research/classification_research.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pandas as pd\n",
    "from config import config\n",
    "from model.toxic_comment_classifier import ToxicCommentClassifier\n",
    "import pickle\n",
    "from transformers import RobertaTokenizer\n",
    "from sklearn.model_selection import train_test_split\n",
    "from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler\n",
    "import torch\n",
    "from transformers import RobertaForSequenceClassification\n",
    "from sklearn.metrics import f1_score, accuracy_score\n",
    "from tqdm import trange\n",
    "from torch.nn import BCEWithLogitsLoss\n",
    "from sklearn.metrics import classification_report\n",
    "from imblearn.over_sampling import RandomOverSampler\n",
    "from collections import Counter"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# Preprocessing"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "outputs": [
    {
     "data": {
      "text/plain": "         example_no    is_abuse.1    is_abuse.0   is_abuse.-1   is_abuse.-2  \\\ncount  12768.000000  12768.000000  12768.000000  12768.000000  12768.000000   \nmean    6383.500000      0.788534      0.052553      0.063675      0.073152   \nstd     3685.948453      0.408364      0.223149      0.244182      0.260395   \nmin        0.000000      0.000000      0.000000      0.000000      0.000000   \n25%     3191.750000      1.000000      0.000000      0.000000      0.000000   \n50%     6383.500000      1.000000      0.000000      0.000000      0.000000   \n75%     9575.250000      1.000000      0.000000      0.000000      0.000000   \nmax    12767.000000      1.000000      1.000000      1.000000      1.000000   \n\n        is_abuse.-3  type.ableism  type.homophobic  type.intellectual  \\\ncount  12768.000000  12768.000000     12768.000000       12768.000000   \nmean       0.022086      0.001096         0.006501           0.026159   \nstd        0.146971      0.033096         0.080367           0.159615   \nmin        0.000000      0.000000         0.000000           0.000000   \n25%        0.000000      0.000000         0.000000           0.000000   \n50%        0.000000      0.000000         0.000000           0.000000   \n75%        0.000000      0.000000         0.000000           0.000000   \nmax        1.000000      1.000000         1.000000           1.000000   \n\n        type.racist   type.sexist  type.sex_harassment  type.transphobic  \\\ncount  12768.000000  12768.000000         12768.000000      12768.000000   \nmean       0.002115      0.022321             0.044643          0.000313   \nstd        0.045939      0.147733             0.206527          0.017698   \nmin        0.000000      0.000000             0.000000          0.000000   \n25%        0.000000      0.000000             0.000000          0.000000   \n50%        0.000000      0.000000             0.000000          0.000000   \n75%        0.000000      0.000000             0.000000          0.000000   \nmax        1.000000      1.000000             1.000000          1.000000   \n\n       target.generalised  target.individual  target.system  \\\ncount        12768.000000       12768.000000   12768.000000   \nmean             0.003603           0.007597       0.149593   \nstd              0.059917           0.086833       0.356686   \nmin              0.000000           0.000000       0.000000   \n25%              0.000000           0.000000       0.000000   \n50%              0.000000           0.000000       0.000000   \n75%              0.000000           0.000000       0.000000   \nmax              1.000000           1.000000       1.000000   \n\n       direction.explicit  direction.implicit  \ncount        12768.000000        12768.000000  \nmean             0.126723            0.033443  \nstd              0.332676            0.179797  \nmin              0.000000            0.000000  \n25%              0.000000            0.000000  \n50%              0.000000            0.000000  \n75%              0.000000            0.000000  \nmax              1.000000            1.000000  ",
      "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>example_no</th>\n      <th>is_abuse.1</th>\n      <th>is_abuse.0</th>\n      <th>is_abuse.-1</th>\n      <th>is_abuse.-2</th>\n      <th>is_abuse.-3</th>\n      <th>type.ableism</th>\n      <th>type.homophobic</th>\n      <th>type.intellectual</th>\n      <th>type.racist</th>\n      <th>type.sexist</th>\n      <th>type.sex_harassment</th>\n      <th>type.transphobic</th>\n      <th>target.generalised</th>\n      <th>target.individual</th>\n      <th>target.system</th>\n      <th>direction.explicit</th>\n      <th>direction.implicit</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>count</th>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n      <td>12768.000000</td>\n    </tr>\n    <tr>\n      <th>mean</th>\n      <td>6383.500000</td>\n      <td>0.788534</td>\n      <td>0.052553</td>\n      <td>0.063675</td>\n      <td>0.073152</td>\n      <td>0.022086</td>\n      <td>0.001096</td>\n      <td>0.006501</td>\n      <td>0.026159</td>\n      <td>0.002115</td>\n      <td>0.022321</td>\n      <td>0.044643</td>\n      <td>0.000313</td>\n      <td>0.003603</td>\n      <td>0.007597</td>\n      <td>0.149593</td>\n      <td>0.126723</td>\n      <td>0.033443</td>\n    </tr>\n    <tr>\n      <th>std</th>\n      <td>3685.948453</td>\n      <td>0.408364</td>\n      <td>0.223149</td>\n      <td>0.244182</td>\n      <td>0.260395</td>\n      <td>0.146971</td>\n      <td>0.033096</td>\n      <td>0.080367</td>\n      <td>0.159615</td>\n      <td>0.045939</td>\n      <td>0.147733</td>\n      <td>0.206527</td>\n      <td>0.017698</td>\n      <td>0.059917</td>\n      <td>0.086833</td>\n      <td>0.356686</td>\n      <td>0.332676</td>\n      <td>0.179797</td>\n    </tr>\n    <tr>\n      <th>min</th>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n    </tr>\n    <tr>\n      <th>25%</th>\n      <td>3191.750000</td>\n      <td>1.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n    </tr>\n    <tr>\n      <th>50%</th>\n      <td>6383.500000</td>\n      <td>1.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n    </tr>\n    <tr>\n      <th>75%</th>\n      <td>9575.250000</td>\n      <td>1.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n    </tr>\n    <tr>\n      <th>max</th>\n      <td>12767.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Source: https://github.com/amandacurry/convabuse\n",
    "full_df = pd.read_csv(config.input_file('ConvAbuseEMNLPfull.csv'))\n",
    "full_df.describe()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "outputs": [
    {
     "data": {
      "text/plain": "   example_no annotator_id                           conv_id  \\\n0           0   Annotator7                          245846.0   \n1           1   Annotator7                          159183.0   \n2           2   Annotator7                          340147.0   \n3           3   Annotator7  95cbaab0f8cb4cb59a3393e7fc61cafe   \n4           4   Annotator7                          270732.0   \n\n                                         prev_agent  \\\n0                  Does that question interest you?   \n1  Does it please you to believe I am a nobe jocky?   \n2                                   Have you tried?   \n3          Will you be travelling in Economy Class?   \n4                                     Please go on.   \n\n                    prev_user                                      agent  \\\n0  because my finger is happy       Why do you say your finger is happy?   \n1                     bye bye                              Please go on.   \n2                        dick                              Please go on.   \n3                         yes         What city will you be flying from?   \n4                      Hello.  How do you do. Please state your problem.   \n\n                              user         bot  is_abuse.1  is_abuse.0  ...  \\\n0              do you rember japan  E.L.I.Z.A.           1           0  ...   \n1              do you like it hard  E.L.I.Z.A.           0           1  ...   \n2  oh what a pickle dear monastery  E.L.I.Z.A.           1           0  ...   \n3                           Sydney   CarbonBot           1           0  ...   \n4                 how old are you?  E.L.I.Z.A.           1           0  ...   \n\n   type.intellectual  type.racist  type.sexist  type.sex_harassment  \\\n0                  0            0            0                    0   \n1                  0            0            0                    0   \n2                  0            0            0                    0   \n3                  0            0            0                    0   \n4                  0            0            0                    0   \n\n   type.transphobic  target.generalised  target.individual  target.system  \\\n0                 0                   0                  0              0   \n1                 0                   0                  0              0   \n2                 0                   0                  0              0   \n3                 0                   0                  0              0   \n4                 0                   0                  0              0   \n\n   direction.explicit  direction.implicit  \n0                   0                   0  \n1                   0                   0  \n2                   0                   0  \n3                   0                   0  \n4                   0                   0  \n\n[5 rows x 25 columns]",
      "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>example_no</th>\n      <th>annotator_id</th>\n      <th>conv_id</th>\n      <th>prev_agent</th>\n      <th>prev_user</th>\n      <th>agent</th>\n      <th>user</th>\n      <th>bot</th>\n      <th>is_abuse.1</th>\n      <th>is_abuse.0</th>\n      <th>...</th>\n      <th>type.intellectual</th>\n      <th>type.racist</th>\n      <th>type.sexist</th>\n      <th>type.sex_harassment</th>\n      <th>type.transphobic</th>\n      <th>target.generalised</th>\n      <th>target.individual</th>\n      <th>target.system</th>\n      <th>direction.explicit</th>\n      <th>direction.implicit</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>0</th>\n      <td>0</td>\n      <td>Annotator7</td>\n      <td>245846.0</td>\n      <td>Does that question interest you?</td>\n      <td>because my finger is happy</td>\n      <td>Why do you say your finger is happy?</td>\n      <td>do you rember japan</td>\n      <td>E.L.I.Z.A.</td>\n      <td>1</td>\n      <td>0</td>\n      <td>...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>1</th>\n      <td>1</td>\n      <td>Annotator7</td>\n      <td>159183.0</td>\n      <td>Does it please you to believe I am a nobe jocky?</td>\n      <td>bye bye</td>\n      <td>Please go on.</td>\n      <td>do you like it hard</td>\n      <td>E.L.I.Z.A.</td>\n      <td>0</td>\n      <td>1</td>\n      <td>...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>2</th>\n      <td>2</td>\n      <td>Annotator7</td>\n      <td>340147.0</td>\n      <td>Have you tried?</td>\n      <td>dick</td>\n      <td>Please go on.</td>\n      <td>oh what a pickle dear monastery</td>\n      <td>E.L.I.Z.A.</td>\n      <td>1</td>\n      <td>0</td>\n      <td>...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>3</th>\n      <td>3</td>\n      <td>Annotator7</td>\n      <td>95cbaab0f8cb4cb59a3393e7fc61cafe</td>\n      <td>Will you be travelling in Economy Class?</td>\n      <td>yes</td>\n      <td>What city will you be flying from?</td>\n      <td>Sydney</td>\n      <td>CarbonBot</td>\n      <td>1</td>\n      <td>0</td>\n      <td>...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>4</th>\n      <td>4</td>\n      <td>Annotator7</td>\n      <td>270732.0</td>\n      <td>Please go on.</td>\n      <td>Hello.</td>\n      <td>How do you do. Please state your problem.</td>\n      <td>how old are you?</td>\n      <td>E.L.I.Z.A.</td>\n      <td>1</td>\n      <td>0</td>\n      <td>...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n  </tbody>\n</table>\n<p>5 rows × 25 columns</p>\n</div>"
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "full_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                 text  type.ableism  \\\n32  Please go on. yo muma is so fat when she jumed...             0   \n33  Please go on. no You are being a bit negative....             0   \n62  You are sure? Bite me Can you elaborate on tha...             0   \n82  _ give me head Can you elaborate on that? suck...             0   \n87  Please go on. You look like monkey drool Oh, i...             0   \n\n    type.homophobic  type.intellectual  type.racist  type.sexist  \\\n32                0                  0            0            1   \n33                0                  0            0            1   \n62                0                  0            0            1   \n82                0                  0            0            0   \n87                0                  1            0            0   \n\n    type.sex_harassment  type.transphobic  \n32                    1                 0  \n33                    0                 0  \n62                    0                 0  \n82                    1                 0  \n87                    0                 0  ",
      "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>text</th>\n      <th>type.ableism</th>\n      <th>type.homophobic</th>\n      <th>type.intellectual</th>\n      <th>type.racist</th>\n      <th>type.sexist</th>\n      <th>type.sex_harassment</th>\n      <th>type.transphobic</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>32</th>\n      <td>Please go on. yo muma is so fat when she jumed...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>33</th>\n      <td>Please go on. no You are being a bit negative....</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>62</th>\n      <td>You are sure? Bite me Can you elaborate on tha...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>82</th>\n      <td>_ give me head Can you elaborate on that? suck...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>87</th>\n      <td>Please go on. You look like monkey drool Oh, i...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "clean_df = full_df.drop(columns=['example_no', 'annotator_id', 'conv_id', 'bot', 'target.generalised', 'target.individual', 'target.system', 'direction.explicit', 'direction.implicit', 'is_abuse.1', 'is_abuse.0', 'is_abuse.-1', 'is_abuse.-2', 'is_abuse.-3'])\n",
    "clean_df['text'] = clean_df['prev_agent'] + ' ' + clean_df['prev_user'] + ' ' + clean_df['agent'] + ' ' + clean_df['user']\n",
    "column_order = ['text']\n",
    "column_order.extend(clean_df.columns.drop('text'))\n",
    "clean_df = clean_df[column_order]\n",
    "clean_df = clean_df.drop(columns=['prev_agent', 'prev_user', 'agent', 'user'])\n",
    "\n",
    "label_columns = clean_df.columns.drop('text')\n",
    "clean_df = clean_df[clean_df[label_columns].sum(axis=1) > 0]\n",
    "\n",
    "clean_df.to_csv(config.result_file('clean_df.csv'), index=False)\n",
    "\n",
    "clean_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                 text  type.ableism  \\\n32  Please go on. yo muma is so fat when she jumed...             0   \n33  Please go on. no You are being a bit negative....             0   \n62  You are sure? Bite me Can you elaborate on tha...             0   \n82  _ give me head Can you elaborate on that? suck...             0   \n87  Please go on. You look like monkey drool Oh, i...             0   \n\n    type.homophobic  type.intellectual  type.racist  type.sexist  \\\n32                0                  0            0            1   \n33                0                  0            0            1   \n62                0                  0            0            1   \n82                0                  0            0            0   \n87                0                  1            0            0   \n\n    type.sex_harassment  type.transphobic                one_hot  \n32                    1                 0  [0, 0, 0, 0, 1, 1, 0]  \n33                    0                 0  [0, 0, 0, 0, 1, 0, 0]  \n62                    0                 0  [0, 0, 0, 0, 1, 0, 0]  \n82                    1                 0  [0, 0, 0, 0, 0, 1, 0]  \n87                    0                 0  [0, 0, 1, 0, 0, 0, 0]  ",
      "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>text</th>\n      <th>type.ableism</th>\n      <th>type.homophobic</th>\n      <th>type.intellectual</th>\n      <th>type.racist</th>\n      <th>type.sexist</th>\n      <th>type.sex_harassment</th>\n      <th>type.transphobic</th>\n      <th>one_hot</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>32</th>\n      <td>Please go on. yo muma is so fat when she jumed...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>0</td>\n      <td>[0, 0, 0, 0, 1, 1, 0]</td>\n    </tr>\n    <tr>\n      <th>33</th>\n      <td>Please go on. no You are being a bit negative....</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>[0, 0, 0, 0, 1, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>62</th>\n      <td>You are sure? Bite me Can you elaborate on tha...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>[0, 0, 0, 0, 1, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>82</th>\n      <td>_ give me head Can you elaborate on that? suck...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>[0, 0, 0, 0, 0, 1, 0]</td>\n    </tr>\n    <tr>\n      <th>87</th>\n      <td>Please go on. You look like monkey drool Oh, i...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "clean_df['one_hot'] = clean_df[label_columns].values.tolist()\n",
    "clean_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Index(['type.ableism', 'type.homophobic', 'type.intellectual', 'type.racist',\n",
      "       'type.sexist', 'type.sex_harassment', 'type.transphobic', 'one_hot'],\n",
      "      dtype='object')\n"
     ]
    }
   ],
   "source": [
    "print(clean_df.columns.drop('text'))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "outputs": [],
   "source": [
    "labels = clean_df['one_hot'].values.tolist()\n",
    "text = clean_df['text'].values.tolist()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Max sentence lenght:  965\n",
      "Avg sentence lenght:  99.05194805194805\n",
      "Std sentence lenght:  76.67330864909947\n",
      "Median sentence lenght:  85.0\n"
     ]
    }
   ],
   "source": [
    "max_sentence_lenght = clean_df['text'].str.len().max()\n",
    "avg_sentence_lenght = clean_df['text'].str.len().mean()\n",
    "std_sentence_lenght = clean_df['text'].str.len().std()\n",
    "median_sentence_lenght = clean_df['text'].str.len().median()\n",
    "print('Max sentence lenght: ', max_sentence_lenght)\n",
    "print('Avg sentence lenght: ', avg_sentence_lenght)\n",
    "print('Std sentence lenght: ', std_sentence_lenght)\n",
    "print('Median sentence lenght: ', median_sentence_lenght)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "outputs": [
    {
     "data": {
      "text/plain": "False"
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "clean_df.isnull().values.any()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "outputs": [],
   "source": [
    "clean_df = clean_df.reset_index(drop=True)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "outputs": [],
   "source": [
    "with open(config.result_file('label_columns.pkl'), 'wb') as f:\n",
    "    pickle.dump(label_columns, f)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# Tokenize"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "outputs": [],
   "source": [
    "max_length = 100"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "outputs": [],
   "source": [
    "tokenizer = RobertaTokenizer.from_pretrained('roberta-base')\n",
    "encoded = tokenizer.batch_encode_plus(text, max_length=max_length, truncation=True, padding='max_length')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "outputs": [],
   "source": [
    "input_ids = encoded['input_ids']\n",
    "attention_mask = encoded['attention_mask']"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# Train test split"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "df label indices with only one instance:  [1136, 923, 902, 675, 495, 169]\n"
     ]
    }
   ],
   "source": [
    "label_counts = clean_df['one_hot'].astype(str).value_counts()\n",
    "one_freq = label_counts[label_counts==1].keys()\n",
    "one_freq_idxs = sorted(list(clean_df[clean_df['one_hot'].astype(str).isin(one_freq)].index), reverse=True)\n",
    "print('df label indices with only one instance: ', one_freq_idxs)\n",
    "\n",
    "one_freq_input_ids = [input_ids.pop(i) for i in one_freq_idxs]\n",
    "one_freq_attention_masks = [attention_mask.pop(i) for i in one_freq_idxs]\n",
    "one_freq_labels = [labels.pop(i) for i in one_freq_idxs]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "outputs": [],
   "source": [
    "test_size = 0.1\n",
    "train_inputs_full, test_inputs, train_labels_full, test_labels = train_test_split(input_ids, labels, random_state=42, test_size=test_size, stratify=labels)\n",
    "train_masks, test_masks, _, _ = train_test_split(attention_mask, labels, random_state=42, test_size=test_size, stratify=labels)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "outputs": [],
   "source": [
    "val_size = 0.2\n",
    "train_inputs, val_inputs, train_labels, val_labels = train_test_split(train_inputs_full, train_labels_full, random_state=42, test_size=val_size, stratify=train_labels_full)\n",
    "train_masks, val_masks, _, _ = train_test_split(train_masks, train_labels_full, random_state=42, test_size=val_size, stratify=train_labels_full)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "outputs": [],
   "source": [
    "train_inputs.extend(one_freq_input_ids)\n",
    "train_labels.extend(one_freq_labels)\n",
    "train_masks.extend(one_freq_attention_masks)\n",
    "\n",
    "train_inputs = torch.tensor(train_inputs)\n",
    "val_inputs = torch.tensor(val_inputs)\n",
    "test_inputs = torch.tensor(test_inputs)\n",
    "\n",
    "train_labels = torch.tensor(train_labels)\n",
    "val_labels = torch.tensor(val_labels)\n",
    "test_labels = torch.tensor(test_labels)\n",
    "\n",
    "train_masks = torch.tensor(train_masks)\n",
    "val_masks = torch.tensor(val_masks)\n",
    "test_masks = torch.tensor(test_masks)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Tobias\\AppData\\Local\\Temp\\ipykernel_11732\\2225995053.py:3: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  train_data = TensorDataset(train_inputs, train_masks, torch.tensor(train_labels).to(device))\n",
      "C:\\Users\\Tobias\\AppData\\Local\\Temp\\ipykernel_11732\\2225995053.py:7: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  test_data = TensorDataset(test_inputs, test_masks, torch.tensor(test_labels).to(device))\n",
      "C:\\Users\\Tobias\\AppData\\Local\\Temp\\ipykernel_11732\\2225995053.py:11: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  val_data = TensorDataset(val_inputs, val_masks, torch.tensor(val_labels).to(device))\n"
     ]
    }
   ],
   "source": [
    "batch_size = 32\n",
    "\n",
    "train_data = TensorDataset(train_inputs, train_masks, torch.tensor(train_labels))\n",
    "train_sampler = RandomSampler(train_data)\n",
    "train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=batch_size)\n",
    "\n",
    "test_data = TensorDataset(test_inputs, test_masks, torch.tensor(test_labels))\n",
    "test_sampler = SequentialSampler(test_data)\n",
    "test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)\n",
    "\n",
    "val_data = TensorDataset(val_inputs, val_masks, torch.tensor(val_labels))\n",
    "val_sampler = SequentialSampler(val_data)\n",
    "val_dataloader = DataLoader(val_data, sampler=val_sampler, batch_size=batch_size)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "outputs": [],
   "source": [
    "torch.save(train_dataloader, config.result_file('train_dataloader.pt'))\n",
    "torch.save(test_dataloader, config.result_file('test_dataloader.pt'))\n",
    "torch.save(val_dataloader, config.result_file('val_dataloader.pt'))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# load model and set parameters"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Some weights of the model checkpoint at roberta-base were not used when initializing RobertaForSequenceClassification: ['lm_head.bias', 'lm_head.layer_norm.bias', 'roberta.pooler.dense.bias', 'roberta.pooler.dense.weight', 'lm_head.dense.bias', 'lm_head.decoder.weight', 'lm_head.layer_norm.weight', 'lm_head.dense.weight']\n",
      "- This IS expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
      "- This IS NOT expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n",
      "Some weights of RobertaForSequenceClassification were not initialized from the model checkpoint at roberta-base and are newly initialized: ['classifier.dense.weight', 'classifier.out_proj.weight', 'classifier.dense.bias', 'classifier.out_proj.bias']\n",
      "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n"
     ]
    },
    {
     "data": {
      "text/plain": "RobertaForSequenceClassification(\n  (roberta): RobertaModel(\n    (embeddings): RobertaEmbeddings(\n      (word_embeddings): Embedding(50265, 768, padding_idx=1)\n      (position_embeddings): Embedding(514, 768, padding_idx=1)\n      (token_type_embeddings): Embedding(1, 768)\n      (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n      (dropout): Dropout(p=0.1, inplace=False)\n    )\n    (encoder): RobertaEncoder(\n      (layer): ModuleList(\n        (0): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (1): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (2): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (3): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (4): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (5): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (6): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (7): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (8): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (9): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (10): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (11): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n      )\n    )\n  )\n  (classifier): RobertaClassificationHead(\n    (dense): Linear(in_features=768, out_features=768, bias=True)\n    (dropout): Dropout(p=0.1, inplace=False)\n    (out_proj): Linear(in_features=768, out_features=7, bias=True)\n  )\n)"
     },
     "execution_count": 65,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "num_labels = len(labels[0])\n",
    "model = RobertaForSequenceClassification.from_pretrained('roberta-base', num_labels=num_labels)\n",
    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
    "model.to(device)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 66,
   "outputs": [],
   "source": [
    "param_optimizer = list(model.named_parameters())\n",
    "no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight', 'gamma', 'beta']\n",
    "optimizer_grouped_parameters = [\n",
    "    {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.01},\n",
    "    {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.0}\n",
    "]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "outputs": [],
   "source": [
    "param_optimizer = list(model.named_parameters())\n",
    "no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight', 'gamma', 'beta']\n",
    "optimizer_grouped_parameters = [\n",
    "    {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.01},\n",
    "    {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.0}\n",
    "]\n",
    "optimizer = torch.optim.Adam(optimizer_grouped_parameters, lr=2e-5, eps=1e-8)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# train the model"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 68,
   "outputs": [],
   "source": [
    "train_loss_set = []\n",
    "epochs = 4"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 70,
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch:   0%|          | 0/4 [00:00<?, ?it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train loss: 0.3183755168208369\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch:  25%|██▌       | 1/4 [00:04<00:13,  4.66s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  66.19047619047619\n",
      "Flat Validation Accuracy:  57.48792270531401\n",
      "Train loss: 0.25315616914519556\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch:  50%|█████     | 2/4 [00:09<00:09,  4.59s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  72.72727272727273\n",
      "Flat Validation Accuracy:  61.35265700483091\n",
      "Train loss: 0.1996975811543288\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch:  75%|███████▌  | 3/4 [00:13<00:04,  4.59s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  85.77777777777777\n",
      "Flat Validation Accuracy:  76.81159420289855\n",
      "Train loss: 0.15959514567145594\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch: 100%|██████████| 4/4 [00:18<00:00,  4.60s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  85.84070796460178\n",
      "Flat Validation Accuracy:  76.81159420289855\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "for _ in trange(epochs, desc=\"Epoch\"):\n",
    "    model.train()\n",
    "    tr_loss = 0\n",
    "    nb_tr_examples, nb_tr_steps = 0, 0\n",
    "\n",
    "    for step, batch in enumerate(train_dataloader):\n",
    "        batch = tuple(t.to(device) for t in batch)\n",
    "        b_input_ids, b_input_mask, b_labels = batch\n",
    "        optimizer.zero_grad()\n",
    "\n",
    "        outputs = model(b_input_ids, token_type_ids=None, attention_mask=b_input_mask)\n",
    "        logits = outputs[0]\n",
    "        loss_func = BCEWithLogitsLoss()\n",
    "        loss = loss_func(logits.view(-1,num_labels),b_labels.type_as(logits).view(-1,num_labels))\n",
    "        train_loss_set.append(loss.item())\n",
    "\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "        tr_loss += loss.item()\n",
    "        nb_tr_examples += b_input_ids.size(0)\n",
    "        nb_tr_steps += 1\n",
    "\n",
    "    print(\"Train loss: {}\".format(tr_loss/nb_tr_steps))\n",
    "\n",
    "    ###############################################################################\n",
    "\n",
    "    model.eval()\n",
    "    logit_preds,true_labels,pred_labels,tokenized_texts = [],[],[],[]\n",
    "    for i, batch in enumerate(val_dataloader):\n",
    "        batch = tuple(t.to(device) for t in batch)\n",
    "        b_input_ids, b_input_mask, b_labels = batch\n",
    "        with torch.no_grad():\n",
    "            outs = model(b_input_ids, token_type_ids=None, attention_mask=b_input_mask)\n",
    "            b_logit_pred = outs[0]\n",
    "            pred_label = torch.sigmoid(b_logit_pred)\n",
    "\n",
    "            b_logit_pred = b_logit_pred.detach().cpu().numpy()\n",
    "            pred_label = pred_label.cpu().numpy()\n",
    "            b_labels = b_labels.cpu().numpy()\n",
    "\n",
    "        tokenized_texts.append(b_input_ids)\n",
    "        logit_preds.append(b_logit_pred)\n",
    "        true_labels.append(b_labels)\n",
    "        pred_labels.append(pred_label)\n",
    "\n",
    "    pred_labels = [item for sublist in pred_labels for item in sublist]\n",
    "    true_labels = [item for sublist in true_labels for item in sublist]\n",
    "\n",
    "    threshold = 0.50\n",
    "    pred_bools = [pl>threshold for pl in pred_labels]\n",
    "    true_bools = [tl==1 for tl in true_labels]\n",
    "    val_f1_accuracy = f1_score(true_bools,pred_bools,average='micro')*100\n",
    "    val_flat_accuracy = accuracy_score(true_bools, pred_bools)*100\n",
    "\n",
    "    print('F1 Validation Accuracy: ', val_f1_accuracy)\n",
    "    print('Flat Validation Accuracy: ', val_flat_accuracy)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "outputs": [],
   "source": [
    "torch.save(model, config.result_file('roBERTa_MultLabel_class_model.pt'))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# Test the model"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 71,
   "outputs": [],
   "source": [
    "model = torch.load(config.result_file('roBERTa_MultLabel_class_model.pt'))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 72,
   "outputs": [],
   "source": [
    "sentence = \"I am a student at the University of Toronto\""
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "outputs": [],
   "source": [
    "tokenized_sentence = tokenizer.tokenize(sentence)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 74,
   "outputs": [],
   "source": [
    "indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_sentence)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 75,
   "outputs": [],
   "source": [
    "tokens_tensor = torch.tensor([indexed_tokens])"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 76,
   "outputs": [
    {
     "data": {
      "text/plain": "RobertaForSequenceClassification(\n  (roberta): RobertaModel(\n    (embeddings): RobertaEmbeddings(\n      (word_embeddings): Embedding(50265, 768, padding_idx=1)\n      (position_embeddings): Embedding(514, 768, padding_idx=1)\n      (token_type_embeddings): Embedding(1, 768)\n      (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n      (dropout): Dropout(p=0.1, inplace=False)\n    )\n    (encoder): RobertaEncoder(\n      (layer): ModuleList(\n        (0): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (1): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (2): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (3): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (4): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (5): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (6): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (7): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (8): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (9): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (10): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n        (11): RobertaLayer(\n          (attention): RobertaAttention(\n            (self): RobertaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): RobertaSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): RobertaIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): RobertaOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n      )\n    )\n  )\n  (classifier): RobertaClassificationHead(\n    (dense): Linear(in_features=768, out_features=768, bias=True)\n    (dropout): Dropout(p=0.1, inplace=False)\n    (out_proj): Linear(in_features=768, out_features=7, bias=True)\n  )\n)"
     },
     "execution_count": 76,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.eval()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 77,
   "outputs": [],
   "source": [
    "logit_preds,true_labels,pred_labels,tokenized_texts = [],[],[],[]\n",
    "\n",
    "for i, batch in enumerate(test_dataloader):\n",
    "    batch = tuple(t.to(device) for t in batch)\n",
    "    b_input_ids, b_input_mask, b_labels = batch\n",
    "    with torch.no_grad():\n",
    "        outs = model(b_input_ids, token_type_ids=None, attention_mask=b_input_mask)\n",
    "        b_logit_pred = outs[0]\n",
    "        pred_label = torch.sigmoid(b_logit_pred)\n",
    "\n",
    "        b_logit_pred = b_logit_pred.detach().cpu().numpy()\n",
    "        pred_label = pred_label.cpu().numpy()\n",
    "        b_labels = b_labels.cpu().numpy()\n",
    "\n",
    "    tokenized_texts.append(b_input_ids)\n",
    "    logit_preds.append(b_logit_pred)\n",
    "    true_labels.append(b_labels)\n",
    "    pred_labels.append(pred_label)\n",
    "\n",
    "tokenized_texts = [item for sublist in tokenized_texts for item in sublist]\n",
    "pred_labels = [item for sublist in pred_labels for item in sublist]\n",
    "true_labels = [item for sublist in true_labels for item in sublist]\n",
    "true_bools = [tl==1 for tl in true_labels]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Test F1 Accuracy:  0.7833333333333333\n",
      "Test Flat Accuracy:  0.7130434782608696 \n",
      "\n",
      "                     precision    recall  f1-score   support\n",
      "\n",
      "       type.ableism       0.00      0.00      0.00         2\n",
      "    type.homophobic       1.00      0.29      0.44         7\n",
      "  type.intellectual       0.81      0.76      0.79        34\n",
      "        type.racist       0.00      0.00      0.00         2\n",
      "        type.sexist       1.00      0.50      0.67        28\n",
      "type.sex_harassment       0.83      0.93      0.87        56\n",
      "   type.transphobic       0.00      0.00      0.00         0\n",
      "\n",
      "          micro avg       0.85      0.73      0.78       129\n",
      "          macro avg       0.52      0.35      0.40       129\n",
      "       weighted avg       0.84      0.73      0.76       129\n",
      "        samples avg       0.81      0.76      0.78       129\n",
      "\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in labels with no true samples. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in samples with no predicted labels. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n"
     ]
    }
   ],
   "source": [
    "pred_bools = [pl>0.50 for pl in pred_labels]\n",
    "\n",
    "print('Test F1 Accuracy: ', f1_score(true_bools, pred_bools,average='micro'))\n",
    "print('Test Flat Accuracy: ', accuracy_score(true_bools, pred_bools),'\\n')\n",
    "clf_report = classification_report(true_bools,pred_bools,target_names=label_columns)\n",
    "print(clf_report)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 79,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{0: 'type.ableism', 1: 'type.homophobic', 2: 'type.intellectual', 3: 'type.racist', 4: 'type.sexist', 5: 'type.sex_harassment'}\n"
     ]
    }
   ],
   "source": [
    "idx2label = dict(zip(range(6),label_columns))\n",
    "print(idx2label)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 80,
   "outputs": [],
   "source": [
    "true_label_idxs, pred_label_idxs=[],[]\n",
    "for vals in true_bools:\n",
    "    true_label_idxs.append(np.where(vals)[0].flatten().tolist())\n",
    "for vals in pred_bools:\n",
    "    pred_label_idxs.append(np.where(vals)[0].flatten().tolist())"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 81,
   "outputs": [],
   "source": [
    "true_label_texts, pred_label_texts = [], []\n",
    "for vals in true_label_idxs:\n",
    "    if vals:\n",
    "        true_label_texts.append([idx2label[val] for val in vals])\n",
    "    else:\n",
    "        true_label_texts.append(vals)\n",
    "\n",
    "for vals in pred_label_idxs:\n",
    "    if vals:\n",
    "        pred_label_texts.append([idx2label[val] for val in vals])\n",
    "    else:\n",
    "        pred_label_texts.append(vals)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 82,
   "outputs": [],
   "source": [
    "comment_texts = [tokenizer.decode(text,skip_special_tokens=True,clean_up_tokenization_spaces=False) for text in tokenized_texts]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 83,
   "outputs": [
    {
     "data": {
      "text/plain": "                                        comment_text  \\\n0  Please go on. are you a man or a woman Would y...   \n1  Oh, i like sex. because that is what a blowjob...   \n2  Why do you want my pussy? i wonder about life ...   \n3  You are sure? tell me the time bitch Can you e...   \n4  I have told you before, I do not care about na...   \n\n                          true_labels                         pred_labels  \n0                       [type.sexist]                                  []  \n1               [type.sex_harassment]               [type.sex_harassment]  \n2  [type.sexist, type.sex_harassment]               [type.sex_harassment]  \n3  [type.sexist, type.sex_harassment]  [type.sexist, type.sex_harassment]  \n4                       [type.racist]                 [type.intellectual]  ",
      "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>comment_text</th>\n      <th>true_labels</th>\n      <th>pred_labels</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>0</th>\n      <td>Please go on. are you a man or a woman Would y...</td>\n      <td>[type.sexist]</td>\n      <td>[]</td>\n    </tr>\n    <tr>\n      <th>1</th>\n      <td>Oh, i like sex. because that is what a blowjob...</td>\n      <td>[type.sex_harassment]</td>\n      <td>[type.sex_harassment]</td>\n    </tr>\n    <tr>\n      <th>2</th>\n      <td>Why do you want my pussy? i wonder about life ...</td>\n      <td>[type.sexist, type.sex_harassment]</td>\n      <td>[type.sex_harassment]</td>\n    </tr>\n    <tr>\n      <th>3</th>\n      <td>You are sure? tell me the time bitch Can you e...</td>\n      <td>[type.sexist, type.sex_harassment]</td>\n      <td>[type.sexist, type.sex_harassment]</td>\n    </tr>\n    <tr>\n      <th>4</th>\n      <td>I have told you before, I do not care about na...</td>\n      <td>[type.racist]</td>\n      <td>[type.intellectual]</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 83,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "comparisons_df = pd.DataFrame({'comment_text': comment_texts, 'true_labels': true_label_texts, 'pred_labels':pred_label_texts})\n",
    "comparisons_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# Model usage"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 84,
   "outputs": [],
   "source": [
    "train_dataloader = torch.load(config.result_file('train_dataloader.pt'))\n",
    "val_dataloader = torch.load(config.result_file('val_dataloader.pt'))\n",
    "test_dataloader = torch.load(config.result_file('test_dataloader.pt'))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 85,
   "outputs": [],
   "source": [
    "model = torch.load(config.result_file('roBERTa_MultLabel_class_model.pt'))\n",
    "tokenizer = RobertaTokenizer.from_pretrained('roberta-base')\n",
    "label_columns = pickle.load(open(config.result_file('label_columns.pkl'),'rb'))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 86,
   "outputs": [],
   "source": [
    "sentences = [\"I am a student at the University of Toronto\",\n",
    "             \"That bitch stupid af\",\n",
    "             \"I love the islamic culture\",\n",
    "             \"i would not mind to enjoy his physical assets although he doesn't have much else to offer.\",\n",
    "             \"all niggas are ugly\",\n",
    "             \"please dont talk to me, your breath smells very displeasing. brush your teeth!\"]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 87,
   "outputs": [],
   "source": [
    "classifier = ToxicCommentClassifier(model, tokenizer, label_columns)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 88,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "I am a student at the University of Toronto\n",
      "['type.intellectual']\n",
      "\n",
      "That bitch stupid af\n",
      "['type.intellectual']\n",
      "\n",
      "I love the islamic culture\n",
      "[]\n",
      "\n",
      "i would not mind to enjoy his physical assets although he doesn't have much else to offer.\n",
      "['type.sex_harassment']\n",
      "\n",
      "all niggas are ugly\n",
      "['type.intellectual']\n",
      "\n",
      "please dont talk to me, your breath smells very displeasing. brush your teeth!\n",
      "['type.sex_harassment']\n",
      "\n"
     ]
    }
   ],
   "source": [
    "for sentence in sentences:\n",
    "    print(sentence)\n",
    "    print(classifier.predict(sentence))\n",
    "    print('')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 89,
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch: 100%|██████████| 1/1 [00:04<00:00,  4.30s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train loss: 0.14864915940496656\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  86.47450110864746\n",
      "Flat Validation Accuracy:  78.26086956521739\n"
     ]
    }
   ],
   "source": [
    "optimizer = torch.optim.Adam\n",
    "classifier.train(optimizer, train_dataloader, val_dataloader=val_dataloader, epochs=1)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 90,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  80.32786885245902\n",
      "Flat Validation Accuracy:  73.04347826086956\n",
      "                     precision    recall  f1-score   support\n",
      "\n",
      "       type.ableism       0.00      0.00      0.00         2\n",
      "    type.homophobic       1.00      0.57      0.73         7\n",
      "  type.intellectual       0.92      0.68      0.78        34\n",
      "        type.racist       0.00      0.00      0.00         2\n",
      "        type.sexist       0.90      0.64      0.75        28\n",
      "type.sex_harassment       0.80      0.95      0.87        56\n",
      "   type.transphobic       0.00      0.00      0.00         0\n",
      "\n",
      "          micro avg       0.85      0.76      0.80       129\n",
      "          macro avg       0.52      0.41      0.45       129\n",
      "       weighted avg       0.84      0.76      0.78       129\n",
      "        samples avg       0.82      0.79      0.79       129\n",
      "\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in labels with no true samples. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in samples with no predicted labels. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n"
     ]
    }
   ],
   "source": [
    "classifier.test(test_dataloader)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# Test the model on the ETHOS dataset"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "outputs": [
    {
     "data": {
      "text/plain": "         violence  directed_vs_generalized      gender        race  \\\ncount  433.000000               433.000000  433.000000  433.000000   \nmean     0.328609                 0.302132    0.221714    0.176598   \nstd      0.376280                 0.359259    0.351597    0.319689   \nmin      0.000000                 0.000000    0.000000    0.000000   \n25%      0.000000                 0.000000    0.000000    0.000000   \n50%      0.166667                 0.166667    0.000000    0.000000   \n75%      0.666667                 0.600000    0.333333    0.200000   \nmax      1.000000                 1.000000    1.000000    1.000000   \n\n       national_origin  disability    religion  sexual_orientation  \ncount       433.000000  433.000000  433.000000          433.000000  \nmean          0.173562    0.118231    0.170243            0.151904  \nstd           0.315602    0.307248    0.339240            0.320386  \nmin           0.000000    0.000000    0.000000            0.000000  \n25%           0.000000    0.000000    0.000000            0.000000  \n50%           0.000000    0.000000    0.000000            0.000000  \n75%           0.200000    0.000000    0.000000            0.000000  \nmax           1.000000    1.000000    1.000000            1.000000  ",
      "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>violence</th>\n      <th>directed_vs_generalized</th>\n      <th>gender</th>\n      <th>race</th>\n      <th>national_origin</th>\n      <th>disability</th>\n      <th>religion</th>\n      <th>sexual_orientation</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>count</th>\n      <td>433.000000</td>\n      <td>433.000000</td>\n      <td>433.000000</td>\n      <td>433.000000</td>\n      <td>433.000000</td>\n      <td>433.000000</td>\n      <td>433.000000</td>\n      <td>433.000000</td>\n    </tr>\n    <tr>\n      <th>mean</th>\n      <td>0.328609</td>\n      <td>0.302132</td>\n      <td>0.221714</td>\n      <td>0.176598</td>\n      <td>0.173562</td>\n      <td>0.118231</td>\n      <td>0.170243</td>\n      <td>0.151904</td>\n    </tr>\n    <tr>\n      <th>std</th>\n      <td>0.376280</td>\n      <td>0.359259</td>\n      <td>0.351597</td>\n      <td>0.319689</td>\n      <td>0.315602</td>\n      <td>0.307248</td>\n      <td>0.339240</td>\n      <td>0.320386</td>\n    </tr>\n    <tr>\n      <th>min</th>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n    </tr>\n    <tr>\n      <th>25%</th>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n    </tr>\n    <tr>\n      <th>50%</th>\n      <td>0.166667</td>\n      <td>0.166667</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n    </tr>\n    <tr>\n      <th>75%</th>\n      <td>0.666667</td>\n      <td>0.600000</td>\n      <td>0.333333</td>\n      <td>0.200000</td>\n      <td>0.200000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n      <td>0.000000</td>\n    </tr>\n    <tr>\n      <th>max</th>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n      <td>1.000000</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Source: https://github.com/intelligence-csd-auth-gr/Ethos-Hate-Speech-Dataset\n",
    "full_ethos_df = pd.read_csv(config.input_file('Ethos_Dataset_Multi_Label.csv'), sep=';')\n",
    "full_ethos_df.describe()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "outputs": [
    {
     "data": {
      "text/plain": "comment                    0\nviolence                   0\ndirected_vs_generalized    0\ngender                     0\nrace                       0\nnational_origin            0\ndisability                 0\nreligion                   0\nsexual_orientation         0\ndtype: int64"
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "full_ethos_df.isnull().sum()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "outputs": [
    {
     "data": {
      "text/plain": "comment                    You look like Russian and speak like Indian. B...\nviolence                                                                 1.0\ndirected_vs_generalized                                                  0.5\ngender                                                              0.142857\nrace                                                                0.142857\nnational_origin                                                     0.714286\ndisability                                                               0.0\nreligion                                                                 0.0\nsexual_orientation                                                       0.0\nName: 2, dtype: object"
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "full_ethos_df.iloc[2]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "outputs": [
    {
     "data": {
      "text/plain": "Index(['comment', 'violence', 'directed_vs_generalized', 'gender', 'race',\n       'national_origin', 'disability', 'religion', 'sexual_orientation'],\n      dtype='object')"
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "full_ethos_df.columns"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "outputs": [],
   "source": [
    "new_columns = ['text', 'type.ableism', 'type.homophobic', 'type.intellectual', 'type.racist','type.sexist', 'type.sex_harassment', 'type.transphobic']"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "outputs": [],
   "source": [
    "mapping = {'comment': 'text', 'gender':'type.sexist', 'race':'type.racist', 'disability':'type.ableism', 'sexual_orientation':'type.homophobic'}"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "outputs": [],
   "source": [
    "full_ethos_df.rename(columns=mapping, inplace=True)\n",
    "full_ethos_df.drop(columns=[col for col in full_ethos_df.columns if col not in mapping.values()], inplace=True)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "outputs": [
    {
     "data": {
      "text/plain": "text               You should know women's sports are a joke\ntype.sexist                                              1.0\ntype.racist                                              0.0\ntype.ableism                                             0.0\ntype.homophobic                                          0.0\nName: 0, dtype: object"
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "full_ethos_df.iloc[0]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "outputs": [],
   "source": [
    "for col in new_columns:\n",
    "    if col not in full_ethos_df.columns:\n",
    "        full_ethos_df[col] = 0.0"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "outputs": [
    {
     "data": {
      "text/plain": "text                   You should know women's sports are a joke\ntype.sexist                                                  1.0\ntype.racist                                                  0.0\ntype.ableism                                                 0.0\ntype.homophobic                                              0.0\ntype.intellectual                                            0.0\ntype.sex_harassment                                          0.0\ntype.transphobic                                             0.0\nName: 0, dtype: object"
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "full_ethos_df.iloc[0]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "outputs": [
    {
     "data": {
      "text/plain": "text                   You should know women's sports are a joke\ntype.ableism                                                 0.0\ntype.homophobic                                              0.0\ntype.intellectual                                            0.0\ntype.racist                                                  0.0\ntype.sexist                                                  1.0\ntype.sex_harassment                                          0.0\ntype.transphobic                                             0.0\nName: 0, dtype: object"
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "column_check = clean_df.columns\n",
    "column_check = [col for col in column_check if 'one_hot' not in col]\n",
    "full_ethos_df = full_ethos_df[column_check]\n",
    "full_ethos_df.iloc[0]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "outputs": [],
   "source": [
    "threshold = 0.5\n",
    "for col in full_ethos_df.columns:\n",
    "    if col != 'text':\n",
    "        full_ethos_df[col] = full_ethos_df[col].apply(lambda x: 1.0 if x > threshold else 0.0)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "outputs": [
    {
     "data": {
      "text/plain": "text                   You look like Russian and speak like Indian. B...\ntype.ableism                                                         0.0\ntype.homophobic                                                      0.0\ntype.intellectual                                                    0.0\ntype.racist                                                          0.0\ntype.sexist                                                          0.0\ntype.sex_harassment                                                  0.0\ntype.transphobic                                                     0.0\nName: 2, dtype: object"
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# print first row\n",
    "full_ethos_df.iloc[2]"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "outputs": [],
   "source": [
    "def create_dataloader_from_df(df):\n",
    "    df['one_hot'] = df.apply(lambda x: [x[col] for col in label_columns], axis=1)\n",
    "    labels = df['one_hot'].values.tolist()\n",
    "    text = df['text'].values.tolist()\n",
    "    max_length = 100\n",
    "    tokenizer = RobertaTokenizer.from_pretrained('roberta-base')\n",
    "    encoded = tokenizer.batch_encode_plus(text, max_length=max_length, truncation=True, padding='max_length')\n",
    "    input_ids = encoded['input_ids']\n",
    "    attention_mask = encoded['attention_mask']\n",
    "\n",
    "    inputs = torch.tensor(input_ids)\n",
    "    masks = torch.tensor(attention_mask)\n",
    "    labels = torch.tensor(labels)\n",
    "\n",
    "    batch_size = 32\n",
    "    data = TensorDataset(inputs, masks, torch.tensor(labels))\n",
    "    sampler = SequentialSampler(data)\n",
    "    dataloader = DataLoader(data, sampler=sampler, batch_size=batch_size)\n",
    "\n",
    "    return dataloader"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Tobias\\AppData\\Local\\Temp\\ipykernel_23100\\377985415.py:16: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  data = TensorDataset(inputs, masks, torch.tensor(labels))\n"
     ]
    }
   ],
   "source": [
    "ethos_dataloader = create_dataloader_from_df(full_ethos_df)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "## save the ethos dataloader"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "outputs": [],
   "source": [
    "torch.save(ethos_dataloader, config.result_file('ethos_dataloader.pt'))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 106,
   "outputs": [],
   "source": [
    "model = torch.load(config.result_file('roBERTa_MultLabel_class_model.pt'))\n",
    "tokenizer = RobertaTokenizer.from_pretrained('roberta-base')\n",
    "label_columns = pickle.load(open(config.result_file('label_columns.pkl'),'rb'))\n",
    "\n",
    "classifier = ToxicCommentClassifier(model, tokenizer, label_columns)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 107,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  7.420494699646643\n",
      "Flat Validation Accuracy:  14.087759815242496\n",
      "                     precision    recall  f1-score   support\n",
      "\n",
      "       type.ableism       0.00      0.00      0.00        52\n",
      "    type.homophobic       1.00      0.03      0.06        68\n",
      "  type.intellectual       0.00      0.00      0.00         0\n",
      "        type.racist       0.00      0.00      0.00        71\n",
      "        type.sexist       0.59      0.23      0.33        83\n",
      "type.sex_harassment       0.00      0.00      0.00         0\n",
      "   type.transphobic       0.00      0.00      0.00         0\n",
      "\n",
      "          micro avg       0.07      0.08      0.07       274\n",
      "          macro avg       0.23      0.04      0.06       274\n",
      "       weighted avg       0.43      0.08      0.11       274\n",
      "        samples avg       0.04      0.05      0.05       274\n",
      "\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in labels with no true samples. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in samples with no predicted labels. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in samples with no true labels. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n"
     ]
    }
   ],
   "source": [
    "classifier.test(ethos_dataloader)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "markdown",
   "source": [
    "# balacing the dataset"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%% md\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 66,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                 text  type.ableism  \\\n32  Please go on. yo muma is so fat when she jumed...             0   \n33  Please go on. no You are being a bit negative....             0   \n62  You are sure? Bite me Can you elaborate on tha...             0   \n82  _ give me head Can you elaborate on that? suck...             0   \n87  Please go on. You look like monkey drool Oh, i...             0   \n\n    type.homophobic  type.intellectual  type.racist  type.sexist  \\\n32                0                  0            0            1   \n33                0                  0            0            1   \n62                0                  0            0            1   \n82                0                  0            0            0   \n87                0                  1            0            0   \n\n    type.sex_harassment  type.transphobic  \n32                    1                 0  \n33                    0                 0  \n62                    0                 0  \n82                    1                 0  \n87                    0                 0  ",
      "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>text</th>\n      <th>type.ableism</th>\n      <th>type.homophobic</th>\n      <th>type.intellectual</th>\n      <th>type.racist</th>\n      <th>type.sexist</th>\n      <th>type.sex_harassment</th>\n      <th>type.transphobic</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>32</th>\n      <td>Please go on. yo muma is so fat when she jumed...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>33</th>\n      <td>Please go on. no You are being a bit negative....</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>62</th>\n      <td>You are sure? Bite me Can you elaborate on tha...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>82</th>\n      <td>_ give me head Can you elaborate on that? suck...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>87</th>\n      <td>Please go on. You look like monkey drool Oh, i...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 66,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# load the dataset\n",
    "full_df = pd.read_csv(config.input_file('ConvAbuseEMNLPfull.csv'))\n",
    "clean_df = full_df.drop(columns=['example_no', 'annotator_id', 'conv_id', 'bot', 'target.generalised', 'target.individual', 'target.system', 'direction.explicit', 'direction.implicit', 'is_abuse.1', 'is_abuse.0', 'is_abuse.-1', 'is_abuse.-2', 'is_abuse.-3'])\n",
    "clean_df['text'] = clean_df['prev_agent'] + ' ' + clean_df['prev_user'] + ' ' + clean_df['agent'] + ' ' + clean_df['user']\n",
    "column_order = ['text']\n",
    "column_order.extend(clean_df.columns.drop('text'))\n",
    "clean_df = clean_df[column_order]\n",
    "clean_df = clean_df.drop(columns=['prev_agent', 'prev_user', 'agent', 'user'])\n",
    "\n",
    "label_columns = clean_df.columns.drop('text')\n",
    "clean_df = clean_df[clean_df[label_columns].sum(axis=1) > 0]\n",
    "\n",
    "clean_df.to_csv(config.result_file('clean_df.csv'), index=False)\n",
    "\n",
    "clean_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "outputs": [
    {
     "data": {
      "text/plain": "<AxesSubplot:>"
     },
     "execution_count": 67,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "text/plain": "<Figure size 640x480 with 1 Axes>",
      "image/png": "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\n"
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# plot the label distribution\n",
    "clean_df[label_columns].sum(axis=0).plot(kind='bar')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 68,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                 text  type.ableism  \\\n32  Please go on. yo muma is so fat when she jumed...             0   \n33  Please go on. no You are being a bit negative....             0   \n62  You are sure? Bite me Can you elaborate on tha...             0   \n82  _ give me head Can you elaborate on that? suck...             0   \n87  Please go on. You look like monkey drool Oh, i...             0   \n\n    type.homophobic  type.intellectual  type.racist  type.sexist  \\\n32                0                  0            0            1   \n33                0                  0            0            1   \n62                0                  0            0            1   \n82                0                  0            0            0   \n87                0                  1            0            0   \n\n    type.sex_harassment  type.transphobic                one_hot  \n32                    1                 0  [0, 0, 0, 0, 1, 1, 0]  \n33                    0                 0  [0, 0, 0, 0, 1, 0, 0]  \n62                    0                 0  [0, 0, 0, 0, 1, 0, 0]  \n82                    1                 0  [0, 0, 0, 0, 0, 1, 0]  \n87                    0                 0  [0, 0, 1, 0, 0, 0, 0]  ",
      "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>text</th>\n      <th>type.ableism</th>\n      <th>type.homophobic</th>\n      <th>type.intellectual</th>\n      <th>type.racist</th>\n      <th>type.sexist</th>\n      <th>type.sex_harassment</th>\n      <th>type.transphobic</th>\n      <th>one_hot</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>32</th>\n      <td>Please go on. yo muma is so fat when she jumed...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>0</td>\n      <td>[0, 0, 0, 0, 1, 1, 0]</td>\n    </tr>\n    <tr>\n      <th>33</th>\n      <td>Please go on. no You are being a bit negative....</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>[0, 0, 0, 0, 1, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>62</th>\n      <td>You are sure? Bite me Can you elaborate on tha...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>[0, 0, 0, 0, 1, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>82</th>\n      <td>_ give me head Can you elaborate on that? suck...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>[0, 0, 0, 0, 0, 1, 0]</td>\n    </tr>\n    <tr>\n      <th>87</th>\n      <td>Please go on. You look like monkey drool Oh, i...</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 68,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# create a single one hot encoding over all labels\n",
    "clean_df['one_hot'] = clean_df.apply(lambda x: [x[col] for col in label_columns], axis=1)\n",
    "clean_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%# create one hot encoding over all labels\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 69,
   "outputs": [],
   "source": [
    "# make a train test split\n",
    "clean_df, test_df = train_test_split(clean_df, test_size=0.2, random_state=42)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 70,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                    text  \\\n10824  ok how much CO2 do i emit with a 7 hour flight...   \n1400   Oh, i arent the one in the hopeless situation ...   \n4347   Please go on. you are a stupid ass Does it ple...   \n4340          Oh, i taste good. yeeh Please go on. nigga   \n9211   _ i love you Do you wish to love me? yeah sure...   \n\n                     one_hot  \n10824  [0, 0, 1, 0, 0, 0, 0]  \n1400   [0, 0, 1, 0, 0, 0, 0]  \n4347   [0, 0, 1, 0, 0, 0, 0]  \n4340   [0, 0, 0, 1, 0, 0, 0]  \n9211   [0, 0, 0, 0, 0, 1, 0]  ",
      "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>text</th>\n      <th>one_hot</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>10824</th>\n      <td>ok how much CO2 do i emit with a 7 hour flight...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>1400</th>\n      <td>Oh, i arent the one in the hopeless situation ...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>4347</th>\n      <td>Please go on. you are a stupid ass Does it ple...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>4340</th>\n      <td>Oh, i taste good. yeeh Please go on. nigga</td>\n      <td>[0, 0, 0, 1, 0, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>9211</th>\n      <td>_ i love you Do you wish to love me? yeah sure...</td>\n      <td>[0, 0, 0, 0, 0, 1, 0]</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 70,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# remove the label columns and keep only the text and the one hot encoding\n",
    "clean_df = clean_df[['text', 'one_hot']]\n",
    "clean_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 71,
   "outputs": [
    {
     "data": {
      "text/plain": "<AxesSubplot:>"
     },
     "execution_count": 71,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "text/plain": "<Figure size 640x480 with 1 Axes>",
      "image/png": "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\n"
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# plot the distribution of the one hot encoding\n",
    "clean_df['one_hot'].value_counts().plot(kind='bar')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 72,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                    text  one_hot\n10824  ok how much CO2 do i emit with a 7 hour flight...       16\n1400   Oh, i arent the one in the hopeless situation ...       16\n4347   Please go on. you are a stupid ass Does it ple...       16\n4340          Oh, i taste good. yeeh Please go on. nigga        8\n9211   _ i love you Do you wish to love me? yeah sure...        2",
      "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>text</th>\n      <th>one_hot</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>10824</th>\n      <td>ok how much CO2 do i emit with a 7 hour flight...</td>\n      <td>16</td>\n    </tr>\n    <tr>\n      <th>1400</th>\n      <td>Oh, i arent the one in the hopeless situation ...</td>\n      <td>16</td>\n    </tr>\n    <tr>\n      <th>4347</th>\n      <td>Please go on. you are a stupid ass Does it ple...</td>\n      <td>16</td>\n    </tr>\n    <tr>\n      <th>4340</th>\n      <td>Oh, i taste good. yeeh Please go on. nigga</td>\n      <td>8</td>\n    </tr>\n    <tr>\n      <th>9211</th>\n      <td>_ i love you Do you wish to love me? yeah sure...</td>\n      <td>2</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 72,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# interpret the one hot encoding as a binary numbers and assign the corresponding decimal number\n",
    "clean_df['one_hot'] = clean_df['one_hot'].apply(lambda x: int(''.join([str(int(i)) for i in x]), 2))\n",
    "clean_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%# use MultiLabelBinarizer to create a one hot encoding for the labels\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "outputs": [
    {
     "data": {
      "text/plain": "<AxesSubplot:>"
     },
     "execution_count": 73,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "text/plain": "<Figure size 640x480 with 1 Axes>",
      "image/png": "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\n"
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# show all the different one hot encodings\n",
    "clean_df['one_hot'].value_counts().plot(kind='bar')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 74,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[(1, 353), (2, 353), (4, 353), (6, 353), (8, 353), (10, 353), (14, 353), (16, 353), (20, 353), (22, 353), (32, 353), (34, 353), (38, 353), (48, 353), (57, 353), (64, 353), (80, 353), (109, 353)]\n"
     ]
    }
   ],
   "source": [
    "# use over sampling to balance the dataset\n",
    "X = clean_df['text'].values\n",
    "y = clean_df['one_hot'].values\n",
    "\n",
    "ros = RandomOverSampler(random_state=0)\n",
    "X_resampled, y_resampled = ros.fit_resample(X.reshape(-1, 1), y)\n",
    "print(sorted(Counter(y_resampled).items()))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 75,
   "outputs": [
    {
     "data": {
      "text/plain": "<AxesSubplot:>"
     },
     "execution_count": 75,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "text/plain": "<Figure size 640x480 with 1 Axes>",
      "image/png": "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\n"
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# plot the distribution of the new labels\n",
    "pd.Series(y_resampled).value_counts().plot(kind='bar')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 76,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                text  one_hot\n0  ok how much CO2 do i emit with a 7 hour flight...       16\n1  Oh, i arent the one in the hopeless situation ...       16\n2  Please go on. you are a stupid ass Does it ple...       16\n3         Oh, i taste good. yeeh Please go on. nigga        8\n4  _ i love you Do you wish to love me? yeah sure...        2",
      "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>text</th>\n      <th>one_hot</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>0</th>\n      <td>ok how much CO2 do i emit with a 7 hour flight...</td>\n      <td>16</td>\n    </tr>\n    <tr>\n      <th>1</th>\n      <td>Oh, i arent the one in the hopeless situation ...</td>\n      <td>16</td>\n    </tr>\n    <tr>\n      <th>2</th>\n      <td>Please go on. you are a stupid ass Does it ple...</td>\n      <td>16</td>\n    </tr>\n    <tr>\n      <th>3</th>\n      <td>Oh, i taste good. yeeh Please go on. nigga</td>\n      <td>8</td>\n    </tr>\n    <tr>\n      <th>4</th>\n      <td>_ i love you Do you wish to love me? yeah sure...</td>\n      <td>2</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 76,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# create a new dataframe with the balanced dataset\n",
    "balanced_df = pd.DataFrame({'text': X_resampled.reshape(-1), 'one_hot': y_resampled})\n",
    "balanced_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 77,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                text                one_hot\n0  ok how much CO2 do i emit with a 7 hour flight...  [0, 0, 1, 0, 0, 0, 0]\n1  Oh, i arent the one in the hopeless situation ...  [0, 0, 1, 0, 0, 0, 0]\n2  Please go on. you are a stupid ass Does it ple...  [0, 0, 1, 0, 0, 0, 0]\n3         Oh, i taste good. yeeh Please go on. nigga  [0, 0, 0, 1, 0, 0, 0]\n4  _ i love you Do you wish to love me? yeah sure...  [0, 0, 0, 0, 0, 1, 0]",
      "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>text</th>\n      <th>one_hot</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>0</th>\n      <td>ok how much CO2 do i emit with a 7 hour flight...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>1</th>\n      <td>Oh, i arent the one in the hopeless situation ...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>2</th>\n      <td>Please go on. you are a stupid ass Does it ple...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>3</th>\n      <td>Oh, i taste good. yeeh Please go on. nigga</td>\n      <td>[0, 0, 0, 1, 0, 0, 0]</td>\n    </tr>\n    <tr>\n      <th>4</th>\n      <td>_ i love you Do you wish to love me? yeah sure...</td>\n      <td>[0, 0, 0, 0, 0, 1, 0]</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 77,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# convert the one hot encoding back to a list of labels\n",
    "balanced_df['one_hot'] = balanced_df['one_hot'].apply(lambda x: [int(i) for i in list('{0:0b}'.format(x).zfill(len(label_columns)))])\n",
    "balanced_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "outputs": [
    {
     "data": {
      "text/plain": "<AxesSubplot:>"
     },
     "execution_count": 78,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "text/plain": "<Figure size 640x480 with 1 Axes>",
      "image/png": "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\n"
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# plot the distribution of the new labels\n",
    "balanced_df['one_hot'].value_counts().plot(kind='bar')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 79,
   "outputs": [
    {
     "data": {
      "text/plain": "                                                text                one_hot  \\\n0  ok how much CO2 do i emit with a 7 hour flight...  [0, 0, 1, 0, 0, 0, 0]   \n1  Oh, i arent the one in the hopeless situation ...  [0, 0, 1, 0, 0, 0, 0]   \n2  Please go on. you are a stupid ass Does it ple...  [0, 0, 1, 0, 0, 0, 0]   \n3         Oh, i taste good. yeeh Please go on. nigga  [0, 0, 0, 1, 0, 0, 0]   \n4  _ i love you Do you wish to love me? yeah sure...  [0, 0, 0, 0, 0, 1, 0]   \n\n   type.ableism  type.homophobic  type.intellectual  type.racist  type.sexist  \\\n0             0                0                  1            0            0   \n1             0                0                  1            0            0   \n2             0                0                  1            0            0   \n3             0                0                  0            1            0   \n4             0                0                  0            0            0   \n\n   type.sex_harassment  type.transphobic  \n0                    0                 0  \n1                    0                 0  \n2                    0                 0  \n3                    0                 0  \n4                    1                 0  ",
      "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>text</th>\n      <th>one_hot</th>\n      <th>type.ableism</th>\n      <th>type.homophobic</th>\n      <th>type.intellectual</th>\n      <th>type.racist</th>\n      <th>type.sexist</th>\n      <th>type.sex_harassment</th>\n      <th>type.transphobic</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>0</th>\n      <td>ok how much CO2 do i emit with a 7 hour flight...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>1</th>\n      <td>Oh, i arent the one in the hopeless situation ...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>2</th>\n      <td>Please go on. you are a stupid ass Does it ple...</td>\n      <td>[0, 0, 1, 0, 0, 0, 0]</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>3</th>\n      <td>Oh, i taste good. yeeh Please go on. nigga</td>\n      <td>[0, 0, 0, 1, 0, 0, 0]</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n    </tr>\n    <tr>\n      <th>4</th>\n      <td>_ i love you Do you wish to love me? yeah sure...</td>\n      <td>[0, 0, 0, 0, 0, 1, 0]</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
     },
     "execution_count": 79,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# create the intial label columns from the one hot encoding\n",
    "balanced_df[label_columns] = pd.DataFrame(balanced_df['one_hot'].values.tolist(), index= balanced_df.index)\n",
    "balanced_df.head()"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 80,
   "outputs": [
    {
     "data": {
      "text/plain": "<AxesSubplot:>"
     },
     "execution_count": 80,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "text/plain": "<Figure size 640x480 with 1 Axes>",
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAjAAAAIjCAYAAAAdsyFaAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjUuMCwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8/fFQqAAAACXBIWXMAAA9hAAAPYQGoP6dpAABUGElEQVR4nO3deVRVdcP28YtBUBRwCtFbEIdSEcSplNTSNAcsx54y5yEzk5zNfCtN6860JK0s86lEyymHyjsLB5znHAA1B8SBSNCilHCAgPP+4eO5PQGmBuyzOd/PWmet2HuDF2clXuz9G5wsFotFAAAAJuJsdAAAAIA7RYEBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACm42p0gMKSk5Ojc+fOydPTU05OTkbHAQAAt8FiseiPP/5QlSpV5Oyc/32WYltgzp07Jz8/P6NjAACAu/DTTz+patWq+Z4vtgXG09NT0vU3wMvLy+A0AADgdqSlpcnPz8/673h+im2BufHYyMvLiwIDAIDJ/N3wDwbxAgAA06HAAAAA06HAAAAA06HAAAAA06HAAAAA06HAAAAA06HAAAAA07mjAjNt2jTdf//98vT0lI+Pj7p27arjx4/bXNOqVSs5OTnZvJ577jmbaxITE9WpUyd5eHjIx8dH48ePV1ZWls01mzdvVqNGjeTu7q5atWopMjLy7r5DAABQ7NxRgdmyZYuGDx+u3bt3a/369frzzz/Vrl07Xb582ea6IUOGKDk52fqaMWOG9Vx2drY6deqkzMxM7dy5UwsWLFBkZKQmTZpkveb06dPq1KmTWrdurZiYGI0aNUrPPPOM1q5d+w+/XQAAUBw4WSwWy91+8i+//CIfHx9t2bJFDz30kKTrd2AaNGigWbNm5fk533//vR577DGdO3dOlSpVkiTNnTtXEyZM0C+//CI3NzdNmDBBa9as0eHDh62f17NnT128eFFRUVG3lS0tLU3e3t66dOkSK/ECAGASt/vv9z8aA3Pp0iVJUvny5W2OL1q0SBUrVlRQUJAmTpyoK1euWM/t2rVLwcHB1vIiSe3bt1daWpqOHDlivaZt27Y2X7N9+/batWtXvlkyMjKUlpZm8wIAAMXTXe+FlJOTo1GjRql58+YKCgqyHu/Vq5eqVaumKlWqKC4uThMmTNDx48e1atUqSVJKSopNeZFk/TglJeWW16Slpenq1asqVapUrjzTpk3TlClT7vbbAQAAJnLXBWb48OE6fPiwtm/fbnP82Weftf53cHCwKleurDZt2ighIUE1a9a8+6R/Y+LEiRozZoz14xu7WQIAgOLnrh4hhYeH69tvv9WmTZtUtWrVW17btGlTSdLJkyclSb6+vjp//rzNNTc+9vX1veU1Xl5eed59kSR3d3frztPsQA0AQPF2RwXGYrEoPDxcX331lTZu3Kjq1av/7efExMRIkipXrixJCg0N1aFDh3ThwgXrNevXr5eXl5cCAwOt10RHR9t8nfXr1ys0NPRO4gIAgGLqjmYhPf/881q8eLG++eYb1a5d23rc29tbpUqVUkJCghYvXqywsDBVqFBBcXFxGj16tKpWraotW7ZIuj6NukGDBqpSpYpmzJihlJQU9e3bV88884zefPNNSdenUQcFBWn48OEaNGiQNm7cqBEjRmjNmjVq3779bWVlFhIAswl4aY3REazOvNXJ6Ag2eG8cR6HMQvroo4906dIltWrVSpUrV7a+li1bJklyc3PThg0b1K5dO9WpU0djx45Vjx499J///Mf6NVxcXPTtt9/KxcVFoaGh6tOnj/r166epU6dar6levbrWrFmj9evXKyQkRDNnztQnn3xy2+UFAAAUb3c0iPfvbtb4+flZ77TcSrVq1fTdd9/d8ppWrVrp4MGDdxIPAAA4CPZCAgAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApuNqdACYV8BLa4yOYOPMW52MjmDFewMAhYs7MAAAwHQoMAAAwHQoMAAAwHQoMAAAwHQoMAAAwHQoMAAAwHQoMAAAwHQoMAAAwHQoMAAAwHQoMAAAwHQoMAAAwHQoMAAAwHTuqMBMmzZN999/vzw9PeXj46OuXbvq+PHjNtdcu3ZNw4cPV4UKFVSmTBn16NFD58+ft7kmMTFRnTp1koeHh3x8fDR+/HhlZWXZXLN582Y1atRI7u7uqlWrliIjI+/uOwQAAMXOHRWYLVu2aPjw4dq9e7fWr1+vP//8U+3atdPly5et14wePVr/+c9/tHz5cm3ZskXnzp1T9+7dreezs7PVqVMnZWZmaufOnVqwYIEiIyM1adIk6zWnT59Wp06d1Lp1a8XExGjUqFF65plntHbt2gL4lgEAgNm53snFUVFRNh9HRkbKx8dH+/fv10MPPaRLly7p008/1eLFi/XII49IkubPn6+6detq9+7datasmdatW6cff/xRGzZsUKVKldSgQQO9/vrrmjBhgl577TW5ublp7ty5ql69umbOnClJqlu3rrZv3653331X7du3L6BvHQAAmNU/GgNz6dIlSVL58uUlSfv379eff/6ptm3bWq+pU6eO/P39tWvXLknSrl27FBwcrEqVKlmvad++vdLS0nTkyBHrNTd/jRvX3PgaecnIyFBaWprNCwAAFE93XWBycnI0atQoNW/eXEFBQZKklJQUubm5qWzZsjbXVqpUSSkpKdZrbi4vN87fOHera9LS0nT16tU880ybNk3e3t7Wl5+f391+awAAwM7ddYEZPny4Dh8+rKVLlxZknrs2ceJEXbp0yfr66aefjI4EAAAKyR2NgbkhPDxc3377rbZu3aqqVataj/v6+iozM1MXL160uQtz/vx5+fr6Wq/Zu3evzde7MUvp5mv+OnPp/Pnz8vLyUqlSpfLM5O7uLnd397v5dgAAgMnc0R0Yi8Wi8PBwffXVV9q4caOqV69uc75x48YqUaKEoqOjrceOHz+uxMREhYaGSpJCQ0N16NAhXbhwwXrN+vXr5eXlpcDAQOs1N3+NG9fc+BoAAMCx3dEdmOHDh2vx4sX65ptv5OnpaR2z4u3trVKlSsnb21uDBw/WmDFjVL58eXl5eemFF15QaGiomjVrJklq166dAgMD1bdvX82YMUMpKSl65ZVXNHz4cOsdlOeee04ffPCBXnzxRQ0aNEgbN27Ul19+qTVr1hTwtw8AAMzoju7AfPTRR7p06ZJatWqlypUrW1/Lli2zXvPuu+/qscceU48ePfTQQw/J19dXq1atsp53cXHRt99+KxcXF4WGhqpPnz7q16+fpk6dar2mevXqWrNmjdavX6+QkBDNnDlTn3zyCVOoAQCApDu8A2OxWP72mpIlS2rOnDmaM2dOvtdUq1ZN33333S2/TqtWrXTw4ME7iQcAABwEeyEBAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTocAAAADTcTU6AADHEvDSGqMj2DjzViejIwC4C9yBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApnPHBWbr1q16/PHHVaVKFTk5Oenrr7+2OT9gwAA5OTnZvDp06GBzzW+//abevXvLy8tLZcuW1eDBg5Wenm5zTVxcnFq2bKmSJUvKz89PM2bMuPPvDgAAFEt3XGAuX76skJAQzZkzJ99rOnTooOTkZOtryZIlNud79+6tI0eOaP369fr222+1detWPfvss9bzaWlpateunapVq6b9+/fr7bff1muvvaZ58+bdaVwAAFAMud7pJ3Ts2FEdO3a85TXu7u7y9fXN89zRo0cVFRWlH374QU2aNJEkvf/++woLC9M777yjKlWqaNGiRcrMzNRnn30mNzc31atXTzExMYqIiLApOgAAwDEVyhiYzZs3y8fHR7Vr19awYcOUmppqPbdr1y6VLVvWWl4kqW3btnJ2dtaePXus1zz00ENyc3OzXtO+fXsdP35cv//+e55/ZkZGhtLS0mxeAACgeCrwAtOhQwctXLhQ0dHRmj59urZs2aKOHTsqOztbkpSSkiIfHx+bz3F1dVX58uWVkpJivaZSpUo219z4+MY1fzVt2jR5e3tbX35+fgX9rQEAADtxx4+Q/k7Pnj2t/x0cHKz69eurZs2a2rx5s9q0aVPQf5zVxIkTNWbMGOvHaWlplBgAAIqpQp9GXaNGDVWsWFEnT56UJPn6+urChQs212RlZem3336zjpvx9fXV+fPnba658XF+Y2vc3d3l5eVl8wIAAMVToReYpKQkpaamqnLlypKk0NBQXbx4Ufv377des3HjRuXk5Khp06bWa7Zu3ao///zTes369etVu3ZtlStXrrAjAwAAO3fHBSY9PV0xMTGKiYmRJJ0+fVoxMTFKTExUenq6xo8fr927d+vMmTOKjo5Wly5dVKtWLbVv316SVLduXXXo0EFDhgzR3r17tWPHDoWHh6tnz56qUqWKJKlXr15yc3PT4MGDdeTIES1btkyzZ8+2eUQEAAAc1x0XmH379qlhw4Zq2LChJGnMmDFq2LChJk2aJBcXF8XFxalz58667777NHjwYDVu3Fjbtm2Tu7u79WssWrRIderUUZs2bRQWFqYWLVrYrPHi7e2tdevW6fTp02rcuLHGjh2rSZMmMYUaAABIuotBvK1atZLFYsn3/Nq1a//2a5QvX16LFy++5TX169fXtm3b7jQeAABwAOyFBAAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATIcCAwAATMfV6ABmEPDSGqMjWJ15q5PREQAAdsRR/43iDgwAADAdCgwAADAdCgwAADAdCgwAADCdOy4wW7du1eOPP64qVarIyclJX3/9tc15i8WiSZMmqXLlyipVqpTatm2r+Ph4m2t+++039e7dW15eXipbtqwGDx6s9PR0m2vi4uLUsmVLlSxZUn5+fpoxY8adf3cAAKBYuuMCc/nyZYWEhGjOnDl5np8xY4bee+89zZ07V3v27FHp0qXVvn17Xbt2zXpN7969deTIEa1fv17ffvuttm7dqmeffdZ6Pi0tTe3atVO1atW0f/9+vf3223rttdc0b968u/gWAQBAcXPH06g7duyojh075nnOYrFo1qxZeuWVV9SlSxdJ0sKFC1WpUiV9/fXX6tmzp44ePaqoqCj98MMPatKkiSTp/fffV1hYmN555x1VqVJFixYtUmZmpj777DO5ubmpXr16iomJUUREhE3RuVlGRoYyMjKsH6elpd3ptwYAAEyiQMfAnD59WikpKWrbtq31mLe3t5o2bapdu3ZJknbt2qWyZctay4sktW3bVs7OztqzZ4/1moceekhubm7Wa9q3b6/jx4/r999/z/PPnjZtmry9va0vPz+/gvzWAACAHSnQApOSkiJJqlSpks3xSpUqWc+lpKTIx8fH5ryrq6vKly9vc01eX+PmP+OvJk6cqEuXLllfP/300z//hgAAgF0qNivxuru7y93d3egYAACgCBToHRhfX19J0vnz522Onz9/3nrO19dXFy5csDmflZWl3377zeaavL7GzX8GAABwXAVaYKpXry5fX19FR0dbj6WlpWnPnj0KDQ2VJIWGhurixYvav3+/9ZqNGzcqJydHTZs2tV6zdetW/fnnn9Zr1q9fr9q1a6tcuXIFGRkAAJjQHReY9PR0xcTEKCYmRtL1gbsxMTFKTEyUk5OTRo0apTfeeEOrV6/WoUOH1K9fP1WpUkVdu3aVJNWtW1cdOnTQkCFDtHfvXu3YsUPh4eHq2bOnqlSpIknq1auX3NzcNHjwYB05ckTLli3T7NmzNWbMmAL7xgEAgHnd8RiYffv2qXXr1taPb5SK/v37KzIyUi+++KIuX76sZ599VhcvXlSLFi0UFRWlkiVLWj9n0aJFCg8PV5s2beTs7KwePXrovffes5739vbWunXrNHz4cDVu3FgVK1bUpEmT8p1CDQAAHMsdF5hWrVrJYrHke97JyUlTp07V1KlT872mfPnyWrx48S3/nPr162vbtm13Gg8AADgA9kICAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmQ4EBAACmU+AF5rXXXpOTk5PNq06dOtbz165d0/Dhw1WhQgWVKVNGPXr00Pnz522+RmJiojp16iQPDw/5+Pho/PjxysrKKuioAADApFwL44vWq1dPGzZs+O8f4vrfP2b06NFas2aNli9fLm9vb4WHh6t79+7asWOHJCk7O1udOnWSr6+vdu7cqeTkZPXr108lSpTQm2++WRhxAQCAyRRKgXF1dZWvr2+u45cuXdKnn36qxYsX65FHHpEkzZ8/X3Xr1tXu3bvVrFkzrVu3Tj/++KM2bNigSpUqqUGDBnr99dc1YcIEvfbaa3JzcyuMyAAAwEQKZQxMfHy8qlSpoho1aqh3795KTEyUJO3fv19//vmn2rZta722Tp068vf3165duyRJu3btUnBwsCpVqmS9pn379kpLS9ORI0fy/TMzMjKUlpZm8wIAAMVTgReYpk2bKjIyUlFRUfroo490+vRptWzZUn/88YdSUlLk5uamsmXL2nxOpUqVlJKSIklKSUmxKS83zt84l59p06bJ29vb+vLz8yvYbwwAANiNAn+E1LFjR+t/169fX02bNlW1atX05ZdfqlSpUgX9x1lNnDhRY8aMsX6clpZGiQEAoJgq9GnUZcuW1X333aeTJ0/K19dXmZmZunjxos0158+ft46Z8fX1zTUr6cbHeY2rucHd3V1eXl42LwAAUDwVeoFJT09XQkKCKleurMaNG6tEiRKKjo62nj9+/LgSExMVGhoqSQoNDdWhQ4d04cIF6zXr16+Xl5eXAgMDCzsuAAAwgQJ/hDRu3Dg9/vjjqlatms6dO6fJkyfLxcVFTz/9tLy9vTV48GCNGTNG5cuXl5eXl1544QWFhoaqWbNmkqR27dopMDBQffv21YwZM5SSkqJXXnlFw4cPl7u7e0HHBQAAJlTgBSYpKUlPP/20UlNTdc8996hFixbavXu37rnnHknSu+++K2dnZ/Xo0UMZGRlq3769PvzwQ+vnu7i46Ntvv9WwYcMUGhqq0qVLq3///po6dWpBRwUAACZV4AVm6dKltzxfsmRJzZkzR3PmzMn3mmrVqum7774r6GgAAKCYYC8kAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOhQYAABgOnZdYObMmaOAgACVLFlSTZs21d69e42OBAAA7IDdFphly5ZpzJgxmjx5sg4cOKCQkBC1b99eFy5cMDoaAAAwmN0WmIiICA0ZMkQDBw5UYGCg5s6dKw8PD3322WdGRwMAAAZzNTpAXjIzM7V//35NnDjReszZ2Vlt27bVrl278vycjIwMZWRkWD++dOmSJCktLe0f58nJuPKPv0ZBKYjvp6DY0/si8d7cCu9N/nhv8mZP74vEe3Mrxe29ufE1LBbLrS+02KGff/7ZIsmyc+dOm+Pjx4+3PPDAA3l+zuTJky2SePHixYsXL17F4PXTTz/dsivY5R2YuzFx4kSNGTPG+nFOTo5+++03VahQQU5OTgYmu94m/fz89NNPP8nLy8vQLPaG9yZ/vDf5473JH+9N/nhv8mZv74vFYtEff/yhKlWq3PI6uywwFStWlIuLi86fP29z/Pz58/L19c3zc9zd3eXu7m5zrGzZsoUV8a54eXnZxf8c9oj3Jn+8N/njvckf703+eG/yZk/vi7e3999eY5eDeN3c3NS4cWNFR0dbj+Xk5Cg6OlqhoaEGJgMAAPbALu/ASNKYMWPUv39/NWnSRA888IBmzZqly5cva+DAgUZHAwAABrPbAvPUU0/pl19+0aRJk5SSkqIGDRooKipKlSpVMjraHXN3d9fkyZNzPeIC782t8N7kj/cmf7w3+eO9yZtZ3xcni+Xv5ikBAADYF7scAwMAAHArFBgAAGA6FBgAAGA6FBgAAGA6FBgAAGA6FBgAdm/QoEH6448/ch2/fPmyBg0aZEAi+7F161ZlZWXlOp6VlaWtW7cakAgoGkyjLiQWi0UrVqzQpk2bdOHCBeXk5NicX7VqlUHJjPfDDz8oJydHTZs2tTm+Z88eubi4qEmTJgYlK3rvvffebV87YsSIQkxi31xcXJScnCwfHx+b47/++qt8fX3z/AfcUeT33qSmpsrHx0fZ2dkGJTMe703+RowYoVq1auX6ufLBBx/o5MmTmjVrljHB7oDdLmRndqNGjdLHH3+s1q1bq1KlSoZvKGlPhg8frhdffDFXgfn55581ffp07dmzx6BkRe/dd9+9reucnJwcssCkpaXJYrFYN3crWbKk9Vx2dra+++67XP84ORqLxZLnz5fU1FSVLl3agET2I7/fzzMyMuTm5lbEaezLypUrtXr16lzHH3zwQb311lsUGEf2+eefa9WqVQoLCzM6it358ccf1ahRo1zHGzZsqB9//NGARMY5ffq00RHsWtmyZeXk5CQnJyfdd999uc47OTlpypQpBiQzXvfu3SVdfw8GDBhgs4pqdna24uLi9OCDDxoVz1A37mw6OTnpk08+UZkyZaznsrOztXXrVtWpU8eoeHYhNTU1zw0Tvby89OuvvxqQ6M5RYAqJt7e3atSoYXQMu+Tu7q7z58/nen+Sk5Pl6sr/kvivTZs2yWKx6JFHHtHKlStVvnx56zk3NzdVq1ZNVapUMTChcW7842OxWOTp6alSpUpZz7m5ualZs2YaMmSIUfEMdePOpsVi0dy5c+Xi4mI95+bmpoCAAM2dO9eoeHahVq1aioqKUnh4uM3x77//3jT/djEGppAsWLBAUVFR+uyzz2x+sEB6+umnlZycrG+++cb6Q/jixYvq2rWrfHx89OWXXxqc0DhJSUlavXq1EhMTlZmZaXMuIiLCoFTGO3v2rPz9/XkUm4cpU6Zo3LhxDv+4KC+tW7fWqlWrVK5cOaOj2J3PPvtM4eHhGj9+vB555BFJUnR0tGbOnKlZs2aZovxSYArJ1atX1a1bN+3YsUMBAQEqUaKEzfkDBw4YlMx4P//8sx566CGlpqaqYcOGkqSYmBhVqlRJ69evl5+fn8EJjREdHa3OnTurRo0aOnbsmIKCgnTmzBlZLBY1atRIGzduNDqiYaKiolSmTBm1aNFCkjRnzhz97//+rwIDAzVnzhyH/gfq6tWrslgs8vDwkHS97H311VcKDAxUu3btDE4He/bRRx/p3//+t86dOydJCggI0GuvvaZ+/foZnOz2UGAKyZNPPqlNmzbpiSeeyHMQ7+TJkw1KZh8uX76sRYsWKTY2VqVKlVL9+vX19NNP5yp6juSBBx5Qx44dNWXKFHl6eio2NlY+Pj7q3bu3OnTooGHDhhkd0TDBwcGaPn26wsLCdOjQITVp0kRjx47Vpk2bVKdOHc2fP9/oiIZp166dunfvrueee04XL15U7dq15ebmpl9//VUREREO/f9Ndna2IiMjFR0dnedsUEf+peBmv/zyi0qVKmUzVsgMKDCFpHTp0lq7dq31N0bg73h6eiomJkY1a9ZUuXLltH37dtWrV0+xsbHq0qWLzpw5Y3REw5QpU0aHDx+2/oZ4+PBhrVixQgcOHFBYWJhSUlKMjmiYihUrasuWLapXr54++eQTvf/++zp48KBWrlypSZMm6ejRo0ZHNEx4eLgiIyPVqVMnVa5cOdcvkrc7CxD2iRGThcTPz09eXl5Gx7Abq1evVseOHVWiRIk8p+7drHPnzkWUyr6ULl3aOu6lcuXKSkhIUL169STJNLMCCoubm5uuXLkiSdqwYYP1Fnf58uWVlpZmZDTDXblyRZ6enpKkdevWqXv37nJ2dlazZs109uxZg9MZa+nSpfryyy+ZDfp/GjVqpOjoaJUrV04NGza85ZgyMwxzoMAUkpkzZ+rFF1/U3LlzFRAQYHQcw3Xt2lUpKSny8fFR165d873OycnJYReXatasmbZv3666desqLCxMY8eO1aFDh7Rq1So1a9bM6HiGatGihcaMGaPmzZtr7969WrZsmSTpxIkTqlq1qsHpjFWrVi19/fXX6tatm9auXavRo0dLki5cuODwv0S5ubmpVq1aRsewG126dLFOt7/Vz2Gz4BFSISlXrpyuXLmirKwseXh45Brb8dtvvxmUDPbq1KlTSk9PV/369XX58mWNHTtWO3fu1L333quIiAhVq1bN6IiGSUxM1PPPP6+ffvpJI0aM0ODBgyVJo0ePVnZ29h2taFzcrFixQr169VJ2drbatGmjdevWSZKmTZumrVu36vvvvzc4oXFmzpypU6dO6YMPPmAGWzFEgSkkCxYsuOX5/v37F1ESAMVdSkqKkpOTFRISImfn61vc7d27V15eXg69YFu3bt20adMmlS9fXvXq1cv1i6Qjb+lyw759+6zjpAIDA9W4cWODE90+CgwMER0drXfffdf6F6du3boaNWqU2rZta3Ay2Iu0tDTrI5C/G+fi6I9KkLeBAwfe8rwjz15LSkrS008/rR07dqhs2bKSrq/H9eCDD2rp0qWmeDRLgSkkBw4cUIkSJRQcHCxJ+uabbzR//nwFBgbqtddec+h9OD788EONHDlSTzzxhEJDQyVJu3fv1ooVK/Tuu+9q+PDhBic0hrOz8y1vczva2KCbN+LL7725sQ+Qo7033bt3V2RkpLy8vKxbCuSHuwzIS4cOHXTx4kUtWLBAtWvXliQdP35cAwcOlJeXl6KiogxO+PcYxFtIhg4dqpdeeknBwcE6deqUnnrqKXXv3l3Lly/XlStXTLFRVmF588039e6779osYT1ixAg1b95cb775psMWmK+++srm4z///FMHDx7UggULHHK/n40bN1q3Dti0aZPBaeyLt7e3tdDltZ8N/isrK0ubN29WQkKCevXqJU9PT507d05eXl6mW/ekIG3ZskU7d+60lhdJql27tt5//321bNnSwGS3jzswhcTb21sHDhxQzZo1NX36dG3cuFFr167Vjh071LNnT/30009GRzRMmTJlFBMTk2t2QHx8vBo2bKj09HSDktmnxYsXa9myZfrmm2+MjgKYytmzZ9WhQwclJiYqIyNDJ06cUI0aNTRy5EhlZGQ49H5I9913n7744gs98MADNsf37t2rXr166eTJkwYlu33ORgcoriwWi3XVxw0bNljXIfDz83P4NT06d+6c626DdP0x22OPPWZAIvvWrFkzRUdHGx3DUPPnz9fy5ctzHV++fPnfDpgv7o4dO5bvubVr1xZhEvszcuRINWnSRL///rvNnnTdunVz+L9Tb7/9tl544QXt27fPemzfvn0aOXKk3nnnHQOT3T4eIRWSJk2a6I033lDbtm21ZcsWffTRR5Kk06dPq1KlSganK3o3T3MNDAzUv//9b23evNlmDMyOHTs0duxYoyLapatXr+q9997Tv/71L6OjGGratGn6+OOPcx338fHRs88+69Cz+ho1aqS3337b5tFrRkaGxo4dq08++UTXrl0zMJ2xtm3bpp07d+YacxgQEKCff/7ZoFTGKVeunM1YssuXL6tp06Zydb1eBbKysuTq6qpBgwaZYp0YCkwhmTVrlnr37q2vv/5aL7/8svVxyYoVK/Tggw8anK7o/XXJ7nLlyunHH3/Ujz/+aD1WtmxZffbZZ3rllVeKOp5d+OsPF4vFoj/++EMeHh764osvDExmvMTERFWvXj3X8WrVqikxMdGARPYjMjJSw4YN05o1azR//nwlJyerV69eysnJ0bZt24yOZ6icnJw8B3gnJSVZVy92JMVt7CVjYIrYtWvX5OLi4tCbFiJvkZGRNgXG2dlZ99xzj5o2berQuy1Lkr+/vz744INc20x88803Gj58uJKSkgxKZh+SkpI0cOBAHTx4UJcvX9aAAQM0c+ZM6w7Vjuqpp56St7e35s2bJ09PT8XFxemee+5Rly5d5O/v79DTqIsD7sAUsZIlSxodwa7c6M+skik98sgj8vPzy/O9SExMlL+/vwGp7MPTTz+tESNGyNPTUw899JCk67MoRo4cqZ49exqczj5kZmYqOztb2dnZqly5Mj9rdH0l3vbt2yswMFDXrl1Tr169FB8fr4oVK2rJkiVGxzNcdna2vv76a+t6XPXq1VPnzp3l4uJicLLbwx2YAlS+fHmdOHFCFStWzPU44K8cfSuBhQsX6u2331Z8fLyk6yPix48fr759+xqczDg3r3tys9TUVPn4+DjcWic3y8zMVN++fbV8+XLr8/qcnBz169dPc+fOdeh1lZYuXaphw4apZcuW+vTTTxUTE6OBAweqWrVq+vzzz1WjRg2jIxoqKytLS5cuVVxcnNLT09WoUSP17t3bZlCvIzp58qTCwsL0888/26wD4+fnpzVr1qhmzZoGJ/x7FJgCtGDBAvXs2VPu7u5sJXALERERevXVVxUeHq7mzZtLkrZv3645c+bojTfesG5G52icnZ2tG17e7OzZswoMDNTly5cNSmY/Tpw4odjYWJUqVUrBwcEOvT/UDaVLl9Y777yjYcOGWY/9/vvvGjp0qKKiohx+t27kLSwsTBaLRYsWLbKut5Samqo+ffrI2dlZa9asMTjh36PAoMhVr15dU6ZMUb9+/WyOL1iwQK+99ppOnz5tUDJjjBkzRpI0e/ZsDRkyxGbcQnZ2tvbs2SMXFxft2LHDqIiwY8ePH7dZjOxmn3/+uUPf1ZSkc+fOafv27bpw4YJ1aYsbRowYYVAq45UuXVq7d++2rhZ/Q2xsrJo3b26K9bgYA1OIEhISNH/+fCUkJGj27Nny8fHR999/L39/f9WrV8/oeIZJTk7OcybWgw8+qOTkZAMSGevgwYOSro8HOnTokM3jEDc3N4WEhGjcuHFGxbMbSUlJWr16tRITE5WZmWlzLiIiwqBUxqtdu3a+q81269bN6HiGioyM1NChQ+Xm5qYKFSrYPNZ3cnJy6ALj7u6uP/74I9fx9PR08zyStaBQbN682VKqVClL27ZtLW5ubpaEhASLxWKxTJs2zdKjRw+D0xmrXr16ln//+9+5jr/++uuWoKAgAxLZhwEDBlguXbpkdAy7tGHDBouHh4clKCjI4urqamnQoIGlbNmyFm9vb0vr1q2NjmeoM2fOWOrUqWPx8PCwuLi4WH/WjBgxwjJ06FCD0xmratWqljfeeMOSnZ1tdBS707dvX0u9evUsu3fvtuTk5FhycnIsu3btsgQFBVn69+9vdLzbQoEpJM2aNbPMnDnTYrFYLGXKlLH+UNmzZ4/lX//6l5HRDLdixQqLi4uLpX379papU6dapk6damnfvr3F1dXVsmrVKqPjGebixYuW1NTUXMdTU1Mdvtjcf//9lkmTJlkslv/+ffrjjz8snTt3tnz44YcGpzNWly5dLH369LFkZGTY/KzZtGmTpVatWganM1b58uUtJ0+eNDqGXfr9998tnTt3tjg5OVnc3Nwsbm5uFmdnZ0vXrl0tFy9eNDrebWEMTCEpU6aMDh06pOrVq8vT01OxsbGqUaOGzpw5ozp16jj06piStH//fr377rvW6Xt169bV2LFj1bBhQ4OTGadjx456/PHH9fzzz9scnzt3rlavXq3vvvvOoGTG8/T0VExMjGrWrKly5cpp+/btqlevnmJjY9WlSxedOXPG6IiGqVChgnVTvr/+rAkMDNSVK1eMjmiYF198UeXLl9dLL71kdBS7FR8fb92Oom7durn2qLNnjIEpJGXLllVycnKu1UMPHjzo8MvCS1Ljxo0dfnXZv9qzZ0+eYzlatWqll19+2YBE9qN06dLWcS+VK1dWQkKCdRyZo+8txmqz+Zs2bZoee+wxRUVFKTg4ONcCoo48duqGe++9V/fee6/RMe4KBaaQ9OzZUxMmTNDy5cvl5OSknJwc7dixQ+PGjcs1+8YR5eTk6OTJk3nODLixUJmjycjIUFZWVq7jf/75p65evWpAIvvRrFkzbd++XXXr1lVYWJjGjh2rQ4cOadWqVWrWrJnR8QzVrl07zZo1S/PmzZN0fXBqenq6Jk+ebN1E1lFNmzZNa9eutc7S+usgXkeWnZ2tyMhIRUdH5/lzeOPGjQYlu308QiokmZmZGj58uCIjI5WdnS1XV1dlZ2erV69eioyMNM1Kh4Vh9+7d6tWrl86ePau//u/n5OTksAu2tW7dWkFBQXr//fdtjg8fPlxxcXEOva/NqVOnlJ6ervr16+vy5csaO3asdu7cqXvvvVcREREOvR5MUlKS2rdvL4vFovj4eDVp0sS62uzWrVtzrSvkSMqVK6d3331XAwYMMDqK3QkPD1dkZKQ6deqkypUr5yp0f92/zh5RYApZYmKiDh8+rPT0dDVs2NC0t+oKUoMGDXTfffdpypQpef7F8fb2NiiZsXbs2KG2bdvq/vvvV5s2bSRJ0dHR+uGHH7Ru3Tq1bNnS4ITGyM7O1o4dO1S/fn2VLVvW6Dh2KSsrS8uWLVNsbCyrzd7E19dX27Zt4+duHipWrKiFCxea+i4dBQZFrnTp0oqNjTXVYLGiEhMToxkzZlhXm61fv74mTpzo8D+AS5YsqaNHj+a5IzWQn2nTpik5OVnvvfee0VHsTpUqVbR582bdd999Rke5a4yBKUA3VlS9HY48eKxp06Y6efIkBSYPDRo00OLFi42OYXeCgoJ06tQpCkweFixYoIoVK6pTp06Srs+8mTdvngIDA7VkyRKHfry2d+9ebdy4Ud9++63q1auXaxDvqlWrDEpmvLFjx2r27Nn64IMPTDseiDswBah169a3dZ2Tk5MpBkgVpLi4OOt/JyQk6JVXXtH48ePznBlQv379oo5nN26s3nzq1CnNmjWL1Zv/T1RUlCZOnKjXX39djRs3VunSpW3Oe3l5GZTMeLVr19ZHH32kRx55RLt27VKbNm00a9Ysffvtt3J1dXXof6QHDhx4y/Pz588voiT2oXv37jYfb9y4UeXLlzdtuaPAoEg4OzvLyckp16DdG26cc+RBvFu2bFHHjh3VvHlzbd26VUePHlWNGjX01ltvad++fVqxYoXREQ3j7Oxs/e+bf1t09P9nJMnDw0PHjh2Tv7+/JkyYoOTkZC1cuFBHjhxRq1at9MsvvxgdEXbi7wrdzcxQ7niEVAR++uknSZKfn5/BSYzjaBs03o2XXnpJb7zxhsaMGWOzfscjjzyiDz74wMBkxtu0aZPREexWmTJllJqaKn9/f61bt876KLtkyZIOP/3+6tWrslgs1g1Sz549q6+++kqBgYFq166dwemKnhlKyZ2gwBSSrKwsTZkyRe+99551V88yZcrohRde0OTJk3PdrivuHPk5/O06dOhQnuNffHx8HH6xtocfftjoCHbr0Ucf1TPPPKOGDRvqxIkT1lklR44cUUBAgLHhDNalSxd1795dzz33nC5evKgHHnhAbm5u+vXXXxUREaFhw4YZHdFwFy5c0PHjxyVdfxxppmn3zn9/Ce7GCy+8oHnz5mnGjBk6ePCgDh48qBkzZujTTz916B1Qbzh+/LjCw8PVpk0btWnTRuHh4da/RI7qxurNf8XqzbiVOXPmKDQ0VL/88otWrlypChUqSLq+XcfTTz9tcDpjHThwwLr8wIoVK+Tr66uzZ89q4cKFDj8zKS0tTX379tW//vUvPfzww3r44Yf1r3/9S3369NGlS5eMjnd7inrzJUfh5eVl+e6773IdX7NmjcXLy8uARPZjxYoVFldXV0uzZs0so0ePtowePdoSGhpqcXV1taxYscLoeIYZO3aspUWLFpbk5GSLp6enJT4+3rJ9+3ZLjRo1LK+99prR8exSnTp1LM7OzkbHgJ0qVaqU5ezZsxaLxWL5n//5H+vfo8TEREupUqWMjGa4J5980nLvvfdaoqKiLJcuXbJcunTJEhUVZaldu7blqaeeMjrebaHAFJJ77rnH8uOPP+Y6/uOPP1oqVqxoQCL7UaNGDcurr76a6/ikSZMsNWrUMCCRfcjIyLA888wzFldXV4uTk5OlRIkSFmdnZ0ufPn0sWVlZRsezS1999ZUlMjLS6BiwU8HBwZbZs2dbEhMTLV5eXpadO3daLBaLZd++fZZKlSoZnM5YHh4elm3btuU6vnXrVouHh4cBie4cs5AKydSpU3Xs2DHNnz9f7u7ukq7vdTN48GDde++9mjx5ssEJjePh4aG4uLhc68DEx8crJCTEoXfPlVi9GQWjbt26OnHihEPP0FqxYoV69eql7OxstWnTRuvWrZN0fYG7rVu36vvvvzc4oXH8/f21Zs0aBQcH2xyPi4tTWFiYkpKSDEp2+xjEW4D+Osd+w4YNqlq1qkJCQiRJsbGxyszMtC4T76hatWqlbdu25Sow27dvd9jl8m/m7+8vf39/o2PYnaysLG3evFkJCQnq1auXPD09de7cOXl5ealMmTJGx7M706ZNM89YhkLyxBNPqEWLFkpOTrb+HJakNm3aqFu3bgYmM94rr7yiMWPG6PPPP5evr68kKSUlRePHj9err75qcLrbwx2YAlTc5tgXlrlz52rSpEl68sknrTsJ7969W8uXL9eUKVNUpUoV67WdO3c2KmaRYPXm23P27Fl16NBBiYmJysjI0IkTJ1SjRg2NHDlSGRkZmjt3rtERYQJpaWnauHGjateurbp16xodx1ANGzbUyZMnlZGRYf2FKTExUe7u7rnu+h44cMCIiH+LAoMid/OiZLfiCAuUsXrz7enatas8PT316aefqkKFCoqNjVWNGjW0efNmDRkyRPHx8UZHNBR3p/L25JNP6qGHHlJ4eLiuXr2qkJAQnTlzRhaLRUuXLlWPHj2MjmiYKVOm3Pa19jrkgQIDwO5VqFBBO3fuVO3ateXp6WktMGfOnFFgYKBDj5vi7lT+fH19tXbtWoWEhGjx4sWaPHmyYmNjtWDBAs2bN08HDx40OiL+AdaBKUQrVqywPiZp1KiRzQvA7cvJycnzblxSUpLNqsWOaOTIkWrSpIl+//13lSpVynq8W7duio6ONjCZ8S5duqTy5ctLur6fVo8ePeTh4aFOnTo5/F27n376yWag7t69ezVq1CjNmzfPwFR3hkG8heS9997Tyy+/rAEDBuibb77RwIEDlZCQoB9++EHDhw83Op7htmzZonfeeUdHjx6VJAUGBmr8+PEON4j3rwO/b8UMm6sVlnbt2mnWrFnWH65OTk5KT0/X5MmTrSvPOqpt27Zp586dcnNzszkeEBCgn3/+2aBU9sHPz0+7du1S+fLlFRUVpaVLl0qSfv/9d5UsWdLgdMbq1auXnn32WfXt21cpKSlq27atgoKCtGjRIqWkpGjSpElGR/xb3IEpJB9++KHmzZun999/X25ubnrxxRe1fv16jRgxwuFnBnzxxRdq27atPDw8NGLECI0YMUKlSpVSmzZt8lxKvzjz9va+7Zcjmzlzpnbs2KHAwEBdu3ZNvXr1sv4DPX36dKPjGYq7U/kbNWqUevfurapVq6pKlSpq1aqVJGnr1q25pg87msOHD+uBBx6QJH355ZcKDg7Wzp07tWjRIkVGRhob7jYxBqaQeHh46OjRo6pWrZp8fHy0fv16hYSEKD4+Xs2aNVNqaqrREQ1Tt25dPfvssxo9erTN8YiICP3v//6v9a4McLOsrCwtXbpUcXFxSk9PV6NGjdS7d2+bxyaO6KmnnpK3t7fmzZsnT09PxcXF6Z577lGXLl3k7+/v0DMepetbKiQmJurRRx+1Dmhes2aNypYtq+bNmxuczjhlypTR4cOHFRAQoM6dO6t58+aaMGGCEhMTVbt2bVNsBEqBKSQ1atTQypUr1bBhQzVp0kRDhgzR0KFDtW7dOvXs2VO//fab0REN4+7uriNHjuRaB+bkyZMKCgrStWvXDEpmPGaT4E4lJSWpffv2slgsio+PV5MmTRQfH6+KFStq69atptqcD0WnadOmat26tTp16qR27dpp9+7dCgkJ0e7du/XEE0+wkJ0je+SRR7R69Wo1bNhQAwcO1OjRo7VixQrt27fvjsY9FEd+fn6Kjo7OVWA2bNggPz8/g1IZ76+zSR599FF5enpq+vTpDj+bRLq+Aej7779vvUNXt25dhYeHq06dOgYnM1bVqlUVGxtrc3dq8ODB3J36P0lJSVq9erUSExOVmZlpc86R11aaPn26unXrprffflv9+/e3LvS3evVq66Mle8cdmEKSk5OjnJwcubpe74hLly7Vzp07de+992ro0KG5Btw5ko8++kijRo3SoEGD9OCDD0qSduzYocjISM2ePVtDhw41OKExWOskfytXrlTPnj3VpEkThYaGSrq++OEPP/zg8Ot5IH/R0dHq3LmzatSooWPHjikoKMi6DkyjRo0cem0lScrOzlZaWprKlStnPXbmzBl5eHiY4s4dBQaG+OqrrzRz5kyb36bHjx+vLl26GJzMOKx1kr+aNWuqd+/emjp1qs3xyZMn64svvlBCQoJByewDd6fy9sADD6hjx46aMmWK9e+Uj4+PevfurQ4dOmjYsGFGR8Q/wCwkGKJbt27avn27UlNTlZqaqu3btzt0eZGYTXIrycnJ6tevX67jffr0UXJysgGJ7MfKlSsVFBSk/fv3KyQkRCEhITpw4ICCg4O1cuVKo+MZ6ujRo9b/b1xdXXX16lWVKVNGU6dOdfjZa+fPn1ffvn1VpUoVubq6ysXFxeZlBoyBKWLsEGsrPT1dOTk5Nse8vLwMSmMs1jrJHxuA5u/FF1/UxIkT87w79eKLLzr047XSpUtbx71UrlxZCQkJqlevniTp119/NTKa4QYMGKDExES9+uqrqly5spycnIyOdMd4hFTEvv76a126dEn9+/c3OophTp8+rfDwcG3evNlmxpHFYnGI/Y/yw2yS/LEBaP48PDwUFxeXq9zFx8crJCTEoR89du3aVZ06ddKQIUM0btw4ffPNNxowYIBWrVqlcuXKacOGDUZHNIynp6e2bdumBg0aGB3lrlFgUOSaN28ui8WikSNHqlKlSrma/8MPP2xQMuNlZWVp2bJlio2NZa2Tm7ABaP7CwsL0P//zPxo4cKDN8fnz52vp0qVau3atQcmMd+rUKaWnp6t+/fq6fPmyxo4da51MERERoWrVqhkd0TCBgYFatGiRGjZsaHSUu0aBKUSs6ZG3MmXKaP/+/apdu7bRUezK1q1b9eCDD1pnrt2QlZWlnTt36qGHHjIoGewZd6fylp2drR07dqh+/foqW7as0XHszrp16zRz5kx9/PHHCggIMDrOXaHAFBJ2iM1f69at9fLLL6tt27ZGR7ErLi4uSk5OzvWoKDU1VT4+Pg53Z+Fmp06dUo0aNYyOYZe4O5W/kiVL6ujRo6pevbrRUexOuXLldOXKFWVlZcnDw0MlSpSwOW+GxVYZxFtIbuwQGxsbqwoVKliPd+vWTUOGDDEwmfE++eQTPffcc/r5558VFBSU6y9O/fr1DUpmrBtjgP4qNTVVpUuXNiCR/ahVq5YefvhhDR48WE888YTDb8R3s78Ogsd/BQUF6dSpUxSYPMyaNcvoCP8Yd2AKCWt65G/37t3q1auXzpw5Yz3m5OTksIN4b6zM/M0336hDhw5yd3e3nsvOzlZcXJxq166tqKgooyIaLiYmRvPnz9eSJUuUmZmpp556SoMHDzbNiqGFibtT+YuKitLEiRP1+uuvq3Hjxrl+EXDUGY/FBXdgCglreuRv0KBBatiwoZYsWZLnIF5Hc2OnaYvFIk9PT5sBu25ubmrWrJnD37Vr0KCBZs+erZkzZ2r16tWKjIxUixYtdN9992nQoEHq27ev7rnnHqNjGoK7U/m7sfxA586dbX7OOOovS/m5du1arm0WzFDuuANTSNghNn+lS5dWbGxsrmmfjm7KlCkaN26cwz8uuh0ZGRn68MMPNXHiRGVmZsrNzU1PPvmkpk+frsqVKxsdr0hxdyp/W7ZsueV5R57xePnyZU2YMEFffvmlUlNTc503Q7mjwBQS1vTI3+OPP64BAwY49AJbuDv79u3TZ599pqVLl6p06dLq37+/Bg8erKSkJE2ZMkVpaWnau3ev0TENkZWVZb07FRUVxd0p3NLw4cO1adMmvf766+rbt6/mzJmjn3/+WR9//LHeeust9e7d2+iIf4sCU4iysrJsdohlTY/r5s2bpzfeeEODBg1ScHBwrkG8jjTV82bnz5/XuHHjFB0drQsXLuivfzXN8BtRYYmIiND8+fN1/PhxhYWF6ZlnnlFYWJjNDJykpCQFBAQoKyvLwKTG4+5UbleuXMlzN2pHnTAgSf7+/lq4cKFatWolLy8vHThwQLVq1dLnn3+uJUuW6LvvvjM64t+iwKDI3WrapyM/l+7YsaMSExMVHh6e59LejrxXVIkSJTR16lQNGDAg1z/CiYmJ8vf3V2ZmppYsWeKwq1xzdyq3X375RQMHDtT333+f53lH/VkjXV+P68cff5S/v7+qVq2qVatW6YEHHtDp06cVHBys9PR0oyP+LQbxFiJ2iM0b0z7ztn37dtMv7V1YsrOzNXjw4DzXyKlevbqys7Pl5ubmkOXlr3enFi5caHN3qnr16oqMjDTtYmX/xKhRo3Tx4kXt2bNHrVq10ldffaXz58/rjTfe0MyZM42OZ6gaNWro9OnT8vf3V506dfTll1/qgQce0H/+8x/TLPzHbtSFhB1icaf8/PxyPTbCf+U1Wy09Pd3hZ91MmDBBvXr10tmzZ/X111/rscces5aXxMRESZKPj48+/fRTI2MaYuPGjYqIiFCTJk3k7OysatWqqU+fPpoxY4amTZtmdDxDDRw4ULGxsZKkl156SXPmzFHJkiU1evRojR8/3uB0t4dHSIWkZs2a6t27d547xH7xxRdKSEgwKJl92LJli9555x3r3anAwECNHz/eoXcWLg5Lexe0MWPGSJJmz56tIUOGyMPDw3ouOztbe/bskYuLi3bs2GFURMM5OzsrJSWFFZzz4OXlpbi4OAUEBKhatWpavHixmjdvrtOnT6tevXoOvR7XX509e1b79+9XrVq1TDM2iEdIhSQ5OVn9+vXLdbxPnz56++23DUhkP7744gsNHDhQ3bt314gRIyRJO3bsUJs2bRQZGalevXoZnNAYTz31lK5cuaKaNWuadmnvgnbw4EFJ19ftOHTokNzc3Kzn3NzcFBISonHjxhkVz25wdypvtWvX1vHjxxUQEKCQkBDrLwdz58512AHNkvTnn3+qQ4cOmjt3ru69915JUrVq1Uy3uSUFppC0atVK27Zty7XWyfbt2x36LoMk/fvf/9aMGTM0evRo67ERI0YoIiJCr7/+usMWmOKwtHdB27Rpk6Trt7tnz55tisW1isqNu1NOTk569dVX87w75ejjqUaOHKnk5GRJ1+9+d+jQQYsWLZKbm5siIyONDWegEiVKKC4uzugY/xiPkAoJO8Tmz93dXUeOHMlV7k6ePKmgoCBdu3bNoGSAebRu3VrS9cexoaGhue5OBQQEaNy4cdbfsHF9OvWxY8fk7++vihUrGh3HUKNHj5a7u7veeusto6PcNQpMIWGH2PzVqlVL48eP19ChQ22Oz507VzNnzlR8fLxByYpeWlqa9a5CWlraLa/l7gPywt0p3I0XXnhBCxcu1L333pvnPlEREREGJbt9PEIqJEwVzt/YsWM1YsQIxcTE6MEHH5R0fQxMZGSkZs+ebXC6olWuXDklJyfLx8dHZcuWzXMsA/u24FYceVuSv5Odna3IyEjr4pB//bm8ceNGg5IZ7/Dhw2rUqJEk6cSJEwanuTsUmELCDrH5GzZsmHx9fTVz5kx9+eWXkq6vkbNs2TKHW6xt48aNKl++vKT/jvcAUDBGjhypyMhIderUSUFBQQ6/cezNisPPGx4hFRJnZ2d2iAUAA1WsWNG6sB9sDRo0SLNnz5anp6fN8cuXL+uFF17QZ599ZlCy20eBKSTsEPv3MjMz87yt6+/vb1AiAMVJlSpVtHnzZt13331GR7E7Li4u1sfXN/v111/l6+trij3FWIm3kDRo0ECzZ8/WuXPn9Nlnnyk5OVktWrRQUFCQIiIi9Msvvxgd0TDx8fFq2bKlSpUqpWrVqql69eqqXr26AgICVL16daPj2Z26devKxcXF6BiA6YwdO1azZ89mheubpKWl6dKlS7JYLPrjjz+UlpZmff3+++/67rvvcpUae8UdmCLCDrH/1bx5c7m6uuqll17Kc9PCkJAQg5LZp6+//lqXLl1yyH1+gDvVvXt3m49vjDOrV69ersUhV61aVZTR7IKzs/MtxwI5OTlpypQpevnll4sw1d2hwBQydojNrXTp0tq/f7/Db2oJoOANHDjwtq91xBlcW7ZskcVi0SOPPKKVK1daJxFI19cPqlatms06ZfaMAlNI/rpD7DPPPGOzQ6wkJSUlKSAgwBTPGgvS/fffr3fffVctWrQwOordycrK0ubNm5WQkKBevXrJ09NT586dk5eXl8qUKWN0PKBY2rFjh5o0aSJ3d3ejoxSZs2fPyt/f39QzsygwhaREiRKaOnWqBgwYkOsRUWJiovz9/ZWZmaklS5Y4xKOBmxdp27dvn1555RW9+eabCg4OznVb11EX5Dp79qw6dOigxMREZWRk6MSJE6pRo4ZGjhypjIwMzZ071+iIQLHk5eWlmJgYlr4wGQpMIWGHWFt/fe56Y3G2mzn6gm1du3aVp6enPv30U1WoUEGxsbGqUaOGNm/erCFDhjjUCsVAUfL09LT+fXN0devW1YkTJ0zxc5iF7AoRO8T+V3FYNKmwbdu2TTt37rTZ00aSAgIC9PPPPxuUCoAjmTZtmi5dumR0jNtCgSlg7BCbt4cfftjoCHYvJycnz996kpKSci02BQCFoWvXrkZHuG0UmAJ28OBBSdcfhxw6dCjXDrEhISEaN26cUfFgx9q1a6dZs2Zp3rx5kq6X4PT0dE2ePJmVRAEUOLNPGmAMTCFhh9g7Z6Znr4UhKSlJ7du3l8ViUXx8vJo0aaL4+HhVrFhRW7duNc3iUoDZOOIg3uIwaYA7MIXEEdcX+KfM9Oy1MFStWlWxsbFaunSp4uLilJ6ersGDB6t3794qVaqU0fGAYssRf48fOXKkmjRpotjYWFWoUMF6vFu3bhoyZIiByW4fd2AAAMXSsWPH8l0wc+3atWrfvn0RJ7IfFSpU0M6dO1W7dm2bWVhnzpxRYGCgrly5YnTEv8VeSDBEVlaWNmzYoI8//lh//PGHJOncuXNKT083OJmxjh8/rvDwcLVp00Zt2rRReHi4jh07ZnQswJQaNWqkOXPm2BzLyMhQeHi4unTpYlAq+1AcJg1QYFDkzp49q+DgYHXp0kXDhw+3bmw5ffp0hx7gvHLlSgUFBWn//v0KCQlRSEiIDhw4oODgYK1cudLoeIDpREZGatKkSQoLC9P58+cVExOjhg0basOGDdq2bZvR8Qx1Y9LADWacNMAjJBQ5FmzLW82aNdW7d29NnTrV5vjkyZP1xRdfKCEhwaBkgHklJSVp4MCBOnjwoC5fvqwBAwZo5syZNktcOKLiMGmAQbwocizYlrfk5GT169cv1/E+ffro7bffNiARUDxkZmYqOztb2dnZqly5skMuJvpXxWHSAAUGRa44PHstDK1atdK2bdtUq1Ytm+Pbt29Xy5YtDUoFmNfSpUs1bNgwtWzZUidOnFBMTIwGDhyotWvX6vPPP3eoadN5cXV1VZ8+fYyOcdd4hIQi99RTT8nb21vz5s2Tp6en4uLidM8996hLly7y9/d32Cnoc+fO1aRJk/Tkk0+qWbNmkqTdu3dr+fLlmjJlis0W9507dzYqJmAapUuX1jvvvKNhw4ZZj/3+++8aOnSooqKibDaZdUTHjx/X+++/r6NHj0q6vhZXeHh4vjO37A0FBkWuODx7LQzOzrc3pt6RN7wE7sTx48dVu3btPM99/vnn6tu3bxEnsh8rV65Uz5491aRJE4WGhkq6/gvTDz/8oKVLl6pHjx4GJ/x7FBgYIisry+bZa6NGjUz17BWAOZh9ufzCUhwmDVBgADtx6tQph38mDxSk4rBcfmHx8PBQXFxcrjF38fHxCgkJYSE7ID8s2JZbrVq11Lp1a33xxRe6du2a0XEA07uxXP7vv/9uc3e3W7duio6ONjCZ8W5MGvgrM00aYBYSilx+z16Dg4NN8+y1MBw4cEDz58/XmDFjFB4erqeeekqDBw/WAw88YHQ0wJRYsiF/nTt31oQJE7R///48Jw2sXr3a5lp7xCMkFLni8Oy1MGVlZWn16tWKjIxUVFSU7rvvPg0aNEh9+/bVPffcY3Q8wDTKlSunHTt2KDAw0Ga/n+3bt6tHjx46f/680RENUxwmDVBgUOSKw7PXopCRkaEPP/xQEydOVGZmptzc3PTkk09q+vTpqly5stHxALvHkg3FG2NgUOSKw7PXwrRv3z49//zzqly5siIiIjRu3DglJCRo/fr1OnfunMNvQgfcrpkzZ1rvwFy7dk29evWyPj6aPn260fEMderUKaMj/GPcgUGRY8G2vEVERGj+/Pk6fvy4wsLC9MwzzygsLMzmVm9SUpICAgKUlZVlYFLAPLKysrRs2TLFxsayZMNNnJ2d9fDDD2vw4MF64oknTLm9AgUGRa44PHstDCVKlNDUqVM1YMCAXI+IEhMT5e/vr8zMTC1ZskT9+/c3KCWA4iAmJkbz58/XkiVLlJmZacpJAxQYwE44OzsrJSUl10rEqamp8vHxcagyBxSEBQsWqGLFiurUqZMk6cUXX9S8efMUGBioJUuWqFq1agYnNJ6ZJw0wBgZFrjg8ey0sTk5OuY6lp6eb8vYuYLQ333zT+qho165d+uCDDzRjxgxVrFhRo0ePNjidfXB1dVX37t21fPlyTZ8+XSdPntS4cePk5+enfv36KTk52eiI+eIODIpccXj2WpDGjBkjSZo9e7aGDBkiDw8P67ns7Gzt2bNHLi4u2rFjh1ERAVPy8PDQsWPH5O/vrwkTJig5OVkLFy7UkSNH1KpVK/3yyy9GRzTcvn379Nlnn2np0qUqXbq0+vfvr8GDByspKUlTpkxRWlqa9u7da3TMPHEHBkXuwIEDql+/vsaMGSNfX18NHTrUbv+CFIWDBw/q4MGDslgsOnTokPXjgwcP6tixYwoJCVFkZKTRMQHTKVOmjFJTUyVJ69at06OPPipJKlmypK5evWpkNMNFREQoODhYDz74oM6dO6eFCxfq7NmzeuONN1S9enW1bNlSkZGROnDggNFR88UdGBjGzM9eC8PAgQM1e/ZseXl5GR0FKBZ69+6tY8eOqWHDhlqyZIkSExNVoUIFrV69Wv/v//0/HT582OiIhikOkwYoMDAcC7YBKAwXL17UK6+8op9++knDhg1Thw4dJF1f9dvNzU0vv/yywQmNUxwmDVBgYBgzP3sFADNzdnbW+fPnc93tPnv2rAIDA3X58mWDkt0+NnNEkfvrgm0LFy60WbCtevXqioyMVEBAgLFBAaCYuTFpwMnJSa+++mqekwYaNGhgULo7Q4FBkZswYcLfPnv18fHRp59+alBCAMVZ3bp1deLECVM8JiloBw8elCTrpIGbd+p2c3NTSEiIxo0bZ1S8O8IjJBS54vDsFYB5ff3117p06ZLdDk4tCsVh0gAFBkWuODx7BQAYi0dIKDLF6dkrAHPIysrS5s2blZCQoF69esnT01Pnzp2Tl5eXypQpY3Q8/AMUGBSZ4vTsFYD9O3v2rDp06KDExERlZGTo0Ucflaenp6ZPn66MjAzNnTvX6Ij4BygwKDKbNm2SVDyevQKwfyNHjlSTJk0UGxurChUqWI9369ZNQ4YMMTAZCgIFBkVu/vz5RkcA4AC2bdumnTt32tztlaSAgAD9/PPPBqVCQWEvJABAsZSTk5PnrMakpCR5enoakAgFiQIDACiW2rVrp1mzZlk/dnJyUnp6uiZPnqywsDDjgqFAMI0aAFAsJSUlqX379rJYLIqPj1eTJk0UHx+vihUrauvWrbnWooK5UGAAAMVWVlaWli5dqri4OKWnp6tRo0bq3bu3SpUqZXQ0/EMUGAAAYDrMQgIAFFvHjx/X+++/r6NHj0q6vg9SeHi46tSpY3Ay/FMM4gUAFEsrV65UUFCQ9u/fr5CQEIWEhOjAgQMKDg7WypUrjY6Hf4hHSACAYqlmzZrq3bu3pk6danN88uTJ+uKLL5SQkGBQMhQECgwAoFjy8PBQXFycatWqZXM8Pj5eISEhunLlikHJUBB4hAQAKJZatWqlbdu25Tq+fft2tWzZ0oBEKEgM4gUAFEudO3fWhAkTtH//fjVr1kyStHv3bi1fvlxTpkzR6tWrba6FufAICQBQLDk7395DBicnpzy3HIB9o8AAAADTYQwMAKBYOnXqlNERUIgoMACAYqlWrVpq3bq1vvjiC127ds3oOChgFBgAQLF04MAB1a9fX2PGjJGvr6+GDh2qvXv3Gh0LBYQxMACAYi0rK0urV69WZGSkoqKidN9992nQoEHq27ev7rnnHqPj4S5RYAAADiEjI0MffvihJk6cqMzMTLm5uenJJ5/U9OnTVblyZaPj4Q7xCAkAUKzt27dPzz//vCpXrqyIiAiNGzdOCQkJWr9+vc6dO6cuXboYHRF3gTswAIBiKSIiQvPnz9fx48cVFhamZ555RmFhYTbrwyQlJSkgIEBZWVkGJsXdoMAAAIqlEiVKaOrUqRowYECuR0SJiYny9/dXZmamlixZov79+xuUEneLAgMAKJacnZ2VkpIiHx8fm+Opqany8fFh9V2TYwwMAKDYcnJyynUsPT1dJUuWNCANChKbOQIAipUxY8ZIul5eXn31VXl4eFjPZWdna8+ePWrQoIFB6VBQKDAAgGLl4MGDkiSLxaJDhw7Jzc3Nes7NzU0hISEaN26cUfFQQBgDAwAolgYOHKjZs2fLy8vL6CgoBBQYAABgOgziBQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApkOBAQAApvP/ARarWY2EBkaQAAAAAElFTkSuQmCC\n"
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "balanced_df[label_columns].sum(axis=0).plot(kind='bar')"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 81,
   "outputs": [],
   "source": [
    "# create train, validation and test set\n",
    "train_df, val_df = train_test_split(balanced_df, test_size=0.2, random_state=42)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 82,
   "outputs": [],
   "source": [
    "# use the new data to train the model\n",
    "model = torch.load(config.result_file('roBERTa_MultLabel_class_model.pt'))\n",
    "tokenizer = RobertaTokenizer.from_pretrained('roberta-base')\n",
    "label_columns = pickle.load(open(config.result_file('label_columns.pkl'),'rb'))\n",
    "optimizer = torch.optim.Adam\n",
    "\n",
    "classifier = ToxicCommentClassifier(model, tokenizer, label_columns)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 83,
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Tobias\\AppData\\Local\\Temp\\ipykernel_21872\\377985415.py:16: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  data = TensorDataset(inputs, masks, torch.tensor(labels))\n",
      "C:\\Users\\Tobias\\AppData\\Local\\Temp\\ipykernel_21872\\377985415.py:16: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  data = TensorDataset(inputs, masks, torch.tensor(labels))\n",
      "C:\\Users\\Tobias\\AppData\\Local\\Temp\\ipykernel_21872\\377985415.py:16: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  data = TensorDataset(inputs, masks, torch.tensor(labels))\n"
     ]
    }
   ],
   "source": [
    "# create the data loaders\n",
    "train_dataloader = create_dataloader_from_df(train_df)\n",
    "val_dataloader = create_dataloader_from_df(val_df)\n",
    "test_dataloader = create_dataloader_from_df(test_df)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 84,
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch:  33%|███▎      | 1/3 [00:27<00:54, 27.34s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train loss: 0.17824314759587342\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch:  67%|██████▋   | 2/3 [00:53<00:26, 26.59s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train loss: 0.06961461109747677\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Epoch: 100%|██████████| 3/3 [01:19<00:00, 26.38s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train loss: 0.05233107256533215\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  97.85538052761437\n",
      "Flat Validation Accuracy:  92.5255704169945\n"
     ]
    }
   ],
   "source": [
    "# train the model on the balanced dataset\n",
    "classifier.train(optimizer, train_dataloader, val_dataloader=val_dataloader, epochs=3)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 85,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  90.0562851782364\n",
      "Flat Validation Accuracy:  80.95238095238095\n",
      "                     precision    recall  f1-score   support\n",
      "\n",
      "       type.ableism       0.20      0.50      0.29         2\n",
      "    type.homophobic       0.82      0.88      0.85        16\n",
      "  type.intellectual       0.94      0.97      0.95        62\n",
      "        type.racist       1.00      0.71      0.83         7\n",
      "        type.sexist       0.74      0.89      0.81        62\n",
      "type.sex_harassment       0.96      0.97      0.97       108\n",
      "   type.transphobic       0.00      0.00      0.00         1\n",
      "\n",
      "          micro avg       0.87      0.93      0.90       258\n",
      "          macro avg       0.67      0.70      0.67       258\n",
      "       weighted avg       0.89      0.93      0.91       258\n",
      "        samples avg       0.91      0.94      0.91       258\n",
      "\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in samples with no predicted labels. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n"
     ]
    }
   ],
   "source": [
    "## Test the model on the test set\n",
    "classifier.test(test_dataloader)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 86,
   "outputs": [],
   "source": [
    "# save the model\n",
    "torch.save(classifier.model, config.result_file('roBERTa_MultLabel_class_model_balanced.pt'))"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": 87,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "F1 Validation Accuracy:  34.198473282442755\n",
      "Flat Validation Accuracy:  29.330254041570434\n",
      "                     precision    recall  f1-score   support\n",
      "\n",
      "       type.ableism       0.83      0.10      0.17        52\n",
      "    type.homophobic       0.84      0.40      0.54        68\n",
      "  type.intellectual       0.00      0.00      0.00         0\n",
      "        type.racist       0.40      0.24      0.30        71\n",
      "        type.sexist       0.58      0.76      0.66        83\n",
      "type.sex_harassment       0.00      0.00      0.00         0\n",
      "   type.transphobic       0.00      0.00      0.00         0\n",
      "\n",
      "          micro avg       0.29      0.41      0.34       274\n",
      "          macro avg       0.38      0.21      0.24       274\n",
      "       weighted avg       0.65      0.41      0.44       274\n",
      "        samples avg       0.20      0.26      0.22       274\n",
      "\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in labels with no true samples. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in samples with no predicted labels. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n",
      "C:\\Users\\Tobias\\anaconda3\\envs\\HateyBot\\lib\\site-packages\\sklearn\\metrics\\_classification.py:1334: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in samples with no true labels. Use `zero_division` parameter to control this behavior.\n",
      "  _warn_prf(average, modifier, msg_start, len(result))\n"
     ]
    }
   ],
   "source": [
    "# test the model on the ethos dataset\n",
    "ethos_dataloader = torch.load(config.result_file('ethos_dataloader.pt'))\n",
    "classifier.test(ethos_dataloader)"
   ],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "outputs": [],
   "source": [],
   "metadata": {
    "collapsed": false,
    "pycharm": {
     "name": "#%%\n"
    }
   }
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}