Census_Income.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "Census-Income.ipynb",
      "provenance": [],
      "collapsed_sections": [],
      "mount_file_id": "1_ZI9HuQ8nqg8llJuEeqxa7fhDQC3j_4w",
      "authorship_tag": "ABX9TyMMtRW00pFxHR/zEypa+Djn",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/finardi/tutos/blob/master/Census_Income.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "DPP9n8qM6zRY",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 357
        },
        "outputId": "d56f8c49-eff3-4570-ee8f-5cb3e4453110"
      },
      "source": [
        "! nvidia-smi"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Sun Aug  2 16:22:03 2020       \n",
            "+-----------------------------------------------------------------------------+\n",
            "| NVIDIA-SMI 450.57       Driver Version: 418.67       CUDA Version: 10.1     |\n",
            "|-------------------------------+----------------------+----------------------+\n",
            "| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\n",
            "| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |\n",
            "|                               |                      |               MIG M. |\n",
            "|===============================+======================+======================|\n",
            "|   0  Tesla K80           Off  | 00000000:00:04.0 Off |                    0 |\n",
            "| N/A   70C    P8    33W / 149W |      0MiB / 11441MiB |      0%      Default |\n",
            "|                               |                      |                 ERR! |\n",
            "+-------------------------------+----------------------+----------------------+\n",
            "                                                                               \n",
            "+-----------------------------------------------------------------------------+\n",
            "| Processes:                                                                  |\n",
            "|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |\n",
            "|        ID   ID                                                   Usage      |\n",
            "|=============================================================================|\n",
            "|  No running processes found                                                 |\n",
            "+-----------------------------------------------------------------------------+\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "GfPyGA4CJkSv"
      },
      "source": [
        "!pip install -q torch_optimizer\n",
        "import torch_optimizer as optim"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "CygUhE-f8BQ1"
      },
      "source": [
        "# Basics\n",
        "import os\n",
        "import math\n",
        "import copy\n",
        "import pandas as pd\n",
        "import numpy as np\n",
        "\n",
        "# PyTorch\n",
        "import torch \n",
        "from torch import nn\n",
        "import torch.nn.functional as F\n",
        "from torch.utils.data import Dataset, TensorDataset, DataLoader\n",
        "\n",
        "# Sklearn\n",
        "from sklearn import preprocessing\n",
        "from sklearn.preprocessing import LabelEncoder\n",
        "from sklearn.preprocessing import quantile_transform\n",
        "from sklearn.preprocessing import power_transform, PowerTransformer\n",
        "from sklearn.metrics import roc_auc_score\n",
        "\n",
        "# Plots\n",
        "from matplotlib import pyplot as plt"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "bNigt0zMS9ig",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "6af5e708-0d24-4e6b-b14a-bcbd93d3250c"
      },
      "source": [
        "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
        "\n",
        "manual_seed = 0\n",
        "def deterministic(rep=True):\n",
        "    if rep:\n",
        "        np.random.seed(manual_seed)\n",
        "        torch.manual_seed(manual_seed)\n",
        "        if torch.cuda.is_available():\n",
        "            torch.cuda.manual_seed(manual_seed)\n",
        "            torch.cuda.manual_seed_all(manual_seed)\n",
        "        torch.backends.cudnn.enabled = False \n",
        "        torch.backends.cudnn.benchmark = False\n",
        "        torch.backends.cudnn.deterministic = True\n",
        "        print(f'Deterministic experiment, seed: {manual_seed}')\n",
        "    else:\n",
        "        print('Random experiment')\n",
        "\n",
        "deterministic()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Deterministic experiment, seed: 0\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "EekDkiU1S9ip"
      },
      "source": [
        "# Dataset - Cencus Income"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "LPuc8NSS18rn",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 255
        },
        "outputId": "3d642cc4-3d9c-4cca-d49e-15d6871ef5e8"
      },
      "source": [
        "column_names = ['age', 'workclass', 'fnlwgt', 'education', 'educational-num','marital-status', 'occupation', 'relationship', 'race', 'gender','capital-gain', 'capital-loss', 'hours-per-week', 'native-country','income']\n",
        "\n",
        "train = pd.read_csv('/content/drive/My Drive/Colab Notebooks/Tabular/adult_data.txt', sep=\",\\s\", header=None, names = column_names, engine = 'python')\n",
        "test = pd.read_csv('/content/drive/My Drive/Colab Notebooks/Tabular/adult_test.txt', sep=\",\\s\", header=None, names = column_names, engine = 'python')\n",
        "test['income'].replace(regex=True,inplace=True,to_replace=r'\\.',value=r'')\n",
        "\n",
        "print(f'train shape {train.shape}')\n",
        "print(f'test shape {test.shape}')\n",
        "\n",
        "adult = pd.concat([test,train])\n",
        "adult.reset_index(inplace = True, drop = True)\n",
        "print(f'dataset shape {adult.shape}')\n",
        "adult.head()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "train shape (32561, 15)\n",
            "test shape (16281, 15)\n",
            "dataset shape (48842, 15)\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "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>age</th>\n",
              "      <th>workclass</th>\n",
              "      <th>fnlwgt</th>\n",
              "      <th>education</th>\n",
              "      <th>educational-num</th>\n",
              "      <th>marital-status</th>\n",
              "      <th>occupation</th>\n",
              "      <th>relationship</th>\n",
              "      <th>race</th>\n",
              "      <th>gender</th>\n",
              "      <th>capital-gain</th>\n",
              "      <th>capital-loss</th>\n",
              "      <th>hours-per-week</th>\n",
              "      <th>native-country</th>\n",
              "      <th>income</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th>0</th>\n",
              "      <td>25</td>\n",
              "      <td>Private</td>\n",
              "      <td>226802</td>\n",
              "      <td>11th</td>\n",
              "      <td>7</td>\n",
              "      <td>Never-married</td>\n",
              "      <td>Machine-op-inspct</td>\n",
              "      <td>Own-child</td>\n",
              "      <td>Black</td>\n",
              "      <td>Male</td>\n",
              "      <td>0</td>\n",
              "      <td>0</td>\n",
              "      <td>40</td>\n",
              "      <td>United-States</td>\n",
              "      <td>&lt;=50K</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>1</th>\n",
              "      <td>38</td>\n",
              "      <td>Private</td>\n",
              "      <td>89814</td>\n",
              "      <td>HS-grad</td>\n",
              "      <td>9</td>\n",
              "      <td>Married-civ-spouse</td>\n",
              "      <td>Farming-fishing</td>\n",
              "      <td>Husband</td>\n",
              "      <td>White</td>\n",
              "      <td>Male</td>\n",
              "      <td>0</td>\n",
              "      <td>0</td>\n",
              "      <td>50</td>\n",
              "      <td>United-States</td>\n",
              "      <td>&lt;=50K</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>2</th>\n",
              "      <td>28</td>\n",
              "      <td>Local-gov</td>\n",
              "      <td>336951</td>\n",
              "      <td>Assoc-acdm</td>\n",
              "      <td>12</td>\n",
              "      <td>Married-civ-spouse</td>\n",
              "      <td>Protective-serv</td>\n",
              "      <td>Husband</td>\n",
              "      <td>White</td>\n",
              "      <td>Male</td>\n",
              "      <td>0</td>\n",
              "      <td>0</td>\n",
              "      <td>40</td>\n",
              "      <td>United-States</td>\n",
              "      <td>&gt;50K</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>3</th>\n",
              "      <td>44</td>\n",
              "      <td>Private</td>\n",
              "      <td>160323</td>\n",
              "      <td>Some-college</td>\n",
              "      <td>10</td>\n",
              "      <td>Married-civ-spouse</td>\n",
              "      <td>Machine-op-inspct</td>\n",
              "      <td>Husband</td>\n",
              "      <td>Black</td>\n",
              "      <td>Male</td>\n",
              "      <td>7688</td>\n",
              "      <td>0</td>\n",
              "      <td>40</td>\n",
              "      <td>United-States</td>\n",
              "      <td>&gt;50K</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>4</th>\n",
              "      <td>18</td>\n",
              "      <td>?</td>\n",
              "      <td>103497</td>\n",
              "      <td>Some-college</td>\n",
              "      <td>10</td>\n",
              "      <td>Never-married</td>\n",
              "      <td>?</td>\n",
              "      <td>Own-child</td>\n",
              "      <td>White</td>\n",
              "      <td>Female</td>\n",
              "      <td>0</td>\n",
              "      <td>0</td>\n",
              "      <td>30</td>\n",
              "      <td>United-States</td>\n",
              "      <td>&lt;=50K</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n",
              "</div>"
            ],
            "text/plain": [
              "   age  workclass  fnlwgt  ... hours-per-week  native-country income\n",
              "0   25    Private  226802  ...             40   United-States  <=50K\n",
              "1   38    Private   89814  ...             50   United-States  <=50K\n",
              "2   28  Local-gov  336951  ...             40   United-States   >50K\n",
              "3   44    Private  160323  ...             40   United-States   >50K\n",
              "4   18          ?  103497  ...             30   United-States  <=50K\n",
              "\n",
              "[5 rows x 15 columns]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 44
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "--u7ZykBv3D3"
      },
      "source": [
        "### Dtypes"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "2TSL6CxLv24Z",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 425
        },
        "outputId": "2c65c49c-4562-446e-cb67-0f9583550eb0"
      },
      "source": [
        "for col in set(adult.columns) - set(adult.describe().columns):\n",
        "    adult[col] = adult[col].astype('category')\n",
        "    \n",
        "print('## 1.1. Columns and their types')\n",
        "print(adult.info())"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "## 1.1. Columns and their types\n",
            "<class 'pandas.core.frame.DataFrame'>\n",
            "RangeIndex: 48842 entries, 0 to 48841\n",
            "Data columns (total 15 columns):\n",
            " #   Column           Non-Null Count  Dtype   \n",
            "---  ------           --------------  -----   \n",
            " 0   age              48842 non-null  int64   \n",
            " 1   workclass        48842 non-null  category\n",
            " 2   fnlwgt           48842 non-null  int64   \n",
            " 3   education        48842 non-null  category\n",
            " 4   educational-num  48842 non-null  int64   \n",
            " 5   marital-status   48842 non-null  category\n",
            " 6   occupation       48842 non-null  category\n",
            " 7   relationship     48842 non-null  category\n",
            " 8   race             48842 non-null  category\n",
            " 9   gender           48842 non-null  category\n",
            " 10  capital-gain     48842 non-null  int64   \n",
            " 11  capital-loss     48842 non-null  int64   \n",
            " 12  hours-per-week   48842 non-null  int64   \n",
            " 13  native-country   48842 non-null  category\n",
            " 14  income           48842 non-null  category\n",
            "dtypes: category(9), int64(6)\n",
            "memory usage: 2.7 MB\n",
            "None\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "zK_Umji74HmO"
      },
      "source": [
        "### NaNs?"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "S8Szd4as4HmS"
      },
      "source": [
        "# NaN == \n",
        "def check_nan(df):\n",
        "    for i, j in zip(df.columns,(df.values.astype(str) == '?').sum(axis = 0)):\n",
        "        if j > 0:\n",
        "            print(str(i) + ': ' + str(j))"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "QiCkSaJe30iy",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 68
        },
        "outputId": "950bf7fe-2c32-4e86-c705-0d473ed4a585"
      },
      "source": [
        "# train\n",
        "check_nan(train)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "workclass: 1836\n",
            "occupation: 1843\n",
            "native-country: 583\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "x8dzVvDg30gD",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 68
        },
        "outputId": "68d25a4f-ca51-4820-a744-b2775a4a0182"
      },
      "source": [
        "# test\n",
        "check_nan(test)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "workclass: 963\n",
            "occupation: 966\n",
            "native-country: 274\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "8A6s4bj6IELD"
      },
      "source": [
        "### Trocar o ? por NaN"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "HYBLJT6eHplD"
      },
      "source": [
        "train_nan = np.where(train.values == '?', np.nan, train.values)\n",
        "train_nan = pd.DataFrame(train_nan, columns=train.columns)\n",
        "\n",
        "test_nan = np.where(test.values == '?', np.nan, test.values)\n",
        "test_nan  = pd.DataFrame(test_nan, columns=train.columns)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "kmbMobE8iREm",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "19244b80-5a28-4e72-9b01-b7c7f584d865"
      },
      "source": [
        "train_nan['workclass'].nunique(), train_nan['occupation'].nunique(), train_nan['native-country'].nunique() "
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(8, 14, 41)"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 50
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ZtPh6v0CIoE6",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 289
        },
        "outputId": "d7fc655a-3bc7-4da9-8369-e545454bf01c"
      },
      "source": [
        "train_nan.isnull().sum()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "age                   0\n",
              "workclass          1836\n",
              "fnlwgt                0\n",
              "education             0\n",
              "educational-num       0\n",
              "marital-status        0\n",
              "occupation         1843\n",
              "relationship          0\n",
              "race                  0\n",
              "gender                0\n",
              "capital-gain          0\n",
              "capital-loss          0\n",
              "hours-per-week        0\n",
              "native-country      583\n",
              "income                0\n",
              "dtype: int64"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 51
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "vzn-WkiVirnU",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "e9754060-433b-4990-85be-59bae771787c"
      },
      "source": [
        "test_nan['workclass'].nunique(), test_nan['occupation'].nunique(), test_nan['native-country'].nunique() "
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(8, 14, 40)"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 52
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "WTF-dlwyIsrz",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 289
        },
        "outputId": "c138e1e4-fe4b-4841-be68-4e0f456a083f"
      },
      "source": [
        "test_nan.isnull().sum()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "age                  0\n",
              "workclass          963\n",
              "fnlwgt               0\n",
              "education            0\n",
              "educational-num      0\n",
              "marital-status       0\n",
              "occupation         966\n",
              "relationship         0\n",
              "race                 0\n",
              "gender               0\n",
              "capital-gain         0\n",
              "capital-loss         0\n",
              "hours-per-week       0\n",
              "native-country     274\n",
              "income               0\n",
              "dtype: int64"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 53
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "NBQxfM-mRMNi",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "72a71ec6-cd04-45c1-b883-6d2f9312e699"
      },
      "source": [
        "train_mask = train_nan.copy()\n",
        "train_mask = np.where(train_mask.isnull() == True, np.nan, 0)\n",
        "\n",
        "test_mask = test_nan.copy()\n",
        "test_mask  = np.where(test_mask.isnull() == True, np.nan, 0)\n",
        "\n",
        "train_mask.shape, np.isnan(train_mask).sum(), test_mask.shape, np.isnan(test_mask).sum()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "((32561, 15), 4262, (16281, 15), 2203)"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 54
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "WdBy0p9FKvlx"
      },
      "source": [
        "### Trocar o np.nan por string NAN para o LabelEncoder() não quebrar, depois volta os np.nan com as Masks criadas\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "UApuQvMcO_7u",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "e5b4cbc5-6d04-4d8c-dd28-37bd482536ed"
      },
      "source": [
        "train_nan.fillna('NAN', inplace=True)\n",
        "test_nan.fillna('NAN', inplace=True)\n",
        "train_nan.isnull().sum().sum(), test_nan.isnull().sum().sum()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(0, 0)"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 55
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "p7dLhu8WjFX0",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 51
        },
        "outputId": "693606b1-962d-4e02-f7e6-32ad6a9b0563"
      },
      "source": [
        "print(train_nan['workclass'].nunique(), train_nan['occupation'].nunique(), train_nan['native-country'].nunique())\n",
        "print(test_nan['workclass'].nunique(), test_nan['occupation'].nunique(), test_nan['native-country'].nunique())"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "9 15 42\n",
            "9 15 41\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "-AT-pob85wZv"
      },
      "source": [
        "### Conferindo os tipos das col (identificar cols. categóricas)"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "vjYXjZXM5wZw",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 289
        },
        "outputId": "c77563ca-87e9-4826-a42e-faba0c262e01"
      },
      "source": [
        "train_nan.dtypes"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "age                 int64\n",
              "workclass          object\n",
              "fnlwgt              int64\n",
              "education          object\n",
              "educational-num     int64\n",
              "marital-status     object\n",
              "occupation         object\n",
              "relationship       object\n",
              "race               object\n",
              "gender             object\n",
              "capital-gain        int64\n",
              "capital-loss        int64\n",
              "hours-per-week      int64\n",
              "native-country     object\n",
              "income             object\n",
              "dtype: object"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 57
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "io0vkgE653tQ"
      },
      "source": [
        "### Separando o dataset em Cols Cat e Cols cont.\n",
        "- Todas as cols object type\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "qoQcfgqBw9tq",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "9f8e7f7d-839f-4ab1-e162-d43f5e2ab47d"
      },
      "source": [
        "cat_cols = ['workclass', 'education', 'marital-status', 'occupation', 'relationship', \n",
        "             'race'     , 'gender'   , 'native-country', 'income']\n",
        "\n",
        "cont_cols = ['age'         , 'fnlwgt'      , 'educational-num', \n",
        "             'capital-gain', 'capital-loss', 'hours-per-week' ]\n",
        "\n",
        "print(len(cat_cols) + len(cont_cols))"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "15\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "J3RsXe9Py-A1"
      },
      "source": [
        "df_cat_train  = train_nan[cat_cols].copy()\n",
        "df_cont_train = train_nan[cont_cols].copy()\n",
        "\n",
        "df_cat_test  = test_nan[cat_cols].copy()\n",
        "df_cont_test = test_nan[cont_cols].copy()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "G2_84Xb730Yt"
      },
      "source": [
        "def transform_cat_label(df):\n",
        "    categorical_dims = {}\n",
        "    \n",
        "    for col in df.columns:\n",
        "        print(col, df[col].nunique())\n",
        "        l_enc = LabelEncoder()\n",
        "        df[col] = l_enc.fit_transform(df[col].values)\n",
        "        categorical_dims[col] = len(l_enc.classes_)\n",
        "\n",
        "    categorical_dims.pop('income')     # target\n",
        "    return list(categorical_dims.values())"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "VG7ix9hk30Uw",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 357
        },
        "outputId": "9dc08f47-5d39-4447-e429-f98183a061a0"
      },
      "source": [
        "cat_train = transform_cat_label(df_cat_train)\n",
        "cat_test = transform_cat_label(df_cat_test)\n",
        "print(cat_train, '\\n',cat_test)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "workclass 9\n",
            "education 16\n",
            "marital-status 7\n",
            "occupation 15\n",
            "relationship 6\n",
            "race 5\n",
            "gender 2\n",
            "native-country 42\n",
            "income 2\n",
            "workclass 9\n",
            "education 16\n",
            "marital-status 7\n",
            "occupation 15\n",
            "relationship 6\n",
            "race 5\n",
            "gender 2\n",
            "native-country 41\n",
            "income 2\n",
            "[9, 16, 7, 15, 6, 5, 2, 42] \n",
            " [9, 16, 7, 15, 6, 5, 2, 41]\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Eqo51RpEgwWg",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "1eff75d4-fe09-48b0-91fe-429ce729c31a"
      },
      "source": [
        "cat_dims = cat_train.copy()\n",
        "len(cat_dims), cat_dims"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(8, [9, 16, 7, 15, 6, 5, 2, 42])"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 62
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "DCkHWoX7AC7X"
      },
      "source": [
        "### YeoJohnson"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "99eLqW6aACnp",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 221
        },
        "outputId": "567ed150-a94c-45bf-9af1-8ebd54c4a043"
      },
      "source": [
        "pt = PowerTransformer(method='yeo-johnson', standardize=True) \n",
        "train_yj = pt.fit(df_cont_train.values)\n",
        "\n",
        "train_yj = pt.transform(df_cont_train.values)\n",
        "df_train_yj = pd.DataFrame(train_yj,\n",
        "                           columns=df_cont_train.columns, \n",
        "                           index=df_cont_train.index)\n",
        "\n",
        "test_yj = pt.fit(df_cont_test.values)\n",
        "\n",
        "test_yj = pt.transform(df_cont_test.values)\n",
        "df_test_yj = pd.DataFrame(test_yj,\n",
        "                          columns=df_cont_test.columns, \n",
        "                          index=df_cont_test.index)\n",
        "print(df_train_yj.shape, df_test_yj.shape)\n",
        "\n",
        "df_train_yj.head()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "(32561, 6) (16281, 6)\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "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>age</th>\n",
              "      <th>fnlwgt</th>\n",
              "      <th>educational-num</th>\n",
              "      <th>capital-gain</th>\n",
              "      <th>capital-loss</th>\n",
              "      <th>hours-per-week</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th>0</th>\n",
              "      <td>0.175803</td>\n",
              "      <td>-1.180293</td>\n",
              "      <td>1.164754</td>\n",
              "      <td>3.317526</td>\n",
              "      <td>-0.22121</td>\n",
              "      <td>-0.035570</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>1</th>\n",
              "      <td>0.891954</td>\n",
              "      <td>-1.089846</td>\n",
              "      <td>1.164754</td>\n",
              "      <td>-0.301425</td>\n",
              "      <td>-0.22121</td>\n",
              "      <td>-2.221546</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>2</th>\n",
              "      <td>0.102709</td>\n",
              "      <td>0.398784</td>\n",
              "      <td>-0.464368</td>\n",
              "      <td>-0.301425</td>\n",
              "      <td>-0.22121</td>\n",
              "      <td>-0.035570</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>3</th>\n",
              "      <td>1.064392</td>\n",
              "      <td>0.562579</td>\n",
              "      <td>-1.200303</td>\n",
              "      <td>-0.301425</td>\n",
              "      <td>-0.22121</td>\n",
              "      <td>-0.035570</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>4</th>\n",
              "      <td>-0.731682</td>\n",
              "      <td>1.340417</td>\n",
              "      <td>1.164754</td>\n",
              "      <td>-0.301425</td>\n",
              "      <td>-0.22121</td>\n",
              "      <td>-0.035570</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n",
              "</div>"
            ],
            "text/plain": [
              "        age    fnlwgt  ...  capital-loss  hours-per-week\n",
              "0  0.175803 -1.180293  ...      -0.22121       -0.035570\n",
              "1  0.891954 -1.089846  ...      -0.22121       -2.221546\n",
              "2  0.102709  0.398784  ...      -0.22121       -0.035570\n",
              "3  1.064392  0.562579  ...      -0.22121       -0.035570\n",
              "4 -0.731682  1.340417  ...      -0.22121       -0.035570\n",
              "\n",
              "[5 rows x 6 columns]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 63
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "zhET0MrBACfL"
      },
      "source": [
        "### Quantile"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "gP5vN0iW0I2L",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 221
        },
        "outputId": "58820dbf-01dd-4f9d-8764-7cd63d5ecfb2"
      },
      "source": [
        "quantile_train = quantile_transform(df_cont_train.values, \n",
        "                                    n_quantiles=10, \n",
        "                                    random_state=manual_seed, \n",
        "                                    copy=True, \n",
        "                                    output_distribution='uniform')\n",
        "\n",
        "quantile_test  = quantile_transform(df_cont_test.values, \n",
        "                                    n_quantiles=10, \n",
        "                                    random_state=manual_seed, \n",
        "                                    copy=True, \n",
        "                                    output_distribution='uniform')\n",
        "\n",
        "df_train_quant = pd.DataFrame(data=quantile_train, \n",
        "                             columns=df_cont_train.columns, \n",
        "                             index=df_cont_train.index)\n",
        "\n",
        "df_test_quant = pd.DataFrame(data=quantile_test, \n",
        "                             columns=df_cont_test.columns, \n",
        "                             index=df_cont_test.index)\n",
        "\n",
        "print(df_train_quant.shape, df_test_quant.shape)\n",
        "\n",
        "df_train_quant.head()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "(32561, 6) (16281, 6)\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "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>age</th>\n",
              "      <th>fnlwgt</th>\n",
              "      <th>educational-num</th>\n",
              "      <th>capital-gain</th>\n",
              "      <th>capital-loss</th>\n",
              "      <th>hours-per-week</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th>0</th>\n",
              "      <td>0.555556</td>\n",
              "      <td>0.125654</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.891304</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>1</th>\n",
              "      <td>0.791667</td>\n",
              "      <td>0.141929</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.055556</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>2</th>\n",
              "      <td>0.527778</td>\n",
              "      <td>0.681646</td>\n",
              "      <td>0.333333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>3</th>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.736893</td>\n",
              "      <td>0.111111</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>4</th>\n",
              "      <td>0.250000</td>\n",
              "      <td>0.890743</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n",
              "</div>"
            ],
            "text/plain": [
              "        age    fnlwgt  ...  capital-loss  hours-per-week\n",
              "0  0.555556  0.125654  ...           0.0        0.500000\n",
              "1  0.791667  0.141929  ...           0.0        0.055556\n",
              "2  0.527778  0.681646  ...           0.0        0.500000\n",
              "3  0.833333  0.736893  ...           0.0        0.500000\n",
              "4  0.250000  0.890743  ...           0.0        0.500000\n",
              "\n",
              "[5 rows x 6 columns]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 64
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "3YnVnut_DYZi"
      },
      "source": [
        "### Concatena as Cols cont_norm + Cols cat."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "W0tKvyJ6DYZk"
      },
      "source": [
        "def concat_df(df1, df2):\n",
        "    value =  df1.columns.to_list() + df2.columns.to_list()\n",
        "    columns = {k:v for k,v in enumerate(value)}\n",
        "    df_concat = pd.concat([df1, df2], axis=1, ignore_index=True)\n",
        "    df_concat.rename(columns=columns, inplace=True)\n",
        "    return df_concat"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "wmseqfyh30IZ",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 204
        },
        "outputId": "18225d0d-8a9f-4e03-8420-47bead349f07"
      },
      "source": [
        "df_train_processed = concat_df(df_train_quant, df_cat_train)\n",
        "df_test_processed  = concat_df(df_test_quant, df_cat_test)\n",
        "df_train_processed.head()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "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>age</th>\n",
              "      <th>fnlwgt</th>\n",
              "      <th>educational-num</th>\n",
              "      <th>capital-gain</th>\n",
              "      <th>capital-loss</th>\n",
              "      <th>hours-per-week</th>\n",
              "      <th>workclass</th>\n",
              "      <th>education</th>\n",
              "      <th>marital-status</th>\n",
              "      <th>occupation</th>\n",
              "      <th>relationship</th>\n",
              "      <th>race</th>\n",
              "      <th>gender</th>\n",
              "      <th>native-country</th>\n",
              "      <th>income</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th>0</th>\n",
              "      <td>0.555556</td>\n",
              "      <td>0.125654</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.891304</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>7</td>\n",
              "      <td>9</td>\n",
              "      <td>4</td>\n",
              "      <td>0</td>\n",
              "      <td>1</td>\n",
              "      <td>4</td>\n",
              "      <td>1</td>\n",
              "      <td>39</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>1</th>\n",
              "      <td>0.791667</td>\n",
              "      <td>0.141929</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.055556</td>\n",
              "      <td>6</td>\n",
              "      <td>9</td>\n",
              "      <td>2</td>\n",
              "      <td>3</td>\n",
              "      <td>0</td>\n",
              "      <td>4</td>\n",
              "      <td>1</td>\n",
              "      <td>39</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>2</th>\n",
              "      <td>0.527778</td>\n",
              "      <td>0.681646</td>\n",
              "      <td>0.333333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4</td>\n",
              "      <td>11</td>\n",
              "      <td>0</td>\n",
              "      <td>5</td>\n",
              "      <td>1</td>\n",
              "      <td>4</td>\n",
              "      <td>1</td>\n",
              "      <td>39</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>3</th>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.736893</td>\n",
              "      <td>0.111111</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4</td>\n",
              "      <td>1</td>\n",
              "      <td>2</td>\n",
              "      <td>5</td>\n",
              "      <td>0</td>\n",
              "      <td>2</td>\n",
              "      <td>1</td>\n",
              "      <td>39</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>4</th>\n",
              "      <td>0.250000</td>\n",
              "      <td>0.890743</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4</td>\n",
              "      <td>9</td>\n",
              "      <td>2</td>\n",
              "      <td>10</td>\n",
              "      <td>5</td>\n",
              "      <td>2</td>\n",
              "      <td>0</td>\n",
              "      <td>4</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n",
              "</div>"
            ],
            "text/plain": [
              "        age    fnlwgt  educational-num  ...  gender  native-country  income\n",
              "0  0.555556  0.125654         0.833333  ...       1              39       0\n",
              "1  0.791667  0.141929         0.833333  ...       1              39       0\n",
              "2  0.527778  0.681646         0.333333  ...       1              39       0\n",
              "3  0.833333  0.736893         0.111111  ...       1              39       0\n",
              "4  0.250000  0.890743         0.833333  ...       0               4       0\n",
              "\n",
              "[5 rows x 15 columns]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 66
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "0LYtajj40Icn"
      },
      "source": [
        "### Volta dos NaNs"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "UgmEeUySPaff"
      },
      "source": [
        "# Volta os NaNs\n",
        "values_with_nan_train = df_train_processed.values + train_mask\n",
        "df_train = pd.DataFrame(values_with_nan_train, columns=df_train_processed.columns)\n",
        "\n",
        "values_with_nan_test  = df_test_processed.values + test_mask\n",
        "df_test = pd.DataFrame(values_with_nan_test, columns=df_test_processed.columns)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "bjP3SaVcQVro",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 289
        },
        "outputId": "ac87707a-b4a7-46eb-dce7-fe1d0e7d16f0"
      },
      "source": [
        "df_train.isnull().sum()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "age                   0\n",
              "fnlwgt             1836\n",
              "educational-num       0\n",
              "capital-gain          0\n",
              "capital-loss          0\n",
              "hours-per-week        0\n",
              "workclass          1843\n",
              "education             0\n",
              "marital-status        0\n",
              "occupation            0\n",
              "relationship          0\n",
              "race                  0\n",
              "gender                0\n",
              "native-country      583\n",
              "income                0\n",
              "dtype: int64"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 68
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "qn4AVxwuQPjg",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 289
        },
        "outputId": "abbb7529-5285-4e9f-e3d3-259491f14882"
      },
      "source": [
        "df_test.isnull().sum()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "age                  0\n",
              "fnlwgt             963\n",
              "educational-num      0\n",
              "capital-gain         0\n",
              "capital-loss         0\n",
              "hours-per-week       0\n",
              "workclass          966\n",
              "education            0\n",
              "marital-status       0\n",
              "occupation           0\n",
              "relationship         0\n",
              "race                 0\n",
              "gender               0\n",
              "native-country     274\n",
              "income               0\n",
              "dtype: int64"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 69
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "R5VKS68E2qyG"
      },
      "source": [
        "### Dataset Processado"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "z5HsWJUo2t4G",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 204
        },
        "outputId": "9f11a6f9-f368-4d6d-a3c5-9f4361925224"
      },
      "source": [
        "df_train.head()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "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>age</th>\n",
              "      <th>fnlwgt</th>\n",
              "      <th>educational-num</th>\n",
              "      <th>capital-gain</th>\n",
              "      <th>capital-loss</th>\n",
              "      <th>hours-per-week</th>\n",
              "      <th>workclass</th>\n",
              "      <th>education</th>\n",
              "      <th>marital-status</th>\n",
              "      <th>occupation</th>\n",
              "      <th>relationship</th>\n",
              "      <th>race</th>\n",
              "      <th>gender</th>\n",
              "      <th>native-country</th>\n",
              "      <th>income</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th>0</th>\n",
              "      <td>0.555556</td>\n",
              "      <td>0.125654</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.891304</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>7.0</td>\n",
              "      <td>9.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>1</th>\n",
              "      <td>0.791667</td>\n",
              "      <td>0.141929</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.055556</td>\n",
              "      <td>6.0</td>\n",
              "      <td>9.0</td>\n",
              "      <td>2.0</td>\n",
              "      <td>3.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>2</th>\n",
              "      <td>0.527778</td>\n",
              "      <td>0.681646</td>\n",
              "      <td>0.333333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4.0</td>\n",
              "      <td>11.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>5.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>3</th>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.736893</td>\n",
              "      <td>0.111111</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>2.0</td>\n",
              "      <td>5.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>2.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>4</th>\n",
              "      <td>0.250000</td>\n",
              "      <td>0.890743</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4.0</td>\n",
              "      <td>9.0</td>\n",
              "      <td>2.0</td>\n",
              "      <td>10.0</td>\n",
              "      <td>5.0</td>\n",
              "      <td>2.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n",
              "</div>"
            ],
            "text/plain": [
              "        age    fnlwgt  educational-num  ...  gender  native-country  income\n",
              "0  0.555556  0.125654         0.833333  ...     1.0            39.0     0.0\n",
              "1  0.791667  0.141929         0.833333  ...     1.0            39.0     0.0\n",
              "2  0.527778  0.681646         0.333333  ...     1.0            39.0     0.0\n",
              "3  0.833333  0.736893         0.111111  ...     1.0            39.0     0.0\n",
              "4  0.250000  0.890743         0.833333  ...     0.0             4.0     0.0\n",
              "\n",
              "[5 rows x 15 columns]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 70
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Ood_8U-Mk1Vy",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "5fd11b80-340f-4957-c127-21d47219e2e4"
      },
      "source": [
        "features = df_train.iloc[:,:-1].shape[1]\n",
        "features, cat_dims, len(cat_dims)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(14, [9, 16, 7, 15, 6, 5, 2, 42], 8)"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 71
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "AO1Lsaik30F0"
      },
      "source": [
        "### Criando o dataset de validação "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "KP-ixae6FW05",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 238
        },
        "outputId": "0482b7c7-e808-4ebc-9ada-e9adbe4438d2"
      },
      "source": [
        "from sklearn.model_selection import train_test_split\n",
        "train_X, valid_X, = train_test_split(df_train.values,\n",
        "                                     shuffle=True, \n",
        "                                     test_size=0.2, \n",
        "                                     random_state=manual_seed)\n",
        "print(train_X.shape, valid_X.shape)\n",
        "\n",
        "df_train = pd.DataFrame(train_X, columns=df_train.columns)\n",
        "df_valid = pd.DataFrame(valid_X, columns=df_train.columns)\n",
        "print(df_train.shape, df_valid.shape, df_test.shape)\n",
        "\n",
        "df_valid.head()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "(26048, 15) (6513, 15)\n",
            "(26048, 15) (6513, 15) (16281, 15)\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "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>age</th>\n",
              "      <th>fnlwgt</th>\n",
              "      <th>educational-num</th>\n",
              "      <th>capital-gain</th>\n",
              "      <th>capital-loss</th>\n",
              "      <th>hours-per-week</th>\n",
              "      <th>workclass</th>\n",
              "      <th>education</th>\n",
              "      <th>marital-status</th>\n",
              "      <th>occupation</th>\n",
              "      <th>relationship</th>\n",
              "      <th>race</th>\n",
              "      <th>gender</th>\n",
              "      <th>native-country</th>\n",
              "      <th>income</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th>0</th>\n",
              "      <td>0.222222</td>\n",
              "      <td>0.492735</td>\n",
              "      <td>0.611111</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.755556</td>\n",
              "      <td>4.0</td>\n",
              "      <td>15.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>1</th>\n",
              "      <td>0.222222</td>\n",
              "      <td>0.684065</td>\n",
              "      <td>0.833333</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4.0</td>\n",
              "      <td>9.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>10.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>2</th>\n",
              "      <td>0.177778</td>\n",
              "      <td>0.789545</td>\n",
              "      <td>0.740741</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.000000</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4.0</td>\n",
              "      <td>7.0</td>\n",
              "      <td>2.0</td>\n",
              "      <td>12.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>3</th>\n",
              "      <td>0.711111</td>\n",
              "      <td>0.360130</td>\n",
              "      <td>0.037037</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.937404</td>\n",
              "      <td>0.500000</td>\n",
              "      <td>4.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>2.0</td>\n",
              "      <td>14.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>0.0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>4</th>\n",
              "      <td>0.688889</td>\n",
              "      <td>0.469147</td>\n",
              "      <td>0.111111</td>\n",
              "      <td>0.0</td>\n",
              "      <td>0.960922</td>\n",
              "      <td>0.945626</td>\n",
              "      <td>4.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>0.0</td>\n",
              "      <td>14.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>4.0</td>\n",
              "      <td>1.0</td>\n",
              "      <td>39.0</td>\n",
              "      <td>1.0</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n",
              "</div>"
            ],
            "text/plain": [
              "        age    fnlwgt  educational-num  ...  gender  native-country  income\n",
              "0  0.222222  0.492735         0.611111  ...     0.0            39.0     0.0\n",
              "1  0.222222  0.684065         0.833333  ...     0.0            39.0     0.0\n",
              "2  0.177778  0.789545         0.740741  ...     1.0            39.0     0.0\n",
              "3  0.711111  0.360130         0.037037  ...     1.0            39.0     0.0\n",
              "4  0.688889  0.469147         0.111111  ...     1.0            39.0     1.0\n",
              "\n",
              "[5 rows x 15 columns]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 72
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "jdVq7l_2FWt5",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "0ee618f2-d6f5-4723-d177-34c18d1b5f4f"
      },
      "source": [
        "class CensusIncome(Dataset):\n",
        "    def __init__(self, df):\n",
        "      super(CensusIncome, self).__init__()\n",
        "\n",
        "      self.data = df.values[:, :-1].copy()\n",
        "      self.target = df['income'].values.copy()\n",
        "      self.target = torch.from_numpy(self.target).type(torch.LongTensor)\n",
        "      \n",
        "    def __len__(self):\n",
        "      return self.target.shape[0]\n",
        "    \n",
        "    def __getitem__(self, index):\n",
        "        target = self.target[index]\n",
        "        x = self.data[index]\n",
        "        xnan = x.copy()\n",
        "        x[np.isnan(x)] = 0.\n",
        "        xnan[~np.isnan(xnan)] = 0.\n",
        "        xnan[np.isnan(xnan)]  = 1.\n",
        "\n",
        "        return torch.tensor(x).type(torch.float), \\\n",
        "               torch.tensor(xnan).type(torch.float), \\\n",
        "               target.squeeze(-1)\n",
        "\n",
        "# Test\n",
        "x = CensusIncome(df_train)\n",
        "x[7][1], len(x[7][1])"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(tensor([0., 1., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0.]), 14)"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 73
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "HRE_7e_YaIzr"
      },
      "source": [
        "### Datasets"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "l_NYtTCJabsy",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "d12b4258-dc92-43b4-c796-ba5b98263513"
      },
      "source": [
        "ds_train = CensusIncome(df_train)\n",
        "ds_debug = CensusIncome(df_train[:1000])\n",
        "ds_valid = CensusIncome(df_valid)\n",
        "ds_test  = CensusIncome(df_test)\n",
        "\n",
        "len(ds_train), len(ds_debug), len(ds_valid), len(ds_test)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(26048, 1000, 6513, 16281)"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 74
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "DlND1RTEabpj"
      },
      "source": [
        "### Dataloaders"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "rAg57Rq2abnb",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "89dbe141-7272-4498-a2a2-679080f237f9"
      },
      "source": [
        "BATCH_SZ = 128\n",
        "\n",
        "# Dataloaders\n",
        "dataloaders = {\n",
        "     'train': DataLoader(ds_train,\n",
        "                         batch_size=BATCH_SZ,\n",
        "                         shuffle=True,\n",
        "                         num_workers=4,\n",
        "                         pin_memory=True),\n",
        "    \n",
        "     'valid': DataLoader(ds_valid,\n",
        "                         batch_size=BATCH_SZ,\n",
        "                         shuffle=True,\n",
        "                         num_workers=4,\n",
        "                         pin_memory=True),\n",
        "    \n",
        "     'test' : DataLoader(ds_test,\n",
        "                         batch_size=BATCH_SZ,\n",
        "                         shuffle=False,\n",
        "                         num_workers=4,\n",
        "                         pin_memory=True),\n",
        "     \n",
        "     'debug': DataLoader(ds_debug,\n",
        "                         batch_size=BATCH_SZ,\n",
        "                         shuffle=False,\n",
        "                         num_workers=4,\n",
        "                         pin_memory=True)\n",
        " }\n",
        "\n",
        "# sanity check\n",
        "dl_sizes = {x: len(dataloaders[x]) for x in dataloaders.keys()}\n",
        "dl_sizes "
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "{'debug': 8, 'test': 128, 'train': 204, 'valid': 51}"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 75
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "U2VUdKov5wLq"
      },
      "source": [
        "###  <span style=\"color:orange\"> Função clones cria N cópias idênticas \n",
        "\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "mbktjEeXekGr"
      },
      "source": [
        "def clones(module, N):\n",
        "    return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "NnLEd-Xk59fZ"
      },
      "source": [
        "#  <span style=\"color:orange\"> Atenção MultiHead\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "1k7JUjXRt3Dq"
      },
      "source": [
        "einsum = torch.einsum\n",
        "sqrt = math.sqrt\n",
        "\n",
        "class MultiHead(nn.Module):\n",
        "  def __init__(self, n_heads, d_model, drop):\n",
        "    super(MultiHead, self).__init__()\n",
        "    assert d_model % n_heads == 0\n",
        "\n",
        "    self.drop = nn.Dropout(drop)\n",
        "    self.n_heads = n_heads\n",
        "    self.d_k = d_model // n_heads; self.d_v = d_model // n_heads \n",
        "    self.W   = clones(nn.Linear(d_model, self.d_k * n_heads, bias=False), 4)\n",
        "      \n",
        "  def forward(self, emb_x): \n",
        "    N, L,_= emb_x.size()\n",
        "\n",
        "    #------------------------------- projeções -------------------------------\n",
        "    q = self.W[0](emb_x).view(N, L, self.n_heads, self.d_k).transpose(1,2) \n",
        "    k = self.W[1](emb_x).view(N, L, self.n_heads, self.d_k).transpose(1,2) \n",
        "    v = self.W[2](emb_x).view(N, L, self.n_heads, self.d_v).transpose(1,2) \n",
        "    \n",
        "    #------------------------------- scores e mask ---------------------------\n",
        "    scores = einsum('bnlq, bniq -> bnli', q, k) / sqrt(self.d_k) \n",
        "    scores = self.drop(F.softmax(scores, dim=-1))\n",
        "    \n",
        "    #------------------------------ value e contexto -------------------------\n",
        "    att = einsum('bnli, bniv -> bnlv', scores, v)         # N, n_neads, L, dv\n",
        "    att = att.transpose(1,2).contiguous().view(N, L, -1)  # N, L, n_heads * dv  \n",
        "    \n",
        "    return self.W[-1](att)  "
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "SpI3NLuEtPBQ"
      },
      "source": [
        "#  <span style=\"color:orange\">  Net"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "tWe3sNZsyUNW"
      },
      "source": [
        "class Net(nn.Module):\n",
        "    def __init__(self, cat_dims, d_model, n_head, drop, out):\n",
        "        super(Net, self).__init__()\n",
        "        assert d_model % n_head == 0\n",
        "\n",
        "        self.cont_feats = features - len(cat_dims)\n",
        "        self.x_nan      = nn.Embedding(2, d_model) # 0 ou 1\n",
        "        self.cat_emb    = nn.ModuleList([nn.Embedding(cat_dim, d_model)\n",
        "                                         for cat_dim in cat_dims])\n",
        "\n",
        "        self.actv = nn.ReLU()\n",
        "        self.drop = nn.Dropout(drop)\n",
        "        \n",
        "        self.d_model    = d_model\n",
        "        self.cont_emb   = nn.Linear(1, d_model)\n",
        "        self.pe   = nn.Embedding(features, d_model)    # Pos. emb. Treinável\n",
        "                \n",
        "        self.mha  = MultiHead(n_head, d_model, drop)\n",
        "        self.norm = nn.LayerNorm(d_model)\n",
        "        \n",
        "        self.ff   = nn.Linear(d_model, d_model//out)\n",
        "        self.fc0  = nn.Linear(d_model//out * features, d_model)\n",
        "        self.fc1  = nn.Linear(d_model, out)\n",
        "\n",
        "        # self.fc  = nn.Linear(d_model * features, out)\n",
        "\n",
        "    def forward(self, x, x_nan):\n",
        "        x_nan  = self.x_nan(x_nan.long())\n",
        "        x_cont = x[:,:self.cont_feats].unsqueeze(-1)\n",
        "        x_cont = self.cont_emb(x_cont)\n",
        "        x_cat  = torch.zeros(x.shape[0], len(cat_dims), self.d_model).to(x.device)\n",
        "       \n",
        "        for i, emb in enumerate(self.cat_emb):\n",
        "            xi = x[:, i + self.cont_feats].long()\n",
        "            xemb = emb(xi)\n",
        "            x_cat[:, i,:] = xemb\n",
        "\n",
        "        x_cont = self.mha(x_cont)\n",
        "        x_cat  = self.mha(x_cat)\n",
        "        # print(x_nan.shape)\n",
        "\n",
        "        idx   = torch.arange(end=features).to(x.device)\n",
        "        pos_e = self.pe(idx).unsqueeze(0)\n",
        "        pos_e = self.mha(x_nan + pos_e)\n",
        "        pos_e = self.norm(pos_e)\n",
        "        o = torch.cat((x_cont, x_cat), dim=1) + pos_e \n",
        "        o = self.norm(o)\n",
        "        o = self.ff(o) \n",
        "        o = self.actv(o)\n",
        "\n",
        "        o = o.view(o.shape[0], -1) \n",
        "        o = self.fc0(o) \n",
        "        o = self.actv(o)\n",
        "        o = self.drop(o)\n",
        "        return self.fc1(o) "
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "JBm8E5ubfaC5",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "8bbd608c-66f8-436a-a53d-633fc46f4fe9"
      },
      "source": [
        "D     = 256\n",
        "H     = 8\n",
        "DROP  = 0.2\n",
        "OUT   = 2\n",
        "model = Net(cat_dims, D, H, DROP, OUT)\n",
        "\n",
        "x, xnan, y = next(iter(dataloaders['debug']))\n",
        "with torch.no_grad():\n",
        "  print(model(x, xnan).shape)   "
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "torch.Size([128, 2])\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "rae5F8qQ4Xh2",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 612
        },
        "outputId": "e5dadffd-5719-4675-c98d-0151b7b3fc90"
      },
      "source": [
        "def count_parameters(model):\n",
        "    return sum(p.numel() for p in model.parameters() if p.requires_grad)\n",
        "\n",
        "print('\\n','#' * 48,f'\\n # The model has {count_parameters(model):,}' \\\n",
        "       ' trainable parameters #\\n', '#' * 48,'\\n' )  \n",
        "model"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "\n",
            " ################################################ \n",
            " # The model has 785,794 trainable parameters #\n",
            " ################################################ \n",
            "\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Net(\n",
              "  (x_nan): Embedding(2, 256)\n",
              "  (cat_emb): ModuleList(\n",
              "    (0): Embedding(9, 256)\n",
              "    (1): Embedding(16, 256)\n",
              "    (2): Embedding(7, 256)\n",
              "    (3): Embedding(15, 256)\n",
              "    (4): Embedding(6, 256)\n",
              "    (5): Embedding(5, 256)\n",
              "    (6): Embedding(2, 256)\n",
              "    (7): Embedding(42, 256)\n",
              "  )\n",
              "  (actv): ReLU()\n",
              "  (drop): Dropout(p=0.2, inplace=False)\n",
              "  (cont_emb): Linear(in_features=1, out_features=256, bias=True)\n",
              "  (pe): Embedding(14, 256)\n",
              "  (mha): MultiHead(\n",
              "    (drop): Dropout(p=0.2, inplace=False)\n",
              "    (W): ModuleList(\n",
              "      (0): Linear(in_features=256, out_features=256, bias=False)\n",
              "      (1): Linear(in_features=256, out_features=256, bias=False)\n",
              "      (2): Linear(in_features=256, out_features=256, bias=False)\n",
              "      (3): Linear(in_features=256, out_features=256, bias=False)\n",
              "    )\n",
              "  )\n",
              "  (norm): LayerNorm((256,), eps=1e-05, elementwise_affine=True)\n",
              "  (ff): Linear(in_features=256, out_features=128, bias=True)\n",
              "  (fc0): Linear(in_features=1792, out_features=256, bias=True)\n",
              "  (fc1): Linear(in_features=256, out_features=2, bias=True)\n",
              ")"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 81
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_naW8wYZvMfm"
      },
      "source": [
        "# TransLayerClassifier"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Q21LagusvQWm",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "5537a2ee-2d22-4a2f-9788-431deda0894c"
      },
      "source": [
        "class TransModel(nn.Module):\n",
        "\n",
        "    def __init__(self, cat_dims, num_layers, out):\n",
        "        super().__init__()\n",
        "\n",
        "        self.n_cont_feats = 6\n",
        "\n",
        "        self.embs = nn.ModuleList([\n",
        "            nn.Embedding(nc, dim_) for nc in cat_dims])\n",
        "\n",
        "        self.cont_emb = nn.Linear(1, dim_, bias=False)\n",
        "\n",
        "        self.x_nan = nn.Embedding(2, dim_)\n",
        "\n",
        "        self.transformer = nn.TransformerEncoder(trans_layer, num_layers=num_layers)\n",
        "      \n",
        "        self.norm1 = nn.BatchNorm1d(features)\n",
        "      \n",
        "        self.classifier1 = nn.Linear(dim_ * features, dim_ * features)\n",
        "        self.relu1      = nn.ReLU()\n",
        "        self.dropout1    = nn.Dropout()\n",
        "        \n",
        "        self.norm2 = nn.BatchNorm1d(dim_ * features)\n",
        "        self.classifier2 = nn.Linear(dim_ * features, out)\n",
        "\n",
        "    def forward(self, x, x_nan):\n",
        "        x_nan = self.x_nan(x_nan.long())\n",
        "\n",
        "        x_cont = x[:,:self.n_cont_feats].unsqueeze(-1)\n",
        "        x_cont = self.cont_emb(x_cont)\n",
        "\n",
        "        x_cat = torch.zeros(x.shape[0], len(cat_dims), dim_).to(x.device)\n",
        "\n",
        "        for i, emb in enumerate(self.embs):\n",
        "            xi = x[:, i + self.n_cont_feats].long()\n",
        "            xemb = emb(xi)\n",
        "            x_cat[:, i,:] = xemb\n",
        "\n",
        "        x = torch.cat((x_cont, x_cat), dim=1) #+ x_nan # N, S, E\n",
        "        x = x.permute(1,0,2).contiguous() #  S, N, E\n",
        "        o = self.transformer(x)\n",
        "        o = o.permute(1,0,2).contiguous() #\n",
        "        o = self.norm1(o) \n",
        "        o = o.view(o.shape[0], -1)\n",
        "        o = self.classifier1(o)\n",
        "        o = self.norm2(o) \n",
        "        o = self.relu1(o)\n",
        "        o = self.dropout1(o)\n",
        "        o = self.classifier2(o)\n",
        "\n",
        "        return o\n",
        "\n",
        "############\n",
        "# TESTANDO\n",
        "############\n",
        "dim_ = 64\n",
        "trans_layer = nn.TransformerEncoderLayer(d_model=dim_, nhead=8, dim_feedforward=128, dropout=0.2, activation='gelu')\n",
        "model = TransModel(cat_dims,num_layers=4 , out=2)\n",
        "\n",
        "x, xnan, y = next(iter(dataloaders['debug']))\n",
        "with torch.no_grad():\n",
        "  print(model(x, xnan).shape)   "
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "torch.Size([64, 2])\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Xf8Uy7qf7NLG",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 102
        },
        "outputId": "b19210d9-6591-4e1d-f9c9-faed93387654"
      },
      "source": [
        "def count_parameters(model):\n",
        "    return sum(p.numel() for p in model.parameters() if p.requires_grad)\n",
        "\n",
        "print('\\n','#' * 48,f'\\n # The model has {count_parameters(model):,}' \\\n",
        "       ' trainable parameters #\\n', '#' * 48,'\\n' )  \n",
        "# model"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "\n",
            " ################################################ \n",
            " # The model has 947,934 trainable parameters #\n",
            " ################################################ \n",
            "\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "yvJtwUD9so_9"
      },
      "source": [
        "def train(model, device, train_loader, loss_fn, optimizer):\n",
        "    model.train()\n",
        "    loss_train = []\n",
        "    for data, x_nan, target in train_loader:\n",
        "        data, x_nan, target = data.to(device), x_nan.to(device), target.to(device)\n",
        "        optimizer.zero_grad()\n",
        "        output = model(data, x_nan)\n",
        "        loss = loss_fn(output, target)\n",
        "        loss_train.append(loss.item())\n",
        "    \n",
        "        loss.backward()\n",
        "        optimizer.step()\n",
        "  \n",
        "    ave_train_loss = sum(loss_train) / len(loss_train)\n",
        "    return ave_train_loss\n",
        "\n",
        "def test(model, device, valid_loader, loss_fn):\n",
        "    model.eval()\n",
        "    preds, trues, loss_test = [],[],[]\n",
        "    for data, x_nan, target in valid_loader:\n",
        "        with torch.no_grad():\n",
        "            data, x_nan, target = data.to(device), x_nan.to(device), target.to(device)\n",
        "            output = model(data, x_nan)\n",
        "\n",
        "        loss = loss_fn(output, target)\n",
        "        loss_test.append(loss.item())\n",
        "      \n",
        "        preds += output.argmax(-1).cpu().numpy().tolist()\n",
        "        trues += target.cpu().numpy().tolist()\n",
        "    \n",
        "    ave_test_loss = sum(loss_test) / len(loss_test)\n",
        "    \n",
        "    preds = np.array(preds)\n",
        "    trues = np.array(trues)\n",
        "    acc = (1.* (preds==trues)).mean() \n",
        "    return ave_test_loss, acc"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "MmEdJmcOraR5"
      },
      "source": [
        "### LabelSmoothing"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "f6YOw8sxrZnp"
      },
      "source": [
        "# Método de regularização\n",
        "class LabelSmoothing(nn.Module):\n",
        "  def __init__(self, smoothing = 0.1):\n",
        "    assert 0.0 <= smoothing < 1.0, f\"got smoothing={smoothing}\"\n",
        "    super(LabelSmoothing, self).__init__()\n",
        "    self.smoothing = smoothing\n",
        "\n",
        "  def forward(self, input, target):\n",
        "    target = target.detach()\n",
        "    log = F.log_softmax(input, dim=1)\n",
        "    oh = torch.ones_like(input).detach_()*(self.smoothing/input.size(1))\n",
        "    oh = oh.scatter_(1, target.unsqueeze(1),\n",
        "                      (1.0 - self.smoothing + self.smoothing/input.size(1)))\n",
        "    return ((-oh*log).sum(dim=1).mean())"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "-6Jl7gDTsXlc"
      },
      "source": [
        "#  <span style=\"color:orange\">  Treino no dataloder debug (poucas amostras) - check de convergência\n",
        "\n",
        "\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "mp6sGapiPMFL",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 272
        },
        "outputId": "aea01c10-7efe-4639-8c6f-eb0405b8869d"
      },
      "source": [
        "start = torch.cuda.Event(enable_timing=True)\n",
        "end = torch.cuda.Event(enable_timing=True)\n",
        "\n",
        "deterministic() \n",
        "N_EPOCHS  = 50\n",
        "D     = 256\n",
        "H     = 8\n",
        "DROP  = 0.2\n",
        "OUT   = 2\n",
        "model = Net(cat_dims, D, H, DROP, OUT).to(device)\n",
        "\n",
        "criterion = nn.CrossEntropyLoss()\n",
        "optimizer = optim.AdaBound(model.parameters())\n",
        "\n",
        "epoch, best_acc = 0,0\n",
        "\n",
        "start.record()\n",
        "for step in range(1, N_EPOCHS+1):\n",
        "    loss_t      = train(model, device, dataloaders['debug'], criterion, optimizer)\n",
        "    _,      acc = test(model, device, dataloaders['debug'], criterion)\n",
        "    if acc >= best_acc:\n",
        "        best_acc = acc\n",
        "        epoch    = step\n",
        "    if step == 1:\n",
        "        print(f'[Epoch [{step}/{N_EPOCHS}] |', end=' ')\n",
        "        print(f'Train Loss: {loss_t:.3f} -- Acc: {acc:.3f}')\n",
        "\n",
        "    if step % 10 == 0:\n",
        "        print(f'[Epoch [{step}/{N_EPOCHS}] |', end=' ')\n",
        "        print(f'Train Loss: {loss_t:.3f} -- Acc: {acc:.3f}')\n",
        "\n",
        "end.record()\n",
        "torch.cuda.synchronize()    \n",
        "\n",
        "print(f'\\nEnd of training. Best Acc: -->[{best_acc:.4f}]<--, at epoch: {epoch}\\n')\n",
        "print(f'Tempo: {start.elapsed_time(end)/1000/60 :.3f} min.')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Deterministic experiment, seed: 0\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "stream",
          "text": [
            "/pytorch/torch/csrc/utils/python_arg_parser.cpp:756: UserWarning: This overload of add_ is deprecated:\n",
            "\tadd_(Number alpha, Tensor other)\n",
            "Consider using one of the following signatures instead:\n",
            "\tadd_(Tensor other, *, Number alpha)\n"
          ],
          "name": "stderr"
        },
        {
          "output_type": "stream",
          "text": [
            "[Epoch [1/50] | Train Loss: 0.580 -- Acc: 0.762\n",
            "[Epoch [10/50] | Train Loss: 0.294 -- Acc: 0.884\n",
            "[Epoch [20/50] | Train Loss: 0.251 -- Acc: 0.866\n",
            "[Epoch [30/50] | Train Loss: 0.255 -- Acc: 0.893\n",
            "[Epoch [40/50] | Train Loss: 0.194 -- Acc: 0.930\n",
            "[Epoch [50/50] | Train Loss: 0.158 -- Acc: 0.941\n",
            "\n",
            "End of training. Best Acc: -->[0.9460]<--, at epoch: 48\n",
            "\n",
            "Tempo: 0.621 min.\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Uzzv2cdH7AqT"
      },
      "source": [
        "# Medindo o tempo com torch.cuda.Event"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "T0g976s-7DvC",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 51
        },
        "outputId": "4b34df6b-8d44-457c-d717-a3f93bb1e676"
      },
      "source": [
        "# Tempos gasto no treino\n",
        "len_train = len(dataloaders['train'])*BATCH_SZ\n",
        "len_valid = len(dataloaders['valid'])*BATCH_SZ\n",
        "start = torch.cuda.Event(enable_timing=True)\n",
        "end = torch.cuda.Event(enable_timing=True)\n",
        "start.record()\n",
        "_ = train(model, device, dataloaders['train'], criterion, optimizer)\n",
        "end.record()\n",
        "torch.cuda.synchronize()    \n",
        "\n",
        "print(f'Tempo no treino em {len_train} exemplos: {start.elapsed_time(end)/1000 :.3f} seg.')\n",
        "\n",
        "# Tempos gasto no teste\n",
        "start.record()\n",
        "_, _ =  test(model, device, dataloaders['valid'], criterion)\n",
        "end.record()\n",
        "torch.cuda.synchronize()    \n",
        "\n",
        "print(f'Tempo no teste em {len_valid} exemplos: {start.elapsed_time(end)/1000 :.3f} seg.')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Tempo no treino em 26048 exemplos: 8.663 seg.\n",
            "Tempo no teste em 6528 exemplos: 1.149 seg.\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ATptaOzQDq79"
      },
      "source": [
        "### Calculando a memória"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "57Fg24TzCLAT",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "54a0edc6-ec40-4376-c8f7-22305bd4482e"
      },
      "source": [
        "!pip install -q memory_profiler\n",
        "%load_ext memory_profiler"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "  Building wheel for memory-profiler (setup.py) ... \u001b[?25l\u001b[?25hdone\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "fS6X4TVoDKYr",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "57047ae5-b090-458a-a7ed-95931dd347e0"
      },
      "source": [
        "%memit train(model, device, dataloaders['train'], criterion, optimizer)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "peak memory: 1805.80 MiB, increment: 0.04 MiB\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "jkaeBN5XDxJA",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "148c190d-0ab6-4c39-8822-eca154b08174"
      },
      "source": [
        "%memit test(model, device, dataloaders['valid'], criterion)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "peak memory: 1805.83 MiB, increment: 0.07 MiB\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "8-ZfIleCESXH"
      },
      "source": [
        "# <span style=\"color:orange\"> Treino completo"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "IJr14D5p_H74",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 782
        },
        "outputId": "85f6ef2d-1573-45ac-db42-0fd6e82f5cb7"
      },
      "source": [
        "start = torch.cuda.Event(enable_timing=True)\n",
        "end = torch.cuda.Event(enable_timing=True)\n",
        "\n",
        "deterministic() \n",
        "N_EPOCHS  = 40\n",
        "D     = 256\n",
        "H     = 16\n",
        "DROP  = 0.2\n",
        "OUT   = 2\n",
        "model = Net(cat_dims, D, H, DROP, OUT).to(device)\n",
        "\n",
        "# dim_ = 32\n",
        "# trans_layer = nn.TransformerEncoderLayer(d_model=dim_, \n",
        "#                                          nhead=8, \n",
        "#                                          dim_feedforward=512, \n",
        "#                                          dropout=0.2, \n",
        "#                                          activation='gelu')\n",
        "\n",
        "# model = TransModel(cat_dims,num_layers=2 , out=2).to(device)\n",
        "\n",
        "\n",
        "# criterion = LabelSmoothing()\n",
        "\n",
        "criterion = nn.CrossEntropyLoss()\n",
        "optimizer = optim.AdaBound(model.parameters())\n",
        "\n",
        "\n",
        "epoch, best_acc = 0,0\n",
        "\n",
        "start.record()\n",
        "for step in range(1, N_EPOCHS+1):\n",
        "    loss_t      = train(model, device, dataloaders['train'], criterion, optimizer)\n",
        "    loss_v, acc =  test(model, device, dataloaders['valid'],   criterion)\n",
        "    if acc >= best_acc:\n",
        "        best_acc = acc\n",
        "        epoch    = step\n",
        "        torch.save(model.state_dict(), os.path.join('', 'epoch-{}.pth'.format(epoch)))\n",
        "    print(f'[Epoch [{step}/{N_EPOCHS}] |', end=' ')\n",
        "    print(f'Train Loss: {loss_t:.3f} -- Valid Loss: {loss_v:.3f} --',end=' ')\n",
        "    print(f'Acc: {acc:.3f}')\n",
        "\n",
        "end.record()\n",
        "torch.cuda.synchronize()    \n",
        "\n",
        "print(f'\\nEnd of training. Best Acc: -->[{best_acc:.4f}]<--, at epoch: {epoch}\\n')\n",
        "print(f'Tempo: {start.elapsed_time(end)/1000/60 :.3f} min.')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Deterministic experiment, seed: 0\n",
            "[Epoch [1/40] | Train Loss: 0.391 -- Valid Loss: 0.344 -- Acc: 0.840\n",
            "[Epoch [2/40] | Train Loss: 0.337 -- Valid Loss: 0.342 -- Acc: 0.841\n",
            "[Epoch [3/40] | Train Loss: 0.335 -- Valid Loss: 0.335 -- Acc: 0.844\n",
            "[Epoch [4/40] | Train Loss: 0.329 -- Valid Loss: 0.335 -- Acc: 0.842\n",
            "[Epoch [5/40] | Train Loss: 0.328 -- Valid Loss: 0.336 -- Acc: 0.845\n",
            "[Epoch [6/40] | Train Loss: 0.327 -- Valid Loss: 0.339 -- Acc: 0.841\n",
            "[Epoch [7/40] | Train Loss: 0.325 -- Valid Loss: 0.333 -- Acc: 0.844\n",
            "[Epoch [8/40] | Train Loss: 0.324 -- Valid Loss: 0.330 -- Acc: 0.842\n",
            "[Epoch [9/40] | Train Loss: 0.325 -- Valid Loss: 0.329 -- Acc: 0.845\n",
            "[Epoch [10/40] | Train Loss: 0.322 -- Valid Loss: 0.329 -- Acc: 0.844\n",
            "[Epoch [11/40] | Train Loss: 0.322 -- Valid Loss: 0.331 -- Acc: 0.844\n",
            "[Epoch [12/40] | Train Loss: 0.324 -- Valid Loss: 0.329 -- Acc: 0.845\n",
            "[Epoch [13/40] | Train Loss: 0.320 -- Valid Loss: 0.328 -- Acc: 0.848\n",
            "[Epoch [14/40] | Train Loss: 0.319 -- Valid Loss: 0.329 -- Acc: 0.848\n",
            "[Epoch [15/40] | Train Loss: 0.319 -- Valid Loss: 0.330 -- Acc: 0.845\n",
            "[Epoch [16/40] | Train Loss: 0.318 -- Valid Loss: 0.327 -- Acc: 0.846\n",
            "[Epoch [17/40] | Train Loss: 0.319 -- Valid Loss: 0.329 -- Acc: 0.846\n",
            "[Epoch [18/40] | Train Loss: 0.317 -- Valid Loss: 0.329 -- Acc: 0.847\n",
            "[Epoch [19/40] | Train Loss: 0.318 -- Valid Loss: 0.327 -- Acc: 0.847\n",
            "[Epoch [20/40] | Train Loss: 0.318 -- Valid Loss: 0.329 -- Acc: 0.846\n",
            "[Epoch [21/40] | Train Loss: 0.316 -- Valid Loss: 0.328 -- Acc: 0.846\n",
            "[Epoch [22/40] | Train Loss: 0.315 -- Valid Loss: 0.330 -- Acc: 0.847\n",
            "[Epoch [23/40] | Train Loss: 0.315 -- Valid Loss: 0.327 -- Acc: 0.847\n",
            "[Epoch [24/40] | Train Loss: 0.315 -- Valid Loss: 0.330 -- Acc: 0.848\n",
            "[Epoch [25/40] | Train Loss: 0.314 -- Valid Loss: 0.331 -- Acc: 0.845\n",
            "[Epoch [26/40] | Train Loss: 0.315 -- Valid Loss: 0.329 -- Acc: 0.847\n",
            "[Epoch [27/40] | Train Loss: 0.315 -- Valid Loss: 0.330 -- Acc: 0.847\n",
            "[Epoch [28/40] | Train Loss: 0.314 -- Valid Loss: 0.327 -- Acc: 0.847\n",
            "[Epoch [29/40] | Train Loss: 0.312 -- Valid Loss: 0.329 -- Acc: 0.848\n",
            "[Epoch [30/40] | Train Loss: 0.312 -- Valid Loss: 0.332 -- Acc: 0.846\n",
            "[Epoch [31/40] | Train Loss: 0.311 -- Valid Loss: 0.335 -- Acc: 0.848\n",
            "[Epoch [32/40] | Train Loss: 0.311 -- Valid Loss: 0.334 -- Acc: 0.846\n",
            "[Epoch [33/40] | Train Loss: 0.310 -- Valid Loss: 0.332 -- Acc: 0.848\n",
            "[Epoch [34/40] | Train Loss: 0.310 -- Valid Loss: 0.331 -- Acc: 0.846\n",
            "[Epoch [35/40] | Train Loss: 0.311 -- Valid Loss: 0.336 -- Acc: 0.847\n",
            "[Epoch [36/40] | Train Loss: 0.309 -- Valid Loss: 0.334 -- Acc: 0.846\n",
            "[Epoch [37/40] | Train Loss: 0.309 -- Valid Loss: 0.337 -- Acc: 0.848\n",
            "[Epoch [38/40] | Train Loss: 0.308 -- Valid Loss: 0.335 -- Acc: 0.845\n",
            "[Epoch [39/40] | Train Loss: 0.308 -- Valid Loss: 0.334 -- Acc: 0.845\n",
            "[Epoch [40/40] | Train Loss: 0.307 -- Valid Loss: 0.332 -- Acc: 0.846\n",
            "\n",
            "End of training. Best Acc: -->[0.8485]<--, at epoch: 24\n",
            "\n",
            "Tempo: 3.132 min.\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "039Qjhr6LUG_"
      },
      "source": [
        "#  <span style=\"color:orange\"> Validação no conj. de teste\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "HlIw4vY5drmf"
      },
      "source": [
        "def logits_and_trues(model, dataloader, device=device):\n",
        "  model.eval()\n",
        "  logits, trues = [],[]\n",
        "  for x, x_nan, y in dataloader:\n",
        "    with torch.no_grad():\n",
        "      y_pred = model(x.to(device), x_nan.to(device))\n",
        "    logits.extend(y_pred.to('cpu').numpy().tolist())\n",
        "    trues.extend(y.to('cpu').numpy().tolist())\n",
        "  return torch.tensor(logits), np.array(trues)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "fY_wyc-UiC4x",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "c133e5d2-2f01-4f23-aa52-535e26f83605"
      },
      "source": [
        "from sklearn.metrics import accuracy_score\n",
        "\n",
        "model = Net(cat_dims, D, H, DROP, OUT).to(device)\n",
        "# model = TransModel(cat_dims,num_layers=2 , out=2).to(device)\n",
        "\n",
        "string = '/content/epoch-24.pth'\n",
        "model.load_state_dict(torch.load(string, map_location=device))\n",
        "\n",
        "logits, trues = logits_and_trues(model, dataloaders['test'], device=device)\n",
        "prob  = F.softmax(logits, dim=1).detach().numpy()\n",
        "probs = prob[:,1]\n",
        "y      = pd.Series(trues)\n",
        "y_pred = pd.Series(probs)\n",
        "print(f'Acc = {accuracy_score(y, y_pred.round())}')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Acc = 0.8504391622136233\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "81gX6QFOiC5M",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 435
        },
        "outputId": "49e0ae7c-b8e6-4a30-cc30-7d6e4341d015"
      },
      "source": [
        "from sklearn.metrics import roc_curve\n",
        "from sklearn.metrics import auc\n",
        "\n",
        "def plot_roc_curve(y, y_pred):\n",
        "  fpr, tpr,_ = roc_curve(y, y_pred)\n",
        "  roc_auc    = auc(fpr, tpr)\n",
        "  acc        = accuracy_score(y, y_pred.round())\n",
        "  \n",
        "  # plot\n",
        "  fig = plt.figure()\n",
        "  plt.figure(figsize=(15, 6))\n",
        "  plt.plot(fpr,tpr,'b',lw=2,label='%s AUC=%0.3f -- ACC=%0.3f'%('Modelo: Emb. Pos. Sinusoidal:',roc_auc, acc))\n",
        "  plt.plot([0, 1], [0, 1], 'm--')\n",
        "  \n",
        "  # decoração\n",
        "  plt.xlim([0.0, 1.0])\n",
        "  plt.xticks(fontsize=16)\n",
        "  plt.yticks(fontsize=16)\n",
        "  plt.ylim([0.0, 1.0])\n",
        "  plt.title('Encoder - MultiHeadAttention', size=18)\n",
        "  plt.xlabel('False Positive Rate', size=16)\n",
        "  plt.ylabel('True Positive Rate', size=16)\n",
        "  plt.legend(loc=0, fontsize=16)\n",
        "  plt.show()\n",
        "\n",
        "plot_roc_curve(y, y_pred)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 0 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 1080x432 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "OS4P6Pg6g-DX"
      },
      "source": [
        "# <span style=\"color:purple\">Fim do notebook\n",
        "\n",
        "\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "bxXZGebzfC5v"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "avk39os_fC2l"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "swhCMOogfCw1"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "7QFqxozsfCuS"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "d0unxkGIfCrb"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "3LAJDFDdfCpI"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "EezMnnCmfCmw"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "CsWVlsFSfCiy"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "_RpChkjlfCgH"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "gRBzF0WdfCdx"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "cSjsQy30fCbN"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Ar_v6kBLabj_"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "HwVzN3O_abh1"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "IUQzAI9Jabf5"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "OjVYEQfSabc0"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "OXr7aUZVabaa"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    }
  ]
}