkOM8sBLiLYJL7tte5+U3LLkr5m1vgSnvzSbG5+fAV3P/8Wf1u5hVsuOZFTRmmsm4iICCTe0vZH4EKgCdgXcxPpssJoq9sDMa1u3/2zWt1EREQgwdmjZvaGu09NYT1Jp9mj6WnvwSa+99gKHnhpPWNLCrjlkpM4ZdTAsMsSERFJqWTOHn3RzE5IQk0iR1SYk8l3P3yo1e2Sn73Ifzy2Qq1uIiLSZyUa2s4ElphZjZktM7PXzWxZKgoTATgjOtbtozNGcteza3jvbc+xZN2OsMsSERHpdol2j47q7Ly7r0taRUmm7tHe44XVW/nqb5exadcBrj5rLF8+dyK5WZGwyxIREUmapHWPRsPZbmAwMCrmJpJyra1ul8W0ur2yXq1uIiLSNyQU2szsauBZ4EngW9H7G5NflkjnCnMy+Y8Pn8AvP3UaBxtbuPi/NdZNRET6hkTHtF0LnAqsc/cq4GSgLulViRzFmRPU6iYiIn1LoqGt3t3rIVho191XAuXJL0vk6DprdfueWt1ERKSXSjS01ZrZAOAPwF/M7I/AxuSXJRK/MyeU8MQXz+LSU0dy57NreJ9a3UREpBdKaPZouxeazQH6A0+4e0NSq0oizR7tW55bVce//XYZ7+yu59Ozx/KluZphKiIi6SNps0ctcKWZfdPdnwGWAtOSUaRIMpw1oZQnvzQ7aHV7Jmh1e1WtbiIi0gsk2j36X8DpwEejx3uAO5JakUgXFeVm8b1/OIH//dQMDjQ0c9F/v8j3HtdYNxERSW+JhrbT3P2fgXoAd98BZCe9KpEkONTqNoI7n1nD+3/6vFrdREQkbSUa2hrNLAI4gJmVAi1Jr0okSYJWtxO5/5Mz2H+wSa1uIiKSthINbbcBvwcGm9l3geeB7yW9KpEkmz3x3a1uSzfsDLssERGRuCU8e9TMJgHnRA//Gl2rrcfS7FHp6Nk367j+4WCG6TWzx/HFuRM0w1RERHqEI80ejSu0mdkjHU9F7x3A3T/YpQpTSKFNOrO7vpH/+PMKHlq0gfFlhdxy8YmcPHJg2GWJiEgfl4zQVgdsAB4EXuJQaAMguvxHj6TQJkfyTLTVbdOues6aUMLHTx9N1aQyIhl29BeLiIgkWTJCWwQ4l2CpjxOBPwMPuvvyZBaaCgptcjS76xu5/8W1/HLhet7ZXc+IQXl8bOZoPlIxgv75WWGXJyIifUiXQ1uHN8shCG+3ADe5+0+7XmLqKLRJvBqbW3hq+WbuW7CWl9/aTm5WBh8+eRgfO300k4f2C7s8ERHpA5IS2qJh7X0EgW008Ahwj7u/naQ6U0KhTY5F9cbd3L9gLX9Y+jb1jS3MGDOIT8wazXlTBpMZSXTStYiISHyS0T16HzAVeBx4yN3fSG6JqaPQJl2xc38Dv168gfsXrKN2xwGG9s/litNGctmMkZQU5oRdnoiI9DLJCG0twL7oYewLDHB377F9RwptkgzNLc7fVm7h/gVreW7VVrIjGbz/pKF8/PTRnDRiQNjliYhIL3Gk0JYZzxu4u/qDpE+LZBjnThnMuVMGs3rLXu5fsJaHl9Tyu1feZtqIAXx81ijee8JQcjK13puIiKRGwhMR0o1a2iRV9tQ38vCSWu5fsI41W/dRUpjN5TNGcvlpoxjSPzfs8kREJA0ldfZoulFok1RraXGeX72V+15cy99qthAx4z1Th/CJWaOpGDUQM635JiIi8ely96iIHF5GhjF7YimzJ5aybts+frlwHb9atIE/L9vElKH9+PisUVw4bZi2yhIRkS5RS5tICuxvaOIPr27kvhfXUrN5DwPys7i0YgRXzhzFiEH5YZcnIiI9lLpHFdokJO7OS29t574X1/JU9WbcnXMmD+YTs0Yza1yxuk5FRKQddY+KhMTMmDm2mJlji9m48wAPvLSOB1/ewF+qNzO+rJCPnz6Kf5g+nIIc/VUUEZEjU0ubSDerb2zm0WWbuO/Ftbz+9i6KcjK56JThfOz0UYwtLQy7PBERCZG6RxXapAdyd17dsJP7XlzLY69vorHZmTOxlI/PGkXlxDIyMtR1KiLS1yi0/f/27jw27vu88/j74X0Nz+Ex4iFSEnWSlnzpSqwmcZyYapBgezpJ27QoYKBXsrttd9v9Yxdom6KLbrtJ0WABr5vdGHXj3XXjxOtKduxcdmKF8iFZpCRalEVKpDS8b1I859s/ZjgkFV2URP74G35eAKGZ3wxnHmIg6cPv8z0U2mSN6xmd5JuNHTzbeJGe0Sk2FmXx6/s38ssPVZKXmep1eSIiskoU2hTaxCemZyO8crqLb7zZztsXB8lMTebfPFDOFw5Us60s4HV5IiKywmodonMAABdPSURBVBTaFNrEh5ovD/PMsXa+c/IKU7MR9m8q5DcPVvPxHaWkJOtkORGRRKTQptAmPjY4Ps1zb3Xwjz+9yOWhq2zIy+Dz+zfy2b1VFGaneV2eiIjcQwptCm2SAOYijtfOdvONN9t584N+0lKS+PTuDXzhQDX1FXlelyciIveAQptCmySY1u5RvnGsnW+9e5mJ6TkeqMrnCweraagLkZai1qmIiF8ptCm0SYIavjrDP7/TyTPH2mnvn6A4kM7n9lbx+X1VlORmeF2eiIgsk0KbQpskuEjE8aPWXr7xZjs/fL+XlCTjcH2ILxzcyANVBTouS0TEJ3SMlUiCS0oyPrqthI9uK6G9b5xnjl3k/73TwYvvXaGuPJffOFDNp+4LkZWmv/IiIn6lkTaRBDU+NcsLJy7zzLF2znWPkZGaxEe2ltBQX8bHtpcQyNCmvSIia43aowptso455zjeNsCRpjBHm7voGZ0iLSWJQ7VBGupCfHxnqU5dEBFZIxTaFNpEgOjct3cvDXKkqYujzWHCw5OkJhsf2hLkcF2Ix3aWUqC930REPLPmQpuZPQ58FUgGnnbO/dU1jx8CvgLcBzzhnHs+dv2jwH9f9NTtsce/faP3UmgTub5IxPFe5xAvN3dxpDlMx8BVkpOMA5uKaKgv45O7ygjmpHtdpojIurKmQpuZJQPngMeATuAt4LPOuTOLnlMN5AJ/BLw4H9queZ1C4DxQ4ZybuNH7KbSJ3JpzjtNXRjjSFOZIU5j2/gmSDPbWFHK4PsQnd5VRqi1ERERW3FpbPboXOO+cuwBgZs8BnwHioc051x57LHKT1/kl4OjNApuI3B4zo648j7ryPP74k9to6RrlaGwO3H/+zmn+y4unebCqgIb6EA11ZWzIz/S6ZBGRdceL0FYOdCy63wnsu4PXeQL42+s9YGZPAk8CVFVV3cFLi6xfZsaOUC47Qrn8+09so7V7lKPNXRxpCvPnL53hz186w57KfA7Xl9FQF6KyMMvrkkVE1gUvQtv1dvlcVo/WzEJAPfDK9R53zj0FPAXR9uhyCxSRBbWlAWpLA3zx0Vra+sY52hzmaFMXf3mkhb880kJdeS4NdSEO14eoCWZ7Xa6ISMLyIrR1ApWL7lcAV5b5Gr8CvOCcm7lnVYnILdUEs/ndj2zhdz+yhY6BCY42hznS1MVfv/I+f/3K+2wvC3A41kKtLQ14Xa6ISELxYiFCCtGFCI8Cl4kuRPicc+70dZ77v4GXrl2IYGY/Bf7UOfeDW72fFiKIrLzLQ1d5ubmLl5vDvH1xEOdgS0kOh+vKaKgPsb0soKO0RERuw5paPQpgZoeJbumRDHzdOfdlM/sz4G3n3Itm9jDwAlAATAJdzrldse+tBn4CVDrnbrZQAVBoE1lt3SOTvHI6OgfueNsAERcdoWuoK+NwfYhdG3IV4EREbmDNhbbVpNAm4p2+sSm+e7qbo81h3vygn7mIo6IgM95C3VOZrwAnIrKIQptCm4jnBsenefVMN0eaw/zkfB8zc44NeRk8Xheiob6MB6sKSEpSgBOR9U2hTaFNZE0ZnpjhtbPdHG3u4vXWXqZnI5QE0nm8LrqNyN6aQpIV4ERkHVJoU2gTWbNGJ2f4fksPR5u6+OG5HiZnIgRz0vjErjIa6srYv6mI1OQkr8sUEVkVCm0KbSK+MDE9yw/f7+VIU5jvt/QwMT1HflYqn9hZSkN9iA9tDpKWogAnIolLoU2hTcR3Jmfm+NG5Xo42hXntbA9jU7MEMlJ4bGcpDXUhHqkNkpGa7HWZIiL3lEKbQpuIr03NzvHj1j6ONnfx3dNdjEzOkp2WzKM7SjlcX8bPbS0hM00BTkT8T6FNoU0kYUzPRjh2oZ+jTWFeOd3F4MQMmanJHNoa5JHaYg7VFlNVpPNQRcSfFNoU2kQS0uxchONtA/xLU5gftPRwZXgSgI1FWTxSGw1xBzYXkZuR6nGlIiK3R6FNoU0k4TnnuNA3zhvnevnx+T6OfdDP+PQcyUnGnsr8WIgLsrsinxStRhWRNUqhTaFNZN2Zno1w4tIgb7T28cb5Pk51DuEcBDJSOLi5SK1UEVmTFNoU2kTWvcHxad78oJ83Wnt5o7WPy0NXAagqXNpKzctUK1VEvKPQptAmIos452jrG4+OwrX2/kwr9cNbghzaqlaqiKw+hTaFNhG5ienZCCc7hnijtZfXWxe1UtNTOLgl2kp9pDbIxqJsr0sVkQSn0KbQJiLLMDQxzU/O9/Pj8728fu56rdQgBzYH1UoVkXtOoU2hTUTu0NJWah/HPuhjfHqOJCO2KrVYrVQRuWcU2hTaROQemZmLcOLSQiu1qXOISKyVemBzEY9sLeaQWqkicocU2hTaRGSFDE0srEq9tpX64dogh9RKFZFlUGhTaBORVeCco71/Ih7grm2lfrg2Ogq3uzKfVLVSReQ6FNoU2kTEAzNzsVWp5xZWpS5ppcb2h9tYlIWZeV2uiKwBCm0KbSKyBiy0Uvt4/VxvvJVaWZgZO6FBrVSR9U6hTaFNRNaYa1upP73Qz9jULEkGu+dXpaqVKrLuKLQptInIGnezVur+WCt1/6Yiakty1EoVSWAKbQptIuIzwxMzvPlBH6/HjtrqHIy2Uguz03i4uoB9NUXsrSlkRyiX5CSFOJFEcbPQlrLaxYiIyK3lZaXSUB+ioT6Ec45LAxM0tg3QeGGA4+39vHK6G4BARgoPVxeyr6aQvTWF1JXnqZ0qkqAU2kRE1jgzY2NRNhuLsvmVhyoBuDJ0leNtAzS29dPYNsD3W3oAyEpL5sGNBbEQV8TuyjzSU5K9LF9E7hG1R0VEEkDP6CRvtQ3S2NbP8bYBWrpGAUhLSeL+ynz2bSpif00h91cVkJmmECeyVmlOm0KbiKwzg+PTvNU+QGPbAMfbBjh9ZZiIg9Rk476KfPbWRFuqD24sIJChLUZE1gqFNoU2EVnnRiZneOfiYHROXFs/pzqHmY04kgx2bchjX00h+zYV8XB1AflZaV6XK7JuKbQptImILDExPcuJS0M0XojOiTvRMcT0bAQz2FYaWBTiCikOpHtdrsi6odCm0CYiclOTM3Oc6hym8UI/x9sHeLt9kKszcwBsLs5mb00R+zdFV6iG8jI9rlYkcWnLDxERuamM1GT2xrYNgehmv82Xh+Nz4l567wrfPH4JgKrCrPicuH01RVQWZmrDX5FVoJE2ERG5pbmI42x4JBbioitUBydmAAjlZcRCXHTD383F2QpxIndI7VGFNhGReyoScZzvHYvPiWtsG6B3dAqAYE7akhC3rTRAkk5tELktCm0KbSIiK8o5R3v/RHROXCzEXR6KHr2Vl5nKw9WF8TlxO0O5pOjUBpHr0pw2ERFZUWZGTTCbmmA2T+ytAqBzcCK2xUj05IbXzkaP3spJT4me2rApOi+uvjyftBSFOJFb0UibiIisiu6RyficuMYLA7T2jAGQkZrEA1UFPLSxgD1V+eyuyKcoR9uMyPqk9qhCm4jImtM/NhU/taHxwgAtXSNEYv8lVRVmsacyP/pVlc/OUC4ZqTp+SxKfQptCm4jImjcxPUtT5zAnO4biX+HhSSB6/NbOUG48xO2pLKC6KEurVCXhKLQptImI+FL3yCQnLs2HuEFOdQ4zMR3d9Dc/K5XdFQujcXsq8inI1hFc4m8KbQptIiIJYS7iaO0Z5eSlhdG4c92j8bZqddHitmoBO0IB0lPUVhX/UGhTaBMRSVhjU4vbqoOc7BiieyS6Z1xachI7N0TbqvdXRcNcVaHaqrJ2KbQptImIrCvh4avx0bgTHUM0dQ7Hz1ItzE5jd0UeeyoL4m3VvKxUjysWidI+bSIisq6E8jIJ1WfSUB8CYHYuwrnuMU50DMbD3A/P9TI/brEpmL1okUM+28tytXecrDkaaRMRkXVpdHKGU7G26vxih76xWFs1JYm6Dbnx0bj7K/OpKMhUW1VWnNqjCm0iInILzjkuD12Nzo2Lhbimy8NMzUaA6Jmq86tV768q4L7KPHIz1FaVe0vtURERkVswMyoKsqgoyOJT920AYGYuwvtdo5yIB7lBvtfSE3s+bC7OWVitWpnP9rKAzlWVFaORNhERkWUYvjrDqc6hJduO9I9PA9EjuerL82IhLtpa3ZCXobaq3Da1RxXaRERkhTjn6By8umQ0rvnKCNOxtmpxIJ09lfnsrshjRyiXHaFcQgpycgNqj4qIiKwQM6OyMIvKwiw+vTvaVp2ejdDSNbJkftyrZ7rj35OXmcr2skAsxEX/3Foa0PmqclMaaRMREVkFo5MznOse5Ux4lLPhEVrCI7R0jcaP5UoyqAlmx0fj5sNcWa5G5dYTjbSJiIh4LJCRyoMbC3lwY2H8WiTiuDQwQUvXSDzMvdc5xEunwvHn5GctHpXLZUdZLrWlORqVW4cU2kRERDySlGRUB7OpDmbzeF0ofn1kcob3u0ZpCS+EueeOd8RPdUhOMjYFs9k+PyJXFg10pbnpGpVLYAptIiIia0xuRioPVxfycPXSUbmLAxPx1uqZ8CjvXhzk/793Jf6cgqxUtpctba9uKdGoXKJQaBMREfGBpCSjJphNTTCbw/VLR+Va5ufJxdqs/3T8IpMz0dWryUnG5uLsnwlzJQGNyvmNQpuIiIiP5WaksremkL01C6NycxHHxf5xzi4Kc+9cHOTFRaNyhdlp7AgFloS5LSU5pKdoVG6tUmgTERFJMMlJxqbiHDYV5/Dz9y2Myg1PzNDSNcLZ8Ahnw6O0dI3wbOPCqFxKkrG5OCca5hatYi3O0ajcWqDQJiIisk7kZaWyb1MR+zYVxa/NRRxtfeNLwtzxtgG+fXJhVK4oOy0e4OZH5raU5JCWoiO7VpNCm4iIyDqWnGRsKclhS0lO/MxVgKGJ6fho3HyYe+bYRaZmF0bltpTksCOUG9+SZGtpQCtYV5BCm4iIiPyM/Kw0Dmwu4sDmhVG52bkI7f3jnAlHtyM5Gx7h2Af9vHDicvw5uRkp1JYG2FqaQ21JgK2x28Va+HDXPDkRwcweB74KJANPO+f+6prHDwFfAe4DnnDOPb/osSrgaaAScMBh51z7jd5LJyKIiIisrMHxac52jXC+Z4xz3aOc6x6jtXuUwYmZ+HPyMlOjQa40wNaSHLaWBqgtDRDMSVOYW2RNnYhgZsnA14DHgE7gLTN70Tl3ZtHTLgG/CfzRdV7iGeDLzrlXzSwHiKxwySIiInITBdlpHNwc5ODmYPyac46+sWlau0ejQa4nGuT+5VSYf7q6EOYKslLjI3NbSwOx0bkcinLSvfhR1jQv2qN7gfPOuQsAZvYc8BkgHtrmR87MbEkgM7OdQIpz7tXY88ZWqWYRERFZBjOjOJBOcSCdg1uWhrne0SnOdUdH5Vp7oiNz3zl5hdHJ2fjzirLT2BIbkYuP0JUGKMxO8+LHWRO8CG3lQMei+53Avtv83q3AkJl9C6gBXgP+xDk3t/hJZvYk8CRAVVXVXRcsIiIi94aZUZKbQUluBh+uXRrmukemYu3VUVq7xzjXM8q3T1xmdGohzAVz0uKjcfNBbmtpDvlZiR/mvAht12tc3+7EuhTgEeB+oi3U/0O0jfoPS17MuaeApyA6p+1OCxUREZHVYWaU5WVQlpfBoa3F8evOOcLDk7TG2qvzc+aef6eT8emFMZviQPqSxQ+1pTlsLQmQl5XqxY+zIrwIbZ1EFxHMqwCu3OC51/veE4taq98G9nNNaBMREZHEYGZsyM9kQ34mP3dNmLsyPBkblVtY/PB/3+5gYlGYKwmkL4S4Ra3W3Az/hTkvQttbQK2Z1QCXgSeAzy3jewvMrNg51wt8DNDSUBERkXXGzCjPz6Q8P5OPbiuJX49EHJeHrsbnys23Wp873sHVmYUwV5ab8TNBrrYkh8AaDnNebflxmOiWHsnA151zXzazPwPeds69aGYPAy8ABcAk0OWc2xX73seAvyHaZn0HeNI5N32j99KWHyIiIhKJODoHr8ZWssbmzHWPcr5nLL5hMMCGvIyFfeZic+a2lOSQk74641w32/LDk9C2mhTaRERE5EbmIo6OgQlae8aWtFrP944xvSjMledn8gsPlPOHn9i2ovWsqX3aRERERNaK5CSjOphNdTCbx3aWxq/PRRyXBiaWBLm8TG9bpwptIiIiItdITjJqgtnUBLP55K4yr8sBIMnrAkRERETk1hTaRERERHxAoU1ERETEBxTaRERERHxAoU1ERETEBxTaRERERHxAoU1ERETEBxTaRERERHxAoU1ERETEBxTaRERERHxAoU1ERETEBxTaRERERHxAoU1ERETEBxTaRERERHxAoU1ERETEB8w553UNK8rMeoGLq/BWQaBvFd5HVoY+P//TZ+h/+gz9T5/h3dvonCu+3gMJH9pWi5m97Zx7yOs65M7o8/M/fYb+p8/Q//QZriy1R0VERER8QKFNRERExAcU2u6dp7wuQO6KPj//02fof/oM/U+f4QrSnDYRERERH9BIm4iIiIgPKLSJiIiI+IBC210ys8fN7H0zO29mf+J1PbI8ZlZpZj8ws7NmdtrMvuR1TbJ8ZpZsZifM7CWva5HlM7N8M3vezFpifxcPeF2TLI+Z/bvYv6HNZvZNM8vwuqZEpNB2F8wsGfga0ADsBD5rZju9rUqWaRb4Q+fcDmA/8Hv6DH3pS8BZr4uQO/ZV4GXn3HZgN/osfcXMyoEvAg855+qAZOAJb6tKTAptd2cvcN45d8E5Nw08B3zG45pkGZxzYefcu7Hbo0T/syj3tipZDjOrAH4eeNrrWmT5zCwXOAT8A4Bzbto5N+RtVXIHUoBMM0sBsoArHteTkBTa7k450LHofif6D9+3zKwauB9o9LYSWaavAP8BiHhdiNyRTUAv8L9iLe6nzSzb66Lk9jnnLgP/DbgEhIFh59x3va0qMSm03R27zjXtoeJDZpYD/DPwb51zI17XI7fHzD4F9Djn3vG6FrljKcADwP9wzt0PjAOaH+wjZlZAtMtUA2wAss3s17ytKjEptN2dTqBy0f0KNCTsO2aWSjSwPeuc+5bX9ciyfAj4tJm1E52e8DEz+0dvS5Jl6gQ6nXPzI9zPEw1x4h8fB9qcc73OuRngW8BBj2tKSAptd+ctoNbMaswsjejEyxc9rkmWwcyM6Fyas865v/W6Hlke59yfOucqnHPVRP/+fd85p9/wfcQ51wV0mNm22KVHgTMeliTLdwnYb2ZZsX9TH0WLSVZEitcF+JlzbtbMfh94hehqma875057XJYsz4eAXweazOxk7Np/cs4d8bAmkfXmD4BnY7/8XgB+y+N6ZBmcc41m9jzwLtEV+SfQcVYrQsdYiYiIiPiA2qMiIiIiPqDQJiIiIuIDCm0iIiIiPqDQJiIiIuIDCm0iIiIiPqDQJiLrmpnNmdnJRV/3bDd+M6s2s+Z79Xoisr5pnzYRWe+uOuf2eF2EiMitaKRNROQ6zKzdzP6rmR2PfW2JXd9oZt8zs1OxP6ti10vN7AUzey/2NX+MT7KZ/U8zO21m3zWzTM9+KBHxNYU2EVnvMq9pj/7qosdGnHN7gb8HvhK79vfAM865+4Bngb+LXf874EfOud1Ez86cPx2lFviac24XMAT84gr/PCKSoHQigoisa2Y25pzLuc71duBjzrkLZpYKdDnnisysDwg552Zi18POuaCZ9QIVzrmpRa9RDbzqnKuN3f+PQKpz7i9W/icTkUSjkTYRkRtzN7h9o+dcz9Si23NoLrGI3CGFNhGRG/vVRX8ei91+E3gidvvzwI9jt78H/A6AmSWbWe5qFSki64N+4xOR9S7TzE4uuv+yc25+2490M2sk+gvuZ2PXvgh83cz+GOgFfit2/UvAU2b220RH1H4HCK949SKybmhOm4jIdcTmtD3knOvzuhYREVB7VERERMQXNNImIiIi4gMaaRMRERHxAYU2ERERER9QaBMRERHxAYU2ERERER9QaBMRERHxgX8FempxbOnNyhIAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 720x1080 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_history(hist9)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_unbeer_keras(df, userid):\n",
    "    beers = df[df['user']==userid]['beer'].unique().tolist() #해당 유저가 평가한 beer\n",
    "    total_beers = df['beer'].unique().tolist() #모든 beer\n",
    "    unbeers = [beer for beer in total_beers if beer not in beers] #해당 유저가 평가안한 beer\n",
    "    return unbeers\n",
    "\n",
    "def recomm_beer_by_keras(userid, model, top_n):\n",
    "    unbeers = get_unbeer_keras(df,userid)\n",
    "    predict =[v[0][0] for v in  [model.predict([[userid],[beer]]) for beer in unbeers]]\n",
    "    ndf= pd.DataFrame({'beer':unbeers,'rating':predict})\n",
    "    final = ndf.sort_values(by='rating', ascending=False)[:top_n]\n",
    "    return final\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Wall time: 5min 9s\n"
     ]
    },
    {
     "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>beer</th>\n",
       "      <th>rating</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>186538</td>\n",
       "      <td>68022</td>\n",
       "      <td>4.930120</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>73822</td>\n",
       "      <td>185846</td>\n",
       "      <td>4.874308</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>210211</td>\n",
       "      <td>122288</td>\n",
       "      <td>4.844587</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>205238</td>\n",
       "      <td>192506</td>\n",
       "      <td>4.801634</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>109872</td>\n",
       "      <td>19011</td>\n",
       "      <td>4.779156</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>85480</td>\n",
       "      <td>86880</td>\n",
       "      <td>4.776610</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>138926</td>\n",
       "      <td>147449</td>\n",
       "      <td>4.762318</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>22538</td>\n",
       "      <td>104253</td>\n",
       "      <td>4.755158</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>191828</td>\n",
       "      <td>52574</td>\n",
       "      <td>4.749087</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>237306</td>\n",
       "      <td>65713</td>\n",
       "      <td>4.742968</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "          beer    rating\n",
       "186538   68022  4.930120\n",
       "73822   185846  4.874308\n",
       "210211  122288  4.844587\n",
       "205238  192506  4.801634\n",
       "109872   19011  4.779156\n",
       "85480    86880  4.776610\n",
       "138926  147449  4.762318\n",
       "22538   104253  4.755158\n",
       "191828   52574  4.749087\n",
       "237306   65713  4.742968"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "unbeers = get_unbeer_keras(df2, 190)\n",
    "recomm_beer_by_keras(190, re9model, 10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(309542, 50)"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "## get weights of the beers embedding matrix\n",
    "beer_embedding_weights = re9model.layers[3].get_weights()[0] #3은 beer-embedding : layer 확인!\n",
    "beer_embedding_weights.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(3, 309542)"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "## import PCA\n",
    "pca = PCA(n_components = 3) ## use 3 components\n",
    "beer_embedding_weights_t = np.transpose(beer_embedding_weights) ## pass the transpose of the embedding matrix\n",
    "beer_pca = pca.fit(beer_embedding_weights_t) ## fit\n",
    "\n",
    "## display the resulting matrix dimensions\n",
    "beer_pca.components_.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([0.5108015 , 0.17140232, 0.05021592], dtype=float32)"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "## display the variance explained by the 3 components\n",
    "beer_pca.explained_variance_ratio_\n",
    " #variance 가 아주 낮으면 좋은 해석을 기대하기 어렵다.\n",
    "    #첫번째 PC 분산은 약 얼마? 78%로, dropout 규제를 넣었는데도 성능이 좋다!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [],
   "source": [
    "from operator import itemgetter\n",
    "\n",
    "## extract first PCA\n",
    "pca0 = beer_pca.components_[0]\n",
    "\n",
    "## get the value (pca0, beer title)\n",
    "beer_comp0 = [(f, beers_dict[i]) for f,i in zip(pca0, list(beers_dict.keys()))]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[(0.038865563, 'Jolly Golden Ale'),\n",
       " (0.03691411, 'G-Spot Jalapeno Ale'),\n",
       " (0.03661457, 'One-Eyed Jack Imperial Pilsner'),\n",
       " (0.034107786, 'Handlebar Hefe'),\n",
       " (0.033302944, 'Winter Sesh'),\n",
       " (0.0328734, 'Pecan Porter'),\n",
       " (0.032648373, 'Full Rosenberry'),\n",
       " (0.031916, 'La Grand Duc (Rousse Au Seigle)'),\n",
       " (0.03170663, \"'Sconnie Badger Brown Ale\"),\n",
       " (0.031124333, 'Cimmerian Schwarzbier')]"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "## beers corresponding to the highest values of pca0\n",
    "sorted(beer_comp0, key = itemgetter(0), reverse = True)[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[(-0.007946341, 'Andre the Brut'),\n",
       " (-0.007806833, 'Stout Impérial Russe Cognac'),\n",
       " (-0.0076717073, 'Mele Kalikimaka Coconut Porter'),\n",
       " (-0.0073953057, 'Oryza IPA'),\n",
       " (-0.0073629706, 'Into Your Arms (Bad Seed/Revolutions)'),\n",
       " (-0.0072802445, 'Dirty Bastard - Bourbon-Soaked Apricots And Ginger'),\n",
       " (-0.006972938, 'Your Highnessness'),\n",
       " (-0.006961703, 'Masham Glory'),\n",
       " (-0.0066309697, 'Farmhand Ale'),\n",
       " (-0.006254611, 'Grid Iron Irish Red')]"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "## beers corresponding to the lowest values of pca0\n",
    "sorted(beer_comp0, key = itemgetter(0))[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}