notebooks/3-rnn.ipynb

{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Recurrent Neural Networks (RNN)\n",
    "\n",
    "### GPU selection\n",
    "\n",
    "Select the GPU id to use."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Select the right device. Change the number to select a different GPU.\n",
    "# If you have used slurm to request a GPU, just keep this line as it is.\n",
    "device_id = 0"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "device(type='cuda', index=0)"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import torch\n",
    "from torch import cuda\n",
    "from multiprocessing import cpu_count\n",
    "\n",
    "# Automatic device selection\n",
    "if cuda.is_available():\n",
    "    torch.set_num_threads(cpu_count() // cuda.device_count())\n",
    "    device = torch.device(f\"cuda:{device_id}\")\n",
    "else:\n",
    "    torch.set_num_threads(cpu_count())\n",
    "    device = torch.device(\"cpu\")\n",
    "device"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Hyperparameters\n",
    "\n",
    "These are the hyperparameters used throughout this chapter. Placed here for\n",
    "easier experiments.\n",
    "\n",
    "There are too many hyperparameters to save as plain variables, so let's group\n",
    "them into one object: the `Params` class. Observe the differences between\n",
    "saving them as a dictionary and as a class instance."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "from typing import Type\n",
    "\n",
    "from torch import nn, optim\n",
    "\n",
    "\n",
    "class Params:\n",
    "    def __init__(self,\n",
    "                 batch_size: int = 128, epochs: int = 20, lr: float = 1e-4,\n",
    "                 embedding_dim: int = 32, hidden_dim: int = 32,\n",
    "                 num_layers: int = 3, dropout: float = 0.2,\n",
    "                 RNN: Type[nn.Module] = nn.GRU,\n",
    "                 readout: str = \"max\",\n",
    "                 Optimizer: Type[optim.Optimizer] = optim.Adam,\n",
    "                 print_every: int = 4):\n",
    "        self.batch_size = batch_size\n",
    "        self.epochs = epochs\n",
    "        self.lr = lr\n",
    "        self.embedding_dim = embedding_dim\n",
    "        self.hidden_dim = hidden_dim\n",
    "        self.num_layers = num_layers\n",
    "        self.dropout = dropout\n",
    "        self.RNN = RNN\n",
    "        self.readout = readout\n",
    "        self.Optimizer = Optimizer\n",
    "        self.print_every = print_every\n",
    "\n",
    "\n",
    "SEED = 42\n",
    "params = Params()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Data preprocessing\n",
    "\n",
    "Now we're going to use a chemistry-oriented dataset: [tdc](https://tdcommons.ai).\n",
    "It is a collection of datasets for drug discovery and is already preprocessed\n",
    "for machine learning. Among them, the `hERG Central` dataset, a part of the\n",
    "`Tox` prediction problem is selected for our task. The set is a collection of\n",
    "molecular structures expressed in SMILES strings and their percent inhibition of\n",
    "the hERG channel. The set is selected because it contains ~300,000 molecules,\n",
    "which is a reasonable size for training a deep learning model.\n",
    "\n",
    "Here, our goal is to predict whether the molecule *is a hERG blocker or not**,\n",
    "i.e., solving a binary classification problem.\n",
    "\n",
    "\\* Whether IC50 of the molecule is less than 10 µM or not."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Found local copy...\n",
      "Loading...\n",
      "Done!\n"
     ]
    }
   ],
   "source": [
    "from tdc.single_pred import Tox\n",
    "from tdc.utils import retrieve_label_name_list\n",
    "\n",
    "label_list = retrieve_label_name_list(\"herg_central\")\n",
    "data = Tox(name='herg_central', label_name=label_list[2])\n",
    "split = data.get_split()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "What's in the dataset?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'train':          Drug_ID                                               Drug  Y\n",
       " 0       26665387        COc1ccc(/C=C2\\SC(=S)N(N3CCOCC3)C2=O)c(OC)c1  0\n",
       " 1       26732361  COc1cc(OC)c2ccc(=O)oc2c1C(CC(=O)N1CCOCC1)c1ccc...  0\n",
       " 2       49735227  COc1cccc(NC(=O)C2C3C=CC4(O3)C2C(=O)N(CCCN2CCCC...  0\n",
       " 3       22413685                            Fc1ccc(Oc2ncccn2)c(F)c1  0\n",
       " 4       24783074     Cc1onc(C(=O)Nc2sc3c(c2C#N)CCCC3)c1[N+](=O)[O-]  0\n",
       " ...          ...                                                ... ..\n",
       " 214820  49824526          COc1ccc(-c2c(-c3ccccc3OC)sc3ccccc23)cc1OC  0\n",
       " 214821  22411411  COC(=O)c1c(NC(=O)CN2CCN(S(=O)(=O)c3ccccc3[N+](...  0\n",
       " 214822   4247425  O=C(Cn1c(=O)n(Cc2ccco2)c(=O)c2cc(Cl)ccc21)NCC1...  0\n",
       " 214823  26728232    CC(=O)CSc1nnc(CN2C(=O)CSc3ccc(C(F)(F)F)cc32)n1C  0\n",
       " 214824  49718253  CCOC(=O)[C@@H]1[C@H]2COc3ccc(Br)cc3[C@H]2N2C(=...  0\n",
       " \n",
       " [214825 rows x 3 columns],\n",
       " 'valid':         Drug_ID                                               Drug  Y\n",
       " 0      85146795  COc1ccc(-c2c(-c3cc(OC)cc(OC)c3)n(C)c3ccc(-c4cc...  0\n",
       " 1      26662044               CCC(=O)N1N=C(c2ccc(Cl)cc2)CC1c1ccco1  0\n",
       " 2      49648442    COc1ccc(-c2noc3c2-c2nc(-c4ccc(Br)cc4)sc2CC3)cc1  0\n",
       " 3      24815923   O=[N+]([O-])c1cccc(-c2ccc3nnc(SCc4ccccn4)n3n2)c1  0\n",
       " 4      49718311  CC(C)Oc1ccc(CN2CCC(n3nccc3NC(=O)CCc3ccccc3)CC2...  1\n",
       " ...         ...                                                ... ..\n",
       " 30684  22405985  COc1cc(/C=C(\\C#N)C(=O)NC2CCCCC2C)ccc1OS(=O)(=O...  0\n",
       " 30685   4257989           CCCOc1ccc(S(=O)(=O)NCCN2CCOCC2)c2ccccc12  0\n",
       " 30686  49649620  CCCOc1ccc(/C(O)=C2\\C(=O)C(=O)N(c3ccccn3)C2c2cc...  0\n",
       " 30687  14729934                          COc1cc2cc(Br)cc(Br)c2nc1C  0\n",
       " 30688  49644065             CN(C)C(=O)N1CCN(c2nc(-c3ccccc3)cs2)CC1  0\n",
       " \n",
       " [30689 rows x 3 columns],\n",
       " 'test':         Drug_ID                                               Drug  Y\n",
       " 0      17508895              Cc1c(C(=O)NCc2ccco2)cccc1[N+](=O)[O-]  0\n",
       " 1      49647072            CC1(C(=O)O)CCC(C(=O)Nc2cccc(F)c2)C1(C)C  0\n",
       " 2      49825891             N=c1oc2ccc(Br)cc2cc1C(=O)NCc1ccc(F)cc1  0\n",
       " 3      49649013          O=C(O)c1cc(-c2ccccc2Cl)n(-c2cccc(Cl)c2)n1  0\n",
       " 4      47195785         COc1cccc(OP(=O)(Nc2ccccc2)Oc2cccc(OC)c2)c1  0\n",
       " ...         ...                                                ... ..\n",
       " 61374  49737086      O=c1ccc(-c2ccc(S(=O)(=O)Nc3ccccc3Cl)o2)n[nH]1  0\n",
       " 61375  22406625            Cc1cc(Br)ccc1SCC(=O)OCC(=O)Nc1ccccc1C#N  0\n",
       " 61376  24821870              COc1ccc(CSC2=CS(=O)(=O)c3ccccc3N2)cc1  0\n",
       " 61377  56463241              O=C(/C=C/C=C/c1ccc2c(c1)OCO2)N1CCCCC1  0\n",
       " 61378   4244597  COc1cccc(-n2nnc3c(=O)n(CC(=O)N4CCCc5ccccc54)cn...  0\n",
       " \n",
       " [61379 rows x 3 columns]}"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "split"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "So the \"split\" is a dictionary of three keys, `train`, `valid`, and `test`.\n",
    "The other two are already familiar with us, but here we have one more: the\n",
    "`test` set. What is it actually?\n",
    "\n",
    "Before getting to the answer directly, think about why we needed a validation\n",
    "set. What does it validate? Well, the model, of course. More precisely, the set\n",
    "validates the model's ability to generalize. The validation set *validates* the\n",
    "model's parameters, that the model indeed performs well on the unseen data.\n",
    "That's why the model is trained on the training set and evaluated on the\n",
    "validation set.\n",
    "\n",
    "Now think about the training procedure. We tune the hyperparameters of the model\n",
    "to maximize the performance on the validation set. Did you get the point here?\n",
    "We have a \"test\" set for the model parameters, the validation set, but we don't\n",
    "have one for the *hyperparameters*. That's why we need the test set. We will\n",
    "use it to evaluate the model's actual performance, with the best hyperparameters\n",
    "(or at least, the best we can get).\n",
    "\n",
    "Next, let's take a closer look at the training set. It is a standard\n",
    "`pandas.DataFrame` object with three columns: PubChem CID of the drug, SMILES\n",
    "string, and the label. The others are also `pandas.DataFrame` objects with the\n",
    "same layout."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Drug_ID</th>\n",
       "      <th>Drug</th>\n",
       "      <th>Y</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>26665387</td>\n",
       "      <td>COc1ccc(/C=C2\\SC(=S)N(N3CCOCC3)C2=O)c(OC)c1</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>26732361</td>\n",
       "      <td>COc1cc(OC)c2ccc(=O)oc2c1C(CC(=O)N1CCOCC1)c1ccc...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>49735227</td>\n",
       "      <td>COc1cccc(NC(=O)C2C3C=CC4(O3)C2C(=O)N(CCCN2CCCC...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>22413685</td>\n",
       "      <td>Fc1ccc(Oc2ncccn2)c(F)c1</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>24783074</td>\n",
       "      <td>Cc1onc(C(=O)Nc2sc3c(c2C#N)CCCC3)c1[N+](=O)[O-]</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>214820</th>\n",
       "      <td>49824526</td>\n",
       "      <td>COc1ccc(-c2c(-c3ccccc3OC)sc3ccccc23)cc1OC</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>214821</th>\n",
       "      <td>22411411</td>\n",
       "      <td>COC(=O)c1c(NC(=O)CN2CCN(S(=O)(=O)c3ccccc3[N+](...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>214822</th>\n",
       "      <td>4247425</td>\n",
       "      <td>O=C(Cn1c(=O)n(Cc2ccco2)c(=O)c2cc(Cl)ccc21)NCC1...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>214823</th>\n",
       "      <td>26728232</td>\n",
       "      <td>CC(=O)CSc1nnc(CN2C(=O)CSc3ccc(C(F)(F)F)cc32)n1C</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>214824</th>\n",
       "      <td>49718253</td>\n",
       "      <td>CCOC(=O)[C@@H]1[C@H]2COc3ccc(Br)cc3[C@H]2N2C(=...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>214825 rows × 3 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "         Drug_ID                                               Drug  Y\n",
       "0       26665387        COc1ccc(/C=C2\\SC(=S)N(N3CCOCC3)C2=O)c(OC)c1  0\n",
       "1       26732361  COc1cc(OC)c2ccc(=O)oc2c1C(CC(=O)N1CCOCC1)c1ccc...  0\n",
       "2       49735227  COc1cccc(NC(=O)C2C3C=CC4(O3)C2C(=O)N(CCCN2CCCC...  0\n",
       "3       22413685                            Fc1ccc(Oc2ncccn2)c(F)c1  0\n",
       "4       24783074     Cc1onc(C(=O)Nc2sc3c(c2C#N)CCCC3)c1[N+](=O)[O-]  0\n",
       "...          ...                                                ... ..\n",
       "214820  49824526          COc1ccc(-c2c(-c3ccccc3OC)sc3ccccc23)cc1OC  0\n",
       "214821  22411411  COC(=O)c1c(NC(=O)CN2CCN(S(=O)(=O)c3ccccc3[N+](...  0\n",
       "214822   4247425  O=C(Cn1c(=O)n(Cc2ccco2)c(=O)c2cc(Cl)ccc21)NCC1...  0\n",
       "214823  26728232    CC(=O)CSc1nnc(CN2C(=O)CSc3ccc(C(F)(F)F)cc32)n1C  0\n",
       "214824  49718253  CCOC(=O)[C@@H]1[C@H]2COc3ccc(Br)cc3[C@H]2N2C(=...  0\n",
       "\n",
       "[214825 rows x 3 columns]"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "split[\"train\"]"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "As usual, we need to wrap those dataframes with a pytorch `Dataset` class. One\n",
    "thing to note here is that we have to convert the SMILES strings to the standard\n",
    "pytorch `Tensor`s before feeding them into the model. Let's do it here, so that\n",
    "we don't have to convert every time retrieving a sample."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[('c', 3379056),\n",
       " ('C', 2261877),\n",
       " ('(', 1259627),\n",
       " (')', 1259627),\n",
       " ('1', 891209),\n",
       " ('O', 877213),\n",
       " ('=', 650423),\n",
       " ('2', 636321),\n",
       " ('N', 531525),\n",
       " ('n', 352423),\n",
       " ('3', 257828),\n",
       " ('S', 123148),\n",
       " ('[', 103232),\n",
       " (']', 103232),\n",
       " ('-', 84076),\n",
       " ('F', 69489),\n",
       " ('l', 63085),\n",
       " ('/', 60403),\n",
       " ('H', 54053),\n",
       " ('o', 50749),\n",
       " ('4', 49028),\n",
       " ('s', 47252),\n",
       " ('@', 34905),\n",
       " ('+', 23122),\n",
       " ('B', 16051),\n",
       " ('r', 16032),\n",
       " ('#', 14263),\n",
       " ('.', 12118),\n",
       " ('\\\\', 11921),\n",
       " ('5', 4649),\n",
       " ('P', 1229),\n",
       " ('I', 565),\n",
       " ('6', 352),\n",
       " ('a', 178),\n",
       " ('7', 54),\n",
       " ('K', 48),\n",
       " ('A', 36),\n",
       " ('%', 18),\n",
       " ('8', 16),\n",
       " ('9', 10),\n",
       " ('0', 10),\n",
       " ('i', 8),\n",
       " ('g', 4),\n",
       " ('t', 3),\n",
       " ('e', 3),\n",
       " ('M', 2),\n",
       " ('L', 2),\n",
       " ('Z', 1),\n",
       " ('d', 1)]"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from collections import Counter\n",
    "\n",
    "mapping = Counter(c for smiles in data.entity1 for c in smiles)\n",
    "mapping.most_common()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Ok, not bad... But what are `l`, `r`, `a`, etc.? We don't have them in\n",
    "chemistry!\n",
    "\n",
    "The problem here is that each character in the SMILES string is **not** a\n",
    "*token*. We need a more sophisticated tokenizing scheme than the simple for\n",
    "loop. For the task, we need to take a look at the SMILES format specification*,\n",
    "and then implement a tokenizer that can split a SMILES string into a list of\n",
    "tokens.\n",
    "\n",
    "\\* See http://opensmiles.org/opensmiles.html"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[('c', 3379056),\n",
       " ('C', 2198789),\n",
       " ('(', 1259627),\n",
       " (')', 1259627),\n",
       " ('1', 891208),\n",
       " ('O', 877213),\n",
       " ('=', 650423),\n",
       " ('2', 636321),\n",
       " ('N', 531349),\n",
       " ('n', 352422),\n",
       " ('3', 257828),\n",
       " ('S', 123141),\n",
       " ('[', 103232),\n",
       " (']', 103232),\n",
       " ('-', 84076),\n",
       " ('F', 69488),\n",
       " ('Cl', 63085),\n",
       " ('/', 60403),\n",
       " ('H', 54051),\n",
       " ('o', 50748),\n",
       " ('4', 49028),\n",
       " ('s', 47215),\n",
       " ('@', 34905),\n",
       " ('+', 23122),\n",
       " ('Br', 16032),\n",
       " ('#', 14263),\n",
       " ('.', 12118),\n",
       " ('\\\\', 11921),\n",
       " ('5', 4648),\n",
       " ('P', 1226),\n",
       " ('I', 565),\n",
       " ('6', 352),\n",
       " ('Na', 175),\n",
       " ('7', 54),\n",
       " ('K', 48),\n",
       " ('As', 36),\n",
       " ('B', 18),\n",
       " ('%', 18),\n",
       " ('8', 16),\n",
       " ('9', 10),\n",
       " ('0', 10),\n",
       " ('Si', 6),\n",
       " ('Pt', 3),\n",
       " ('Mg', 2),\n",
       " ('Li', 2),\n",
       " ('Hg', 2),\n",
       " ('Ca', 2),\n",
       " ('se', 1),\n",
       " ('Ba', 1),\n",
       " ('Zn', 1),\n",
       " ('Nd', 1),\n",
       " ('Co', 1),\n",
       " ('Fe', 1),\n",
       " ('15', 1),\n",
       " ('Se', 1)]"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import re\n",
    "\n",
    "# Q) Why do we need Cl and Br first?\n",
    "token_re = re.compile(r\"(\\[[^\\]]+\\]|Cl|Br|[BCNOPSFIbcnops]|.)\")\n",
    "# Q) Why do we need separate rules for \"[A-Z][a-z]*\" and \"[a-z]*\"?\n",
    "#    What would happen if we merge them together into \"[A-Z]?[a-z]*\"?\n",
    "bracket_re = re.compile(r\"(\\d+|[A-Z][a-z]*|[a-z]+|.)\")\n",
    "\n",
    "\n",
    "def smi_to_tok(smiles: str):\n",
    "    for tok in token_re.finditer(smiles):\n",
    "        tok = tok.group()\n",
    "        if tok.startswith(\"[\"):\n",
    "            for t in bracket_re.finditer(tok):\n",
    "                yield t.group()\n",
    "        else:\n",
    "            yield tok\n",
    "\n",
    "\n",
    "mapping = Counter(tok for smiles in data.entity1 for tok in smi_to_tok(smiles))\n",
    "mapping.most_common()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Great! Now we made all symbols \"chemical,\" so the next step will be converting\n",
    "them into indices."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Q) Why do we start the index from 1?\n",
    "tok_to_idx = {\n",
    "    tok: idx for idx, (tok, _) in enumerate(mapping.most_common(), 1)}"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Finally we have the dataset class for our model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "from torch.utils.data import Dataset, DataLoader\n",
    "\n",
    "\n",
    "class HergCentralDataset(Dataset):\n",
    "    def __init__(self, data: pd.DataFrame):\n",
    "        # Q) Why do we need a list of Tensors, not a 2D Tensor?\n",
    "        self.smiles = [\n",
    "            torch.tensor([tok_to_idx[tok] for tok in smi_to_tok(smiles)])\n",
    "            for smiles in data[\"Drug\"]]\n",
    "        self.y = torch.tensor(data[\"Y\"].values, dtype=torch.long)\n",
    "\n",
    "    def __len__(self):\n",
    "        return len(self.smiles)\n",
    "\n",
    "    def __getitem__(self, idx):\n",
    "        x = self.smiles[idx]\n",
    "        y = self.y[idx]\n",
    "        return x, y"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "And the actual datasets for future use."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "datasets = {k: HergCentralDataset(v) for k, v in split.items()}"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### RNN model\n",
    "\n",
    "A SMILES string is basically a list of tokens, so it is natural to use a\n",
    "model that could handle sequences by design: the *RNN*s. RNNs, or Recurrent\n",
    "Neural Networks, were proposed for tasks related to sequential data. The basic\n",
    "idea is that the model has a *memory* that stores the information about the\n",
    "previous inputs.\n",
    "\n",
    "The memory is implemented as a hidden state, which is, in turn, a vector of\n",
    "numbers. The hidden state is updated along the model's output at each time step\n",
    "and used as a part of the input for the next time step.\n",
    "\n",
    "#### Main model\n",
    "\n",
    "So, this is our main model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "class RNN(nn.Module):\n",
    "    def __init__(self, params: Params):\n",
    "        super().__init__()\n",
    "        self.params = params\n",
    "\n",
    "        if self.params.num_layers < 1:\n",
    "            raise ValueError(\"num_layers must be >= 1\")\n",
    "\n",
    "        # Embedding layer\n",
    "        self.embedding = nn.Embedding(\n",
    "            len(tok_to_idx) + 1, self.params.embedding_dim, padding_idx=0)\n",
    "\n",
    "        # RNN layers\n",
    "        self.middle = self.params.RNN(\n",
    "            self.params.embedding_dim, self.params.hidden_dim,\n",
    "            num_layers=self.params.num_layers,\n",
    "            dropout=self.params.dropout)\n",
    "\n",
    "        # All readout functions make (L, N, hidden_dim) -> (N, hidden_dim)\n",
    "        if self.params.readout == \"sum\":\n",
    "            def readout(x: torch.Tensor):\n",
    "                return x.sum(dim=0)\n",
    "        elif self.params.readout == \"mean\":\n",
    "            def readout(x: torch.Tensor):\n",
    "                return x.mean(dim=0)\n",
    "        elif self.params.readout == \"max\":\n",
    "            def readout(x: torch.Tensor):\n",
    "                return x.max(dim=0)[0]\n",
    "        elif self.params.readout == \"last\":\n",
    "            def readout(x: torch.Tensor):\n",
    "                return x[-1]\n",
    "        else:\n",
    "            raise ValueError(f\"Unknown readout: {self.params.readout}\")\n",
    "\n",
    "        self.readout = readout\n",
    "\n",
    "        # Linear projection to the output\n",
    "        # Q) Why do we need 2 outputs?\n",
    "        self.out = nn.Linear(self.params.hidden_dim, 2)\n",
    "\n",
    "    def forward(self, X: torch.Tensor):\n",
    "        # Shape: (L, N) -> (L, N, embedding_dim)\n",
    "        h = self.embedding(X)\n",
    "        # Shape: (L, N, embedding_dim) -> (L, N, hidden_dim)\n",
    "        h, _ = self.middle(h)\n",
    "        # Shape: (L, N, hidden_dim) -> (N, hidden_dim)\n",
    "        h = self.readout(h)\n",
    "        # Shape: (N, hidden_dim) -> (N, 2)\n",
    "        y_hat = self.out(h)\n",
    "        return y_hat"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Training and validation code\n",
    "\n",
    "They are mostly imported from the previous chapter. Still, there are some\n",
    "significant changes."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "from sklearn import metrics\n",
    "from torch.nn import functional as F\n",
    "from tqdm import tqdm, trange\n",
    "\n",
    "\n",
    "class Callbacks:\n",
    "    def __init__(self, model: nn.Module):\n",
    "        self.model = model\n",
    "\n",
    "    def on_epoch_begin(self):\n",
    "        pass\n",
    "\n",
    "    def on_batch_begin(self):\n",
    "        pass\n",
    "\n",
    "    def on_batch_end(self, loss: torch.Tensor):\n",
    "        pass\n",
    "\n",
    "\n",
    "class TrainingCallback(Callbacks):\n",
    "    def __init__(self, model: nn.Module, optim: torch.optim.Optimizer):\n",
    "        super().__init__(model)\n",
    "        self.optim = optim\n",
    "\n",
    "    def on_epoch_begin(self):\n",
    "        self.model.train()\n",
    "\n",
    "    def on_batch_begin(self):\n",
    "        self.optim.zero_grad()\n",
    "\n",
    "    def on_batch_end(self, loss: torch.Tensor):\n",
    "        loss.backward()\n",
    "        self.optim.step()\n",
    "\n",
    "\n",
    "class ValidationCallback(Callbacks):\n",
    "    def on_epoch_begin(self):\n",
    "        self.model.eval()\n",
    "\n",
    "\n",
    "def run(model: nn.Module, criterion: nn.Module,\n",
    "        loader: DataLoader, callbacks: Callbacks):\n",
    "    callbacks.on_epoch_begin()\n",
    "\n",
    "    epoch_loss = []\n",
    "    epoch_true = []\n",
    "    epoch_pred = []\n",
    "\n",
    "    for x, y in loader:\n",
    "        callbacks.on_batch_begin()\n",
    "\n",
    "        # Here, we move the tensors to a GPU one batch at a time (if available).\n",
    "        # It is indeed slower than moving the entire dataset to the GPU at once,\n",
    "        # but it's also much more memory efficient.\n",
    "        # For larger models with larger datasets, this is a good trade-off.\n",
    "        # We want good models that might take long time to train, not bad\n",
    "        # models that train fast.\n",
    "        x = x.to(device)\n",
    "        y = y.to(device)\n",
    "\n",
    "        # Shape: (L, N) -> (N, 2)\n",
    "        y_hat = model(x)\n",
    "        # calculate loss\n",
    "        loss = criterion(y_hat, y)\n",
    "\n",
    "        callbacks.on_batch_end(loss)\n",
    "        epoch_loss.append(loss.item())\n",
    "\n",
    "        epoch_true.append(y.cpu().numpy())\n",
    "        epoch_pred.append(y_hat.argmax(dim=1).cpu().numpy())\n",
    "\n",
    "    epoch_true = np.concatenate(epoch_true)\n",
    "    epoch_pred = np.concatenate(epoch_pred)\n",
    "    epoch_accu = metrics.accuracy_score(epoch_true, epoch_pred)\n",
    "\n",
    "    return np.mean(epoch_loss), epoch_accu\n",
    "\n",
    "\n",
    "def train(model: nn.Module, criterion: nn.Module,\n",
    "          loader: DataLoader, optimizer: optim.Optimizer):\n",
    "    callbacks = TrainingCallback(model, optimizer)\n",
    "    return run(model, criterion, loader, callbacks)\n",
    "\n",
    "\n",
    "@torch.no_grad()\n",
    "def validate(model: nn.Module, criterion: nn.Module, loader: DataLoader):\n",
    "    callbacks = ValidationCallback(model)\n",
    "    return run(model, criterion, loader, callbacks)\n",
    "\n",
    "\n",
    "def train_rnn(model: RNN, criterion: nn.Module, train_loader, valid_loader):\n",
    "    optimizer = params.Optimizer(model.parameters(), lr=params.lr)\n",
    "    stats = []\n",
    "\n",
    "    for epoch in trange(1, params.epochs + 1):\n",
    "        train_loss, train_accu = train(\n",
    "            model, criterion, train_loader, optimizer)\n",
    "        valid_loss, valid_accu = validate(model, criterion, valid_loader)\n",
    "\n",
    "        if not epoch % params.print_every or epoch in (1, params.epochs):\n",
    "            tqdm.write((f\"{train_loss = :.4f}, {train_accu = :.2%}, \"\n",
    "                        f\"{valid_loss = :.4f}, {valid_accu = :.2%}\"))\n",
    "\n",
    "        stats.append((epoch, train_loss, train_accu, valid_loss, valid_accu))\n",
    "\n",
    "    stats = pd.DataFrame.from_records(stats, columns=[\n",
    "        \"epoch\", \"train_loss\", \"train_accu\", \"valid_loss\", \"valid_accu\"])\n",
    "    return stats"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "And here are some utility functions, also imported from the previous chapter:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "import random\n",
    "\n",
    "import seaborn as sns\n",
    "from matplotlib import pyplot as plt\n",
    "\n",
    "sns.set_theme(style=\"whitegrid\", palette=\"muted\")\n",
    "\n",
    "\n",
    "def seed_everything(seed=SEED):\n",
    "    \"\"\"Seed all the random number generators for reproducibility.\"\"\"\n",
    "    random.seed(seed)\n",
    "    np.random.seed(seed)\n",
    "    torch.manual_seed(seed)\n",
    "\n",
    "\n",
    "def to_dataloader(dataset: HergCentralDataset, shuffle: bool):\n",
    "    return DataLoader(dataset, params.batch_size,\n",
    "                      shuffle=shuffle,\n",
    "                      generator=torch.Generator().manual_seed(SEED))\n",
    "\n",
    "\n",
    "def get_dataloaders():\n",
    "    return (to_dataloader(v, k == \"train\") for k, v in datasets.items())\n",
    "\n",
    "\n",
    "def print_model_summary(model: nn.Module):\n",
    "    param_cnt = sum(param.numel()\n",
    "                    for param in model.parameters() if param.requires_grad)\n",
    "\n",
    "    print(model)\n",
    "    print(f\"Model has {param_cnt:,} trainable parameters\")\n",
    "\n",
    "\n",
    "def plot_stats(stats: pd.DataFrame):\n",
    "    stats_long = stats.melt(\n",
    "        id_vars=\"epoch\", var_name=\"type\", value_name=\"stat\")\n",
    "    stats_long[\"run\"] = [tp.split(\"_\")[0] for tp in stats_long[\"type\"]]\n",
    "    stats_long[\"type\"] = [tp.split(\"_\")[1] for tp in stats_long[\"type\"]]\n",
    "\n",
    "    epochs = params.epochs\n",
    "    fg = sns.relplot(x=\"epoch\", y=\"stat\", data=stats_long,\n",
    "                     hue=\"run\", col=\"type\", marker=\"o\", kind=\"line\",\n",
    "                     col_order=[\"loss\", \"accu\"],\n",
    "                     facet_kws={\"sharey\": False, \"xlim\": (1, epochs + 1)})\n",
    "    for ax in fg.axes[0, 1:]:\n",
    "        ax.set_ylim(0, 1)\n",
    "\n",
    "    xticks = list(range(1, epochs + 1, params.print_every))\n",
    "    if epochs != xticks[-1]:\n",
    "        xticks.append(epochs)\n",
    "    fg.set(xticks=xticks)\n",
    "    plt.show()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Model training and hyperparameter optimization: first try\n",
    "\n",
    "Now we're ready to train the model. Let's try!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "RNN(\n",
      "  (embedding): Embedding(56, 32)\n",
      "  (middle): GRU(32, 32, num_layers=3, dropout=0.2)\n",
      "  (out): Linear(in_features=32, out_features=2, bias=True)\n",
      ")\n",
      "Model has 20,866 trainable parameters\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "  0%|          | 0/20 [00:00<?, ?it/s]\n"
     ]
    },
    {
     "ename": "RuntimeError",
     "evalue": "stack expects each tensor to be equal size, but got [31] at entry 0 and [63] at entry 1",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mRuntimeError\u001b[0m                              Traceback (most recent call last)",
      "Cell \u001b[0;32mIn[15], line 18\u001b[0m\n\u001b[1;32m     13\u001b[0m     plot_stats(stats)\n\u001b[1;32m     15\u001b[0m     \u001b[39mreturn\u001b[39;00m model\n\u001b[0;32m---> 18\u001b[0m model \u001b[39m=\u001b[39m main()\n",
      "Cell \u001b[0;32mIn[15], line 12\u001b[0m, in \u001b[0;36mmain\u001b[0;34m()\u001b[0m\n\u001b[1;32m      9\u001b[0m \u001b[39m# Visualize the model\u001b[39;00m\n\u001b[1;32m     10\u001b[0m print_model_summary(model)\n\u001b[0;32m---> 12\u001b[0m stats \u001b[39m=\u001b[39m train_rnn(model, criterion, train_loader, valid_loader)\n\u001b[1;32m     13\u001b[0m plot_stats(stats)\n\u001b[1;32m     15\u001b[0m \u001b[39mreturn\u001b[39;00m model\n",
      "Cell \u001b[0;32mIn[13], line 97\u001b[0m, in \u001b[0;36mtrain_rnn\u001b[0;34m(model, criterion, train_loader, valid_loader)\u001b[0m\n\u001b[1;32m     94\u001b[0m stats \u001b[39m=\u001b[39m []\n\u001b[1;32m     96\u001b[0m \u001b[39mfor\u001b[39;00m epoch \u001b[39min\u001b[39;00m trange(\u001b[39m1\u001b[39m, params\u001b[39m.\u001b[39mepochs \u001b[39m+\u001b[39m \u001b[39m1\u001b[39m):\n\u001b[0;32m---> 97\u001b[0m     train_loss, train_accu \u001b[39m=\u001b[39m train(\n\u001b[1;32m     98\u001b[0m         model, criterion, train_loader, optimizer)\n\u001b[1;32m     99\u001b[0m     valid_loss, valid_accu \u001b[39m=\u001b[39m validate(model, criterion, valid_loader)\n\u001b[1;32m    101\u001b[0m     \u001b[39mif\u001b[39;00m \u001b[39mnot\u001b[39;00m epoch \u001b[39m%\u001b[39m params\u001b[39m.\u001b[39mprint_every \u001b[39mor\u001b[39;00m epoch \u001b[39min\u001b[39;00m (\u001b[39m1\u001b[39m, params\u001b[39m.\u001b[39mepochs):\n",
      "Cell \u001b[0;32mIn[13], line 83\u001b[0m, in \u001b[0;36mtrain\u001b[0;34m(model, criterion, loader, optimizer)\u001b[0m\n\u001b[1;32m     80\u001b[0m \u001b[39mdef\u001b[39;00m \u001b[39mtrain\u001b[39m(model: nn\u001b[39m.\u001b[39mModule, criterion: nn\u001b[39m.\u001b[39mModule,\n\u001b[1;32m     81\u001b[0m           loader: DataLoader, optimizer: optim\u001b[39m.\u001b[39mOptimizer):\n\u001b[1;32m     82\u001b[0m     callbacks \u001b[39m=\u001b[39m TrainingCallback(model, optimizer)\n\u001b[0;32m---> 83\u001b[0m     \u001b[39mreturn\u001b[39;00m run(model, criterion, loader, callbacks)\n",
      "Cell \u001b[0;32mIn[13], line 50\u001b[0m, in \u001b[0;36mrun\u001b[0;34m(model, criterion, loader, callbacks)\u001b[0m\n\u001b[1;32m     47\u001b[0m epoch_true \u001b[39m=\u001b[39m []\n\u001b[1;32m     48\u001b[0m epoch_pred \u001b[39m=\u001b[39m []\n\u001b[0;32m---> 50\u001b[0m \u001b[39mfor\u001b[39;00m x, y \u001b[39min\u001b[39;00m loader:\n\u001b[1;32m     51\u001b[0m     callbacks\u001b[39m.\u001b[39mon_batch_begin()\n\u001b[1;32m     53\u001b[0m     \u001b[39m# Here, we move the tensors to a GPU one batch at a time (if available).\u001b[39;00m\n\u001b[1;32m     54\u001b[0m     \u001b[39m# It is indeed slower than moving the entire dataset to the GPU at once,\u001b[39;00m\n\u001b[1;32m     55\u001b[0m     \u001b[39m# but it's also much more memory efficient.\u001b[39;00m\n\u001b[1;32m     56\u001b[0m     \u001b[39m# For larger models with larger datasets, this is a good trade-off.\u001b[39;00m\n\u001b[1;32m     57\u001b[0m     \u001b[39m# We want good models that might take long time to train, not bad\u001b[39;00m\n\u001b[1;32m     58\u001b[0m     \u001b[39m# models that train fast.\u001b[39;00m\n",
      "File \u001b[0;32m~/anaconda3/envs/ldld/lib/python3.10/site-packages/torch/utils/data/dataloader.py:628\u001b[0m, in \u001b[0;36m_BaseDataLoaderIter.__next__\u001b[0;34m(self)\u001b[0m\n\u001b[1;32m    625\u001b[0m \u001b[39mif\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_sampler_iter \u001b[39mis\u001b[39;00m \u001b[39mNone\u001b[39;00m:\n\u001b[1;32m    626\u001b[0m     \u001b[39m# TODO(https://github.com/pytorch/pytorch/issues/76750)\u001b[39;00m\n\u001b[1;32m    627\u001b[0m     \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_reset()  \u001b[39m# type: ignore[call-arg]\u001b[39;00m\n\u001b[0;32m--> 628\u001b[0m data \u001b[39m=\u001b[39m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49m_next_data()\n\u001b[1;32m    629\u001b[0m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_num_yielded \u001b[39m+\u001b[39m\u001b[39m=\u001b[39m \u001b[39m1\u001b[39m\n\u001b[1;32m    630\u001b[0m \u001b[39mif\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_dataset_kind \u001b[39m==\u001b[39m _DatasetKind\u001b[39m.\u001b[39mIterable \u001b[39mand\u001b[39;00m \\\n\u001b[1;32m    631\u001b[0m         \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_IterableDataset_len_called \u001b[39mis\u001b[39;00m \u001b[39mnot\u001b[39;00m \u001b[39mNone\u001b[39;00m \u001b[39mand\u001b[39;00m \\\n\u001b[1;32m    632\u001b[0m         \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_num_yielded \u001b[39m>\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_IterableDataset_len_called:\n",
      "File \u001b[0;32m~/anaconda3/envs/ldld/lib/python3.10/site-packages/torch/utils/data/dataloader.py:671\u001b[0m, in \u001b[0;36m_SingleProcessDataLoaderIter._next_data\u001b[0;34m(self)\u001b[0m\n\u001b[1;32m    669\u001b[0m \u001b[39mdef\u001b[39;00m \u001b[39m_next_data\u001b[39m(\u001b[39mself\u001b[39m):\n\u001b[1;32m    670\u001b[0m     index \u001b[39m=\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_next_index()  \u001b[39m# may raise StopIteration\u001b[39;00m\n\u001b[0;32m--> 671\u001b[0m     data \u001b[39m=\u001b[39m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49m_dataset_fetcher\u001b[39m.\u001b[39;49mfetch(index)  \u001b[39m# may raise StopIteration\u001b[39;00m\n\u001b[1;32m    672\u001b[0m     \u001b[39mif\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_pin_memory:\n\u001b[1;32m    673\u001b[0m         data \u001b[39m=\u001b[39m _utils\u001b[39m.\u001b[39mpin_memory\u001b[39m.\u001b[39mpin_memory(data, \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_pin_memory_device)\n",
      "File \u001b[0;32m~/anaconda3/envs/ldld/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py:61\u001b[0m, in \u001b[0;36m_MapDatasetFetcher.fetch\u001b[0;34m(self, possibly_batched_index)\u001b[0m\n\u001b[1;32m     59\u001b[0m \u001b[39melse\u001b[39;00m:\n\u001b[1;32m     60\u001b[0m     data \u001b[39m=\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39mdataset[possibly_batched_index]\n\u001b[0;32m---> 61\u001b[0m \u001b[39mreturn\u001b[39;00m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49mcollate_fn(data)\n",
      "File \u001b[0;32m~/anaconda3/envs/ldld/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:265\u001b[0m, in \u001b[0;36mdefault_collate\u001b[0;34m(batch)\u001b[0m\n\u001b[1;32m    204\u001b[0m \u001b[39mdef\u001b[39;00m \u001b[39mdefault_collate\u001b[39m(batch):\n\u001b[1;32m    205\u001b[0m \u001b[39m    \u001b[39m\u001b[39mr\u001b[39m\u001b[39m\"\"\"\u001b[39;00m\n\u001b[1;32m    206\u001b[0m \u001b[39m        Function that takes in a batch of data and puts the elements within the batch\u001b[39;00m\n\u001b[1;32m    207\u001b[0m \u001b[39m        into a tensor with an additional outer dimension - batch size. The exact output type can be\u001b[39;00m\n\u001b[0;32m   (...)\u001b[0m\n\u001b[1;32m    263\u001b[0m \u001b[39m            >>> default_collate(batch)  # Handle `CustomType` automatically\u001b[39;00m\n\u001b[1;32m    264\u001b[0m \u001b[39m    \"\"\"\u001b[39;00m\n\u001b[0;32m--> 265\u001b[0m     \u001b[39mreturn\u001b[39;00m collate(batch, collate_fn_map\u001b[39m=\u001b[39;49mdefault_collate_fn_map)\n",
      "File \u001b[0;32m~/anaconda3/envs/ldld/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:143\u001b[0m, in \u001b[0;36mcollate\u001b[0;34m(batch, collate_fn_map)\u001b[0m\n\u001b[1;32m    140\u001b[0m transposed \u001b[39m=\u001b[39m \u001b[39mlist\u001b[39m(\u001b[39mzip\u001b[39m(\u001b[39m*\u001b[39mbatch))  \u001b[39m# It may be accessed twice, so we use a list.\u001b[39;00m\n\u001b[1;32m    142\u001b[0m \u001b[39mif\u001b[39;00m \u001b[39misinstance\u001b[39m(elem, \u001b[39mtuple\u001b[39m):\n\u001b[0;32m--> 143\u001b[0m     \u001b[39mreturn\u001b[39;00m [collate(samples, collate_fn_map\u001b[39m=\u001b[39mcollate_fn_map) \u001b[39mfor\u001b[39;00m samples \u001b[39min\u001b[39;00m transposed]  \u001b[39m# Backwards compatibility.\u001b[39;00m\n\u001b[1;32m    144\u001b[0m \u001b[39melse\u001b[39;00m:\n\u001b[1;32m    145\u001b[0m     \u001b[39mtry\u001b[39;00m:\n",
      "File \u001b[0;32m~/anaconda3/envs/ldld/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:143\u001b[0m, in \u001b[0;36m<listcomp>\u001b[0;34m(.0)\u001b[0m\n\u001b[1;32m    140\u001b[0m transposed \u001b[39m=\u001b[39m \u001b[39mlist\u001b[39m(\u001b[39mzip\u001b[39m(\u001b[39m*\u001b[39mbatch))  \u001b[39m# It may be accessed twice, so we use a list.\u001b[39;00m\n\u001b[1;32m    142\u001b[0m \u001b[39mif\u001b[39;00m \u001b[39misinstance\u001b[39m(elem, \u001b[39mtuple\u001b[39m):\n\u001b[0;32m--> 143\u001b[0m     \u001b[39mreturn\u001b[39;00m [collate(samples, collate_fn_map\u001b[39m=\u001b[39;49mcollate_fn_map) \u001b[39mfor\u001b[39;00m samples \u001b[39min\u001b[39;00m transposed]  \u001b[39m# Backwards compatibility.\u001b[39;00m\n\u001b[1;32m    144\u001b[0m \u001b[39melse\u001b[39;00m:\n\u001b[1;32m    145\u001b[0m     \u001b[39mtry\u001b[39;00m:\n",
      "File \u001b[0;32m~/anaconda3/envs/ldld/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:120\u001b[0m, in \u001b[0;36mcollate\u001b[0;34m(batch, collate_fn_map)\u001b[0m\n\u001b[1;32m    118\u001b[0m \u001b[39mif\u001b[39;00m collate_fn_map \u001b[39mis\u001b[39;00m \u001b[39mnot\u001b[39;00m \u001b[39mNone\u001b[39;00m:\n\u001b[1;32m    119\u001b[0m     \u001b[39mif\u001b[39;00m elem_type \u001b[39min\u001b[39;00m collate_fn_map:\n\u001b[0;32m--> 120\u001b[0m         \u001b[39mreturn\u001b[39;00m collate_fn_map[elem_type](batch, collate_fn_map\u001b[39m=\u001b[39;49mcollate_fn_map)\n\u001b[1;32m    122\u001b[0m     \u001b[39mfor\u001b[39;00m collate_type \u001b[39min\u001b[39;00m collate_fn_map:\n\u001b[1;32m    123\u001b[0m         \u001b[39mif\u001b[39;00m \u001b[39misinstance\u001b[39m(elem, collate_type):\n",
      "File \u001b[0;32m~/anaconda3/envs/ldld/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:163\u001b[0m, in \u001b[0;36mcollate_tensor_fn\u001b[0;34m(batch, collate_fn_map)\u001b[0m\n\u001b[1;32m    161\u001b[0m     storage \u001b[39m=\u001b[39m elem\u001b[39m.\u001b[39mstorage()\u001b[39m.\u001b[39m_new_shared(numel, device\u001b[39m=\u001b[39melem\u001b[39m.\u001b[39mdevice)\n\u001b[1;32m    162\u001b[0m     out \u001b[39m=\u001b[39m elem\u001b[39m.\u001b[39mnew(storage)\u001b[39m.\u001b[39mresize_(\u001b[39mlen\u001b[39m(batch), \u001b[39m*\u001b[39m\u001b[39mlist\u001b[39m(elem\u001b[39m.\u001b[39msize()))\n\u001b[0;32m--> 163\u001b[0m \u001b[39mreturn\u001b[39;00m torch\u001b[39m.\u001b[39;49mstack(batch, \u001b[39m0\u001b[39;49m, out\u001b[39m=\u001b[39;49mout)\n",
      "\u001b[0;31mRuntimeError\u001b[0m: stack expects each tensor to be equal size, but got [31] at entry 0 and [63] at entry 1"
     ]
    }
   ],
   "source": [
    "criterion = nn.CrossEntropyLoss()\n",
    "train_loader, valid_loader, test_loader = get_dataloaders()\n",
    "\n",
    "\n",
    "def main():\n",
    "    seed_everything()\n",
    "\n",
    "    model = RNN(params).to(device)\n",
    "    # Visualize the model\n",
    "    print_model_summary(model)\n",
    "\n",
    "    stats = train_rnn(model, criterion, train_loader, valid_loader)\n",
    "    plot_stats(stats)\n",
    "\n",
    "    return model\n",
    "\n",
    "\n",
    "model = main()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Something is definitely wrong here. Could you find out what?"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Introduction to real-world problems\n",
    "\n",
    "#### Collate function\n",
    "\n",
    "The problem here is that the sequences (SMILES strings) have different lengths.\n",
    "The `DataLoader` class requires that all samples in a batch are of the same\n",
    "length. In the previous chapters, all figures (happened to) had the same size,\n",
    "so we didn't have to worry about it. But in real-world problems like this, the\n",
    "inputs might have different lengths.\n",
    "\n",
    "So we need to *pad* the shorter sequences with some special tokens. Here, we'll\n",
    "use `0` as the padding index, which is the most common choice (now you get why\n",
    "we started the token indices from `1`).\n",
    "\n",
    "The `collate_fn` argument of the `DataLoader` class is responsible for stacking\n",
    "all the batch samples into a single pytorch `Tensor`. The default implementation\n",
    "simply assumes all the samples are of the same length and try to stack them. But\n",
    "here, we need to pad them before stacking, so we need to implement one on our\n",
    "own."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "from typing import List, Tuple\n",
    "\n",
    "from torch.nn.utils.rnn import pad_sequence\n",
    "\n",
    "\n",
    "def collate_fn(batch: List[Tuple[torch.Tensor, torch.Tensor]]):\n",
    "    # Q) What does this line do?\n",
    "    x, y = zip(*batch)\n",
    "    # Shape: N (L_i, ) sized tensors -> (max(L_i), N)\n",
    "    x = pad_sequence(x)\n",
    "    # Shape: N * (, ) -> (N, )\n",
    "    y = torch.stack(y)\n",
    "    return x, y"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Also don't forget to re-define the `to_dataloader` function."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "def to_dataloader(dataset: HergCentralDataset, shuffle: bool):\n",
    "    return DataLoader(dataset, params.batch_size,\n",
    "                      shuffle=shuffle, collate_fn=collate_fn,\n",
    "                      generator=torch.Generator().manual_seed(SEED))"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Are we done? No, not yet. Since we've padded the sequences, the model should\n",
    "ignore the padding indices. However, the current model assumes every token is\n",
    "valid, and that could lead to incorrect training.\n",
    "\n",
    "Before actually fixing the model, first we need some utility functions."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_mask(X: torch.Tensor):\n",
    "    return X != 0\n",
    "\n",
    "\n",
    "def masked_sum(X: torch.Tensor, mask: torch.Tensor):\n",
    "    return (X * mask).sum(dim=0)\n",
    "\n",
    "\n",
    "def masked_mean(X: torch.Tensor, mask: torch.Tensor):\n",
    "    x_sum = masked_sum(X, mask)\n",
    "    mask_sum = mask.sum(dim=0)\n",
    "    return x_sum / mask_sum\n",
    "\n",
    "\n",
    "def masked_max(X: torch.Tensor, mask: torch.Tensor):\n",
    "    return (X * mask).max(dim=0)[0]\n",
    "\n",
    "\n",
    "def masked_last(X: torch.Tensor, mask: torch.Tensor):\n",
    "    # Q) Why the following line works?\n",
    "    last_idx = mask.sum(dim=0) - 1\n",
    "    return X[last_idx, torch.arange(mask.shape[1], device=X.device)]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "class RNN(nn.Module):\n",
    "    def __init__(self, params: Params):\n",
    "        super().__init__()\n",
    "        self.params = params\n",
    "\n",
    "        if self.params.num_layers < 1:\n",
    "            raise ValueError(\"num_layers must be >= 1\")\n",
    "\n",
    "        # Embedding layer\n",
    "        self.embedding = nn.Embedding(\n",
    "            len(tok_to_idx) + 1, self.params.embedding_dim, padding_idx=0)\n",
    "\n",
    "        # RNN layers\n",
    "        self.middle = self.params.RNN(\n",
    "            self.params.embedding_dim, self.params.hidden_dim,\n",
    "            num_layers=self.params.num_layers,\n",
    "            dropout=self.params.dropout)\n",
    "\n",
    "        # All readout functions make (L, N, hidden_dim) -> (N, hidden_dim)\n",
    "        if self.params.readout == \"sum\":\n",
    "            self.readout = masked_sum\n",
    "        elif self.params.readout == \"mean\":\n",
    "            self.readout = masked_mean\n",
    "        elif self.params.readout == \"max\":\n",
    "            self.readout = masked_max\n",
    "        elif self.params.readout == \"last\":\n",
    "            self.readout = masked_last\n",
    "        else:\n",
    "            raise ValueError(f\"Unknown readout: {self.params.readout}\")\n",
    "\n",
    "        # Linear projection to the output\n",
    "        self.out = nn.Linear(self.params.hidden_dim, 2)\n",
    "\n",
    "    def forward(self, X: torch.Tensor):\n",
    "        # Q) Why do we unsqueeze the mask?\n",
    "        mask = get_mask(X).unsqueeze_(-1)\n",
    "\n",
    "        # Shape: (L, N) -> (L, N, embedding_dim)\n",
    "        h = self.embedding(X)\n",
    "        # Shape: (L, N, embedding_dim) -> (L, N, hidden_dim)\n",
    "        h, _ = self.middle(h)\n",
    "        # Shape: (L, N, hidden_dim) -> (N, hidden_dim)\n",
    "        h = self.readout(h, mask)\n",
    "        # Shape: (N, hidden_dim) -> (N, 2)\n",
    "        y_hat = self.out(h)\n",
    "        return y_hat"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now we can train the model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "RNN(\n",
      "  (embedding): Embedding(56, 32, padding_idx=0)\n",
      "  (middle): GRU(32, 32, num_layers=3, dropout=0.2)\n",
      "  (out): Linear(in_features=32, out_features=2, bias=True)\n",
      ")\n",
      "Model has 20,866 trainable parameters\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "  5%|▌         | 1/20 [00:10<03:24, 10.74s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.2559, train_accu = 91.21%, valid_loss = 0.1788, valid_accu = 95.68%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 20%|██        | 4/20 [00:41<02:43, 10.21s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.1821, train_accu = 95.48%, valid_loss = 0.1744, valid_accu = 95.67%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 40%|████      | 8/20 [01:22<02:03, 10.30s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.1531, train_accu = 95.48%, valid_loss = 0.1490, valid_accu = 95.67%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 60%|██████    | 12/20 [02:01<01:19,  9.88s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.1481, train_accu = 95.48%, valid_loss = 0.1462, valid_accu = 95.67%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 80%|████████  | 16/20 [02:37<00:37,  9.34s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.1449, train_accu = 95.49%, valid_loss = 0.1436, valid_accu = 95.67%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 20/20 [03:12<00:00,  9.62s/it]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.1424, train_accu = 95.50%, valid_loss = 0.1405, valid_accu = 95.66%\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1091.88x500 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "train_loader, valid_loader, test_loader = get_dataloaders()\n",
    "model = main()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Imbalanced dataset\n",
    "\n",
    "It seems that the model performs really well, but is it *really* good? Let's\n",
    "take a deeper look into the dataset."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.04515303298830986"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "datasets[\"train\"].y.float().mean().item()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The mean of the labels is around 0.05, which means the dataset is *imbalanced*.\n",
    "In such case, if a model predicts all samples as \"not a hERG blocker,\" it would\n",
    "still get a 95% accuracy. Indeed, it is happening here: let's draw a confusion\n",
    "matrix of the trained model on the validation set."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(0.04324024894913487, 0.0005213594447521913)"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "@torch.no_grad()\n",
    "def draw_confusion_matrix():\n",
    "    model.eval()\n",
    "\n",
    "    y_true = []\n",
    "    y_pred = []\n",
    "    for X, y in valid_loader:\n",
    "        y_true.append(y.numpy())\n",
    "\n",
    "        y_hat = model(X.to(device))\n",
    "        y_pred.append(y_hat.argmax(dim=1).cpu().numpy())\n",
    "\n",
    "    y_true = np.concatenate(y_true)\n",
    "    y_pred = np.concatenate(y_pred)\n",
    "\n",
    "    metrics.ConfusionMatrixDisplay.from_predictions(y_true, y_pred)\n",
    "    return y_true, y_pred\n",
    "\n",
    "\n",
    "y_true, y_pred = draw_confusion_matrix()\n",
    "y_true.mean(), y_pred.mean()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The model merely predicts everything (>99.9%) to be \"not a hERG blocker.\" This\n",
    "is not what we're happy with. An important implication from this result is that\n",
    "you should never blindly trust the numbers that models produce. Always take a\n",
    "look at the details.\n",
    "\n",
    "Well, actually, there *is* another metric that is much more useful than the\n",
    "accuracy for this case: the F1 score. It is defined as a harmonic mean of the\n",
    "precision and recall, and it is more robust to the imbalanced dataset (still,\n",
    "this does not invalidate the claim that we've made in the last paragraph).\n",
    "\n",
    "The model now performs terribly in terms of the F1 score."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.008935219657483246"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "metrics.f1_score(y_true, y_pred)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "There are many ways to deal with the imbalanced dataset problem. One of the\n",
    "easiest ways is to adjust weights of the loss function, so that the model could\n",
    "take account the imbalance. The weights are calculated as the inverse of the\n",
    "class frequencies.\n",
    "\n",
    "Note that we need to have the sum of the weight equal to the count of the\n",
    "classes, so that the loss does not (accidentally) scale down the learning rate."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_set = datasets[\"train\"]\n",
    "\n",
    "total = len(train_set)\n",
    "pos_freq = train_set.y.sum().item() / total\n",
    "\n",
    "w = torch.tensor([pos_freq, 1 - pos_freq])\n",
    "w *= len(w)\n",
    "\n",
    "criterion = nn.CrossEntropyLoss(weight=w).to(device)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Model training and hyperparameter optimization: second try\n",
    "\n",
    "Let's see the differences."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "RNN(\n",
      "  (embedding): Embedding(56, 32, padding_idx=0)\n",
      "  (middle): GRU(32, 32, num_layers=3, dropout=0.2)\n",
      "  (out): Linear(in_features=32, out_features=2, bias=True)\n",
      ")\n",
      "Model has 20,866 trainable parameters\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "  5%|▌         | 1/20 [00:08<02:45,  8.71s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.6418, train_accu = 66.39%, valid_loss = 0.5840, valid_accu = 68.20%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 20%|██        | 4/20 [00:38<02:31,  9.48s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.5245, train_accu = 77.13%, valid_loss = 0.5253, valid_accu = 74.65%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 40%|████      | 8/20 [01:15<01:53,  9.46s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.4999, train_accu = 78.22%, valid_loss = 0.5056, valid_accu = 81.49%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 60%|██████    | 12/20 [01:56<01:20, 10.02s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.4861, train_accu = 78.59%, valid_loss = 0.4955, valid_accu = 81.79%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      " 80%|████████  | 16/20 [02:35<00:39,  9.83s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.4757, train_accu = 79.06%, valid_loss = 0.4875, valid_accu = 81.80%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 20/20 [03:15<00:00,  9.78s/it]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_loss = 0.4705, train_accu = 79.56%, valid_loss = 0.4799, valid_accu = 79.51%\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1091.88x500 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "train_loader, valid_loader, test_loader = get_dataloaders()\n",
    "model = main()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Yes, the model seems to be performing worse than the previous one in terms of\n",
    "accuracy, but the more important thing here is that the model has now started to\n",
    "consider the \"hERG blocker\" class. Check out the **much** better F1 score.\n",
    "\n",
    "Also, the training loss seems not to be saturated, so this model *might* perform\n",
    "better than the previous one eventually."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(0.22793183225259864, 0.04324024894913487, 0.24441240086517665)"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAhoAAAG2CAYAAAA5jy8uAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjYuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8o6BhiAAAACXBIWXMAAA9hAAAPYQGoP6dpAABISklEQVR4nO3deVxUZfs/8M8wMCCOI6LIpiiioCgqVqKZ4JqlkJm5PZn7BprWE4/hN/X5mbi1SOJuUrmVSampqeVSaKWWYu4LIokiIos47DPMnN8fPEyOoAPOHEc5n/frNS+bc+5zn3uQ4PK67vs+MkEQBBARERGJwMbaAyAiIqKai4EGERERiYaBBhEREYmGgQYRERGJhoEGERERiYaBBhEREYmGgQYRERGJhoEGERERicbW2gMQw8mTJyEIAuzs7Kw9FCIiqiatVguZTIbAwEDR7nHp0iVoNBqL9KVQKODn52eRvmqiGhloCIIAQdACulvWHookCJBDq6sPO3k2ZNBZeziScTNLZe0hSIbcRgYnlT1y1SXQ6bmZstjqqeyhF/nrrNFooNEUQSHPMq8fXQMLjajmqpGBhp2dHaC7hdZu06w9FEko1DTFpcyFaOb8MRwVf1t7OJIRPmeCtYcgGY3clPjPuI5YG38aN27lW3s4Nd6syZ0fy30U8iy0cptqVh8XbsUCaGyZAdVQnKNBREREoqmRGQ0iIiJTBAA6QW92HzKLjKbmYqBBRESSpQfn3IiNgQYREUmUAD3My2iAgYpJnKNBREREomFGg4iIJKlsjoZ5GQnO0TCNgQYREUkW52iIj6UTIiIiEg0zGkREJFk6ZjREx0CDiIgkSYD5pROGKaaxdEJERESiYUaDiIgky9xVJ2QaAw0iIpIsc7frItNYOiEiIiLRMKNBRESSJMD8VScsvJjGQIOIiCRLx0hBdAw0iIhIsjhHQ3yco0FERESiYUaDiIgkqWyOhnmPRGPlxTQGGkREJE0CoDc3UmCkYRJLJ0RERCQaZjSIiEiyzC2dkGkMNIiISJI4R+PxYOmEiIiIRMOMBhERSZZeYOlEbAw0iIhIsjhHQ3wsnRAREZFomNEgIiJJEiCD3sx/b8uYETGJgQYREUmWuXM05BYaR03GQIOIiCTL3DkaDDRM4xwNIiIiEg0zGkREJEkCAJ1g3r+3uWGXaQw0iIhIssydDEqm8StMREREomFGg4iIJEpmgQ27uLzVFAYaREQkSZyj8XiwdEJERESiYUaDiIgkS8/Sh+gYaBARkSQJAHRmJvZZOjGNpRMiIiISDTMaREQkWeZOBiXTGGgQEZFEmf/0Vi5vNY2BBhERSVLZ8lbzAgXO0TCNOSMiIiISDTMaREQkWeauOiHTGGgQEZE0CYDe3MmgrJ2YxFCOiIjISvbs2YOIiAiEhISgffv2CAsLw1dffQW9Xm/ULiEhAa+++ioCAgLQu3dvbNq0qdL+4uLi0KNHDwQEBGDgwIE4duxYhTb5+fmYPXs2goKCEBgYiEmTJiEtLa1Cu5SUFIwdOxbt27dH586dER0djeLi4mp/RgYaREQkSQJk0MHGrJdg5qqTL774AgqFAtOnT8eqVavQq1cvzJs3Dx999JGhzcmTJxEREQF/f3989tlnGDBgAKKjoxEfH2/UV1xcHGJiYvDGG29gzZo1aNKkCcaPH49Lly4ZtXv33Xdx8OBBzJo1CzExMbh9+zZGjx5tFESo1WqMHDkSBQUFiI2NxXvvvYedO3di5syZ1f6MLJ0QEZFkmbvqxFyrVq2Cs7Oz4X2nTp1QWFiITZs24Z133oFCocDy5cvh7++P+fPnG9qkp6djyZIlGDhwIGxsbKDRaLBy5UqMGDECY8eOBQB07NgRYWFhWLVqFWJiYgAAp06dwi+//II1a9YgJCQEAODr64vevXtj27ZtGDZsGABg8+bNUKvV2L59u2F8crkckZGRCA8Ph4+PT5U/IzMaREREVnJvkFGuVatWKCkpQW5uLjQaDY4ePYp+/foZtQkLC0NmZibOnz8PAEhMTEReXh5CQ0MNbeRyOfr27YuEhAQIQtlkkoSEBKhUKgQHBxvaeXh4oEOHDkhISDAcO3ToEDp37mw0vj59+kChUBi1qwoGGkREJFl62Jj1EsOJEyfg5OSE+vXrIzU1FVqtFs2aNTNq07x5cwBAcnKy0Z/3t/Px8UFBQQEyMjIM7by9vSGTGWdymjdvbuijvN39WQuFQgEvLy+jdlXB0gkREUlS2YZd5j9ULT09HW+//fYD2xw4cKDK/Z05cwZbt27F5MmTIZfLcffuXQCASqUyalf+vvy8Wq2GQqGAg4ODUbu6desCAHJzc+Hm5ga1Wo06depUuK9KpTL0Vd7f/fesrF1VMKNBRET0BMjMzMTUqVMREBCA8ePHG527PwNR2fHK2pSXTEy1e9jx+/urSrt7MaNBRESSpbfAs0rc3d2rlbWoTF5eHsaPHw8HBwesXLkSdnZ2AP7JSNyfRVCr1QD+yWyoVCqUlJSgpKQE9vb2FdqV96NSqZCenl7h/vdnMFQqleHa+8dZnYmgADMaREQkWTLoBBuzXpZ4qFpJSQnCw8ORlZWFtWvXol69eoZzXl5esLOzw9WrV42uuXLlCgAYfumX/3n//Ink5GTUrl0brq6uhnYpKSmGTMe9/d0bQPj4+FToS6PRIDU1lYEGERFRVQiABfbRME9paSmmTZuGixcvYu3atfD09DQ6r1Ao0KlTJ+zZs8fo+K5du+Di4gJ/f38AQIcOHVCnTh3s3r3b0Ean02HPnj0ICQkxlDtCQkKgVqtx+PBhQ7v09HQkJiYalrsCQHBwMI4ePYo7d+4Yju3btw8ajcaoXVWwdEJERGQlH3zwAX7++Wf85z//QXFxMf766y/DuebNm0OpVGLy5MkYPnw4Zs6cibCwMCQmJiI+Ph4ffPABbGzK8gUKhQLh4eGIiYmBs7Mz/P39ER8fj+vXr2Px4sWGPtu1a4du3brh/fffR1RUFJRKJZYsWQJPT08MGDDA0G7o0KHYuHEjIiIiEBERgezsbCxcuBBhYWHVzmgw0CAiIsnSW3nDrl9//RUAjHYCLbd+/XrDNuErVqzA4sWLsX37dri5uWHmzJkYNGiQUfsxY8ZAEARs2LABWVlZ8PX1xZo1a+Dn52fU7pNPPsGiRYswZ84caLVaBAUFYenSpUYrVlQqFdatW4fo6Gi89dZbcHBwQGhoKCIjI6v9GRloEBGRZFn76a0HDx6sUruQkBCTJQuZTIZx48Zh3LhxD22nVCoxd+5czJ0796HtvL29ERcXV6XxPQznaBAREZFomNEgIiJJEiAz+zHx5j5UTQoYaBARkWTpGCiIjqUTIiIiEg0zGkREJFnmlk7INAYaREQkSWUbdplXOjF3wy4pYChHREREomFGg4iIJIulE/Ex0CAiImkSZP97MJp5fdDDMdAgIiJJEmD+Y+I5R8M05oyIiIhINMxoEBGRZJldOiGTGGgQEZFkWfvprVLAUI6IiIhEw4wGERFJUtmGXeY+VI1MYaBBREQSJbNA6YSlF1NYOiEiIiLRMKNBRESSpee/t0XHQIOIiCRJAKAzs3TCORqmMZQjIiIi0TCjQUREksV9NMTHQIOIiCSLT28VHwMNIiKSJAEy6Mx+qBozIqYwlCMiIiLRMKMhAX/9qsSB7+rh/PHayLxpB2VdHVq0LcLwf99Ci7ZFhnbb1zbAz9vr4WaKPYoKbODUoBT+zxbgX+9koKlfsaFdcaENFv+7MZLP1kJ2hh30ehvUdVuLXq/ZY0i4DRwc9cb3/02JzbGuuHreASVFNnBvosFLw7IRNjoLcnnlYy4pkiG8tx/Srjpg3Kw0DArPFOVrQzVX+xY3EfvvHyo9N+nDV3A+xfV/7wQM7H4Orwafh3v9POQV2uPwqaZY8/1zyC+0N1zj7JiAr/676oH3W73tOWz6qb3hfaDvTQx/6S80b5QNe7tSpGepsOt3P2z7xZ/p+icI52iI74kINFJSUhAdHY0TJ06gVq1a6NevHyIjI+Hg4GDtodUIu9Y3gPqOHK+Oy0QT32LczbbFd6sbYlqoL+Z/lYz2L+QDANR3bPFcdzWaTS6G0qkU6dfssWVZQ0zr1wLL9l5G4+YlAIBSrQyCALw2MRNujTXQ6F1x6Ocu2BKbifPHvLFoS7Lh3omHlHj/Xz5o0ykfb390HQ6Oehz9qS5Wzm6E9Gv2CJ+bVumY133ojuJC/jAm863e/hxOXnY3OpZy09nw35MHHsPrPc7im30BOH7RE03dczEm7DhaNslE+If9odOXfR+qiwMxe+1Q3M4pNOprbOgJPOefhkOnmhqOPdMyDR+/tQenktzw0cauKNLYoUvba5g2+Ag8G6gRG/+8eB+Yqk7gPhqPg9UDDbVajZEjR8LDwwOxsbHIycnBggULkJubi48//tjaw6sRpsy/AacGpUbHnu2eh9HPt8LXsa6GQGPEf24ZtWnbuQCtninA+JBWOLi1HkZOLzuvrKvD+6uvGdoVaurDwS8YdW0PYtvqOki/poB7Ew0AYN8WZ8jtBMxdn2LIdHQIzseNZHv8tMW50kDj4klH7PiiAd5bdg3RE7wt94UgSbpxW3VP9sJYg7oFGNj9LLYl+GPV9iAAwPGLjXAnzwH/HfszXu58Gbt+awkAKNWrcCXNHTdu5Ruud1Bo0brZbZy64orrGU6G4y93uoxSnQ2iVvRBscYOAHDioie8XHPxUufLDDRIUqweaGzevBlqtRrbt2+Hs3PZvzLkcjkiIyMRHh4OHx8fK4/w6Xd/kAEAtWrr4eVbjKybdg+9tq5z2bVyW9Pb0qjql7WRy/9pK7cVYGcnQOFgXE6prdJBYW98DAC0GhkW/7sxwkZloUW7ogrniSypdbPbsJULOHq2sdHx3894AQBCAlMMgUZlejxzFY4OWvxwX5tSnQ1KS21QojX+EZtfaA+N1uo/dukeek7mFJ3Vc0aHDh1C586dDUEGAPTp0wcKhQIJCQlWHFnNVqC2wZUzjmhyz9yLcjodoCmRITXJHjGRXnBqoMWLQ3IqtBMEQFcKFOYBl36/ih2f2aPbq3fQsJHW0CZ0RDa0WhlWzPRE9i1b5N+VY/+39fD73roYFHG7Qp+bYlxRXGhjyJ4Qmeudob/j4LK12LP4S3z81m4E+PzzvWUr1wEAtKXGk4V0Ohvo9YCPZ8Xv+3v163IJ+UV2+PlEM6Pj3x9uBVtbPaYN/h316xZAWasEL3ZMQtf2f+Prn9pa6JORucp3BjXnxZ1BTbN6aJ2cnIyBAwcaHVMoFPDy8kJycvIDriJzLfu/RigutMGwaRkVzvVv3hbakrIYtFGzYnz03RU09NRWaJfwvRMWRDT937st6DFIi8hPrhm1admhEIu2XMG8iU2x80sXAICNXMCYGTfx+iTjCZ7JZ2shfkVDzFn3vzJLtvmfk6SroEiB+INtcPKyO9QF9vB0UWNY79NY8s4uvLe8D/680Bh/p9cDAAT43MLJyx6Ga9v4ZMDGBlDVrhiIl/NyzUWATwa+P9SyQubiwt8N8c6n/TBn/H681u08AKBUJ8Oa75/DNwcYaJC0WD3QUKvVUKlUFY6rVCrcvXv3kfsVIEehpqkZI6u5vvrEAQe3OmD8nEJ4tnRFocb4/ILvClCqAW6l2mBnnD3+M9APc77Kh5evcanDv4sMH+3Iw927DfDnkSAcWn8I/73TGlFrCmDzv1xZ8hk55o6tDd/2Okyclw+HWsCZI7b48kMPFBQ6Y/DUsgmmulLg43eU6BJaCv/n66NQUx/FmrJOtDpnFGpqi/51edo0clNaewhPtKJSJbYdbmp4n5MPJK9rjUXhGzB18HHMWN0KJTolLvztiX+9eAYFJW44k+wFT5ccTHr1d+j0MgAyNHJTomF9RwAw/AkAQ3snAgD+uBhY4e/C2z0D0/+1H1fS3PDl7gCUaO3Q2vs6xr9yHC715Nh2qJPon/9pZiuXoVT3eHIFZq8AYuXFJKsHGg8iCAJkskf/G9Tq6uNS5kILjqhm2P/Zr9i/5lf0iQiGT9/ncamyVaP/mzfn1hgY0b4EH722GquifTHyk9crtnUHlO5A95aAs2c9fP3+99j27Wi06e4HAFg+Yz1qOWkxIHo0bORl/0M/0wLIKTqMzZ/+jkZd30H9Rk44tOEYbl77Ha/NHYW/rtYCAOTeUgP4HDdzXsRfV5+BvaPC0AcB/xln7RE8nfTyJDRx24/p49pDgAKlNn7QYhWmDSpbCqsXbJGZ3xclpWcgtynEf8Z1NFw7ckCb//1XKdq4xaFQ0wSv93ulwj18XWbBRlYfNsq56BXyz/dsVmE8Xu+2Df6thkCjq3yCKpXJuvM45mjJLLC8lZGGKVYPNFQqFdRqdYXjeXl5Zk0EtZNno5kzV63c65tP7bF/TS0MfbsIQ97eAWCH6YtcAK/mStxNy4OfS1SlTYq1HriWOxWdOv+IrwHIsr+Cn0tZpuLW5bro+ooGrdz+z+gadWdbHFirhOLOx/ALLMWPaY4ozlfg49dWV+h/36rD2LfqMBb/kAfv1rrqfuwaa9yC16w9hKfSmH4Z6PUsEPPFcWh15T8Cu0PlGIS6ygJk3VVBo7XF6ul7ceRMC6zZ8Qca1nfEyAFtsG7bWdzOLsSzLa8gcIgam358Bj/9+UeFe3z5fgqOnPHD6h3HjY4HttDhP/8S8MPBn/FXUrMK11GZ8YMfT3lJgPmTQTlHwzSrBxo+Pj4V5mJoNBqkpqZWmLtRHTLo4Kj428zR1RybYlyx+VMn/OvtW9WaaHk3W47Uy63Q+tl8k1/Py8dzADiiSfNbcFSUlb3qu7XC1bN62Mv/Ntqc6+opNwBKeDa+AUdFMf411R4vDzNeAXPnti0WRDRFvxFZCHklF818C1FLUXGlilTdu8ySqkbpWIK2Pldx+Xp9pKRVNv+iFgAtBnY/CXs7LTbu8TX6Ot/OLsSNW/mY8tpfKNHIseVAY+QXVvx7yMp1ROOG6biZoTZKzb/UsWwO04Wrcv79PcTjKpvQ42H1QCM4OBgrV67EnTt3UK9e2cSsffv2QaPRICQkxMqjqxm+XeWC9R+549nuanTsqcaFE45G51s9U4gCtQ2ihvqg+4BceHqXwN5BjxtX7bF9rQu0JTIMf/ef4OSHDfVx9lhtdAjJg4uHFnfVtvgt4Wcc2VIL/s/m4/k+/8yteW18JlbMaoT/jmyGvsOzYF9Lj79+rYPvVjdEYNc8+LQu+2Hv1aIEXi1KjMZ167oCAODepATtnucPZaqeWWMO4naOEhevNcDdfAc0aqjGkF6n4awqxIL1//xsCe1yEQBwM6sOlLU0CGp9Hf2ev4TPvn8Ol683qNBv/boF6Oh/Az+faGa0c+i9thxog7eHHMHCiJ+w43BLFGts8UzLmxjS6zT+vOCJ5LT64nxoqjbuDCo+qwcaQ4cOxcaNGxEREYGIiAhkZ2dj4cKFCAsL4x4aFnJsX10AwPGfVTj+c8WJtz/e/At29gKa+Rdjz8b6yLxpB02JDeq5aNHu+XzMWpuBJr7/BAFNWxbh2D4VvljgDnWOLWzkMjg3SsHrk4sxJDwZ8nu+q/qPzUJ9dy22rnHBp5FeKCmWwbWxBsP/fQuvTeC24iSeqzec0f3Zq3il6wXUstcir9AeZ664Yd6X3XHxmouhnUwmYFCPs3B1zocgyJB0vT5mru6NX083rbTflztfhq1ceOj+Glt/aYOs3NoY1PMMpg8/DIVdKW7l1MGXP3TAlgMBlv6oZAZuBy8+mSAIVs9R3bsFuYODA0JDQ83agvzMmTOA7jpau02z8EipMoWapriUuRB+LlEsVz1GweETrD0EyWjkpsR/xnXER2v/YMnjMZg1uTMAoE+vF0S7x5kzZ3C7JAtrtN+Y1c8EuyFoaN8AAQEMIB/E6hkNAPD29kZcXJy1h0FERBLD0on4nohAg4iI6HHjqpPHg8UpIiIiEg0zGkREJE2CBTbsYunFJAYaREQkWZyjIT6WToiIiEg0zGgQEZFkMaMhPgYaREQkSQLMDzS46sQ0BhpERCRZ5i5vJdM4R4OIiIhEw4wGERFJFudoiI+BBhERSRLnaDweLJ0QERGRaJjRICIiyWLpRHwMNIiISKIssAU5V62YxNIJERERiYYZDSIikiYBEMx+qJplhlKTMdAgIiLJ4oZd4mPphIiIiETDjAYREUkS99F4PBhoEBGRZJk9R4NMYqBBRESSxX00xMc5GkRERCQaZjSIiEiyWDoRHwMNIiKSKO4M+jiwdEJERESiYUaDiIgkSQAgmLk+lctbTWOgQUREksWdQcXH0gkRERGJhhkNIiKSJj5U7bFgRoOIiCRLL8jMepnr2rVrmD17Nvr37w9/f3+EhoZWaBMVFQU/P78Kr0OHDlVoGxcXhx49eiAgIAADBw7EsWPHKrTJz8/H7NmzERQUhMDAQEyaNAlpaWkV2qWkpGDs2LFo3749OnfujOjoaBQXF1f7MzKjQUREZCVJSUlISEhAu3btoNfrITxgdmrjxo3x8ccfGx3z8fExeh8XF4eYmBi888478Pf3R3x8PMaPH4/4+Hj4+fkZ2r377rs4d+4cZs2aBaVSidjYWIwePRo7duyAg4MDAECtVmPkyJHw8PBAbGwscnJysGDBAuTm5lYYhykMNIiISLLMXXVirh49eqBXr14AyjIXZ8+erbSdg4MD2rdv/8B+NBoNVq5ciREjRmDs2LEAgI4dOyIsLAyrVq1CTEwMAODUqVP45ZdfsGbNGoSEhAAAfH190bt3b2zbtg3Dhg0DAGzevBlqtRrbt2+Hs7MzAEAulyMyMhLh4eEVgpyHYemEiIgkqWx5q8y8l5ljsLGxzK/hxMRE5OXlGZVe5HI5+vbti4SEBEOmJCEhASqVCsHBwYZ2Hh4e6NChAxISEgzHDh06hM6dOxuCDADo06cPFAqFUbuqYKBBRESSZW6g8bikpqbi2WefRZs2bfDaa69h//79RueTk5MBAM2aNTM67uPjg4KCAmRkZBjaeXt7QyYzHnvz5s0NfZS3uz9roVAo4OXlZdSuKlg6ISIiMkN6ejrefvvtB54/cOCAWf23atUKAQEBaN68OfLy8vD1119j8uTJWLJkCV566SUAZXMqFAqFYY5Fubp16wIAcnNz4ebmBrVajTp16lS4h0qlwt27dw3v1Wo1VCqVyXZVwUCDiIgk62l4TPzIkSON3vfo0QNDhw5FbGysIdAAUCFLAcBQMrn3XGXtHnb8/v6q0u5eDDSIiEiyLDEZ1N3d3eysRXXY2NjgxRdfxEcffYTi4mI4ODhApVKhpKQEJSUlsLe3N7RVq9UA/slsqFQqpKenV+jz/gyGSqUyXHuvvLy8ak0EBThHg4iI6Klz/zLY8l/+98+fSE5ORu3ateHq6mpol5KSUuH6K1euGAUQPj4+FfrSaDRITU1loEFERFRVT8tk0Hvp9Xr8+OOPaNGihWFORocOHVCnTh3s3r3b0E6n02HPnj0ICQkxlDtCQkKgVqtx+PBhQ7v09HQkJiYalrsCQHBwMI4ePYo7d+4Yju3btw8ajcaoXVWwdEJERNJkiWDBzOuLiooMy0XT0tKQn5+PvXv3AijbB6OoqAhRUVEIDQ2Fl5cX7t69i6+//hpnz57F0qVLDf0oFAqEh4cjJiYGzs7Ohg27rl+/jsWLFxvatWvXDt26dcP777+PqKgoKJVKLFmyBJ6enhgwYICh3dChQ7Fx40ZEREQgIiIC2dnZWLhwIcLCwqqd0WCgQUREZCXZ2dmYNm2a0bHy9+vXr4efnx+USiWWL1+OnJwc2NnZoU2bNvjss8/QtWtXo+vGjBkDQRCwYcMGZGVlwdfXF2vWrDHaFRQAPvnkEyxatAhz5syBVqtFUFAQli5darRiRaVSYd26dYiOjsZbb70FBwcHhIaGIjIystqfkYEGERFJlrWfidaoUSNcunTpoW1WrlxZpb5kMhnGjRuHcePGPbSdUqnE3LlzMXfu3Ie28/b2RlxcXJXu/TAMNIiISJLKdwY1tw96OE4GJSIiItEwo0FERNLFlIToGGgQEZFkWWuJqpQw0CAiIsmy9mPipaBKgcaMGTOq3KFMJsP8+fMfeUBERERUc1Qp0Dh27FiVO6zuw1aIiIishaUT8VUp0Dh48KDY4yAiInr8GGiIjstbiYiISDSPPBn08OHD+OOPP3Dnzh1ERETAw8MDp0+fRqNGjeDs7GzJMRIREVmeYIHJoJxMalK1A42ioiJERETgyJEjhvkYw4YNg4eHBz7//HO4u7vjvffes/hAiYiILI6BguiqXTqJiYkxPDXu+PHjRs+079KlC37//XeLDpCIiIieXtXOaOzduxfTpk1D7969odPpjM55eHggPT3dYoMjIiISE1ediK/agUZOTg6aN29e6TkbGxsUFxebPSgiIqLHgqUT0VW7dOLq6orLly9Xeu7SpUto1KiR2YMiIiKimqHagcaLL76IVatW4fz584ZjMpkMaWlp+PLLL/HSSy9ZdIBERERiEQSZWS8yrdqlk8mTJ+PIkSMYNGgQWrRoAZlMhhkzZiA1NRXe3t6YMGGCGOMkIiKyPJZORFftjIZSqcTmzZsxbdo0ODo6wsvLC7Vq1cLEiROxadMmODg4iDFOIiIiEcjMfJEpj7Rhl4ODAyZMmMDsBRERET3UI+8MWlJSgnPnziE3NxdOTk5o3bo17O3tLTk2IiIicbF0IrpHCjS++OILrFixAvn5+RAEATKZDLVr10ZERATGjBlj6TESERGJg4GG6KodaGzYsAGLFi1Cly5dEBoaigYNGiArKws7d+7ERx99BFtbW4wYMUKMsRIREdFTptqBxrp16/DKK6/gww8/NDo+YMAAREZGYv369Qw0iIjoySfIzH9MPJe4mlTtVSe3b99GWFhYpef69++P27dvmz0oIiIisQkoe3qrWS9rf4inQLUDjaZNmyI7O7vSc5mZmWjSpInZgyIiIqKaodqBxtSpUxEbG1thG/KLFy9i2bJlmDp1qsUGR0REJCrBzBeZVKU5GpMmTTJ6r9Pp8Oqrr6J58+ZwcXFBZmYmrly5goYNG2Lr1q3o3bu3KIMlIiKyKM6xEF2VAo37sxdyuRxubm7Iz89Hfn4+AMDNza3StkRERCRdVQo0Dh48KPY4iIiIHisZAJmZ5Q/mQ0x75J1BiYiInnqcZyE6swKNnJwcFBcXVzju4eFhTrdERESPB+doiO6RAo0VK1Zgw4YNyM3NrfT8hQsXzBkTERER1RDVXt767bff4rPPPsObb74JQRAwceJETJgwAW5ubmjSpAmio6PFGCcREZFlmbu0lUtcq6TagcZXX32FiRMnYuLEiQCA3r1745133sGePXtQu3Zt3Llzx+KDJCIiEgWDDNFVO9C4du0a2rVrBxubsku1Wi0AwMHBAWPGjMGWLVssO0IiIiJ6alV7joatbdklMpkMSqUSt27dMpyrV68eMjIyLDc6IiIiMTErIbpqZzSaNGliCC4CAgIQHx8PrVYLnU6Hb775Bp6enhYfJBERkSjKn+D6qC8yqdqBRnBwMP78808AwIQJE3D06FE899xz6NixI3766SeMHz/e4oMkIiKip1O1SydTpkwx/Hfnzp3x9ddfY/fu3ZDJZAgJCUGnTp0sOkAiIiKxmLszKJlm9s6gbdu2Rdu2bS0xFiIioseLgYboql06ISIiIqqqKmU0RowYUeUOZTIZ1q1b98gDIiIiopqjSoGGIFQ9t1SdtkRERNbEORriq1KgsWHDBrHHYXHpqfb4d4f21h6GJHj4uWLqRmByHz/cvORk7eFIRm3Hs9YegmTUauECjOuIWkeTUTsp09rDqfFsxjzz+G7GJaqi4xwNIiIiEo3Zq06IiIieSpZ4XglLLyYx0CAiIulioCA6lk6IiIhINMxoEBGRZHHVifgYaBARkXQx0BDdIwcaycnJ+PPPP3Hnzh28/vrrcHFxQUZGBurWrQsHBwdLjpGIiIieUtUONHQ6HWbNmoVt27ZBEATIZDIEBwfDxcUF//3vf9GqVStMmzZNjLESERFZFjMaoqv2ZNCVK1di165dmD59Onbt2mW0E2jXrl1x+PBhiw6QiIhILDLBvBeZVu2MxrZt2xAREYHRo0dDp9MZnWvUqBFu3LhhscERERHR063agUZGRgbat29f6Tl7e3sUFBSYOyYiIqLHg1uQi67apZP69evj+vXrlZ5LSUmBm5ub2YMiIiISnWChFz1UtQONkJAQrFq1ChkZGYZjMpkMeXl52LBhA7p3727RARIREYmFczTEV+3SydSpU3Ho0CH07dsXQUFBkMlkWLx4MZKSkmBra4uIiAgxxklERERPoWpnNBo0aIBvv/0W/fr1w7lz5yCXy3Hx4kUEBwdj8+bNcHJyEmGYREREImDZRHSPtGFXgwYN8MEHH1h6LERERI8Vyx/i40PViIiISDTVzmjMmDHjoedlMhnmz5//yAMiIiJ6bJjREF21A41jx45VOJabm4vCwkKoVCrUqVPHIgMjIiISHQMN0VU70Dh48GClx48cOYI5c+ZgyZIlZg+KiIiIagaLzdHo3Lkzhg8fjnnz5lmqSyIiItHIYP4+GtxX1DSLTgb18fHBmTNnLNklERERPcUsGmj8+eefqFevniW7JCIioqdYtedoLFu2rMIxrVaLS5cu4dChQxg7dqxFBkZERCQ6K08GvXbtGuLi4nDq1CkkJSWhWbNm2LVrV4V2CQkJiImJQXJyMtzc3DBq1Ci88cYbFdrFxcVh06ZNyMzMhK+vL6ZPn46goCCjNvn5+fjwww/x448/QqPRICgoCLNmzYKnp6dRu5SUFERHR+PEiROoVasW+vXrh8jISDg4OFTrM1ok0FAoFPD09MTUqVMZaBAR0dPBEs8rMfP6pKQkJCQkoF27dtDr9RCEih2ePHkSERER6N+/P6KiopCYmIjo6GgoFAoMGjTI0C4uLg4xMTF455134O/vj/j4eIwfPx7x8fHw8/MztHv33Xdx7tw5zJo1C0qlErGxsRg9ejR27NhhCCLUajVGjhwJDw8PxMbGIicnBwsWLEBubi4+/vjjan3GagcaFy9erO4lRERETyYrZzR69OiBXr16AQCioqJw9uzZCm2WL18Of39/wx5VnTp1Qnp6OpYsWYKBAwfCxsYGGo0GK1euxIgRIwz/4O/YsSPCwsKwatUqxMTEAABOnTqFX375BWvWrEFISAgAwNfXF71798a2bdswbNgwAMDmzZuhVquxfft2ODs7AwDkcjkiIyMRHh4OHx+fKn/Gas3RKC4uxrvvvovjx49X5zIiIiKqhI3Nw38NazQaHD16FP369TM6HhYWhszMTJw/fx4AkJiYiLy8PISGhhrayOVy9O3bFwkJCYZMSUJCAlQqFYKDgw3tPDw80KFDByQkJBiOHTp0CJ07dzYEGQDQp08fKBQKo3ZV+ozVaezg4IADBw5UmtohIiJ66jzhD1VLTU2FVqtFs2bNjI43b94cAJCcnGz05/3tfHx8UFBQgIyMDEM7b29vyGTGC3ObN29u6KO83f1ZC4VCAS8vL6N2VVHt0knLli1x+fJlPPfcc9W9lIiI6IliiYeqpaen4+23337g+QMHDjxy33fv3gUAqFQqo+Pl78vPq9VqKBSKChM169atC6BsB283Nzeo1epKd/BWqVSGvsr7u/+elbWrimovb42MjERcXBz++OOP6l5KREREj+D+DERlxytrU16BMNXuYcfv768q7e5VpYzGn3/+CX9/f9SuXRtz5sxBQUEBRo4cCZVKhYYNG1YY6I4dO6o1CCIiIquwQEbD3d3drKzFw5RnJO7PIqjVagD/ZDZUKhVKSkpQUlICe3v7Cu3K+1GpVEhPT69wn/szGCqVynDtvfLy8qo1ERSoYkZjxIgRhpqMk5MTfH198eyzz8LX1xdOTk5Gr/IPQ0RE9KQzdwtysXl5ecHOzg5Xr141On7lyhUAMPzSL//z/vkTycnJqF27NlxdXQ3tUlJSKsy1vHLlilEA4ePjU6EvjUaD1NTUagcaVcpo3DugDRs2VOsGRERE9GgUCgU6deqEPXv2YNSoUYbju3btgouLC/z9/QEAHTp0QJ06dbB7927DMZ1Ohz179iAkJMRQ7ggJCcHy5ctx+PBhw8qT9PR0JCYmYubMmYb+g4ODsXLlSty5c8ew4/e+ffug0WgMy2KrqtqTQYmIiGoMKy+iLCoqMiwXTUtLQ35+Pvbu3QugbB8MZ2dnTJ48GcOHD8fMmTMRFhaGxMRExMfH44MPPjAsj1UoFAgPD0dMTAycnZ0NG3Zdv34dixcvNtyvXbt26NatG95//31ERUVBqVRiyZIl8PT0xIABAwzthg4dio0bNyIiIgIRERHIzs7GwoULERYWJk5Gg4iIqEaycqCRnZ2NadOmGR0rf79+/XoEBQUhMDAQK1aswOLFi7F9+3a4ublh5syZRruCAsCYMWMgCAI2bNiArKws+Pr6Ys2aNUa7ggLAJ598gkWLFmHOnDnQarUICgrC0qVLjVasqFQqrFu3DtHR0Xjrrbfg4OCA0NBQREZGVvszVjnQGDlyZJVmmspkMpw4caLaAyEiIpKaRo0a4dKlSybbhYSEmCxZyGQyjBs3DuPGjXtoO6VSiblz52Lu3LkPbeft7Y24uDiTYzOlyoFGeQqHiIioRngCnnUiBVUONCZPnoy2bduKORYiIqLHi4GC6DhHg4iIpIuBhuiqvTMoERERUVUxo0FERJL1ODbdkroqBRoXL14UexxERESPHwMN0bF0QkRERKJh6YSIiCSLpRPxMdAgIiLpYqAhOpZOiIiISDTMaBARkTQJMD+jwYyISQw0iIhIskw/wYvMxdIJERERiYYZDSIiki6WPkTHQIOIiCRJBvOXt7L0YhoDDSIiki5mNETHORpEREQkGmY0iIhIupjREB0DDSIikixuQS4+lk6IiIhINMxoEBGRdDGjIToGGkREJFksnYiPpRMiIiISDTMaREQkTXyo2mPBQIOIiCSLpRPxsXRCREREomFGg4iIpIsZDdEx0CAiIulioCE6BhpERCRZnKMhPs7RICIiItEwo0FERNLFjIboGGgQEZFkyQRGGmJj6YSIiIhEw4wGERFJE3cGfSwYaBARkWRx1Yn4WDohIiIi0TCjQURE0sWMhugYaBARkSTJYH7pRGaRkdRsLJ0QERGRaJjRICIi6WLpRHQMNIiISLK46kR8DDSIiEi6GGiIjnM0iIiISDTMaBARkWSxdCI+BhpERCRNAgBzH6rGQMUklk6IiIhINMxoEBGRZLF0Ij4GGkREJF0MNETH0gkRERGJhhkNMmjWugij3kuHd6ti1HUuhabYBteT7bHzywY4uLWeoV3rjvnoPfgOmrcpQhO/YijsBZy7FfrAfl8Zk4lXRmXDtbEGORl2+GlLPWyOdYWulE8JIPHUqq3DsMk34ONfAB//AtR1LsXG2EbYFNu4Qluf1vkYOz0VLdvnQaeT4dSRuli7sAluXXcwatfr1WR4NU2Hj38B3BqX4PQxFd57o3WF/hZtOoe2QeoHju1fnZ7BnSyF+R+SzCbTW3sENZ/VA41r164hLi4Op06dQlJSEpo1a4Zdu3ZZe1iSpFTpkHVTgV+210P2LTs4OOrR/bU7eG9ZKlwba/D1ElcAQPsX8hHYNQ/JZ2uhMM8G7boUPLDPYVMzMGL6LWxZ1hAnEurAt30hRk6/hQZuWiyZXvEHPpGl1HHS4uWhGUi56Ijf9znj5SG3K23XqFkRPtx0HsnnHbFgqi/s7PV48+3r+Ojrs5jySjvczbEztO3W9xry7wo4dVQFB8c7D7z38v96w1GpMzpm76DH3M8v4Mq52gwyniQsnYjO6oFGUlISEhIS0K5dO+j1egjmLjWiR3b6iBKnjyiNjh3br4JbYw36Ds82BBpfxbhi02I3AMDrk24/MNCoU68Uw6ZlYM8mZ3yx0N1wD1tbASPfu4Vtn7kgNcmh0muJzHU7zR6DOjwHQAZVPe0DA403374OrUaG/zehJQrzy34kXjmrxNr9JzFw7E18/lETQ9tZE7sj7XIWAGDl7r8eeO/UK44VjvUacBt2CgF7tzR89A9F9BSy+hyNHj16ICEhAbGxsWjdumIKkqxPnSM3KnMIQtVKHs92y4N9LQE/feNsdPynb5xhYwM8/9Jdi46TyJgMph7ibSMX0LH7Hfy6t74hyACA2zftcfqYCp1fzDFqX9Xv/cq8OOg2CvNtcOiHBo/cB1meTDDvRaZZPdCwsbH6EOg+MpkAG7mAus6lCB2ZhWe65WHL8ur/K6xpy2IAQMoF46xFzm075GbLDeeJrMXdqxgOtfRIuVQxA5Fy0REeTYphpzC/iO/RpAgBHfOQ8EMDFBfKze6PLKR8wy6zXtb+EE8+q5dO6Mnz1oI09BuRDQDQlMiwcpYndm+sX+1+VPVKoSmWoaSo4g/WvFxb1KlXavZYicyhcir7HszPrfijMO+uHWxsAGVd879P+wwuK9v8FM+yyZOGWQnx1dhAQy63gYefq7WH8VQ6+IMKx3/TQFVPg3adMzF53g00bFIbe7c0rdBW5VIEIB0A4NLUOBhxdMoCZLmV/j3Y2iXB3tGef0dmsHHg/JaqUqpKAAB1nB3h0cLFcLxB47Ig2MlNZXQcAFT1cwEArt71YV+nbNWVi9c/q6/sFLZQ1LKrcN39bGz0eHFQItL+rgN1QTN4tDD749R4cjs5dFqd6Yb0VKixgUZd1zqYunG0tYdRI2QXxWHwpJ/h/+q7KNWrjM41VO4CcBkAMCz6FaNz7qqvobDfgWmb/gVBsDc618DtCOwb+PHviB4LuY0awI/o+EoAvLq/bjhub3sTwK/oF94WQSNeNLrGQ7UJgnARIz96AwLKVokMmf2S4bxzw0So3OtgytphD723yuE4nJx3Id9mMKas7Wuxz1TT5dx8THO4mNEQXY0NNO5m5GF95HfWHkaN8MJL6Rj7ng7b58bh6gUno3MvDf4bQ8LL/vvrmTuQ+Xe24VynnumYOBPYGb3C6DpVvRIs2ZqHg1/dxg+bvngMn6BmYkaj6pSqEiz5Bvhjxxns2KQ1HLex0WPZd3Jc+e0gNi7LNrrm7bnHILg5Yun47+DiVQ9DZr+Ebz7Yi8zUsmWtH6y8izx1EZa99/VD7z1l9h/QPmuDDydkoiDv4W2pzJsLwh7bvVg6EV+NDTR0Oj1uXsqw9jBqhCYRN6HTAWcOFeBuTonROXVmnuG/M//ONvqa78vQY9S7MgQGJePX7Y0Mx1+YkgG9Hvhxgxw3k/h39KhsHCtOYKTKqeqVBRd5OYW4mZRpdO7oASe0D7qJpf/niqKCslKKi3sJWrbNwrYv3I3aZ6beMbzXakqhKZJV6O9e9RpoEPBsBn77yRlJiVxlVVUsm9QsNTbQoOqb9uF1FObLcemkI+5k2qKucym6ht1Ft/652LLCBXdzyr5d6jqXIqBzPoB/VpaoHP7Cs8G3kFKvAGeOlu3FkZdri6+XuGLE9FvIy5XjREId+LUvxJv/zsDer5y5hwaJ7tngO3Bw1KNW7bJfXF7Ni/DCS2WZiz9/cUJJsRwblzTGkq2nMeezC9iyyhMKez2Gv30dd3NssTXOw6i/Ji1y0cyn7HpHpQ4yGQz9XT6txO2bxiXCXq9lwtZOwI/cO+PJxb2bRGf1QKOoqAgJCQkAgLS0NOTn52Pv3r0AgI4dO8LZ2flhl5MFXThRGy8OyUGvQTlQqnQoKpDj6nkHLJriZbQFeRO/Ysz67JrRtY2dPsfkOcCp32tj+uvNDce/jnVFYYENwkZlY+CkTNzJtMU3yxsaNv8iEtOUD1Lg2uifLFxw32wE9y0LDEaGBOJ2mhw3rtbCe2+0xpjpqXh/2eX/bUGuwtyFTY12BQWAnmEp6NL7utGx95eVzVH6ZLoP9m81DihefP02bl23x8nf6orx8chMMphfOuGDFEyzeqCRnZ2NadOmGR0rf79+/XoEBQVZY1iS9NM3zhU216rM6SNK9PFoZ3jv4eeKqRtHI3b4F5WWq76Pc8H3cQ+fmU8khlHdOlSp3ZVzSvzfSH+T7T5fHIjo8EYm25Ub/2JgldsS1VRWDzQaNWqES5cuWXsYREQkRayciM7qgQYREZG1cNWJ+Lj/NxEREYmGGQ0iIpImAYDezJQGMyImMdAgIiLpYqAgOpZOiIhIsqz9mPitW7fCz8+vwuvjjz82apeQkIBXX30VAQEB6N27NzZt2lRpf3FxcejRowcCAgIwcOBAHDt2rEKb/Px8zJ49G0FBQQgMDMSkSZOQlpZm/od5AGY0iIiIrGzt2rWoU6eO4b2r6z97DZ08eRIRERHo378/oqKikJiYiOjoaCgUCgwaNMjQLi4uDjExMXjnnXfg7++P+Ph4jB8/HvHx8fDz8zO0e/fdd3Hu3DnMmjULSqUSsbGxGD16NHbs2AEHER5twECDiIgkSrDAzqCWqb20bt36gRtULl++HP7+/pg/fz4AoFOnTkhPT8eSJUswcOBA2NjYQKPRYOXKlRgxYgTGjh0LoGzTy7CwMKxatQoxMTEAgFOnTuGXX37BmjVrEBISAgDw9fVF7969sW3bNgwb9vCHBD4Klk6IiEiyrF06MUWj0eDo0aPo16+f0fGwsDBkZmbi/PnzAIDExETk5eUhNDTU0EYul6Nv375ISEiA8L+AKiEhASqVCsHBwYZ2Hh4e6NChg2GXbktjoEFERGRloaGhaNWqFXr27InVq1dDpyt7Pk9qaiq0Wi2aNWtm1L5587JHPSQnJxv9eX87Hx8fFBQUICMjw9DO29sbMpnx5unNmzc39GFpLJ0QEZF0WSArkZ6ejrfffvuB5w8cOPDAcy4uLnjrrbfQrl07yGQyHDx4EJ9++ikyMjIwe/Zs3L1b9tRflUpldF35+/LzarUaCoWiwhyLunXLnrOTm5sLNzc3qNVqo7kg9/ZX3pelMdAgIiJpEgCZuXM0zLy8a9eu6Nq1q+H9Cy+8AHt7e6xbtw6TJk0yHL8/A1HZ8cralJdMTLV72HFzMdAgIiIyg7u7+0OzFtX18ssv4/PPP8eFCxfg6ekJABWyDWq1GsA/mQ2VSoWSkhKUlJTA3t6+QrvyzIZKpUJ6enqFe6rV6gpZE0vhHA0iIpIuvZkvkXl5ecHOzg5Xr141On7lyhUAZXMw7v3z/nkWycnJqF27tmG5rI+PD1JSUgyZjnv7K+/D0hhoEBGRZMkEwayXGHbv3g25XA5/f38oFAp06tQJe/bsMWqza9cuuLi4wN/fHwDQoUMH1KlTB7t37za00el02LNnD0JCQgxlkZCQEKjVahw+fNjQLj09HYmJiYblrpbG0gkREZGVjB07Fp06dYKvry+AsomjW7ZswYgRI+Di4gIAmDx5MoYPH46ZM2ciLCwMiYmJiI+PxwcffAAbm7J8gUKhQHh4OGJiYuDs7GzYsOv69etYvHix4X7t2rVDt27d8P777yMqKgpKpRJLliyBp6cnBgwYIMpnZKBBRETSZeVnnXh7e+Pbb7/FrVu3oNfr0bRpU/zf//0f3nzzTUObwMBArFixAosXL8b27dvh5uaGmTNnGu0KCgBjxoyBIAjYsGEDsrKy4OvrizVr1hjtCgoAn3zyCRYtWoQ5c+ZAq9UiKCgIS5cuFWVXUICBBhERSZlI5Y+qmjlzZpXahYSEmCxtyGQyjBs3DuPGjXtoO6VSiblz52Lu3LlVHqc5GGgQEZFkPY7dPaWOk0GJiIhINMxoEBGRdFm5dCIFDDSIiEiaBEBm7l4YjFNMYumEiIiIRMOMBhERSRdLJ6JjoEFERNLFOEN0LJ0QERGRaJjRICIiSZLB/MfEi/Ng9ZqFgQYREUkX52iIjqUTIiIiEg0zGkREJF3m7qNBJjHQICIiaRIEs+dosPRiGgMNIiKSLgYKouMcDSIiIhINMxpERCRdzGiIjoEGERFJFyeDio6lEyIiIhINMxpERCRZZq86IZMYaBARkXQx0BAdSydEREQkGmY0iIhImgSYn9FgQsQkBhpERCRdLJ2IjoEGERFJF5e3io5zNIiIiEg0zGgQEZFkcXmr+BhoEBGRRAkWmKPBQMUUlk6IiIhINMxoEBGRdOmZkRAbAw0iIpIm7qPxWLB0QkRERKJhRoOIiKSLq05Ex0CDiIiki4GG6Fg6ISIiItEwo0FERNLFVSeiY6BBREQSJQCCuQ87YaBiCgMNIiKSLs7REB3naBAREZFomNEgIiJpEmD+HA0mRExioEFERNLF0onoWDohIiIi0TCjQURE0sWMhugYaBARkXQx0BAdSydEREQkGmY0iIhIuvTmbthFpjDQICIiiRIsUDph6cUUlk6IiIhINMxoEBGRNAkwP6PBhIZJDDSIiEi6+PRW0THQICIiyRLMfnormcI5GkRERCQaZjSIiEi6WDoRHQMNIiKSLu4MKjqWToiIiEg0zGgQEZE0CYL5O4MyI2ISAw0iIpIuBgqiY+mEiIiIRMOMBhERSZbAh6qJjoEGERFJF0snomPphIiIiETDjAYREUkXN+wSHQMNIiKSJkEAzH3WCUsvJjHQICIiyRKY0RAd52gQERGRaJjRICIi6eJj4kXHQIOIiCRJgPmlExZeTGPphIiIiEQjE4SaN2U2MTERpdpS5GbkWXsokmBrJ0fdhnVw93YeSrU6aw9HMmQymbWHIBlyOznquihxNzMfOn6Pi65uwzrQ6/To9HyQaPc4c+YMigqLkZuhNqsfJ1cVajk6ICAgwEIjq3lqZOlEJpPB1s4W7t4NrT0USXFpVN/aQyASlYuns7WHIAlarRYyhbi/nhQKBQCglreDxfqiytXIjAYRERE9GThHg4iIiETDQIOIiIhEw0CDiIiIRMNAg4iIiETDQIOIiIhEw0CDiIiIRMNAg4iIiETDQIOIiIhEw0CDiIiIRMNAg4iIiETDQIOIiIhEw0CDiIiIRMNAgx5ZSkoKxo4di/bt26Nz586Ijo5GcXGxtYdFZDHXrl3D7Nmz0b9/f/j7+yM0NNTaQyJ66tTIx8ST+NRqNUaOHAkPDw/ExsYiJycHCxYsQG5uLj7++GNrD4/IIpKSkpCQkIB27dpBr9eDD7smqj4GGvRINm/eDLVaje3bt8PZ2RkAIJfLERkZifDwcPj4+Fh5hETm69GjB3r16gUAiIqKwtmzZ608IqKnD0sn9EgOHTqEzp07G4IMAOjTpw8UCgUSEhKsODIiy7Gx4Y9IInPx/yJ6JMnJyRWyFgqFAl5eXkhOTrbSqIiI6EnDQIMeiVqthkqlqnBcpVLh7t27VhgRERE9iRhokEUJggCZTGbtYRAR0ROCgQY9EpVKBbVaXeF4Xl5epZkOIiKSJgYa9Eh8fHwqzMXQaDRITU3lihMiIjJgoEGPJDg4GEePHsWdO3cMx/bt2weNRoOQkBArjoyIiJ4k3EeDHsnQoUOxceNGREREICIiAtnZ2Vi4cCHCwsKY0aAao6ioyLBcOy0tDfn5+di7dy8AoGPHjkbLu4mocjKBW93RI0pJSUF0dDROnDgBBwcHhIaGIjIyEg4ODtYeGpFF3LhxAz179qz03Pr16xEUFPSYR0T09GGgQURERKLhHA0iIiISDQMNIiIiEg0DDSIiIhINAw0iIiISDQMNIiIiEg0DDSIiIhINAw0iIiISDQMNeupt3boVfn5+hpe/vz+Cg4MxY8YMZGRkPJYx9OjRA1FRUYb3x44dg5+fH44dO1atfhITE7F06dJKH1hnrqioKPTo0cNkuzfffBNvvvnmI92jR48emDhx4iNd+7A+7/3aEtHThVuQU42xYMECNGvWDMXFxTh+/DhWr16NP/74Azt37oSjo+NjHUvr1q3xzTffoHnz5tW67uTJk1i2bBkGDBjAp+ASUY3AQINqjBYtWiAgIAAA0KlTJ+h0OqxYsQL79+/HK6+8Uuk1RUVFqFWrlsXHolQq0b59e4v3S0T0tGHphGqs8l/0N2/eBFBWOggMDMSlS5cwZswYBAYGYtSoUQDKHnG/YsUKvPTSS2jTpg06deqEGTNmICcnx6hPrVaLDz/8EF26dEG7du0wbNgwnD59usK9H1Q6OXXqFCZNmoSgoCAEBASgV69emDdvHgBg6dKl+PDDDwEAPXv2NJSC7u1j9+7dGDJkCNq3b4/AwECMHTsW58+fr3D/rVu3ok+fPmjTpg1efvllbN++/ZG+huWWLVuGQYMGoWPHjujQoQMGDBiA+Ph4POgJBvv27UNYWBgCAgLQs2dPrF+/vkKb/Px8LFq0CD169ECbNm3QtWtXzJs3D4WFhWaNlYieLMxoUI117do1ADB6wqZWq0V4eDiGDh2K8ePHQ6fTQa/XIyIiAidOnMDYsWPRoUMHpKWlYenSpTh9+jS+++47w4PiZs2ahe3bt2PMmDHo0qULkpKSMGXKFBQUFJgcz+HDhxEeHo5mzZohKioK7u7uSEtLw2+//QYAGDRoEO7evYsNGzZg2bJlcHFxAQBD+WXVqlX49NNP8dprryE8PBxarRZxcXF44403EB8fb2i3detWzJgxAz179kRUVBTy8vKwbNkyaDQa2Ng82r8t0tLSMGTIEHh4eAAA/vrrL0RHRyMjIwNTpkwxanvhwgXMnz8fU6ZMQYMGDbBz507MmzcPWq0WY8eOBVCWSRo+fDhu3bqFSZMmwc/PD0lJSYiNjcXly5fx5ZdfQiaTPdJYiegJIxA95b777jvB19dX+OuvvwStVivk5+cLP//8s9CpUychMDBQyMzMFARBEN577z3B19dX+Pbbb42u37Vrl+Dr6yv8+OOPRsdPnz4t+Pr6Cps2bRIEQRCuXLki+Pr6CvPnzzdqt2PHDsHX11d47733DMeOHj0q+Pr6CkePHjUc69Wrl9CrVy+huLj4gZ9l7dq1gq+vr3D9+nWj4zdv3hT8/f2FuXPnGh3Pz88XunTpIkybNk0QBEHQ6XTCCy+8IAwYMEDQ6/WGdjdu3BBat24tdO/e/YH3Ljd8+HBh+PDhDzyv0+kErVYrLFu2TOjYsaPRfbp37y74+fkJFy5cMLpm9OjRQocOHYTCwkJBEARh9erVQsuWLYXTp08btdu7d6/g6+sr/PLLL0Z93vu1JaKnC0snVGMMHjwYrVu3RocOHTBx4kQ0aNAAn332GRo0aGDUrk+fPkbvf/75Z6hUKnTv3h2lpaWGV6tWreDi4oI//vgDAAwljLCwMKPrX375ZdjaPjw5mJKSgtTUVLz++uuwt7ev9mf79ddfUVpaiv79+xuN0d7eHs8995xhjCkpKbh9+zZCQ0ONMgKenp4IDAys9n3LHTlyBKNGjcIzzzyDVq1aoXXr1oiNjUVubi6ys7ON2rZo0QItW7Y0OhYaGor8/HycO3cOQNnXvEWLFmjVqpXR53nhhRcgk8kMn4eInn4snVCNsWjRIvj4+MDW1hb169dHw4YNK7SpVasWlEql0bHs7Gyo1Wq0adOm0n7v3LkDAMjNzQUAQ0mjnK2tLZycnB46tvK5Hq6urlX5KBVkZWUBAF5//fVKz5eXRMrHen9wVX4sLS2t2vc+ffo0xo4di44dO2Lu3Llwc3ODnZ0d9u/fj1WrVqG4uLjCfSq7N/DP1zA7OxvXrl1D69atK71n+ecgoqcfAw2qMXx8fAyrTh6ksrp/vXr14OTkhLVr11Z6Te3atQHAEExkZmYaBQylpaWGX6APUj5P5FH39ahXrx4AIDY21jBP4mHtygOTe1V2rCp++OEH2NraYvXq1UbZmP3791fa/mH3Lv8a1qtXD/b29pg/f36lfZR/DiJ6+jHQIMnr1q0bfvjhB+j1erRr1+6B7YKCggAAO3fuNMp+7NmzB6WlpQ+9h7e3N7y8vPDdd99h9OjRUCgUlbYrP15SUmJ0/IUXXoCtrS1SU1MrlH7uv4+Liwt27dqF0aNHGwKrtLQ0nDx5stIsjykymQxyudxoImlxcTF27NhRafukpCRcvHjRqHyya9cu1K5d25DB6NatG1avXg0nJyc0bty42mMioqcHAw2SvH79+mHnzp2YMGEC3nzzTbRt2xZ2dna4desWjh07hp49e6J3797w8fHBK6+8gnXr1sHW1hbPP/88kpKSEBcXV6EcU5nZs2cjPDwcgwcPxqhRo+Du7o709HQcPnwYn3zyCQDA19cXALBu3ToMGDAAtra28Pb2RqNGjTB16lR8+umnuH79OoKDg6FSqZCVlYUzZ86gVq1amDp1KmxsbDBt2jTMnDkTkydPxuDBg6FWq7Fs2bJKSxpVERISgi+++ALvvvsuhgwZgtzcXMTFxT0wWGrYsCHCw8MxZcoUuLi4YMeOHfjtt98QGRlp2LNk5MiR+OmnnzB8+HCMGjUKfn5+0Ov1SE9Px6+//ooxY8Y8NOgjoqcHAw2SPLlcjpUrV2L9+vX4/vvvsWbNGsjlcri5ueG5554z/PIHgHnz5qFBgwbYtm0bNmzYgFatWmHp0qX497//bfI+Xbt2xcaNG7F8+XJER0ejpKQEbm5uRtuCBwUFYeLEidi2bRvi4+Oh1+uxfv16w3EfHx+sX78eP/zwAzQaDVxcXNCmTRsMGzbM0MegQYMAAGvXrsWUKVPg6emJiRMn4s8//3ykSZadO3fG/Pnz8dlnn2HSpElwdXXF4MGD4ezsjPfff79C+1atWuG1117D0qVL8ffff6Nhw4aYMWOGYc8SAHB0dMSmTZuwZs0afPPNN7hx4wYcHBzg7u6O559/Hp6entUeJxE9mWSC8IAdd4iIiIjMxOWtREREJBoGGkRERCQaBhpEREQkGgYaREREJBoGGkRERCQaBhpEREQkGgYaREREJBoGGkRERCQaBhpEREQkGgYaREREJBoGGkRERCQaBhpEREQkmv8PiINAIqKnM5AAAAAASUVORK5CYII=",
      "text/plain": [
       "<Figure size 640x480 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "y_true, y_pred = draw_confusion_matrix()\n",
    "y_pred.mean(), y_true.mean(), metrics.f1_score(y_true, y_pred)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Model training and hyperparameter optimization: last try\n",
    "\n",
    "How do we make the model good in both metrics: accuracy and the F1 score? Well,\n",
    "good luck, and it's your turn. This lecture is intended to demonstrate the RNN\n",
    "model with a good quality code, not to provide a complete solution to this\n",
    "problem. So, you need to figure out how to fix the model yourself.\n",
    "\n",
    "Adjusting the hyperparameters at the beginning of this notebook might help.\n",
    "Or you may need to change the model architecture, training procedure, etc. This\n",
    "is the essence of deep learning research: **YOU figure out how to make the model\n",
    "work**."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Model evaluation\n",
    "\n",
    "Once satisfied with the model, you can evaluate the final model on the test set.\n",
    "This is your *score* on this problem. Never be tempted to evaluate the model on\n",
    "the test set beforehand: it's cheating."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(0.4715857890745004, 0.7991332540445429)"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "validate(model, criterion, test_loader)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Summary\n",
    "\n",
    "We have learned about a real-world cheminformatics problem in this chapter.\n",
    "Specifically, we:\n",
    "\n",
    "- Learned about the SMILES format and how to tokenize it.\n",
    "- Could now handle datasets with heterogeneous lengths.\n",
    "- Discussed the imbalanced dataset problem.\n",
    "\n",
    "And most importantly, we have a lesson: **never blindly trust the \"accuracy\" or\n",
    "any other metrics**. Do you really want a model that considers everything,\n",
    "*literally everything*, safe as a drug?"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.8"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}