{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# 리뷰 데이터를 통해 제품 성공 예측하기 \n", "\n", "_**매출이 \"Hit\"임계치를 초과할지를 예측하기 위해 XGBoost 이용하기 [(원본)](https://github.com/jihys/sagemaker-workshop-0809/blob/master/module2-video-game-sales-xgboost.ipynb)**_\n", "\n", "---\n", "\n", "## 목차\n", "\n", "1. [배경](#배경)\n", "1. [설정](#설정)\n", "1. [데이터](#데이터)\n", "1. [훈련](#훈련)\n", "1. [호스트](#호스트)\n", "1. [평가](#평가)\n", "1. [확장](#확장)\n", "1. [정리](#정리)\n", "\n", "\n", "## 배경\n", "\n", "사용자 리뷰, 비평 리뷰, 소셜 미디어 댓글 등의 입소문은 종종 제품이 궁극적으로 성공할것인지에 대한 인사이트를 제공할 수 있습니다. 특히 비디오 게임 업계에서는 리뷰와 평가가 게임의 성공에 큰 영향을 줄 수 있습니다. 그러나 나쁜 리뷰의 게임이 모두 실패하는 것이 아니고 또한 좋은 리뷰의 게임이 hit가 되는 것 또한 아닙니다. hit게임을 예측하기 위해서는 기계학습 알고리즘은 리뷰와 함께 잠재적인 다양한 관련 데이터의 특성을 이용할 수 있습니다. \n", "\n", "이 노트북에서는 Kaggle의 [Video Game Sales with Ratings](https://www.kaggle.com/rush4ratio/video-game-sales-with-ratings) 데이터셋을 사용할 것입니다. 이 데이터셋은 [Metacritic](http://www.metacritic.com/browse/games/release-date/available)와 user review, 뿐만 아니라 critic review와 ESRB ratings과 다른 것들을 포함하고 있습니다. user review와 critic review는 rating score 형식으로서 0에서 10, 또는 0에서 100까지 척도로 되어있습니다. 이것들은 편리하지만 데이터셋의 중대한 문제는 상대적으로 작다는 것입니다.\n", "\n", "이와 같이 작은 데이터셋을 다루는 것은 기계학습에서 공통적인 문제입니다. 이 문제는 종종 작은 데이터셋의 클래스간의 불균형으로 인해 악화되기도 합니다. 이러한 상황해서 앙상블 학습을 사용하는 것은 좋은 선택입니다. 이 노트북에서는 앙상블 Learner인 XGBoost를 사용하여, 게임의 hit 여부를 판단하기 위한 분류기를 만드는데 초점을 맞출 것입니다. \n", "\n", "\n", "## 설정\n", "\n", "다음 설정부터 시작합니다. \n", "\n", "- 실습을 위해 필요한 다양한 Python 라이브러리들을 import 합니다 \n", "- 노트북 내의 다양한 작업을 위해 SageMaker의 세션객체를 생성하고 AWS 리전 정보를 가져옵니다. \n", "- 학습과 모델 데이터 사용을 위한 S3 버킷과 Prefix를 설정합니다. \n", "- S3 데이터 접근을 위한 정의한 IAM role을 SageMaker 노트북 인스턴스에서 가져옵니다. \n" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [], "source": [ "import timeit\n", "start_time = timeit.default_timer()" ] }, { "cell_type": "code", "execution_count": 34, "metadata": { "isConfigCell": true }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Bucket: \n", "role: arn:aws:iam::415373942856:role/service-role/AmazonSageMaker-ExecutionRole-20191024T194435\n" ] } ], "source": [ "import numpy as np \n", "import pandas as pd \n", "import matplotlib.pyplot as plt \n", "from IPython.display import Image \n", "from IPython.display import display \n", "from sklearn.datasets import dump_svmlight_file \n", "from time import gmtime, strftime \n", "import sys \n", "import math \n", "import json\n", "import boto3\n", "import sagemaker\n", "\n", "session = sagemaker.Session()\n", "region = session.boto_region_name\n", "\n", "#bucket = session.default_bucket()\n", "bucket=''\n", "prefix = 'sagemaker/videogames-xgboost'\n", "role = sagemaker.get_execution_role()\n", "\n", "print('Bucket: {}'.format(bucket))\n", "print('role: {}'.format(role))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "---\n", "## 데이터\n", "\n", "먼저 public S3 버킷에서 이 노트북 인스턴스로 데이터셋을 다운로드합니다. 이 데이터는 이 노트북과 같은 디렉토리에 나타날 것입니다. \n", "우리는 그 데이터를 먼저 먼저 살펴보겠습니다. " ] }, { "cell_type": "code", "execution_count": 8, "metadata": {}, "outputs": [ { "data": { "text/html": [ "

\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "

	Name	Platform	Year_of_Release	Genre	Publisher	NA_Sales	EU_Sales	JP_Sales	Other_Sales	Global_Sales	Critic_Score	Critic_Count	User_Score	User_Count	Developer	Rating
0	Wii Sports	Wii	2006.0	Sports	Nintendo	41.36	28.96	3.77	8.45	82.53	76.0	51.0	8	322.0	Nintendo	E
1	Super Mario Bros.	NES	1985.0	Platform	Nintendo	29.08	3.58	6.81	0.77	40.24	NaN	NaN	NaN	NaN	NaN	NaN
2	Mario Kart Wii	Wii	2008.0	Racing	Nintendo	15.68	12.76	3.79	3.29	35.52	82.0	73.0	8.3	709.0	Nintendo	E
3	Wii Sports Resort	Wii	2009.0	Sports	Nintendo	15.61	10.93	3.28	2.95	32.77	80.0	73.0	8	192.0	Nintendo	E
4	Pokemon Red/Pokemon Blue	GB	1996.0	Role-Playing	Nintendo	11.27	8.89	10.22	1.00	31.37	NaN	NaN	NaN	NaN	NaN	NaN
5	Tetris	GB	1989.0	Puzzle	Nintendo	23.20	2.26	4.22	0.58	30.26	NaN	NaN	NaN	NaN	NaN	NaN
6	New Super Mario Bros.	DS	2006.0	Platform	Nintendo	11.28	9.14	6.50	2.88	29.80	89.0	65.0	8.5	431.0	Nintendo	E
7	Wii Play	Wii	2006.0	Misc	Nintendo	13.96	9.18	2.93	2.84	28.92	58.0	41.0	6.6	129.0	Nintendo	E
8	New Super Mario Bros. Wii	Wii	2009.0	Platform	Nintendo	14.44	6.94	4.70	2.24	28.32	87.0	80.0	8.4	594.0	Nintendo	E
9	Duck Hunt	NES	1984.0	Shooter	Nintendo	26.93	0.63	0.28	0.47	28.31	NaN	NaN	NaN	NaN	NaN	NaN
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
16709	15 Days	PC	2009.0	Adventure	DTP Entertainment	0.00	0.01	0.00	0.00	0.01	63.0	6.0	5.8	8.0	DTP Entertainment	NaN
16710	Men in Black II: Alien Escape	GC	2003.0	Shooter	Infogrames	0.01	0.00	0.00	0.00	0.01	NaN	NaN	tbd	NaN	Atari	T
16711	Aiyoku no Eustia	PSV	2014.0	Misc	dramatic create	0.00	0.00	0.01	0.00	0.01	NaN	NaN	NaN	NaN	NaN	NaN
16712	Woody Woodpecker in Crazy Castle 5	GBA	2002.0	Platform	Kemco	0.01	0.00	0.00	0.00	0.01	NaN	NaN	NaN	NaN	NaN	NaN
16713	SCORE International Baja 1000: The Official Game	PS2	2008.0	Racing	Activision	0.00	0.00	0.00	0.00	0.01	NaN	NaN	NaN	NaN	NaN	NaN
16714	Samurai Warriors: Sanada Maru	PS3	2016.0	Action	Tecmo Koei	0.00	0.00	0.01	0.00	0.01	NaN	NaN	NaN	NaN	NaN	NaN
16715	LMA Manager 2007	X360	2006.0	Sports	Codemasters	0.00	0.01	0.00	0.00	0.01	NaN	NaN	NaN	NaN	NaN	NaN
16716	Haitaka no Psychedelica	PSV	2016.0	Adventure	Idea Factory	0.00	0.00	0.01	0.00	0.01	NaN	NaN	NaN	NaN	NaN	NaN
16717	Spirits & Spells	GBA	2003.0	Platform	Wanadoo	0.01	0.00	0.00	0.00	0.01	NaN	NaN	NaN	NaN	NaN	NaN
16718	Winning Post 8 2016	PSV	2016.0	Simulation	Tecmo Koei	0.00	0.00	0.01	0.00	0.01	NaN	NaN	NaN	NaN	NaN	NaN

\n", "

16719 rows × 16 columns

\n", "

" ], "text/plain": [ " Name Platform \\\n", "0 Wii Sports Wii \n", "1 Super Mario Bros. NES \n", "2 Mario Kart Wii Wii \n", "3 Wii Sports Resort Wii \n", "4 Pokemon Red/Pokemon Blue GB \n", "5 Tetris GB \n", "6 New Super Mario Bros. DS \n", "7 Wii Play Wii \n", "8 New Super Mario Bros. Wii Wii \n", "9 Duck Hunt NES \n", "... ... ... \n", "16709 15 Days PC \n", "16710 Men in Black II: Alien Escape GC \n", "16711 Aiyoku no Eustia PSV \n", "16712 Woody Woodpecker in Crazy Castle 5 GBA \n", "16713 SCORE International Baja 1000: The Official Game PS2 \n", "16714 Samurai Warriors: Sanada Maru PS3 \n", "16715 LMA Manager 2007 X360 \n", "16716 Haitaka no Psychedelica PSV \n", "16717 Spirits & Spells GBA \n", "16718 Winning Post 8 2016 PSV \n", "\n", " Year_of_Release Genre Publisher NA_Sales EU_Sales \\\n", "0 2006.0 Sports Nintendo 41.36 28.96 \n", "1 1985.0 Platform Nintendo 29.08 3.58 \n", "2 2008.0 Racing Nintendo 15.68 12.76 \n", "3 2009.0 Sports Nintendo 15.61 10.93 \n", "4 1996.0 Role-Playing Nintendo 11.27 8.89 \n", "5 1989.0 Puzzle Nintendo 23.20 2.26 \n", "6 2006.0 Platform Nintendo 11.28 9.14 \n", "7 2006.0 Misc Nintendo 13.96 9.18 \n", "8 2009.0 Platform Nintendo 14.44 6.94 \n", "9 1984.0 Shooter Nintendo 26.93 0.63 \n", "... ... ... ... ... ... \n", "16709 2009.0 Adventure DTP Entertainment 0.00 0.01 \n", "16710 2003.0 Shooter Infogrames 0.01 0.00 \n", "16711 2014.0 Misc dramatic create 0.00 0.00 \n", "16712 2002.0 Platform Kemco 0.01 0.00 \n", "16713 2008.0 Racing Activision 0.00 0.00 \n", "16714 2016.0 Action Tecmo Koei 0.00 0.00 \n", "16715 2006.0 Sports Codemasters 0.00 0.01 \n", "16716 2016.0 Adventure Idea Factory 0.00 0.00 \n", "16717 2003.0 Platform Wanadoo 0.01 0.00 \n", "16718 2016.0 Simulation Tecmo Koei 0.00 0.00 \n", "\n", " JP_Sales Other_Sales Global_Sales Critic_Score Critic_Count \\\n", "0 3.77 8.45 82.53 76.0 51.0 \n", "1 6.81 0.77 40.24 NaN NaN \n", "2 3.79 3.29 35.52 82.0 73.0 \n", "3 3.28 2.95 32.77 80.0 73.0 \n", "4 10.22 1.00 31.37 NaN NaN \n", "5 4.22 0.58 30.26 NaN NaN \n", "6 6.50 2.88 29.80 89.0 65.0 \n", "7 2.93 2.84 28.92 58.0 41.0 \n", "8 4.70 2.24 28.32 87.0 80.0 \n", "9 0.28 0.47 28.31 NaN NaN \n", "... ... ... ... ... ... \n", "16709 0.00 0.00 0.01 63.0 6.0 \n", "16710 0.00 0.00 0.01 NaN NaN \n", "16711 0.01 0.00 0.01 NaN NaN \n", "16712 0.00 0.00 0.01 NaN NaN \n", "16713 0.00 0.00 0.01 NaN NaN \n", "16714 0.01 0.00 0.01 NaN NaN \n", "16715 0.00 0.00 0.01 NaN NaN \n", "16716 0.01 0.00 0.01 NaN NaN \n", "16717 0.00 0.00 0.01 NaN NaN \n", "16718 0.01 0.00 0.01 NaN NaN \n", "\n", " User_Score User_Count Developer Rating \n", "0 8 322.0 Nintendo E \n", "1 NaN NaN NaN NaN \n", "2 8.3 709.0 Nintendo E \n", "3 8 192.0 Nintendo E \n", "4 NaN NaN NaN NaN \n", "5 NaN NaN NaN NaN \n", "6 8.5 431.0 Nintendo E \n", "7 6.6 129.0 Nintendo E \n", "8 8.4 594.0 Nintendo E \n", "9 NaN NaN NaN NaN \n", "... ... ... ... ... \n", "16709 5.8 8.0 DTP Entertainment NaN \n", "16710 tbd NaN Atari T \n", "16711 NaN NaN NaN NaN \n", "16712 NaN NaN NaN NaN \n", "16713 NaN NaN NaN NaN \n", "16714 NaN NaN NaN NaN \n", "16715 NaN NaN NaN NaN \n", "16716 NaN NaN NaN NaN \n", "16717 NaN NaN NaN NaN \n", "16718 NaN NaN NaN NaN \n", "\n", "[16719 rows x 16 columns]" ] }, "execution_count": 8, "metadata": {}, "output_type": "execute_result" } ], "source": [ "raw_data_filename = 'Video_Games_Sales_as_at_22_Dec_2016.csv'\n", "data_bucket = 'sagemaker-workshop-pdx'\n", "\n", "s3 = boto3.resource('s3')\n", "s3.Bucket(data_bucket).download_file(raw_data_filename, 'raw_data.csv')\n", "\n", "data = pd.read_csv('./raw_data.csv')\n", "pd.set_option('display.max_rows', 20) \n", "data" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "계속 진행하기 전에, 예측할 타켓을 지정해야 합니다. 비디오게임개발 에산은 수천만달러에 이르기 때문에, 비용을 회수하고 수익을 얻기위해 게임 퍼블리셔가 \"hit\"게임을 퍼블리싱하는 것은 매우 중요합니다. \"hit\"게임은 글로벌 매출이 100만 이상인 것을 타켓으로 지정할 것입니다. " ] }, { "cell_type": "code", "execution_count": 12, "metadata": {}, "outputs": [], "source": [ "data['y'] = (data['Global_Sales'] > 1)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "이제 목표가 정의되었으므로 \"hit\"와 \"not a hit\" 클래스의 불균형을 살펴보도록 하겠습니다.: " ] }, { "cell_type": "code", "execution_count": 10, "metadata": {}, "outputs": [ { "data": { "image/png": "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\n", "text/plain": [ "

" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "plt.bar(['not a hit', 'hit'], data['y'].value_counts())\n", "plt.show()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "놀라울 것도 없이, 단지 일부의 게임만이 \"hits\"로 간주되어집니다. 다음으로 타켓에 예측력이 높은 피처를 선택할 것입니다. 리뷰 스코어들과 글로벌 매출을 플로팅을 시작하여 이러한 스코어들이 매출에 영향을 미치는지 확인합니다. 명확성을 위해 로그스케일을 사용합니다. " ] }, { "cell_type": "code", "execution_count": 17, "metadata": {}, "outputs": [ { "data": { "image/png": "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\n", "text/plain": [ "

" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "image/png": "iVBORw0KGgoAAAANSUhEUgAAAY8AAAEPCAYAAAC6Kkg/AAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDMuMC4zLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvnQurowAAIABJREFUeJztnXmcFPWZ/z9PVR9zAMMw3MxwmAFcUEAdRQPy80iMUdTsiiYa1/2tiSa7McnGRE3WNV6/HEbNYTSb4LEJ0dUQkhVBchl0EYLHqMMZxAkoMyDXMAzM1Uf19/dHdzXV1VXdVd1VXdU9z/v18iVdU/X9PlU9833q+5wkhADDMAzD2EHyWgCGYRim/GDlwTAMw9iGlQfDMAxjG1YeDMMwjG1YeTAMwzC2YeXBMAzD2IaVB8MwDGMbVh4MwzCMbVh5MAzDMLZh5cEwDMPYJuC1AG4xevRoMXXqVK/FYBiGKRvefPPNw0KIMVbOrVjlMXXqVLS2tnotBsMwTNlARO9bPbcslAcRfQLApQBGAHhCCPFHj0ViGIYZ0njm8yCiJ4noIBFt1R2/mIjeIaJ2Ivo6AAghnhNC3Ajg8wA+6YW8DMMwzAm8dJj/HMDF2gNEJAN4FMDHAcwCcA0RzdKc8h+pnzMMwzAe4pnyEEKsA3BEd/gsAO1CiF1CiCiAZwFcQUnuB/A7IcRbpZaVYRiGycRvobqTAHRoPnemjn0RwEcALCGiz5tdTEQ3EVErEbUeOnTIXUkZhmGGMGXhMBdCPAzgYQvnLQWwFABaWlq4RSLDMJ7R1RtBZ/cAGuur0TAs7LU4juM35bEXQJPmc2PqmGWI6DIAlzU3NzspF8MwjGVWtu3F7b/ZjKAkIZZI4HtXzsHl8yZ5LZaj+M1s9QaA6UQ0jYhCAD4F4Hk7AwghVgkhbqqrq3NFQIZhmFx09UZw+282YzCWwPFIHIOxBG77zWZ09Ua8Fs1RvAzVfQbARgAziaiTiD4jhIgDuBnAHwD8FcByIcQ2r2RkGIaxS2f3AIJS5tIalCR0dg94JJE7eGa2EkJcY3J8DYA1hY7LZiuGYbyksb4asUQi41gskUBjfbVHErmD38xWRcNmK4ZhvKRhWBjfu3IOqoIShocDqApK+N6VcyrOae43hznDMEzZc/m8SVjQPJqjrcoJNlsxDOMHGoaFK1JpqLDZimEYhrFNxSkPhmEYxn0qTnkQ0WVEtLSnp8drURiGYSqWilMebLZiGIZxn4pTHgzDMIz7sPJgGIZhbFNxyoN9HgzDMO5TccqDfR4MwzDuU3HKg2EYhnEfVh4MwzCMbVh5MAzDlIiu3gg2dRytiN4eXNuKYRimBFRad8GK23mww5xhGD+g3WVUYnfBitt5MAzDeI1+l/GF85oRlCQM4kSTKLW7YLlW3mXlwTAM4yDaXYaqLB55qR2AyDgvqijoGYiiqzdSlgqk4sxWDMMwuXDbaW3UwzwkS7j5/OmoCkqoDcmQCVASAl94+m0suH8tnm/b64osbsI7D4ZhhgylcFqb9TC/dv5kjKoN4Z5V26EIAAI4HokDAG77zWYsaB5dVjuQitt5cHkShmGMKJXT2qyHOQDc98J2RJVE1jWq/6OcqLidhxBiFYBVLS0tN3otC8Mw/kE1J5XCaW3Uw3xTx9Gs+VViiQQa66sdlcFtKk55MAzDGGFmTnJr0db3MDeaHwDCAcL3rpxTViYroALNVgzDMEaYmZOcWLStOOH184cDEv7l/5yEx65vwYLm0UXLUGpICJH/rDKkpaVFtLa2ei0GwzAe0dUbyTAb5Tte6Jh2nfDqGFv39uC+F7b7KuOciN4UQrRYOZfNVgzDVBy5FnS9OamYMRc0j87K6TCKnDJSOp9cutH0umIUXKlg5cEwTEVhlKRXbCis2ZhL/7ElywkuE+GlHQdx/slj0TAsbKh0pjTUmjrv17cfLosaWOzzYBimojBK0is2FNZsTEBkOcH7ogruXrUNC+5fi6dfe98wPLg2JBs672tDctnUwGLlwTBMReFGVJXZmLMn1qWd4LUhOf2z3oiCwVgC96zaDpko47qgJKEvqhg67/uiiuOKzy0qzmzFJdkZZmijRjXdpjP9FOM7yDWmmtPx0o6DuHvVNvRGlPR1QZkQixsrsrlNI7NyQbp6IyUNJy4GjrZiGKYiccPpnGvMrt4IFty/FoOxE4t/OEC4YcE0PLlhN0KybMmH8Xzb3iwlVSqfh51oK1YeDMMwDqFd+AdicRARqgIyokoCN5/fjGvnTzZUZHql5FW0FSsPsPJgGKY4Cl3Au3oj2LbvGG5c1oqIxmRVFZSw4fYLACBjXD91GOQ8D4Ypc8ohzt/vFPMMi1nQG4aFUVcdREiWMpRHUJLw9Gt78JOX29Pj3rl4Fu5bvd3RsOJSwcqDYXyGn95E7VDIYu2WkizmGTqRJ2IUnRVVFDz6Ujsi8RPj3rNqO4JSdjRWOXQY5FBdhvEAs1pI5drremXbXiy4fy2ue/w1y82NCrnGCsU+QyfyRIzqaN18/nSE5MxxJQKicSXjmF+jq/TwzoNhSkyut+JSlg13ikLe1N3IAlfp7B6ASGT6chUlgW37jmHRjDF5r3cqT0Rblr02JGNfzwAi8XjGOYOxBIIyISAJVAcDSVPWpbPSisqv3znAOw+GKSn53opLXTbcCQp5U3cjC1ylNiQjomQqj1gCuHFZq6XdTaHVd412kw3Dwnivqw+LH1mPG3/RiqiSfV1MEZCI8NWLZuCWj8zAfS9sd3w35ga882CYEpJvZ+FGgpvbFKLw3FSSfVEFVUEpI98CACLxhOXdjVEzp1yY7Sa1Lwu5iCoC9//+HQzEktqlHJznrDwYpoRYWTTtLlxeU4jCc1NJ5lJAdkyAVqvvGpngvvbrTZg1YUS63IhR90A9quLQIhLCtybLslAeRHQSgDsA1AkhlngtD8MUitVFs9Cy4V5RiMJzS0mqz/hrv27LMhM5tbtpP3AcbR1HMa9ppKGCiCoCl/x4Pe66bJZh90CrRBSRUTPLT3imPIjoSQCLARwUQpyiOX4xgB8BkAE8LoT4rhBiF4DPENEKb6RlKp1S5lWU287CKoUoPLeUpABAJCEkA1ElgbBMIKmwdq/6341vPrcFy17dk/751S2NhgoiGk/gvtXbceels3DfC9uRUBKI2tQjVcFkEUU/4uXO4+cAHgGwTD1ARDKARwF8FEAngDeI6HkhxHZPJGSGBF7kVZTbzqKcUM1I2gQ9QYQXbl6I5nHDLV2vKgttb42oouCaMydnKA4AWN7aiSvmTsDKTR9kjSUSAqdMqsOG2y/ASzsO4q7nt9lSBkIkdx6bOo767kXDM+UhhFhHRFN1h88C0J7aaYCIngVwBQBWHowruBkyyniDUVBCWLb2Bq99kYgqCSiJBOKJEw7sn2983/C6NVv3Gx5XzU4Nw8I4/+Sx+I+VW23dSzwBXPyjdekwXj8ljPotVHcSgA7N504Ak4iogYh+CuA0IvqG2cVEdBMRtRJR66FDh9yWlakAjHICVCflUMUsgbFcKDSSSx9GHYknFYcV9Ml/KlqzU8OwMK4+ozHj55eeMh5Xt2Qe0yacKwmBeAK+TBgtC4e5EKILwOctnLcUwFIgWRjRbbmY8scoJ8DPTkq3KdfSKFpUh/mtKzZBJgmKsBbJZbRjscLVLZPw27fM8zEa66tTxRJ78PRrmSav32/bj4BO8SRyrFx+Shj1m/LYC6BJ87kxdcwy3AyKsYNRToCfnZRuUkkmvOT6SwABEJm1o8yCI4x2LEGZQBCGyX1Xnj4J15zZhP6YYqg8woGkg171mxAIuvcUKAII2Khs7qeEUb8pjzcATCeiaUgqjU8BuNbOAEKIVQBWtbS03OiCfEyFYfaH6Jc/0FLiVmkUJyLZ7Ixh5DBXlaDWAa7fWZmFUS9oHo3HX9mF//zfXRnzrGzbi1Wb9oGIssxbNSEZP73udMyeWJfVIEqPfucrUVJphWQZA7E4hACqgjIUIXyVMOplqO4zAM4DMJqIOgHcJYR4gohuBvAHJEN1nxRCbPNKRqbyKceMbrdwI+vbCTOY3THMlODGv3Xh1l9vQlQRpjsrszDqi0+ZgF+++n5Gi9mkwhBQ9zlaEkJgYl01XtpxMKuHeT4kAu65bDZef68bqzbtRTCQdN7ffflsX5kQK64ZlMZsdeO7777rtThMmcD9M5I42QLVqC2r2hDJTsl2u2MYXROQAIkIUd1b/vBwAE99dj7mNo3MK8f8b7+Y14FeHZAgCLj6jEYsf7MTAYkyFE4xBCTgtX//iKu/n0O6GRSbrZhC4LyLJE4mMDphBitkDP1uMqooSAhkKQ7AeR/CDQun4u9Pa8TiR9ZnmapqQzLiiQQ++ndjsWbLAZtu+eROx2pl4FJQccqDYRh/4IQZrNAxtEqw40gfbluxBTElcwcQkjMzznP1EX/6tT2WwnZ/tm4XEgJZCq8mJOGKeRPxqzc6sHrLAYt3n83e7v6Cr3WailMeHG3FMIXjZKiuE/6kQsNu1WvXtx/GbSsynecAEApIWPPFExnn+vtWzU5qsmDUYsJHPAH85//ugqxzc/RHE/jVGx1Z0VZ2OdIX8U22ecX5PFRaWlpEa2ur12IwTNnghI/CbNxizGAr2/bithWbIUsEJSHwwBJrCs3ofoBkCO0DS+amx2g/cByXPPyKoVmrUFRZnSYckBCS3cvBsePz8FuGOcMwHuFWg6aGYWHMbRpZkOLQht32R5V0Tw6jLGt9ZrzR/dSEZDx2fUt60V3ZtheX/Hi9o4oDAAIEhF3IM43E/dOemM1WDMMA8L6LodEOxarD3MjctqB5dNb9qCG0mzqOIhZX0qG7TqPP3XADr7PNK27nIYRYJYS4qa6uzmtRGMZX5KtZVWj7VSdY2bYXC+5fm9V+1YpCM2vtCyDjfsIBwgUzx+DSH7+CTy7diCU/e9VQcQSlpNnJ73idbV5xOw+GYbKx6gj3otdIvrIo+ZzuuYpbqvfz9Gt78OhL7VizVY10Mt4ZhAISvn/VXHx1eRv8WqCmNuSPbHNWHgxT4ditWVXqnJdci3/DsHBehWaluOVPXm7PirjSE5IJDy6Zg5PHDy+J2akQgjLhP687A7MnjvA82qrizFZEdBkRLe3p6fFaFIYpGCfLorvlCNdTqMxWFv9cTne1uKUWbXHLbfuOQUJuM1QoIGHNl87F5fMmGY7nF6oCMuqqg54rDqACdx6cYc6UO06XRbfqN8hnqsp1jl7mOy+dhVMm1VkyfRVb2djI7p9IJJWPGuZrtuvQtqdVcz78XBQzEld80y6g4pQHw5QzbpRFz+c3sKKscp1jJPMdz21FdVBCQoiMnAojnKhs/IXzmvHjtTshBBBLAJJEuOThV6AkhGFi3qWnjMdXPjoDfVElS8GpTZv07Wb9ABFh8SPrfdFnhZUHw/iIQmo5Wdk1XD5vEmZNGIG2jqOY1zQy/ZZtRVnlO8esidJAaidxy/K2nMqvmEx0VamJRGbPjVwl0AHgz+8cxFc+OiPjmPocY3EFz77RYXKlt6g7KD/0WWHlwTA+wm6uhVUTl9l5VpRVvnOMZNZipaBfIVFeWqVmF5EQuOTH6xGWT5Qj+VVrB4RJAUW/4XWOB8AOc4bxFXZyLczyG/QOa6Pzbl2xCet2HkRtSM6rrPIpNK3MVQGzJSX/gmzkFNc74bWfjQIBrBJRBKKabO1lr+5BJC7KQnEA3ud4ABW482CHOVPuWH0Lt2riMjovEhf4/FNvISEErm5pxPLWTlOTkRWzkirztn3H8JlfvIGYZhEOyoTZE60n7aqKYeveHtz3wvYTxQp1ct65eFbOHY8RYZkgkPSJFLJj8QNqe1uvI664MCLDlClWCxmaFQjUXrP65oWGzmOtPwWA4b/1i9jzbXtxawGFDIET5jUrTZRCMuFrF83E91/cCZEQlnIzAjLhZ58+HZ976k1LJdZlAohg6dxS8C//5yR89tyTXFMcrjSDIqLvAfh/AAYA/B7AHABfEUI8VZCUDMMUhZUdgbr433npLNz3wnZIIPTHMhfloJQMi9V30zPzk+TzsxSapd5+4LitWlNRReDBP+3EXZfNwikT6xCLK7j2iddzlk+PK8kdl2JBGQQlQBGwdG4pIAB/N8H75EAVyzsPImoTQswjor8HsBjALQDWCSHmuilgofDOgxkqmEVbZeVeLJ6Fpvoa3LisNSPvQb/zAIBt+3pS54ms8/Rd8oop2641Ud2zervlvhlatPOrbXSt7kTMCEgEImSY3/yATMDrd7jXitatNrTquZcC+LUQoodsNnZnGMZ5jMqJGIXX3rd6OzbcfgEeWJK5W7n6jGTb1KAkYTCuQAiBUEDOUBxAcofS1nG06NayKlZNVFUBCQmRQEwxdrurpUwAYEpDLZ664Sxc+8TrsOKkNyPuQi8OJ1AEsPFvXVg8d6LXothSHquJaAeSZqt/IaIxAAbdEatwuCQ7Uw4U2yApH7mc6VqzUm1ITu8ktOfGDbK7Y4kE5jWNdKRsu5UwWwmALBMkiRCPm6uCiCLw6q4u/ODFnQhKEiJKAlShvlwgmTeTEMLzJEHLcW5CiK8D+DCAFiFEDEA/gCvcEqxQuCQ743fMyo87iZXw2rlNI9EXVfKGu9YE5XTIcPO44ZZCifPVubISZptA0mzUH1VyOqxlAh760850KHI0nvBtYUMniCrC80ZQgD2HeQ2AfwUwGcBNACYCmAlgtTuiMUxloI9Ycrr8iBFWs7bzJfjJBPz0HzOruOZziFtJXMw3rx0UAVRJQNSR0coDPyQJ2jFb/ReAN5HcfQDAXgC/BisPZohjp2DgF85rdsxnkA8rUU9aJSMTZRUjlCQyLP9tVrbdam2ufPPaIRyQ0Bf1SUhUiSi3JMEPCSE+SUTXAIAQop/YY84McewWDHzkpXborfduLgRWenOoSualHQdx96ptGc7rqoBsS7HZqc2Va147JBzawZQDVTKQAOHOS2d5HrJrJ7c/SkTVSP3mE9GHAHhrdGMYD8lXHsTIrh+SJdx8/nRPWr3momFYGOefPDYrysiuYjMyR0UVBT0D0QwbffuB41jR2oHuvijOP3ksBmPWFcfMcbUZn/9uwgjL15Y7g0oyY/++F7a74iuzg52dx11IJgc2EdHTABYA+L9uCMUw5UAhBQNjiQSunT8Z186fnGFOcjv6ygrFVLfVjqEvZx5PAF94+u30eK3vHcn4+dUtjUgaMaw5ud850Jfxefu+Y5blqwRUE53XlXUtKw8hxJ+I6C0AZyOZ7PhlIcRh1yRjGJ9jtWCg2WJsp5+G05gpK7vZ4fpxunojWP5mZ8Y5SkLgeCQOAPiaQQb58tZO1ASlghPy4pUbWJUTmchTp3le5UFEp+sOfZD6/2QimiyEeMt5sRjG/9gpGGi2GLvR/Ckfye56myCTBEUkspo1af0kuWpbPf3aHjz60rsIyXK6e2A4IEHO4QqNmyiIqF9qgJQRMcVbp7mVncdDOX4mAFzgkCyOwEmCTD6cNBFZjWgqpjKuk/J29Ubw1eVtqbyJpJ/BrFmTdkc0EIuDiFAVkDEYV5DQdOiLxJO7ijue24rakJwzcspMRdz2sZNx/+93GHb9KxVBCSCSykaR3XXZbH+H6gohzi+FIE7BJdn9hx/s+SpumIisRDSZkc/0ZUVeO893275jWQl3Rs2ajHZEgEBMieccv5CQ20tOHYfacACyJEHxcOEWIul/GTs8hB/8ud0zOaxw+dwJ+PTZUzyVwVY/DyI6BcAsAFXqMSHEMqeFYioHL+z5ZrhlIjKy+1t1hucyfVmR1/7zNX61PzYQxaaOo2kZzVrLOsnpk0di694erNt5GGu2HDA85/K5E7Bmy/6S1JqKC+Dp1/cgJPs/A+H5TR9g/knv49PzvVMgdjLM7wJwHpLKYw2AjwNYD4CVB2OIF/b8XBTSHzwf+sVb37Do6jMasfzNzpyLu5npK5+8hTzf2RPrEJQpwzktEXDL8k0ISCd8IAuaRzuWAW7GW3uOAkiG8uqRCfjRp05D06gavLTjUNrhXgrKpZvgPau24+LZ4z3bzdvJ81gC4EIA+4UQ/wxgLgAuIMWYYpTnoC5+XmC3P3i++kxGeR7LNu7J/PzqnrxtYgHjFqz55C3k+TYMC+Ohq+YiHJBQE5LTb9lRRaA/piASF7hleRsApGtYVZu0lq0OSAgHJHz1ozPwrb8/BeEA5WhDaw9FAONHJMOde0uoOMqJoEye/S0B9sxWA0KIBBHFiWgEgIMAmlySi6kA7C7WVijGf6KaiG7VRBqZ5TFYMQcVYtqxs9PJF81V6PPV7nQ6jvTh5mfaMn6u+kDU81Zt2oe7V23PGuf2j5+My+ZORMOwMFa27QWQrIAblAlCCIQDuZ3n+bjm8ddw60UziyisXtkoCeFptJWd14RWIhoJ4DEka1y9BWCjK1IxFYG6+DmVTe1ENdrkQkTJTCUY27bzZY6rFFLcz67yvHzeJKy+eSHuumwWVt+8MCukttDnq+50RlQHDX++69BxdPVG0DAsjMvmTkRQ5wcIypRWHOrzisQT6I8qiCkCsiThX8/7kOHYVjcnMUXgO7/bYe3kIURVAL6oTGAnSfBfU//8KRH9HsAIIcRmd8RiKoVcoax2dhFO+E+0i5yK0RhWfSNGO4Msn4fus90/eLdavqoY+UAA4ME/7sR3f/9Oer6Hrpqb1Zc8l28mIBGODhjXub3x3JPwX395D0EpGRa75PRGrHirM+N7UeFdRzZxQbjtohme9/OwkiQ4BcBRIURP6vP5AD4B4H0i2iGEGEqVkJkCMApltRsl5ISz2+oYdsxBRov3ly+ckfOzVexUqC30DVT1gdy6YjMkAgZSzZnUIoXqfPr7BJCOzjJ6Xn1RBf/9WofhnBeePBZXnt6Ito6jmNc0EvW1IfyqdY/huUw2cUXg22uSO7KbFhnv7kqBlZ3HcgB/D6CHiOYhWYb9O0g6zH8C4LPuicdUIoXsIpzwn1gdw26NJ/3ine+zFbp6I3hpx8GsbG2nyrdrd325qttq51P/M1L8RuXVzfwdS372KmSJUBOUEVUU3LBwGszSOwhAKCAZ7kqGOt9eswO14YBn4bpWlEe1EGJf6t/XAXhSCPEQEUkA2nJcxzCGFJJV7VTRvlw5Fdr5ijUHFYO2t7d+AXaifLvZru/8k8fiP1ZuzThXXxF3274e3LZiEyJxkf7+vvbrTVjzpXOx4fYL8NKOg7jjf7bk7eSnrXf1ny/vMj0vIFsvmDgU8TJc14ry0L76XADgGwCQirxyRagsAYhqkdzlRAG8LIR4uiQTM65QaFa1Ewu60Rhm8xVjDioUs97etSEZihBFO0nNdn2zJoxAX1TBnZfOwn0vbE+XJIkrAjf+4k3EEwkQAaGAjIiuEmFUEbjkx+vx4JI5mNc00tkWsAmBGRPqsGVfj3NjVhABCZ4VR7SiPNYS0XIkCyLWA1gLAEQ0AUV0fiSiJwEsBnBQCHGK5vjFAH4EQAbwuBDiuwD+AcAKIcQqIvoVAFYeZUwxWdVOLOj6wn9+SGRUdz49A7GsXVltWMY9l83G+SePLbp8u9GuTySSi39YTn4Xdy6ehab6GvzTk69DACdKhgggbmKKisaTEWlL/7EFVUEpS/kVSkyAFUcO4op34bpWlMe/AfgkgAkAFgohYqnj4wHcUcTcPwfwCDQZ6kQkA3gUwEcBdAJ4g4ieB9AIYEvqtMIDxxnfUGhWtdPYnc+NOl3anU9USUDR7cqUhEgrjmLLvRjt+pI7BYFoyq9w76pt+MzCabaNRcmERTYxlRIviyNaKYwoADxrcPxt7Wci2iiEOMfqxEKIdUQ0VXf4LADtQohdqTGfBXAFkoqkEUkfizMprIznGO0iGuurMRDLzCgeiMVde7uy44h3sk6XqoRqQ3LWzicoE8IBpEud26l1lQ+jZk0BCRnFEiNxgcde2Z13rOzr4th1qBe3fGQGvv/iTkgg9NvoEMjY498vOdnT4oi2CiPmoSr/KXmZBEAb39cJYD6AhwE8QkSXAlhldjER3QTgJgCYPHmyA+IwXqDvKuemb82qI95J85ZWCUWURFYlWQnAQ1fNxWAsgXlNI9E8bjgAZ3ZlRs2ajAKZ8jVmCkjAly+cgUdfbkdQktAbiSOqAHev+iuAZEHDmeOG44E/7rQkF2OP2rCM+dMaPJXBSeXh2n5VCNEH4J8tnLcUwFIAaGlp4f1zGdLZPYCqgJxR+rsqILvqFLTiiHfKnGZc6jyTiCLw1V9vRkjO3OE4Ea7c2T0AoatQW0g8UzwBfPyU8bh2/mRs/FsXbn4mwxCB5zd9gJC83+aojFW8Lk0C+M8EtBeZ9bIaU8cYn5OviKBV3KiHZQWjwoRuyGVUzNCISDy7NIoT5V5qQ3JWNFQ+xSEbiBsOENo6klVxB01MU+VSnbbcCMrwvDQJ4OzOwwnbwhsAphPRNCSVxqcAXGtLCO4kWHKc9AU4kc/hBnbk0jvV2w8cT2dTF1IPS7vDsdLWNtcOqi+q2I6GMkrgi8QF7l61DXc8twXXnMkm4lJBAH73pUVpU6ansiT94Q4MRHSKEGJr/jPT5z+DZH+Q0QAOALhLCPEEEV0C4IdIhuo+KYT4ViHytLS0iNbW1kIuZWzQ1RvBgvvXZixGVUEJG26/wNEGS34hn1x6RXrmlHq80t6V/vn150xGy5RRGUqoZUo91mvOkQjQWpasPk+rXQf135cdqmTCoMGOQiaC4tBawphz5WkT8e+XznLtb4KI3hRCtFg6N5/yIKLjMN7ZEpLBWCPsi+gemp3Hje+++67X4pQNhS7WmzqO4rrHX8to1jM8HMBTn52PuU0j3RDVt1hdmF/8yiLU14bS0VYX/2hdhtNaomTElTbaKt9OzooSV7/jrXt70omARsorFzeeOw3L/vKeaSJgQCJHu/6dPa0eg7EE2jo51wMAhoVlxBPCtY6cdpSHlVBd7/dHNuAe5vYpxuzklY/CT+RK8DOireMolrQ0oWFYGOt2HsqKdkqIZLRV06hay8o8n0Nf/x3fuXgWTplYh9qQjMWPrLd0nxIBH5s1LmcYr9PtYt94rztVooQBsgtWlkVJdhUiGovMHuZcDrOMKTYE1a8+ilKRL8HnTlGaAAAgAElEQVTPiHmaHdne7n7Dc44Pxk13bka7xFxK3Og7vm/1dmy4/QJbDa0kAl7bfSTveU6iCCBUUDxYZSEBGd+QTORZWRIVOz3MLwfwEICJSHYRnALgrwBmuyNaYbDD3B5OhKB6WUTQS4wW5aS5SSAgyVBEAmdNHZXl89A6OyNx40gls+Mr2/bi1l+3gSBBIIEHr5qXrsNlpsQ3dRxFTLe9icUT6e/LqgM/ngAeXttu6VwniZiV3B1C6J9AX1TB1n09npqG7ew87gNwNoAXhRCnpfp6XOeOWIXDZit7OGV2crqIoF8d5lqMFG+yQEeqU6EgXNXShC9fOB3r3j2MRdNHo0WX2LWweQyS72AwOJ5JV28E//ZsW+odPDnnl59tSxc1XNA8Or2b0D63WFyB3kWhiORxrdLRllM3w6w0+pcvbMbP1u0yHEMmZM1vi6G96TDlnue3oqm+GrMn1nnyN2InzyMmhOgCIBGRJIR4CYAlxwrjX5xuFesETrSbLQVmdaKiikB/VEEknsAty9vw6Sdex39teA/XPfl61r00jxuO68/JDHW9/pzJqK8NZeXNbPxbV9Y6KgB8/OFX0s/q91uzE/Pe6zI2janHL583CRtuvwD3XznH8LwqCz6H6WOHY/XNC3Hrx2ZmxewX6wbhfYcxUQX4/FNvefY3YmfncZSIhgF4BcDTRHQQQJ87YhUOm63s4yezk1+q3FpBbyqKxBUkhMhwgMcTQDyRSL+xG93LvVeciuvPnprOBdn2wTEsuH9tVgDD4d5BQzliikhn5N/x3NaMiJwFzaNRFZQNr9P6XhqGhU2T/f7xw1OxbOP7ORsy/X7rPnxtxSEQyFDBMe7QH/XOgW5HeVwBYBDJKrufBlAH4F43hCoGNlsVhhe9K4wodVXdYtEq3lhcwZKfvZrzfLN7aR43HM3jhudUnmYmLj1qRM4ty9sgESEgSVm5I3rfS1dvBHET38LHZo3DqZPqcNtvNkMC0G8Qirx6y4G8cjHu4cXfiGXlIYToI6LxSFa+PQLgDykzFsM4RjmG/qqKd1PH0bzZ2/nuJZfynNs0EgubGyznZCQ3CgJR5cSOIhxIWqpnjhuR7kG+vv0wbv/NZkgGBSgDUtK8df7JY7Hh9gsM61gx3uPF34idaKvPAvgmks2gCMCPieheIcSTbgnHDD3KOfTX6I83KBMkyi6vnmuMXCG3r+4q7n1NNT3d8dxWVAclKAmRZWrTEk8Ad67cisRzAt9cPBvvdfnOUj1kCcpAVSDg2d+IHbPVrQBOU3cbRNQA4C8AfKU82OdR/jjpg3E6aktbp0pfX8hM8dm5F6N+G1e3NJomFBpRE5QRSyjI10pjwGKJEvW8O57bmt65MN4SlAm/+9K56Isqnvkp7SiPLgDHNZ+Pp475CvZ5uEOpQ2ed8ME4WbARAL753JaMRf36cybj3itOzTjHSPHZqTRs1G9jeWsnvnzhDFh1PX9vyRycPH44Pv7wK3n7ctgll9OcKR3XnNXkeXHEvMqDiG5J/bMdwGtEtBLJ3+IrAGx2UTbGJ1gtuOeHaC2tPE5GbbUfOJ6hOABg2cY9uP7sqYY7EHUOs2dn9rxy+TxmT6zLcnzrqQpKaBpVg+Zxw/HQVXNx64rNkMj6LoMpD9QXCr9nmKt/GX9L/aey0nlxGL9hZRF2+g3fCZyO2lJ7VxgdN3sDNHt2xwfjGYUJtc8rX8CALBESeXYT6rmXz5uEWRNG4PlN+/DE+l3oi7qvQGTJuIQ74yx+iEC0UhjxHu3nVK4HhBC9bgnF+AejznMiIdK/uH7Ny3Aiaku7O5jaUGN4jtlxwPjZJRIC96zejmjc+HkZ+U3uvHRWuvBiVpfFoIREQiAckDPOBYD17Ydx24pNkIjy7jyCEiBLEgaLNEux4igNfohAtBNtdQqAXwIYlfp8GMD1QohtLslWEKV2mPvNXFMoZvdh1HkuogjUhpKJZ37Nyyg2amtl217ctmITZJKgiARuWDDN8DyjnAcVo2cXVQRqQ4So5pj+eWn9Jtry6WaFF9ekHKdb9/bg3tXbIJOEeCKBmCJSXpITMlQHJAzoFAQh2Seek/nKg6BUfp0ElwK4JVWWBER0HoDHAHzYBbkKppQOcz+aawoh130YdZ6rCkrp+kVu5WU4oZQLjdrq6o3gq8vbUpFNyftcum6XydnZS662RLv+2YUD2XkgA7G44fPqGYji3tXbEImLjMKL4UBm6G99bQj7enrwzZVbUzWkzMOsblg4FY++nHkvAmrLWFYf5UAsAbS+f8TztcaO8qhVFQcACCFeJqJaF2QqC/xqrrFLvvswUwLqcTfyMpxua2tXlm37jmWFxCoiu8NfUCbMnlhnKrvxTkFAkgiKZkdCuuQ8dQyJCJF45oJeFZDx6KdPQ111KJ3gt+D+tSCQpeKDNSHjUiVMeWEWrFFK7CiPXUR0J5KmKyBZUdfsdazi8au5xi757kNVDrdqTDh65eB0XoaTStnqDkZ7ntkb+OcWnYQnN7wHWSIoCYEHlmQ+B7MS7dqdwhfOa8bSdbsy/RYBGdv2HUNddRC1ITk9hhGxRAI1QRnvHjiOWFzJea6eoEyY0jDM0rmM/8kVrFEK7CiPGwDcA+C3qc+vpI4NScqxjIYRVu4juZSeKDNuhFO1sfI56O1gdQdj1GUvKFNGjkRQJnz23JPw2XNPMlVGRrJLAB67viW9UwCAR1/O7IkxEIvjxmWtCMkSIkoCZNAauiYoIwGBlin1GfWzchW8pZTcAVlKK7tzPtSQdW+yRAhIgEwS+vNlFjK+wet8TcvTCyG6hRBfEkKcnvrvy0KIbjeF8zN+LGVeCPnuQ32bjsQT6TLjt/1ms63ENzvkc9BbRbsLOB6JYzBmLLfRefet3o67L5uNcEBCTUhGOCDhoavmphXk3KaRht+zmewT66rT1+ifdzhAICJE4sn5o/FE1hjhgISf/uMZeOqGs7LqWuUyVREB//3Z+bj38tl44YsL002jrjmzKeO8T89vwgtfPBe3XTzTVo+GfIRlgkzeL3KVyq2/2eJpuwIrSYKrkMOTJoS43FGJygg/lTIvhlz3UWrzXD4HvVWsym123imT6vCXr2c3VnJCdu3z7hmI4gtPv51hxgpIyPC5fPLMRiyaMQY/32DeO9yIhACuffx1hAOZpVL++/XMZMdfbtyDX73RiaBMjvbOSCCpwDh81x1iivDUz2rFbPWgwTFVmfiuM32pQ3X9Usq8WMzuo9TmuXwOejvjWJE713lGzySXD8WO7OrYXb2RrPn1zvpnX9+Dk0bXFlRqJKokEFVO+I4eXDI3a3yBZNmRSDz7+mJwujQKk42XflYrG8qRAE4RQvyvEOJ/ATwA4BcAfg5grIuyFYQQYpUQ4qa6urr8JzN5KbV5zqn5rI6T67yu3khGN798HQ7VooZa1KKGueTUX6M380QV4O5Vf8W31uyw9Qz0SABe/Gt2p0GmfJB0r+te+llJGDjnMk4g2gDgU0KIjtTnNgAXAqgF8F9CiAtdl7IAWlpaRGtrq9diVAylToZ0ar5Coq0ahoUNnej3rd6eZZLacPsFGf6hBfevzXmO0bz6axjGjEXNDXj9/W7XcsuI6E0hhKX24lbMViFVcaRYnyrL3jWU8zyGGqU2zzk1n9VxtOcZhdzes2o7grrXPr3JwMx/oobhmkVo6a8hcLoeY8wr7V3401cWeVqKXcWK8qjXfhBC3Kz5OMZZcRg38GsJlVLIVcgchkpAJsTixr4RdY7akJzlv9CG4Rq9KTbWV2MglulsYMXBmCEA7OsZxKIZ3i+9VpTHa0R0oxDiMe1BIvocgNfdEYtxCr+WUMknlxOKpdB7N3KiKwmBuy6bna4dpSZLqi1c1TmubmnE8tbOVIa5goRQndHmCY/JDHNrKmPCiBA+OHaiMta5zQ3J7oJErjmog1JSukQCCMiUKmXCeMWxgWj+k0qAFeXxFQDPEdG1AN5KHTsDQBjAJ9wSjCkev5ZQySeXEwrP7r3rlZVRyRV9suTxSDztB1HnWN7aiUevOQ2bOnswoa4Kdz+fWTdUTXgEYFopNxfdA3Gs+NzZeK+rH/OaRmLZxvfwSnsX3NyvKOJEWRZWHN4zojrotQgArJVkPwjgw0R0AYDZqcMvCCHWuioZUzR+LaGSSy4Ajig8O5nqZspKm/sCAAvuX5vRSc/IDxKLJ/CZZW+ayhVRBF7d1YUfvLjTtP5Vrn1IUJIQDMhY0tJk2KDKDXI1n2JKz8Q6f1SxsFyeJKUsfK8wuIf5CZzM0XDSP5FLLqcUntVM9Xw7FHXOTR1HLflBrLyYP/jHdxBTMivlhmSBgCQjnlAgYG6C0n5/Zg2qmMpmX8+g5y1oARvlScoFv+d56HMH3MSpnIl8+Q1OyuWUwlOzvbUYZXurykqLdhekYuTYjsQU3HXZbIQDhJqgnLPOVAa68PhEQiBtDgPl/KM8c0p9+vub1zTS8JygBITkZI/1cIBQxfVBKgx/bAXtFEZkisQL53WxJVTc8puYyeVUifdc2d7aXZQdZaV3bKul1IUAFAhIlH/nEZCS/Ri0KAJQFIGokr8EyyvtXWjd3YVgQEZjfTUWNjdk1btSx999qA8AQZIob+9zpnwoO7MVUxxeOq+LyZlw029iJpcTNcPMlJA+Oup7V86xpKw6uweydgQEaBowWVuZPzZrPP78zsGcSYEBCZCIQCBEDApDXfP4a6gKyIgqiQwfjJ5XdEqFqQz8YrZi5VEi/Oq8zodXpeedSBI0c3rrFfiG2y/A6psXoq3jKOY1jTT8wzRrKWuXF3ccRD5FE08AT1x/Ot7c042f6Lr+AcmaUVajs5jK49VdhzF74ggA8DR/i5VHiSjX/h9udApUcdIJbzaW3umdlegXT+Dp1/bgJy+35zQn9kWVrGq3hZiCQrKEmxadhEdT80XiCgQyiwgGJOBf/vttBPSFjBgGwE9e3oWf/u8uyBIlw7w9yt9i5VEi3FyE3caN0vNO+n+sjhWLK1k+CUUAP/7zTsQSyGlOrA3JWdVoE8J+KZGBWBzXzp+Ma+dPTmelL35kfYbyiCcAJBLwRypYJtUBwkCcnSdekxBAQrMD9SJ/i8MwSsjl8yZhw+0X4KnPzseG2y/wRaa3VXI1QQLsRZFZbdRkZZ58Y2nPf6+r33BMifJHW+3rGTS5Nq/IGej7ldfXhjIiz0IyZUWJ6QnJSEepndvcYE+AIvmHM5owLJy/MReBF5dSYvQ76za88ygxVm35fq1HZYTdXYSVBL72A8ezfBBG80xpqDX1Jemd47d8ZIahPAJWzInGb9uhgIQBGxVxJcDQTKb6XKY21OC6J3NX/fn3S/4Op00ehcb6avzoxZ0ldYwTgFiO7k5quLIi/BJQOjTwwgTOysOH+LUelRGFRJHlS+D75nNbMjKnrz9nMr584QzDeVbfvNDQl1QbkrPOf+hPOw3l+dpFJ+P7qYxvM3OiWXikPjs8HxFF4JG17yKqSRK8ZXkbZElKF0+cPKoaOw/0mY5xysQ6zG0aWbIMcy3/8/ZeJERSSRjFC3D1ktJSG5KhCOGJCZyVh8/waz0qMwqJIjMzAanH9Qviso17cNbUBsN5+qKKoS+pL6pknW/GqNoQNtyeu+WsWYvZL5zXnOH8jiki5xt3SJYQkCgjpyOeAOKJE2G3uRQHAOw/ljTFrX/3UN57cxo1yTLAvnzPqQlJuOfy2Tj/5LEcbWUGEZ0E4A4AdUKIJV7L4yblFtJbWBSZ2fIqTEtuHO4dNJ1nbtPILIe+UXtXM+alfDm5nm9jfTUiOvNUJJYwdH7nyuFQEoksv4ddblnehnBAxmDcXl93J2Gfuff0RxOIKAnP1gXXfVpE9CQRHSSirbrjFxPRO0TUTkRfzzWGEGKXEOIzduaNJ0TJyoA4id3FuFTlTszmUaPI1BId4QDl3UKbmYAm1lWbltxY2DwmZ6kVvUPfqATKA0vm4PpzJmeMe/05ky0lXHX3RbNUnkgdV9E7v8MBgqzzqEsS4ZuLZ2sc5HmnziKqCByPxLlHOIP7Vm/3bI0rxc7j5wAeAbBMPUBEMoBHAXwUQCeAN4joeQAygO/orr8hVdnXFjv2H8N1j7/me5+BHjshvaXyjeSbR1+qPB9mJqC+qIK5TSNx/TmTsWxjps+jedxwNI8bbitk2CjE+PJ5k3D92VNzJgQaYbYjenLDbvz27b0Zz0Y1gfUMRHHjL1qh3R9IAE6ZVJdxzk2/fLPgNrRVQQlKQiAhBHL4sZkKxUurhOvKQwixjoim6g6fBaBdCLELAIjoWQBXCCG+A2CxM/MCxyPexUAXg5W8ilL5RvLNo/5cWyYjnxy56k4BwL1XnGq6wNvNPDc6X1VEdjDbEf3mrb2IxLMz1lWHtllggCqXE2+Nv/vSudix/zhufubtosdS+ft5E/A/bR84Nh7jDl4mGnsVij0JgLYvemfqmCFE1EBEPwVwGhF9I8d5NxFRKxG1Kv096eN2YqDdNgNZHT9fXoXVarDFkm+eQuSwUu23edxwLGlpcqWGTyHfcfO44Vkmr0tOHYeQbH7vuSr7qjIAyHgWVirzEoCwLCEkJ69tHjcci+dOzJLPcpVfA45HlKKuZ9xHJniaaFwWDnMhRBeAz1s4bymApQAQnjA9/cpnVTu7bQZycnyjEuEDsbjjbyH5fDCFll1xI2vdCsV8B2dMGYVfvdGZzipf0DwGa3dkRjzpn01cZ0uKKwls3duDTy7daGjqOnx8MGczKaTmVgsmtr5/5IT8OhdIMVV0X/yrbUsxU2KKDbwoFq92HnsBNGk+N6aOOQYRbPWwKCbr2QpujK//5XHjl0ndJYTkzLfdXI5pv5ZdKTazXTXPDaZ6kt+3ejvuvHRW1r0DyTpa3X1Rw+/o3tXbsmTYfagX7x44jrrqoK2s9WUb96B1dxdWb9qXFeIscCIbPSgTAhKymmEBqTYiTNkRTwjcusK5NcouXu083gAwnYimIak0PgXgWicGVjsJTjvpQ3jqs/Mtv9W6HSLr9Pid3QNZva+rArIrzrPW944gGd5v8LaLwnYRXiRCFvMdmGXFa53fjfXVWN9+GAvuX5vM+1ASqUZPmcR0ca6DsQSW/OzV9Ge75qJrHn8NssmLgzYbHQBe2nEQd6/aht7ICTe+lT4kjD+RJfLMYV6KUN1nAGwEMJOIOonoM0KIOICbAfwBwF8BLBdCbHNiPrWT4Kj6kTl9Bnrcrnrr9PilqtJrlMW8bOMetB84nnEsn49Gi9u7PDOKeWa5suLVeweQcV/ReCJrUY4p+mIo2dhdyGOKwKBJX4+FzWPS30vDsDDOP3ks4jqFxoqjfFESonId5kKIa4QQE4QQQSFEoxDiidTxNUKIGUKIDwkhvuXUfER0GREt7enpyX+yBrfNL06PX0h+RSGYhajqj9txQpfK2Q8kld+K1g60Hzhu+zvQXmulra3RfemxWmZdJqAmKCMgnfh3OEBZhRCDObYp+hwWtV7anZfOQlACghJZMlnlWyTY7OUNEgEPLGGHuWMIIVYBWNXS0nKj3WvdduI6Pb7d/IpCMAtR1R63a4Iq1a7JqEbWvVecauk70F97dUujYYa5Vmaj+9Jj1Z8hS1LyqyWCECL9HU8bXYvX3+sGIekQz1WkMNmGNon2O+qNxFO/O9a2HPl2Srxx8YblN52Nlmmlraqshasm67BjfvFyfK0Dtz+qIBJ3x/RjFKKqfaMtxARVCid7LnNbvu/A6NrlrZ15M8yN7uv6cyZnfL7rstlZWed6fSJLhKiS/F5jioAikP6Ol726J+20jyqJnAu32u9c/x25sdh/bPZYF0ZlzJAlwrQxwzyVoeJ2HqrDvLm52WtRXKWUNbByJe0VKofbu7xc5rZ8uSNm11oZz+i+vnzhjPTnzu4B1ATldAIrAAwLB/CdfzgVgzEFVUEZ3/jtloyfF8O6dw8jGJAtF4kslJnjRuDWi07Gd3+/g8N8XUDfdKwm6E5wjB0qbuehOsxrh48oy9pWKvl8CHZMP04kPpol7VmRI1ddLLd2ebnMbVpfhp1rzc6183zNntf4EclnMH5E2HJBRyssmj7akjmtWGqDEt7v6kPC5XmGKrpcVF+0sCYhKtNiWTVxuph+4yNlV9sKsO5DeL5tb1YNLP15pQiJzSWHl71Jrnv8VazXNEo6t7kB00bXGvpB9Fz0g5czSqPPHFeL+Sc1ZNXcOmPKqIz7u/qMRix/s/PE55ZGLG/tzLh/ABnPq2VKfYacM8fV4h2Tsux1VTJ6Bk846SeMCOGDY8YNa2eOq8UfvnKe4f1o4day/icoJ31f1cGAq39HRPSmEKLF0rmVqjzCE6aLCf/0QwDJqJgNt1/gy8Q1PV29ESy4f21W0UAz+XN1HLQ7VrFy6+Uo5fxG8ujnDsmEqEFc6otfWZSxo2o/cBwf+cE6w/MApM139bWhrDnyod4/kDT5xeJKRo6Hk6hz7T7U69gc15zZiGfe6LR07uzxw7Btf6/hz2QA3hWUL0/CAQmPXd+C2RNHuPb3Y0d5VJzZSg3VTQyeeMvyor9vodgNY81l+illSKyRHKWcX4/R3GYZ+HofRz5/iWq+sxKaq0frD5rbNNK0r7oTqHOte/ewY2Me6jXe5Rix/7i5Gc+l4MCKJiRLqKsO+uYluOKUh+rzkKpq08fctA86XUjRyTBWJ8cq5D7zze9mEUqjuc122XofRy5/iVbmQnwJ+udvx79iF7Ud78S6KsfG7B2IWT63KmC+vFSowcNVooqCjiP9WLfzkC98uRWnPFTs1rYqhJVte7Hg/rW47vHXsOD+tXi+rfjyXE6GsTo1VqH3mWt+N55dvrkfvGqupWZQzeOGY6EuGe/c5gZs++BYhswb2g9nzaG/bua42ozPV7c0ZlUP1stUV2XeISqoe2OvztEPtr46gMWPrMe31uzImcinn09K/e0Y8ep73TlGymRvT46dh+VRGJVIXODmZ97G9U++jvnfftHxvxm7VKzPY97pZ4hfrPyzaxVb3bbn5/JllHIsJ+5TP7/Xvpj2A8dzNoMyki8ckAAIRDSOZb3/wkobWrP7VGUKSMC/Ld9czC0XTTgg4WsXzcC31uzwVA4mN+EA4S9fv9DRvxk7Po+KzvOY66JJwO08C7tNj9way4n71M9fyhyVQppBGcknS5Qy1J9w86oyq76eTR1H8+ZTmN2nKtP3//iOzTt0npAsof2gsaOb8Q8yeddFEKhAs5Xq86irqyvoeqt2+FKV2CgVZvftxn0WM2a+HA0nMJJPSQgoIrfMVnwg+e5z0fTRBUjsLAPROPaVSYDJUEYR3q43Fac8isGOHb6c+ljkI9d9NwwL4+ozGjPO19vt7VLos/vmc1vwkR+sw9dWbMZHfrAO31y5pWAZ7Mr3wJI5OHPqqIzzzpxSnyGz0bPK5/PQ0zKtIav4oRa9f2LCiJDpuUFC+h5yofebxAXwyt+6TM5m/EBAAh5YMtfT9aZifR4tLS2itbXV8vmF2uGd9E14Qb77dtM/YefZ5cq9cKNVrV6+7r5o3vmNnpUeq8+udXcX/qdtH55+bU/Wz3549RzsOtyPuY11+MIzb+ec74dXz0FAlh3tb854R1Am/ODqeTjnQw2urDec51FASfZCcxLcLqToNm70KLeKnWdntTS8k2jlszK/lbwPq8+uZVoDTjPx2cUTwC0XzcTo4VV559u46wgGY5yOVykEZQlNo2p8sd5UnPIo1OdRaT4Mq7jVo9xprJSG93p+J3wedua0Mt9Fs8aV7Bkx7uNl8yc9Fac8CqWSfBh2yHfffnku+UrD+2F+KyXZ7Ty7fHPq59Mzc1wtLpw13nAcSZMHlcvHooVgnP9BsN8QqpBrhjpBmTxt/qSHfR46yt2HUSj57tsvzyVfjoYf5jfKaynm2VnJS1HHb9vTjT9uP4CLZo3DhbPGm45TXxvKkEn7s/e7+tJjTGmoNbxm96FerHv3MOY21mH08KqMY2rEmPrvkTUhtHUcRX1NEN39sYyxDh8fxKbOHtPzjvZHse7dwyCRwLb9vRhZJeP97kHMmzQC1eEgFk0fjZ6BGP64/QBGhGXsPjKA8cND2H88imoZaO8awBlNdRg1rApHegfxZkdP+nNtUMKurn70D0ax5YNeTK2vQl9coL5KRkdPBM0N1RhQgGmjqnEsoiAoAS/tPARFAc6cUocBBem5Dh/rx/b9vairCuIT8ybiWETBiLCMP/31IHojcTSODONAbwx1YRl7ugdQEwrg9KY67Njfi7qaID6zcBreev8oXt3dhTmTRiAuCLMnjsD5M8diX88gAIHZE+tc/9sb0nkexeJkfoURflmE9eS7bzeeSyHPor42hOnjhqO+1jzKKB9Gi7FVWfLliADWnlU+BaP9nA/tfPMm16cXcz1H+6PYc6QfUxtq0ovyoumj0TAsnPVcz5w6ClMaajPu98/b96eVSl11EADwwdGB9OLfMxDDgWOD6BmIpX8OnPjODh8fTM+vKqiTGmrS523dexS/37Yfs8cPgyAJUxtqMLImhMmjatDR1YuegRgkkUBvJI79PQMYUAZw0uga9A7GsetwH2oDhPe6B9E/GEL3oIKmujCGhQM4eCypNPSfZ48fBlmWEYknEIkriCoJAIRoPIGBqILjg3H0xQWAaowbUYXW3YexvyeCAAGbOntwsDeKqaOqMbwmjDOnNuB4JIG9R/vxyruHQLKM+ioZvZE4hoUDGDu8CkcHExhVG0L3QBwBiTB2RBU6ugdw4FgEazbvw96eCIISYV5TPQbiCWzbdwxjhoUwalgYAKFtT3f6WasdBL1cT3jnUUK8LE/uNwp5Fk48P6PWtPqy6k5+L0YyCyCzjLuubLv2c6+u859ZCXmzudT70Jen1zJzXC3eP5J09vdF40hoJlTny1XS3Qr6ZkZMcZzb3IAlLU2O/95ySXb4T3l4WZ7cbxTyLJx4fmbhvvpS7U6GIlspc6EX+8kAAA6NSURBVGIXo/DkXM+n2JLs/+/yWfiP57cXfD3jDkEJ0EZpO/F7y6G6BYTquo2X5cn9RiHPwonnZxZuqy/V7tT3YiSzLBFkKu7Pzug+cj2fYkuyr9z8QVHXM27hzu+tVSpOeRRbnsQt/BLy6gcKeRZOPD+zkFX97tup78VqmRO7GN1HrudTbMmTK+ZMKOp6xi3c+b21SsUpD7/il5BXP1DIs3Di+ZmFvj541VxXvhezMicPLJmbM5xX+1kfzmoWnpzr+eQreTJzXG36Okk34fXnTMZ1H56WVWbFLhyW6yznNjfgoavnebqesM+jxPg12soLCnkWTjy/YqKtCsGsRa/VaKvuvqjl8ORc99G6u8swlLZlWoOl+fTRVuvePYyxw0I42BvNCJnV/nzR9NGYNmZYVliueu5JDTXoiyWwaPpodHb3Y/WWE9FW2vDdjq5e/GV3d1YY7eJTx6N3MI6Vmz9IR1tNHHEi2qp7UElfo/+sRlt19w5iywfHMW1UNWKC0uOoobtqWHDr7sN47b2jCBAwbkQ4I9pq3qQRWPvOIew92o+po2rS0VY7D/ZhWDiAGWNrsfNQPyaOCOFvh/sRkAjnzxyDN9/vRldfDPMaRyT7nwiBa86anI62Om/G6HS0VSyuuB5txQ5z+Fd5MAzD+JUh7TBnGIZh3IeTBMscv5jBCpGjkGxxu/M4nZGuNf2opoN8Mloxk+W6r3zjaT9rk/9apjVkmJo+ODqAlZs/wBVzJuDsD402HUObYa7NUjfKUN996Dg27jqSrKE1ud7Q9KXKdNLoGsQTyDimZnlr5dNmj6smr1/+ZRde2nkYTSPDiCQow2y1ZvM+rGvvQvPoGoyvr0U8Fsc7h/owc0wtAsEAJJHAzkP9GBYkfHA8mh5DKAo6jg6iNijhWETByeNqMXPCSHR09WLbB8dRVxVAz2Ac0xqqEQwGcbw/gp0H+1EdJCz4UAO6BxUcPjaAjqODmDgijL6ognhCYFpDNQ70xlAblNDZPYiasIyWySPR0RNBLBbHnu5BfGhMDT591hQsb+3Ekf4omkfX4G+H+zGyJoALZo7FOwf7MLGuCguaR2PH/mTvmo/NHu9JVQUz2GxVxvgl6bAQOYyS9cyS3wqdp5A5cqFPtDu3uQG//OzZOWVsmVKfcY1RUqI+SVB7X/rxzpxSj1c0480cV4t3TJL3qgOEAQv5JLnGmDmuFn/4ynlZz1KWCImEyIj3IQDDwgH0xxQoCWfWFYkAh4aqCIr9Hc4H+zxQ+crDL0mHhchRSG8Ou/M43f+jdXeXYaLdis+dneG8zNfPAwBCMhDNUSVd2xvdynhuw0mC/sLNHjZD2ufh1yRBp/FL0mEhchTSm8PuPE73/zBLtNMet9LPAwAoz5+del9Wx3MbThL0F272sLGD97+ZDuPXJEGn8UvSYSFyFNKbw+48Tvf/MEu00x630l8DAASs9fywOp7bcJKgv/BLf5aKUx5DBb8kHRYiRyG9OezO43T/D6NEu3ObGzKc5kYy6q9JJiXOy5kkqN6XlfFyJe/pe5ObkWuMmeNqcd2Hp2U9S1mirMQ/td+HrM80LAIHh6oIStnDJh/s8yhzONrK+TlywdFWHG1VydFW7DDH0FEeDMMwTjGkHeYMwzCM+7DyYBiGYWzDyoNhGIaxDSsPhmEYxjasPBiGYRjbsPJgGIZhbFMWVXWJ6BMALgUwAsATQog/eiwSwzDMkMZ15UFETwJYDOCgEOIUzfGLAfwIgAzgcSHEd83GEEI8B+A5IqoH8CAAVh4p/JIk6DROJQMW83ysdAC0cp2VxEK7cmgxSt6zkmxo9jNtgp92vHwdDZ97qwOrt+zH4lPH49wZY9PXAcm6X7G4gve6+jOSBLWdCNWugV3HB/HGnqMZ46hdCEkksG1/LxafOh6fOL3J8L5UOaaNqsaxiIJzThqFaWOGpxMaT504AlWhAAISsOtwP470DuLNjh6c0VSHUcOq0omLQQl452AfZo6tRSwBjB8ews5DfRg/ogoBAtbvOoJxtUEc6Ith9vhhONIfR21IRl11AG929GBcbRA7D/ahOizjtMY6vN89iJlja7G3ZxDDQgHMnzYKOw70oq46gNMn12P15n1492AvPtXShOs+PC3ru6kNydjXM4C93QOIxBUsbB7jaba560mCRLQIQC+AZaryICIZwE4AHwXQCeANANcgqUi+oxviBiHEwdR1DwF4WgjxVr55h0KSoF9KsjuNU6XXi3k+RtcKIO94+usmj6rGTk25c6My7sU8C6NS6TVBOW9pd7Of6cupq+P1ReMZpdH1pcHP/vafsP9YNEP24eEABmJxEBGEEIgXWKaLABitUhNGhPD1S2Zl3FdYJvQM5ihZXCbUVcnYdPfF6e8GgGF1ZadLtPsuw5yIpgJYrVEe5wC4WwjxsdTnbwCAEEKvONTrCcB3AfxJCPGilTkrXXn4pSS70zhVen3F587GdU++XtDzMZIhHCAAhEjcfDyrJdm1ZdztyqGd0+ze9ecC2aXdnSj7rpYGf+6tDvzb8s22r3eCAAEWWpaUJd/42Az8YG173u/GyRLt5ZBhPglAh+ZzZ+qYGV8E8BEAS4jo82YnEdFNRNRKRK2HDh1yRlKf4peS7E7jVOn1de8eLvj5GMkgk5RV8E8/ntUS6mbl3a3IoZ0zX2nuXKXdnSj7rs6/esv+gq53ggrVGwCA5zZ/YOm78apEe1lEWwkhHhZCnCGE+LwQ4qc5zlsqhGgRQrSMGTOmlCKWHL+UZHcap0qvL5o+uuDnYySDIhJZ3fH041ktoW5W3t2KHNo585XmzlXa3Ymy7+r8i08dn+dM96jkorufmDPB0nfjVYl2r5THXgBNms+NqWNFM1SaQfmlJLvTOFV6vWVaQ8HPx0iGB5bMxQNLco9ndJ2+3Lm+jHsxz8Lo3mWJLJV2N/uZfneljqcvja4tDf6J05swYUQoS/7h4aRTOigTAkWsNGYKYsKIEL7/yczy9nVVcuET+Yi6KhmfO396+rupCho/QC9LtHvl8wgg6TC/EEml8QaAa4UQ25yas9J9HiocbZWEo6042oqjrYrHVw5zInoGwHkARgM4AOAuIcQTRHQJgB8iGWH1pBDiW07OO1SUB8MwjFPYUR6u53kIIa4xOb4GwBqn5yOiywBc1tzc7PTQDMMwTIqycJjbYaj0MGcYhvGSilMeQ8VhzjAM4yUVpzx458EwDOM+Fac8GIZhGPcpSaiuFxDRIQBHARRivxoNwFoaMFMsdSjsO/I7fr0vr+Rye16nx3dqvGLGKfTaYtavKUIISxnWFas8AICIlgohbirgular4WpMcRT6Hfkdv96XV3K5Pa/T4zs1XjHj+H39qnSz1SqvBWDyUqnfkV/vyyu53J7X6fGdGq+Ycfz6OwSgwncehcI7D4ZhyhXeeXjLUq8FYBiGKZCSrF+882AYhmFswzsPhmEYxjasPBiGYRjbsPJgGIZhbMPKIw9EVEtEvyCix4jo017LwzAMYwciOomIniCiFU6OOySVBxE9SUQHiWir7vjFRPQOEbUT0ddTh/8BwAohxI0ALi+5sAzDMDrsrGFCiF1CiM84LcOQVB4Afg7gYu0BIpIBPArg4wBmAbiGiGYh2SK3I3WaUkIZGYZhzPg5rK9hrjAklYcQYh2AI7rDZwFoT2npKIBnAVwBoBNJBQIM0efFMIy/sLmGuQIvhieYhBM7DCCpNCYB+C2AK4noP+HzcgEMwwxpDNcwImogop8COI2IvuHUZK63oS13hBB9AP7ZazkYhmEKQQjRBeDzTo/LO48T7AXQpPncmDrGMAxTDpR0DWPlcYI3AEwnomlEFALwKQDPeywTwzCMVUq6hg1J5UFEzwDYCGAmEXUS0WeEEHEANwP4A4C/AlguhNjmpZwMwzBG+GEN48KIDMMwjG2G5M6DYRiGKQ5WHgzDMIxtWHkwDMMwtmHlwTAMw9iGlQfDMAxjG1YeDMMwjG1YeTAMwzC2YeXBMACIaKpBb4S7iehrLswlEdHDRLSViLYQ0RtENM3peRjGTbgwIsO4BBEFUlm/ej4JYCKAOUKIBBE1AuhzaS6GcQXeeTBMHojoS0S0nYg2E9GzqWO1qW5urxPR20R0Rer4/yWi54loLYA/mww5AcAHQogEAAghOoUQ3anrLyait4hoExH9OXVsFBE9l5r/VSKakzp+NxH9kog2APglEclE9EBqJ7OZiD7n7pNhhjK882CY/HwdwDQhRISIRqaO3QFgrRDihtSx14noxdTPTkdyV6Fv1qOyHMB6IjoXSQXzlBDibSIaA+AxAIuEELuJaFTq/HsAvC2E+AQRXQBgGYB5qZ/NArBQCDFARDcB6BFCnElEYQAbiOiPQojdzj0KhknCyoNhkpgVeRMANgN4moieA/Bc6vhFAC7X+ESqAExO/ftPORQHhBCdRDQTwAWp//5MRFcBqAGwTl3sNWMsBHBl6tjaVHOfEamfPS+EGNDINIeIlqQ+1wGYDoCVB+M4rDwYJkkXgHrdsVFILryXAlgE4DIAdxDRqQAIwJVCiHe0FxDRfFjwXwghIgB+B+B3RHQAwCcA/LEAubVzEYAvCiH+UMA4DGML9nkwDAAhRC+AD1JmIaRMRhcDWA+gSQjxEoDbkXybH4Zk2esvEhGlzj/N6lxEdDoRTUz9WwIwB8D7AF4FsEiNvNKYrV4B8OnUsfMAHBZCHDMY+g8A/oWIgqlzZxBRreWHwDA24J0Hw5zgegCPEtH3U5/vAbAHwEtEVIfkm/3DQoijRHQfgB8C2JxSALsBLLY4z1gAj6X8EgDwOoBHhBCDKb/Fb1NjHgTwUQB3A3iSiDYD6AfwTybjPg5gKoC3UkrtEJI7GoZxHO7nwTAMw9iGzVYMwzCMbdhsxTAukXKs/1J3OCKEmO+FPAzjJGy2YhiGYWzDZiuGYRjGNqw8GIZhGNuw8mAYhmFsw8qDYRiGsQ0rD4ZhGMY2/x/kVASR9wn5EwAAAABJRU5ErkJggg==\n", "text/plain": [ "

" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "viz = data.filter(['User_Score','Critic_Score', 'Global_Sales'], axis=1)\n", "viz['User_Score'] = pd.Series(viz['User_Score'].apply(pd.to_numeric, errors='coerce'))\n", "viz['User_Score'] = viz['User_Score'].mask(np.isnan(viz[\"User_Score\"]), viz['Critic_Score'] / 10.0)\n", "viz.plot(kind='scatter', logx=True, logy=True, x='Critic_Score', y='Global_Sales')\n", "viz.plot(kind='scatter', logx=True, logy=True, x='User_Score', y='Global_Sales')\n", "plt.show()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "리뷰 스코어들과 매출 사이의 관계에 대한 우리의 직관이 타당해보입니다. 우리는 또한 데이터셋으로부터 다른 관련 피처들을 추출할 수 있습니다. 예를 들면, ESRB raing은 'E'가 있는 게임이 일반적으로 성인 등급에 대한 연령 제한이 있는 \"M\"게임보다 더 넓은 사용자에 도달하기 때문에 영향이 더 미치지만, 다른 피처들, 장르 (shooter혹은 action)에 따라 M-rating 게임들 또한 큰 히트를 칠 수 있습니다. 모델은 이러한 관계들과 다른 것들을 배우게 될 것입니다. \n", "\n", "다음으로, 데이터셋의 피처들을 살펴보면 제외해야 할 몇개의 컬럼을 식별할 수 있습니다. 예를 들면 매출 숫자를 지정하는 5개의 컬럼이 있습니다: 이 숫자는 우리가 예측할 타켓과 직접적으로 연관이 있기 때문에 삭제해야 합니다. 게임 이름과 같은 다른 피처들도 관련이 없을 수 있습니다. " ] }, { "cell_type": "code", "execution_count": 18, "metadata": {}, "outputs": [], "source": [ "data = data.drop(['Name', 'Year_of_Release', 'NA_Sales', 'EU_Sales', 'JP_Sales', 'Other_Sales', 'Global_Sales', 'Critic_Count', 'User_Count', 'Developer'], axis=1)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "컬럼 숫자를 줄였으니 이제 얼마나 많은 컬럼에 데이터가 누락되었는지 체크해 보겠습니다:" ] }, { "cell_type": "code", "execution_count": 19, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "Platform 0\n", "Genre 2\n", "Publisher 54\n", "Critic_Score 8582\n", "User_Score 6704\n", "Rating 6769\n", "y 0\n", "dtype: int64" ] }, "execution_count": 19, "metadata": {}, "output_type": "execute_result" } ], "source": [ "data.isnull().sum()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "이 데이터셋의 Kaggle의 개요에서 언급한 것과 같이, 많은 리뷰 rating이 누락되어 있습니다. 불행하게도 그것들은 우리가 예측의 위해 의존하는 중요한 피처들이고 그들을 대체하기 위한 방법이 없기 때문에 해당 피처가 누락된 행은 삭제할 필요가 있습니다. " ] }, { "cell_type": "code", "execution_count": 20, "metadata": {}, "outputs": [], "source": [ "data = data.dropna()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "이제 User_Score 컬럼에 있는 문제를 해결할 것입니다: 그것은 'tbd'라는 문자열을 포함하고 있기 때문에 명백히 숫자라고 할 수 없습니다. User_Score는 범주형 피처보다는 숫자타입이 더 적합하기 때문에 문자열 타입을 숫자형으로 변환할 필요가 있고 일시적으로 tbd값들을 NaN으로 채울 것입니다. 그 다음으로 새로운 NaN 값들을 어떻게 해야할지 결정해야만 합니다. 우리는 이미 많은 행들을 버렸기 때문에 이 행들을 복구해야 합니다. 첫번째 근사치로서 Critic_Score 컬럼을 이용할 것이며, User Score는 Critic Score를 따라가는 경향이 있기 때문에 Critic Score를 10으로 나눈 값을 사용할 것입니다. \n" ] }, { "cell_type": "code", "execution_count": 21, "metadata": {}, "outputs": [], "source": [ "data['User_Score'] = data['User_Score'].apply(pd.to_numeric, errors='coerce')\n", "data['User_Score'] = data['User_Score'].mask(np.isnan(data[\"User_Score\"]), data['Critic_Score'] / 10.0)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "최종 데이터 전처리로서 범주형 피처들을 one-hot encdoing 메서드를 사용하여 숫자로 변환하는 것을 포함합니다." ] }, { "cell_type": "code", "execution_count": 23, "metadata": {}, "outputs": [ { "data": { "text/html": [ "

\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "

	Critic_Score	User_Score	Platform_3DS	Platform_DC	Platform_DS	Platform_GBA	Platform_GC	Platform_PC	Platform_PS	Platform_PS2	...	Publisher_inXile Entertainment	Rating_AO	Rating_E	Rating_E10+	Rating_K-A	Rating_M	Rating_RP	Rating_T	y_no	y_yes
0	76.0	8.0	0	0	0	0	0	0	0	0	...	0	0	1	0	0	0	0	0	0	1
2	82.0	8.3	0	0	0	0	0	0	0	0	...	0	0	1	0	0	0	0	0	0	1
3	80.0	8.0	0	0	0	0	0	0	0	0	...	0	0	1	0	0	0	0	0	0	1
6	89.0	8.5	0	0	1	0	0	0	0	0	...	0	0	1	0	0	0	0	0	0	1
7	58.0	6.6	0	0	0	0	0	0	0	0	...	0	0	1	0	0	0	0	0	0	1

\n", "

5 rows × 333 columns

\n", "

" ], "text/plain": [ " Critic_Score User_Score Platform_3DS Platform_DC Platform_DS \\\n", "0 76.0 8.0 0 0 0 \n", "2 82.0 8.3 0 0 0 \n", "3 80.0 8.0 0 0 0 \n", "6 89.0 8.5 0 0 1 \n", "7 58.0 6.6 0 0 0 \n", "\n", " Platform_GBA Platform_GC Platform_PC Platform_PS Platform_PS2 ... \\\n", "0 0 0 0 0 0 ... \n", "2 0 0 0 0 0 ... \n", "3 0 0 0 0 0 ... \n", "6 0 0 0 0 0 ... \n", "7 0 0 0 0 0 ... \n", "\n", " Publisher_inXile Entertainment Rating_AO Rating_E Rating_E10+ \\\n", "0 0 0 1 0 \n", "2 0 0 1 0 \n", "3 0 0 1 0 \n", "6 0 0 1 0 \n", "7 0 0 1 0 \n", "\n", " Rating_K-A Rating_M Rating_RP Rating_T y_no y_yes \n", "0 0 0 0 0 0 1 \n", "2 0 0 0 0 0 1 \n", "3 0 0 0 0 0 1 \n", "6 0 0 0 0 0 1 \n", "7 0 0 0 0 0 1 \n", "\n", "[5 rows x 333 columns]" ] }, "execution_count": 23, "metadata": {}, "output_type": "execute_result" } ], "source": [ "if data['y'].dtype == bool:\n", " data['y'] = data['y'].apply(lambda y: 'yes' if y == True else 'no')\n", "model_data = pd.get_dummies(data)\n", "\n", "model_data.head()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "모델의 오퍼피팅을 막기 위해 데이터를 3개의 그룹으로 랜덤하게 나눌 것입니다. 구체적으로 이 모델은 데이터의 70%에 대해 학습할 것입니다. 그 다음 데이터의 20%에 대해 평가하여 새로운 데이터의 정확도를 추정합니다. 최종 테스트 데이터셋으로서 나머지 10%를 남겨둘 것입니다. " ] }, { "cell_type": "code", "execution_count": 24, "metadata": {}, "outputs": [], "source": [ "train_data, validation_data, test_data = np.split(model_data.sample(frac=1, random_state=1729), [int(0.7 * len(model_data)), int(0.9 * len(model_data))]) " ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Amazon SageMaker의 XGBoost는 CSV와 libSVM의 입력 데이터를 지원합니다. 우리는 여기서 피쳐와 타켓 변수를 별도의 파라미터로 제공하는 libSVM을 사용할 것입니다. 무작위 순서변경으로 인한 오정렬 문제를 피하기 위해서 위의 셀에서 분리가 완료 후 수행합니다. 학습 전에 마지막 단계로, Amazon SageMaker에서의 학습을 위한 입력으로서, 결과 파일들을 S3에 복사할 것입니다. " ] }, { "cell_type": "code", "execution_count": 25, "metadata": {}, "outputs": [], "source": [ "dump_svmlight_file(X=train_data.drop(['y_no', 'y_yes'], axis=1), y=train_data['y_yes'], f='train.libsvm')\n", "dump_svmlight_file(X=validation_data.drop(['y_no', 'y_yes'], axis=1), y=validation_data['y_yes'], f='validation.libsvm')\n", "dump_svmlight_file(X=test_data.drop(['y_no', 'y_yes'], axis=1), y=test_data['y_yes'], f='test.libsvm')\n", "\n", "s3.Bucket(bucket).Object(prefix + '/train/train.libsvm').upload_file('train.libsvm')\n", "s3.Bucket(bucket).Object(prefix + '/validation/validation.libsvm').upload_file('validation.libsvm')\n", "\n", "s3_input_train = sagemaker.s3_input(s3_data='s3://{}/{}/train'.format(bucket, prefix), content_type='libsvm')\n", "s3_input_validation = sagemaker.s3_input(s3_data='s3://{}/{}/validation/'.format(bucket, prefix), content_type='libsvm')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "---\n", "## 훈련\n", "\n", "우리의 데이터는 XGBoost 모델을 학습하기 위한 준비가 되었습니다. XGBoost 알고리즘은 많은 튜닝 가능한 하이퍼파라미터를 가지고 있습니다. 이러한 하이퍼파라미터의 일부는 다음과 같습니다. 처음에는 몇 가지만 사용하도록 할 것입니다. \n", "\n", "- `max_depth`: 트리의 최대 깊이. 주의사항으로, 이 값이 너무 작으면 데이터가 언더피트의 가능성이 있고, 반면 증가하면 모델이 더욱 복잡해지고 오버피트의 가능성이 있습니다.(즉, 고전적인 Bias-Variance의 트레이드오프)\n", "- `eta`: 오버피트 방지를 위한 업데이트에서 사용되는 단계 크기 축소 \n", "- `eval_metric`: 검증 데이터에 대한 평가 메트릭. 불균형이 있는 이와 같은 데이터는 데이터의 경우는 AUC 매트릭을 사용합니다. \n", "- `scale_pos_weight`: 양과 음의 가중치의 균형을 제어하여 불균형한 클래스를 갖는 데이터셋에 유용합니다.\n", "\n", "먼저 우리는 Amazon SageMaker Estimator 객체를 위한 파라미터를 와 알고리즘 자체의 하이퍼파라미터를 설정할 것입니다. Amazon SageMaker Python SDK의 Estimator 객체는 최소한의 코드로 Training job을 설정할 수 있는 편리한 방법입니다. " ] }, { "cell_type": "code", "execution_count": 27, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Training job videogames-xgboost-2019-11-04-12-49-17\n" ] } ], "source": [ "job_name = 'videogames-xgboost-' + strftime(\"%Y-%m-%d-%H-%M-%S\", gmtime())\n", "print(\"Training job\", job_name)\n", "\n", "from sagemaker.amazon.amazon_estimator import get_image_uri\n", "\n", "container = get_image_uri(region, 'xgboost', '0.90-1')\n", "\n", "xgb = sagemaker.estimator.Estimator(container,\n", " role, \n", " base_job_name=job_name,\n", " train_instance_count=1, \n", " train_instance_type='ml.c5.xlarge',\n", " output_path='s3://{}/{}/output'.format(bucket, prefix),\n", " sagemaker_session=session)\n", "\n", "xgb.set_hyperparameters(max_depth=3,\n", " eta=0.1,\n", " subsample=0.5,\n", " eval_metric='auc',\n", " objective='binary:logistic',\n", " scale_pos_weight=2.0,\n", " num_round=100)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "다음으로, 우리는 Trainig job을 실행할 것입니다. Training job을 위한 하드웨어는 노트북 인스턴스와 별개이며 Amazon SageMaker에서 관리합니다. Amazon SageMaker는 훈련 Cluster를 설정 및 작업 완료 시에 해제하는 것과 같은 무거운 작업을 수행합니다. 다음 한줄의 코드로 Training job이 시작됩니다. " ] }, { "cell_type": "code", "execution_count": 28, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "2019-11-04 13:07:06 Starting - Starting the training job...\n", "2019-11-04 13:07:08 Starting - Launching requested ML instances......\n", "2019-11-04 13:08:37 Starting - Preparing the instances for training......\n", "2019-11-04 13:09:31 Downloading - Downloading input data\n", "2019-11-04 13:09:31 Training - Downloading the training image...\n", "2019-11-04 13:09:59 Uploading - Uploading generated training model\n", "2019-11-04 13:09:59 Completed - Training job completed\n", "\u001b[31mINFO:sagemaker-containers:Imported framework sagemaker_xgboost_container.training\u001b[0m\n", "\u001b[31mINFO:sagemaker-containers:Failed to parse hyperparameter eval_metric value auc to Json.\u001b[0m\n", "\u001b[31mReturning the value itself\u001b[0m\n", "\u001b[31mINFO:sagemaker-containers:Failed to parse hyperparameter objective value binary:logistic to Json.\u001b[0m\n", "\u001b[31mReturning the value itself\u001b[0m\n", "\u001b[31mINFO:sagemaker-containers:No GPUs detected (normal if no gpus installed)\u001b[0m\n", "\u001b[31mINFO:sagemaker_xgboost_container.training:Running XGBoost Sagemaker in algorithm mode\u001b[0m\n", "\u001b[31m[13:09:49] 5614x331 matrix with 33684 entries loaded from /opt/ml/input/data/train\u001b[0m\n", "\u001b[31m[13:09:49] 1604x331 matrix with 9624 entries loaded from /opt/ml/input/data/validation\u001b[0m\n", "\u001b[31mINFO:root:Single node training.\u001b[0m\n", "\u001b[31mINFO:root:Train matrix has 5614 rows\u001b[0m\n", "\u001b[31mINFO:root:Validation matrix has 1604 rows\u001b[0m\n", "\u001b[31m[0]#011train-auc:0.778801#011validation-auc:0.765479\u001b[0m\n", "\u001b[31m[1]#011train-auc:0.79049#011validation-auc:0.776439\u001b[0m\n", "\u001b[31m[2]#011train-auc:0.794468#011validation-auc:0.779476\u001b[0m\n", "\u001b[31m[3]#011train-auc:0.800483#011validation-auc:0.788066\u001b[0m\n", "\u001b[31m[4]#011train-auc:0.807711#011validation-auc:0.792379\u001b[0m\n", "\u001b[31m[5]#011train-auc:0.809722#011validation-auc:0.793961\u001b[0m\n", "\u001b[31m[6]#011train-auc:0.810795#011validation-auc:0.792471\u001b[0m\n", "\u001b[31m[7]#011train-auc:0.812122#011validation-auc:0.794129\u001b[0m\n", "\u001b[31m[8]#011train-auc:0.814009#011validation-auc:0.795064\u001b[0m\n", "\u001b[31m[9]#011train-auc:0.814834#011validation-auc:0.795198\u001b[0m\n", "\u001b[31m[10]#011train-auc:0.81763#011validation-auc:0.799186\u001b[0m\n", "\u001b[31m[11]#011train-auc:0.819472#011validation-auc:0.800404\u001b[0m\n", "\u001b[31m[12]#011train-auc:0.821013#011validation-auc:0.802302\u001b[0m\n", "\u001b[31m[13]#011train-auc:0.82369#011validation-auc:0.80253\u001b[0m\n", "\u001b[31m[14]#011train-auc:0.824144#011validation-auc:0.803563\u001b[0m\n", "\u001b[31m[15]#011train-auc:0.826624#011validation-auc:0.806042\u001b[0m\n", "\u001b[31m[16]#011train-auc:0.828072#011validation-auc:0.808865\u001b[0m\n", "\u001b[31m[17]#011train-auc:0.831149#011validation-auc:0.80937\u001b[0m\n", "\u001b[31m[18]#011train-auc:0.832595#011validation-auc:0.808052\u001b[0m\n", "\u001b[31m[19]#011train-auc:0.83325#011validation-auc:0.808756\u001b[0m\n", "\u001b[31m[20]#011train-auc:0.834194#011validation-auc:0.808151\u001b[0m\n", "\u001b[31m[21]#011train-auc:0.836241#011validation-auc:0.810824\u001b[0m\n", "\u001b[31m[22]#011train-auc:0.837778#011validation-auc:0.813709\u001b[0m\n", "\u001b[31m[23]#011train-auc:0.838762#011validation-auc:0.814457\u001b[0m\n", "\u001b[31m[24]#011train-auc:0.839512#011validation-auc:0.813747\u001b[0m\n", "\u001b[31m[25]#011train-auc:0.840195#011validation-auc:0.814684\u001b[0m\n", "\u001b[31m[26]#011train-auc:0.842162#011validation-auc:0.816531\u001b[0m\n", "\u001b[31m[27]#011train-auc:0.843906#011validation-auc:0.81694\u001b[0m\n", "\u001b[31m[28]#011train-auc:0.84562#011validation-auc:0.817372\u001b[0m\n", "\u001b[31m[29]#011train-auc:0.846485#011validation-auc:0.818337\u001b[0m\n", "\u001b[31m[30]#011train-auc:0.848618#011validation-auc:0.820157\u001b[0m\n", "\u001b[31m[31]#011train-auc:0.849204#011validation-auc:0.820631\u001b[0m\n", "\u001b[31m[32]#011train-auc:0.851065#011validation-auc:0.821782\u001b[0m\n", "\u001b[31m[33]#011train-auc:0.851897#011validation-auc:0.822088\u001b[0m\n", "\u001b[31m[34]#011train-auc:0.853238#011validation-auc:0.822793\u001b[0m\n", "\u001b[31m[35]#011train-auc:0.854162#011validation-auc:0.822355\u001b[0m\n", "\u001b[31m[36]#011train-auc:0.855103#011validation-auc:0.822992\u001b[0m\n", "\u001b[31m[37]#011train-auc:0.856192#011validation-auc:0.82483\u001b[0m\n", "\u001b[31m[38]#011train-auc:0.856972#011validation-auc:0.825236\u001b[0m\n", "\u001b[31m[39]#011train-auc:0.857321#011validation-auc:0.826683\u001b[0m\n", "\u001b[31m[40]#011train-auc:0.857793#011validation-auc:0.827152\u001b[0m\n", "\u001b[31m[41]#011train-auc:0.85975#011validation-auc:0.828857\u001b[0m\n", "\u001b[31m[42]#011train-auc:0.861199#011validation-auc:0.828673\u001b[0m\n", "\u001b[31m[43]#011train-auc:0.861739#011validation-auc:0.829101\u001b[0m\n", "\u001b[31m[44]#011train-auc:0.862806#011validation-auc:0.829095\u001b[0m\n", "\u001b[31m[45]#011train-auc:0.863004#011validation-auc:0.829155\u001b[0m\n", "\u001b[31m[46]#011train-auc:0.864234#011validation-auc:0.829926\u001b[0m\n", "\u001b[31m[47]#011train-auc:0.865389#011validation-auc:0.830902\u001b[0m\n", "\u001b[31m[48]#011train-auc:0.866072#011validation-auc:0.83073\u001b[0m\n", "\u001b[31m[49]#011train-auc:0.866688#011validation-auc:0.83173\u001b[0m\n", "\u001b[31m[50]#011train-auc:0.867109#011validation-auc:0.832104\u001b[0m\n", "\u001b[31m[51]#011train-auc:0.867848#011validation-auc:0.832753\u001b[0m\n", "\u001b[31m[52]#011train-auc:0.868071#011validation-auc:0.832541\u001b[0m\n", "\u001b[31m[53]#011train-auc:0.869171#011validation-auc:0.83358\u001b[0m\n", "\u001b[31m[54]#011train-auc:0.870338#011validation-auc:0.834622\u001b[0m\n", "\u001b[31m[55]#011train-auc:0.87096#011validation-auc:0.835333\u001b[0m\n", "\u001b[31m[56]#011train-auc:0.871977#011validation-auc:0.836434\u001b[0m\n", "\u001b[31m[57]#011train-auc:0.872843#011validation-auc:0.83723\u001b[0m\n", "\u001b[31m[58]#011train-auc:0.872993#011validation-auc:0.837513\u001b[0m\n", "\u001b[31m[59]#011train-auc:0.873447#011validation-auc:0.837555\u001b[0m\n", "\u001b[31m[60]#011train-auc:0.873873#011validation-auc:0.836535\u001b[0m\n", "\u001b[31m[61]#011train-auc:0.874091#011validation-auc:0.836567\u001b[0m\n", "\u001b[31m[62]#011train-auc:0.8745#011validation-auc:0.836862\u001b[0m\n", "\u001b[31m[63]#011train-auc:0.874914#011validation-auc:0.837149\u001b[0m\n", "\u001b[31m[64]#011train-auc:0.875699#011validation-auc:0.83781\u001b[0m\n", "\u001b[31m[65]#011train-auc:0.876221#011validation-auc:0.838345\u001b[0m\n", "\u001b[31m[66]#011train-auc:0.8768#011validation-auc:0.838519\u001b[0m\n", "\u001b[31m[67]#011train-auc:0.87719#011validation-auc:0.838754\u001b[0m\n", "\u001b[31m[68]#011train-auc:0.877336#011validation-auc:0.838593\u001b[0m\n", "\u001b[31m[69]#011train-auc:0.877831#011validation-auc:0.839167\u001b[0m\n", "\u001b[31m[70]#011train-auc:0.878327#011validation-auc:0.839834\u001b[0m\n", "\u001b[31m[71]#011train-auc:0.878621#011validation-auc:0.839761\u001b[0m\n", "\u001b[31m[72]#011train-auc:0.878942#011validation-auc:0.839909\u001b[0m\n", "\u001b[31m[73]#011train-auc:0.879695#011validation-auc:0.841542\u001b[0m\n", "\u001b[31m[74]#011train-auc:0.879939#011validation-auc:0.841853\u001b[0m\n", "\u001b[31m[75]#011train-auc:0.880311#011validation-auc:0.841607\u001b[0m\n", "\u001b[31m[76]#011train-auc:0.880835#011validation-auc:0.842055\u001b[0m\n", "\u001b[31m[77]#011train-auc:0.881272#011validation-auc:0.842345\u001b[0m\n", "\u001b[31m[78]#011train-auc:0.882109#011validation-auc:0.842463\u001b[0m\n", "\u001b[31m[79]#011train-auc:0.882501#011validation-auc:0.843188\u001b[0m\n", "\u001b[31m[80]#011train-auc:0.882929#011validation-auc:0.843293\u001b[0m\n", "\u001b[31m[81]#011train-auc:0.883472#011validation-auc:0.84352\u001b[0m\n", "\u001b[31m[82]#011train-auc:0.883691#011validation-auc:0.843611\u001b[0m\n", "\u001b[31m[83]#011train-auc:0.884096#011validation-auc:0.843169\u001b[0m\n", "\u001b[31m[84]#011train-auc:0.884764#011validation-auc:0.842866\u001b[0m\n", "\u001b[31m[85]#011train-auc:0.885542#011validation-auc:0.842436\u001b[0m\n", "\u001b[31m[86]#011train-auc:0.885901#011validation-auc:0.842554\u001b[0m\n", "\u001b[31m[87]#011train-auc:0.886121#011validation-auc:0.842114\u001b[0m\n", "\u001b[31m[88]#011train-auc:0.886387#011validation-auc:0.842419\u001b[0m\n", "\u001b[31m[89]#011train-auc:0.886605#011validation-auc:0.843055\u001b[0m\n", "\u001b[31m[90]#011train-auc:0.886894#011validation-auc:0.842851\u001b[0m\n", "\u001b[31m[91]#011train-auc:0.886822#011validation-auc:0.843002\u001b[0m\n", "\u001b[31m[92]#011train-auc:0.887083#011validation-auc:0.843142\u001b[0m\n", "\u001b[31m[93]#011train-auc:0.887122#011validation-auc:0.842454\u001b[0m\n", "\u001b[31m[94]#011train-auc:0.88751#011validation-auc:0.842648\u001b[0m\n", "\u001b[31m[95]#011train-auc:0.887956#011validation-auc:0.843361\u001b[0m\n", "\u001b[31m[96]#011train-auc:0.888247#011validation-auc:0.843686\u001b[0m\n", "\u001b[31m[97]#011train-auc:0.88874#011validation-auc:0.845334\u001b[0m\n", "\u001b[31m[98]#011train-auc:0.889659#011validation-auc:0.844977\u001b[0m\n", "\u001b[31m[99]#011train-auc:0.889981#011validation-auc:0.84512\u001b[0m\n", "Training seconds: 43\n", "Billable seconds: 43\n" ] } ], "source": [ "xgb.fit({'train': s3_input_train, 'validation': s3_input_validation})" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "---\n", "## 호스트\n", "\n", "이제 우리의 데이터로 XGBoost 알고리즘으로 훈련이 되었으므로, 훈련된 모델을 Amazon SageMaker에서 호스팅된 endpoint에 다음과 같은 간단한 코드 한줄로 배포가 가능합니다. " ] }, { "cell_type": "code", "execution_count": 29, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "---------------------------------------------------------------------------------------!" ] } ], "source": [ "xgb_predictor = xgb.deploy(initial_instance_count=1,\n", " instance_type='ml.m5.xlarge')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "---\n", "\n", "## 평가\n", "\n", "이제 우리가 호스팅하는 endpoint가 있으므로, 우리는 이것으로 예측을 생성할 수 있습니다. 보다 구체적으로, 우리의 모델이 아직 보지 못한 데이터를 얼마나 잘 생성하는지 이해하기 위해서 테스트 데이터셋에서 예측을 생성해 보기로 하겠습니다. \n", "\n", "기계학습 모델의 성ㅇ능 비교하기 위한 많은 방법이 있습니다. 우리는 간단히 게임이 \"hit\" (`1`) 이거나 혹은 아니인지(`0`)에 대한 실제와 예측 값을 비교할 것입니다. 그 다음 혼동 행렬을 만들어서 모델이 각 범주에서 얼마나 많은 테스트 데이터의 포인트 수를 예측했는지와 얼마나 많은 테스트 데이터의 포인트가 실제 각 범주에 속했는지를 볼 것입니다. " ] }, { "cell_type": "code", "execution_count": 30, "metadata": {}, "outputs": [], "source": [ "xgb_predictor.content_type = 'text/x-libsvm'\n", "xgb_predictor.deserializer = None\n", "\n", "def do_predict(data):\n", " payload = '\\n'.join(data)\n", " response = xgb_predictor.predict(payload).decode('utf-8')\n", " result = response.split(',')\n", " preds = [float((num)) for num in result]\n", " preds = [round(num) for num in preds]\n", "\n", " return preds\n", "\n", "def batch_predict(data, batch_size):\n", " items = len(data)\n", " arrs = []\n", " \n", " for offset in range(0, items, batch_size):\n", " if offset+batch_size < items:\n", " results = do_predict(data[offset:(offset+batch_size)])\n", " arrs.extend(results)\n", " else:\n", " arrs.extend(do_predict(data[offset:items]))\n", " sys.stdout.write('.')\n", " return(arrs)" ] }, { "cell_type": "code", "execution_count": 31, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ ".........\n", "error rate=0.144458\n", "CPU times: user 44.3 ms, sys: 505 µs, total: 44.8 ms\n", "Wall time: 353 ms\n" ] } ], "source": [ "%%time\n", "import json\n", "\n", "with open('test.libsvm', 'r') as f:\n", " payload = f.read().strip()\n", "\n", "labels = [int(line.split(' ')[0]) for line in payload.split('\\n')]\n", "test_data = [line for line in payload.split('\\n')]\n", "preds = batch_predict(test_data, 100)\n", "\n", "print ('\\nerror rate=%f' % ( sum(1 for i in range(len(preds)) if preds[i]!=labels[i]) /float(len(preds))))" ] }, { "cell_type": "code", "execution_count": 32, "metadata": {}, "outputs": [ { "data": { "text/html": [ "

\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "

col_0	0	1
row_0
0	616	55
1	61	71

\n", "

" ], "text/plain": [ "col_0 0 1\n", "row_0 \n", "0 616 55\n", "1 61 71" ] }, "execution_count": 32, "metadata": {}, "output_type": "execute_result" } ], "source": [ "pd.crosstab(index=np.array(labels), columns=np.array(preds))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "우리의 매트릭에 의해 테스트셋에서 \"hits\"인 것은 132개의 게임이었고, 모델은 정확히 70개가 넘게 식별했습니다. 반면 대략적인 에러율은 13%입니다. False Negative와 True Positive의 양은 하이퍼파라미터의 scale_pos_weight 값을 증가시킴으로서 True Positive로 유리하게 이동될 수 있습니다. 물론 이러한 증가는 정확도 감소와 에러율 증가, 그리고 추가적인 False Positive를 초래할 수 있습니다. 이러한 trade-off를 궁극적으로 어떻게 할것인가는 False Positive와 False Negative 등의 상대적 비용에 기반한 비즈니스상의 결정입니다. " ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "---\n", "## 확장\n", "\n", "이 XGBoost 모델은 게임이 리뷰와 다른 특성을 기반으로 하여 hit가 될지 여부를 에측하기 위한 출발점일 뿐입니다. 모델 성능을 향상하기위한 몇개의 가능한 방법이 있습니다. 첫째로, 우선 더 많은 데이터를 모아야 하고 가능하면 기존 누락 필드를 실제 정보로 채워야만 합니다. 또다른 가능성은 Amazon SageMaker의 Automatic Model Tuning기능을 이용한 추가적인 하이퍼파라미터 튜닝입니다. 이 기능의 예제는 [hyperparameter tuning directory of the SageMaker Examples GitHub repository](https://github.com/awslabs/amazon-sagemaker-examples/tree/master/hyperparameter_tuning) 와 Amazon SageMaker 노트북 인스턴스의 **SageMaker Examples** 탭에서 찾을 수 있습니다. 또한 앙상블 학습은 종종 불균형 데이터셋을 잘 처리하지만 다운샘플링이나 합성 데이터증강이나 다른 접근법과 같은, 불균형을 완화할 수 있는 기술을 살펴볼 가치가 있을 수 있습니다. \n", ". \n", "\n", "---\n", "## 정리\n", "\n", "이 노트북이 끝나면 아래 셀을 실행시켜 주세요. 이것은 생성했던 endpoint를 제거하여 남겨진 유휴인스턴스에 요금이 발생되지 않도록 하기 위합니다.\n" ] }, { "cell_type": "code", "execution_count": 33, "metadata": {}, "outputs": [], "source": [ "session.delete_endpoint(xgb_predictor.endpoint)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "conda_python3", "language": "python", "name": "conda_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": "3.6.5" }, "notice": "Copyright 2017 Amazon.com, Inc. or its affiliates. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the \"License\"). You may not use this file except in compliance with the License. A copy of the License is located at http://aws.amazon.com/apache2.0/ or in the \"license\" file accompanying this file. This file is distributed on an \"AS IS\" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License." }, "nbformat": 4, "nbformat_minor": 4 }