Exp_06_Lab_Machine Lerning.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 5,
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.6.12"
    },
    "colab": {
      "name": "Exp. 06 - Lab. Introdução ao Processamento Digital de Imagens e Visão Computacional sem Resp.ipynb",
      "provenance": [],
      "collapsed_sections": []
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "P_YPhGn6DqRA"
      },
      "source": [
        "<img src=\"https://www.anped.org.br/sites/default/files/images/ufcg-lateral.png\" width=\"780\" height=\"240\" align=\"center\"/>"
      ],
      "id": "P_YPhGn6DqRA"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "3y0RikX2DqRE"
      },
      "source": [
        "## Centro de Engenharia Elétrica e Informática\n",
        "## Departamento de Engenharia Elétrica\n",
        "## Disciplina: Int. ao Processamento de Imagem Digital e Visão Computacional\n",
        "## Professora: Luciana Veloso\n",
        "## Aluno(a): William Henrique A. Martins"
      ],
      "id": "3y0RikX2DqRE"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "iD7o7jYQDqRF"
      },
      "source": [
        "# Experimento 06: Redes Neurais Artificiais"
      ],
      "id": "iD7o7jYQDqRF"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "xd7HHr2JDqRF"
      },
      "source": [
        "import os                                             # Operational System para manipulação de arquivos.\n",
        "import cv2                                            # OpenCV para manipulação de imagens.\n",
        "import numpy as np                                    # Numpy para manipulação de matrizes e arrays.\n",
        "import matplotlib.pyplot as plt                       # Pyplot para plotagem de gráficos e imagens.\n",
        "from sklearn.model_selection import train_test_split  # Função para particionamento dos dados\n",
        "\n",
        "from tensorflow.keras.models import Sequential        # Classe de modelos sequenciais para construir as redes neurais.\n",
        "from tensorflow.keras.layers import Dense, Input      # Camada de neurônios densamente conectados.\n",
        "from tensorflow.keras.optimizers import SGD           # Otimizador \"Descida do Gradiente com Momento\".\n",
        "from tensorflow.keras.datasets import boston_housing  # Dataset utilizado nesse experimento."
      ],
      "id": "xd7HHr2JDqRF",
      "execution_count": 1,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9DABnOD-DqRG"
      },
      "source": [
        "## 1. Banco de Dados\n",
        "\n",
        "* Vamos utilizar um banco de dados do catálogo de datasets do Keras, que é disponibilizado como uma função pronta;\n",
        "\n",
        "* Cada instância do banco de dados corresponde a um conjunto de 13 valores referentes a características de subúrbios de Boston na década de 1970, a exemplo de taxa de crimes, imposto sobre propriedade, etc;\n",
        "\n",
        "* Esses valores serão utilizados para calcular o valor mediano das residências no respectivo subúrbio em um problema de regressão, de modo que iremos mapear um vetor de entrada com 13 valores em um vetor de saída com 1 único elemento;\n",
        "\n",
        "* O banco de dados contém 506 valores divididos em 404 instâncias de treino e 102 de teste;\n",
        "\n",
        "* Uma descrição mais detalhada dos valores de entrada pode ser vista em http://lib.stat.cmu.edu/datasets/boston, onde os valores são descritos na ordem que aparecem;"
      ],
      "id": "9DABnOD-DqRG"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "NKNQBIRGDqRH"
      },
      "source": [
        "### 1.1 Organização do banco de dados\n",
        "\n",
        "* Comumente os dados utilizados no treinamento de modelos de inteligência artificial reservam a primeira dimensão para controlar a amostra e espalham o tipo de dado utilizado nas demais dimensões do tensor. \n",
        "\n",
        "\n",
        "* Nesse caso, os nossos dados são vetores de características unidimensionais (1D), de modo que são organizados em tensores bidimensionais (2D) com formato: **dados.shape = (amostras, características)**\n",
        "    * O i-ésimo exemplo pode ser acessado a partir de: **exemplo = dados[i]**\n",
        "    * A j-ésima característica dos exemplos pode ser acessada a partir de: **caracteristica = dados[:, j]**\n",
        "    * Porções do tensor podem ser acessadas utilizando fatiamento, por exemplo:\n",
        "        * primeiros_5_exemplos = dados[:5]\n",
        "        * caracteristicas_9a13 = dados[8:13]\n",
        "        \n",
        "\n",
        "* Lembrem-se que a contagem de índices em Python começa em 0 e só é inclusiva no primeiro elemento:\n",
        "    * **:5** produz os índices **0, 1, 2, 3, 4** \n",
        "    * **8:13** produz os índices **8, 9, 10, 11, 12** \n",
        "        "
      ],
      "id": "NKNQBIRGDqRH"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "648vZLTsDqRI"
      },
      "source": [
        "* O banco de dados pode ser carregado utilizando:\n",
        "    * **(train_data, train_targets), (test_data, test_targets) = boston_housing.load_data()**\n",
        "    * train_data é um tensor com as entradas do conjunto de treino;\n",
        "    * test_data é um tensor com as entradas do conjunto de teste;\n",
        "    * train_targets é um tensor com os gabaritos do conjunto de treino;\n",
        "    * test_targets é um tensor com os gabaritos do conjunto de teste;"
      ],
      "id": "648vZLTsDqRI"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "7bG5dAZwDqRI"
      },
      "source": [
        "**a. O código abaixo carrega o banco de dados, acesse alguns exemplos de treinamento e de teste e veja suas dimensões.**\n",
        "\n",
        "* Use índices entre **[0, 404)** para os dados de treino e entre **[0, 102)** para os de teste."
      ],
      "id": "7bG5dAZwDqRI"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ZGIplvfqDqRJ",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "b2bd0ce7-b9d2-4f7d-cb32-1d14d973b7da"
      },
      "source": [
        "(train_data, train_targets), (test_data, test_targets) = boston_housing.load_data()"
      ],
      "id": "ZGIplvfqDqRJ",
      "execution_count": 2,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/boston_housing.npz\n",
            "57344/57026 [==============================] - 0s 0us/step\n",
            "65536/57026 [==================================] - 0s 0us/step\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ePv4PGGODqRJ"
      },
      "source": [
        "**b. Verifique as dimensões e as faixas de valores de cada característica nos vetores de entrada dos conjuntos de treino e de teste. A faixa de valores das características são semelhantes? E os dados de treino e de teste?**\n",
        "\n",
        "* Encontre os valores mínimo/máximo além da média (np.mean) e da variância (np.var) de cada uma das 13 características.\n",
        "* Sugestão: leia sobre o parâmetro \"axis\" na documentação das funções np.min, np.max e np.mean."
      ],
      "id": "ePv4PGGODqRJ"
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "0_0cQJcjDqRJ",
        "outputId": "92fd0d30-90f9-4b88-ae09-eb8cc2466b65"
      },
      "source": [
        "train_data.shape\n",
        "# 404 instâncias\n",
        "# 13 caractéristicas\n",
        "test_data.shape\n",
        "# 102 instâncias\n",
        "# 13 caractéristicas\n",
        "\n",
        "print('nº de caratéristicas: \\n\\t Train data: {} \\n\\t Test Data: {} '.format(train_data.shape[1], test_data.shape[1]))\n",
        "\n",
        "\n",
        "min = np.min(train_data, axis = 0)\n",
        "max = np.max(train_data, axis = 0)\n",
        "mean = np.mean(train_data, axis = 0)\n",
        "std = np.std(train_data, axis = 0)\n",
        "print('\\nCONJUNTO DE TREINO\\n')\n",
        "for i in range(0, test_data.shape[1]): \n",
        "  print('{} - Limite: [{:.2f},{:.2f}], com Média {:.2f} e Variância {:.2f}'.format(f'caractéristica {i}', min[i], max[i], mean[i], std[i]))\n",
        "\n",
        "\n",
        "min = np.min(test_data, axis = 0)\n",
        "max = np.max(test_data, axis = 0)\n",
        "mean = np.mean(test_data, axis = 0)\n",
        "std = np.std(test_data, axis = 0)\n",
        "print('\\nCONJUNTO DE TESTE')\n",
        "for i in range(0, test_data.shape[1]): \n",
        "  print('{} - Limite: [{:.2f},{:.2f}], com Média {:.2f} e Variância {:.2f}'.format(f'caractéristica {i}', min[i], max[i], mean[i], std[i]))"
      ],
      "id": "0_0cQJcjDqRJ",
      "execution_count": 3,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "nº de caratéristicas: \n",
            "\t Train data: 13 \n",
            "\t Test Data: 13 \n",
            "\n",
            "CONJUNTO DE TREINO\n",
            "\n",
            "caractéristica 0 - Limite: [0.01,88.98], com Média 3.75 e Variância 9.23\n",
            "caractéristica 1 - Limite: [0.00,100.00], com Média 11.48 e Variância 23.74\n",
            "caractéristica 2 - Limite: [0.46,27.74], com Média 11.10 e Variância 6.80\n",
            "caractéristica 3 - Limite: [0.00,1.00], com Média 0.06 e Variância 0.24\n",
            "caractéristica 4 - Limite: [0.39,0.87], com Média 0.56 e Variância 0.12\n",
            "caractéristica 5 - Limite: [3.56,8.72], com Média 6.27 e Variância 0.71\n",
            "caractéristica 6 - Limite: [2.90,100.00], com Média 69.01 e Variância 27.91\n",
            "caractéristica 7 - Limite: [1.13,10.71], com Média 3.74 e Variância 2.03\n",
            "caractéristica 8 - Limite: [1.00,24.00], com Média 9.44 e Variância 8.69\n",
            "caractéristica 9 - Limite: [188.00,711.00], com Média 405.90 e Variância 166.17\n",
            "caractéristica 10 - Limite: [12.60,22.00], com Média 18.48 e Variância 2.20\n",
            "caractéristica 11 - Limite: [0.32,396.90], com Média 354.78 e Variância 93.99\n",
            "caractéristica 12 - Limite: [1.73,37.97], com Média 12.74 e Variância 7.25\n",
            "\n",
            "CONJUNTO DE TESTE\n",
            "caractéristica 0 - Limite: [0.01,25.05], com Média 3.09 e Variância 5.35\n",
            "caractéristica 1 - Limite: [0.00,90.00], com Média 10.90 e Variância 21.47\n",
            "caractéristica 2 - Limite: [1.22,27.74], com Média 11.26 e Variância 7.05\n",
            "caractéristica 3 - Limite: [0.00,1.00], com Média 0.10 e Variância 0.30\n",
            "caractéristica 4 - Limite: [0.39,0.87], com Média 0.54 e Variância 0.11\n",
            "caractéristica 5 - Limite: [4.88,8.78], com Média 6.35 e Variância 0.67\n",
            "caractéristica 6 - Limite: [6.00,100.00], com Média 66.85 e Variância 28.89\n",
            "caractéristica 7 - Limite: [1.47,12.13], com Média 4.01 e Variância 2.37\n",
            "caractéristica 8 - Limite: [1.00,24.00], com Média 9.98 e Variância 8.73\n",
            "caractéristica 9 - Limite: [187.00,711.00], com Média 417.50 e Variância 176.52\n",
            "caractéristica 10 - Limite: [13.00,21.20], com Média 18.37 e Variância 2.02\n",
            "caractéristica 11 - Limite: [24.65,396.90], com Média 364.16 e Variância 78.75\n",
            "caractéristica 12 - Limite: [1.92,31.99], com Média 12.31 e Variância 6.66\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Cxgy8FoJbKTp"
      },
      "source": [
        "<font color=yellow>\n",
        "Os conjuntos são divididos randomicamente, mas percebe-se uma certa concordancia entre os dados dos conjuntos de treino  e teste. Ou seja, as faixas de valores distoa um pouco para algumas caractéristicas, mas possuem outros parâmentros concordantes.\n",
        "</font>"
      ],
      "id": "Cxgy8FoJbKTp"
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "-6RlfkvTbcIZ",
        "outputId": "d5908c23-cdcf-4c58-c071-1e286b59202d"
      },
      "source": [
        "\n",
        "min = np.min(train_targets)\n",
        "print(min)"
      ],
      "id": "-6RlfkvTbcIZ",
      "execution_count": 4,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "5.0\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "D2puf8XTDqRK"
      },
      "source": [
        "**c. Verifique as dimensões e as faixas de valores dos gabaritos dos conjuntos de treino e de teste. Os valores encontrados são semelhantes?**"
      ],
      "id": "D2puf8XTDqRK"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "q5txP0WBDqRK",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "2eac8b0e-a318-4689-e018-bb5dd0ce03b3"
      },
      "source": [
        "print('nº de caratéristicas: \\n\\t Train data: {} \\n\\t Test Data: {} '.format(len(train_targets), len(test_targets)))\n",
        "\n",
        "min = np.min(train_targets, axis = 0)\n",
        "max = np.max(train_targets, axis = 0)\n",
        "mean = np.mean(train_targets, axis = 0)\n",
        "std = np.std(train_targets, axis = 0)\n",
        "print('\\nGABARITO DE TREINO\\n')\n",
        "print('{} - Limite: [{:.2f},{:.2f}], com Média {:.2f} e Variância {:.2f}'.format('train_targets', min, max, mean, std))\n",
        "\n",
        "\n",
        "min = np.min(test_targets, axis = 0)\n",
        "max = np.max(test_targets, axis = 0)\n",
        "mean = np.mean(test_targets, axis = 0)\n",
        "std = np.std(test_targets, axis = 0)\n",
        "print('\\GABARITO DE TESTE')\n",
        "print('{} - Limite: [{:.2f},{:.2f}], com Média {:.2f} e Variância {:.2f}'.format('test_targets', min, max, mean, std))"
      ],
      "id": "q5txP0WBDqRK",
      "execution_count": 5,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "nº de caratéristicas: \n",
            "\t Train data: 404 \n",
            "\t Test Data: 102 \n",
            "\n",
            "GABARITO DE TREINO\n",
            "\n",
            "train_targets - Limite: [5.00,50.00], com Média 22.40 e Variância 9.20\n",
            "\\GABARITO DE TESTE\n",
            "test_targets - Limite: [5.60,50.00], com Média 23.08 e Variância 9.12\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "NbeBq7PPdFCx"
      },
      "source": [
        "<font color='yellow'>\n",
        "Os números de instâncias dos data target e do train data são iguais a seus respectivos conjuntos de treino\n",
        "</font>"
      ],
      "id": "NbeBq7PPdFCx"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "mhBtLK3FDqRK"
      },
      "source": [
        "### 1.2 Pre-processamento dos dados\n",
        "\n",
        "* Dados cujas características assumem diferentes faixas de valores muitas vezes são problemáticos para o aprendizado dos modelos e podem reduzir a velocidade de convergência ou até mesmo limitar as capacidades do modelo final.\n",
        "\n",
        "* Nesse sentido, uma prática comum é a normalização dos dados antes do treinamento, que geralmente é feito por característica em forma da subtração da média e divisão pelo desvio padrão, o que faz com que os dados resultantes tenham média 0 e variância 1.\n",
        "\n",
        "* Um ponto muito importante é que a normalização deve ser feita partir dos mesmos valores em todos os conjuntos, ou seja, os dados são normalizados segundo informações do conjunto de treino.\n",
        "\n",
        "* Na prática não temos como calcular a média e variância real, mas se os dados de treino são significativos os seus valores são suficientes."
      ],
      "id": "mhBtLK3FDqRK"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "aRtJdTRQDqRL"
      },
      "source": [
        "# Calcula a média do conjunto de treino\n",
        "mean = train_data.mean(axis=0)\n",
        "\n",
        "# Calcula o desvio padrão do conjunto de treino\n",
        "std = train_data.std(axis=0)\n",
        "\n",
        "# Normaliza os dados de treino\n",
        "train_data -= mean\n",
        "train_data /= std\n",
        "\n",
        "# Normaliza os dados de teste\n",
        "test_data -= mean\n",
        "test_data /= std"
      ],
      "id": "aRtJdTRQDqRL",
      "execution_count": 6,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "GTQS4AzrDqRL"
      },
      "source": [
        "**d. Repita o item b para os dados normalizados. O que se observa quantos aos valores de média e variância para os dados de treino e teste? Comente a sua interpretação sobre as diferenças observadas.**"
      ],
      "id": "GTQS4AzrDqRL"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "-ppS_7EADqRL"
      },
      "source": [
        "nmean = train_data.mean(axis = 0)\n",
        "nstd = train_data.std(axis = 0)"
      ],
      "id": "-ppS_7EADqRL",
      "execution_count": 7,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "EYa9f4XDDqRL"
      },
      "source": [
        "Agora vamos criar uma partição de validação a partir do conjunto de treino para realizar uma validação cruzada. Vamos utilizar a função **train_test_split**, que separa dados e os seus respectivos gabaritos segundo uma fração especificada."
      ],
      "id": "EYa9f4XDDqRL"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "kAqrcD5bDqRL",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "bda683ff-e213-4eea-9b71-f5e338aade89"
      },
      "source": [
        "# Fração escolhida para separar o mesmo número de instâncias do conjunto de testes\n",
        "data_frac = test_data.shape[0] / train_data.shape[0]\n",
        "\n",
        "# Criação do conjunto de validação\n",
        "train_data, val_data, train_targets, val_targets = train_test_split(train_data,             # Dados de treino\n",
        "                                                                    train_targets,          # Gabaritos de treino\n",
        "                                                                    test_size = data_frac,  # Proporção de dados p/ validação\n",
        "                                                                    random_state=42)        # Semente para garantir consistencia nos resultados\n",
        "\n",
        "print(\"treino:\", train_data.shape, train_targets.shape)\n",
        "print(\"validação:\", val_data.shape, val_targets.shape)"
      ],
      "id": "kAqrcD5bDqRL",
      "execution_count": 8,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "treino: (302, 13) (302,)\n",
            "validação: (102, 13) (102,)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Wp15lAZeDqRM"
      },
      "source": [
        "### 1.2 Construindo o modelo\n",
        "\n",
        "Para construir o modelo usaremos a classe **Sequential**, que possibilita a construção de modelos sequenciais de forma bastante simples.\n",
        "* A construção do modelo é feita a partir do seu instanciamento como objeto da classe seguido de chamadas à função **add()** para adicionar camadas.\n",
        "* Como estamos construindo apenas Redes Neurais Artificiais por enquanto, vamos utilizar apenas as camadas **Dense** e **Input**.\n",
        "    * A camada Input cria a entrada da rede com **Input( shape = None )**\n",
        "        * shape corresponde ao formato do tensor de entrada, no nosso caso será o número de características do nosso banco de dados (13);\n",
        "    * A camada Dense pode ser chamada com **Dense( n_unidades, activation = 'linear' )**\n",
        "        * n_unidades corresponde ao número de neurônios da camada;\n",
        "        * activation corresponde à função de ativação utilizada na camada;\n",
        "* Algumas funções de ativação disponíveis são:\n",
        "    * \"linear\"\n",
        "    * \"relu\"\n",
        "    * \"sigmoid\"\n",
        "    * \"softmax\"\n",
        "    * \"tanh\"\n",
        "\n",
        "\n",
        "* Mais informações sobre a camada dense podem ser vistas em **https://keras.io/api/layers/core_layers/dense/**\n",
        "* Mais informações sobre as ativações disponíveis podem ser vistas em **https://keras.io/api/layers/activations/**"
      ],
      "id": "Wp15lAZeDqRM"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "eTZULVYJDqRM"
      },
      "source": [
        "**e. A função abaixo constroi um modelo de rede neural e utiliza a função summary() para apresentar um resumo das informações da rede neural produzida. Comente o que faz cada linha do código.**\n",
        "\n",
        "* Modifique parâmetros como o número de unidades de cada camada e/ou o formato do tensor de entrada e/ou o número de saídas. \n",
        "* Explique como o número de parâmetros de cada camada é calculado."
      ],
      "id": "eTZULVYJDqRM"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Rc0UTacsDqRM",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "87dadbc9-f762-41a6-b9d1-5d307897d9bf"
      },
      "source": [
        "def build_model( n_inputs, n_outputs ):\n",
        "    rede = Sequential() # Cria um modelo sequencial \n",
        "    rede.add( Dense( 64, activation = \"relu\", input_shape = (n_inputs, ) ) ) # adiciona uma camada densa com a função relu com 13 entradas\n",
        "    rede.add( Dense( 64, activation = \"relu\") ) # adiciona uma camada densa com a função relu com entradas referentes as saídas da camada anterior\n",
        "    rede.add( Dense( n_outputs ) ) # adiciona uma camada densa que possuí n saídas e entradas referentes anteriores\n",
        "    return rede\n",
        "\n",
        "model = build_model( 13, 1 ) # monta o modelo\n",
        "model.summary()"
      ],
      "id": "Rc0UTacsDqRM",
      "execution_count": 9,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Model: \"sequential\"\n",
            "_________________________________________________________________\n",
            "Layer (type)                 Output Shape              Param #   \n",
            "=================================================================\n",
            "dense (Dense)                (None, 64)                896       \n",
            "_________________________________________________________________\n",
            "dense_1 (Dense)              (None, 64)                4160      \n",
            "_________________________________________________________________\n",
            "dense_2 (Dense)              (None, 1)                 65        \n",
            "=================================================================\n",
            "Total params: 5,121\n",
            "Trainable params: 5,121\n",
            "Non-trainable params: 0\n",
            "_________________________________________________________________\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "c2PyRVutDqRN"
      },
      "source": [
        "**f. Nesse caso utilizamos uma saída com ativação linear. Qual seria a desvantagem de utilizar esse tipo de ativação nas demais camadas?**"
      ],
      "id": "c2PyRVutDqRN"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "HGkZgL2AfOZR"
      },
      "source": [
        "<font color='yellow'>\n",
        "Se utilizarmos uma função de ativação linear nas demais camadas, o no conjunto seria combinação linear da camada anterior, o que poderia ser representa como uma rede de unica camada.\n",
        "</font>"
      ],
      "id": "HGkZgL2AfOZR"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "v20NhFGFfKaB"
      },
      "source": [
        ""
      ],
      "id": "v20NhFGFfKaB"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ZLzRptFYDqRN"
      },
      "source": [
        "Após a construção do modelo ele deve ser compilado antes que os parâmetros sejam treinados. Isso é feito utilizando a função **compile**:\n",
        "\n",
        "* **model.compile( optimizer = opt, loss = fperdas, metrics = [] )** \n",
        "    * O optimizer é o algoritmo otimizador utilizado no lugar da descida do gradiente, o Keras oferece diversas opções;\n",
        "        * Para treinar a partir da descida do gradiente utilizaremos o SGD (descida do gradiente com momento), mas setaremos esse parâmetro para 0.\n",
        "        * **opt = SGD( learning_rate = taxa_de_aprendizagem, momentum = 0 )**\n",
        "    * A função de perdas pode ser definida a partir do parâmetro loss, como este é um problema de regressão utilizaremos o erro médio quadrático: \n",
        "        * **loss = \"mse\"**\n",
        "    * Podemos passar uma lista de métricas a serem computadas durante o treinamento, nesse caso utilizaremos o erro médio absoluto:\n",
        "        * **metrics = [\"mae\"]**\n",
        "        * Note que estamos passando uma lista com uma única métrica, mas outras poderiam ser adicionadas à lista.\n",
        "         \n",
        "         \n",
        "\n",
        "* Algoritmos otimizadores populares são o Adam (**https://keras.io/api/optimizers/adam/**) e o RMSprop(**https://keras.io/api/optimizers/rmsprop/**)\n",
        "* Mais informações sobre os otimizadores disponíveis podem ser vistas em **https://keras.io/api/optimizers/**\n",
        "* Mais informações sobre as funções de perdas disponíveis podem ser vistas em **https://keras.io/api/losses/**\n",
        "* Mais informações sobre as métricas disponíveis podem ser vistas em **https://keras.io/api/metrics/**\n",
        "\n"
      ],
      "id": "ZLzRptFYDqRN"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "DuHI5EyQDqRN",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "676f0532-4fa7-4eb6-fb5b-1853552178a0"
      },
      "source": [
        "model = build_model( 13, 1 )\n",
        "model.compile( optimizer = SGD(learning_rate=0.001, momentum=0.0), loss = \"mse\", metrics=[\"mae\"] )\n",
        "model.summary()"
      ],
      "id": "DuHI5EyQDqRN",
      "execution_count": 10,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Model: \"sequential_1\"\n",
            "_________________________________________________________________\n",
            "Layer (type)                 Output Shape              Param #   \n",
            "=================================================================\n",
            "dense_3 (Dense)              (None, 64)                896       \n",
            "_________________________________________________________________\n",
            "dense_4 (Dense)              (None, 64)                4160      \n",
            "_________________________________________________________________\n",
            "dense_5 (Dense)              (None, 1)                 65        \n",
            "=================================================================\n",
            "Total params: 5,121\n",
            "Trainable params: 5,121\n",
            "Non-trainable params: 0\n",
            "_________________________________________________________________\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "O4C8MJZKDqRO"
      },
      "source": [
        "### 1.3 Treinando o modelo\n",
        "\n",
        "O treinamento é realizado a partir da função **fit**, que recebe dados de treino e de validação além de hiperparâmetros como o número de épocas e o tamanho dos lotes de dados (batchsize).\n",
        "\n",
        "* **hist = model.fit( x = None, y = None, epochs = 1, batchsize = None, validation_data = None, verbose = \"auto\")**\n",
        "    * x corresponde aos dados de treino;\n",
        "    * y corresponde aos gabaritos de treino;\n",
        "    * epochs corresponde ao número de épocas de treinamento;\n",
        "    * batchsize corresponde ao tamanho dos lotes entregues à rede de cada vez;\n",
        "    * validation_data corresponde a uma tupla ( val_data, val_targets ) com os dados de validação;\n",
        "    * verbose indica como a função deve reportar os resultados:\n",
        "        * 0: modo silencioso, nenhum retorno em formato de texto;\n",
        "        * 1: retorno a cada época e barra de progresso;\n",
        "        * 2: retorno a cada época sem barra de progresso;\n",
        "    * hist é um dicionário de retorno com os valores de loss e das métricas computadas para treino e validação;"
      ],
      "id": "O4C8MJZKDqRO"
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "Gmhr-wwaDqRO",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "6532860c-4566-4efd-87d0-fbf8b0c70713"
      },
      "source": [
        "hist = model.fit( x = train_data, \n",
        "                  y = train_targets, \n",
        "                  epochs = 100, \n",
        "                  batch_size = 32, \n",
        "                  validation_data = ( val_data, val_targets ), \n",
        "                  verbose = 1)"
      ],
      "id": "Gmhr-wwaDqRO",
      "execution_count": 11,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch 1/100\n",
            "10/10 [==============================] - 1s 20ms/step - loss: 524.5849 - mae: 21.0093 - val_loss: 291.0752 - val_mae: 15.4410\n",
            "Epoch 2/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 173.6250 - mae: 10.6190 - val_loss: 73.6277 - val_mae: 6.9674\n",
            "Epoch 3/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 61.0815 - mae: 5.6440 - val_loss: 39.2404 - val_mae: 4.9275\n",
            "Epoch 4/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 40.3301 - mae: 4.5378 - val_loss: 25.9560 - val_mae: 3.8117\n",
            "Epoch 5/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 31.0991 - mae: 3.8394 - val_loss: 20.1330 - val_mae: 3.3033\n",
            "Epoch 6/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 26.4128 - mae: 3.5602 - val_loss: 18.9914 - val_mae: 3.1228\n",
            "Epoch 7/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 23.3863 - mae: 3.2375 - val_loss: 16.1864 - val_mae: 3.0027\n",
            "Epoch 8/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 21.0159 - mae: 3.1108 - val_loss: 14.9882 - val_mae: 2.7704\n",
            "Epoch 9/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 19.5917 - mae: 2.9435 - val_loss: 13.4647 - val_mae: 2.6603\n",
            "Epoch 10/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 17.7870 - mae: 2.8038 - val_loss: 12.4472 - val_mae: 2.5820\n",
            "Epoch 11/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 17.0233 - mae: 2.7971 - val_loss: 11.7524 - val_mae: 2.4797\n",
            "Epoch 12/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 16.5762 - mae: 2.7436 - val_loss: 11.7478 - val_mae: 2.5131\n",
            "Epoch 13/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 15.9979 - mae: 2.7127 - val_loss: 10.9575 - val_mae: 2.3727\n",
            "Epoch 14/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 15.2319 - mae: 2.6480 - val_loss: 10.5445 - val_mae: 2.3195\n",
            "Epoch 15/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 14.8978 - mae: 2.6021 - val_loss: 10.6717 - val_mae: 2.3741\n",
            "Epoch 16/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 14.1487 - mae: 2.5801 - val_loss: 11.3767 - val_mae: 2.4875\n",
            "Epoch 17/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 13.5730 - mae: 2.5399 - val_loss: 9.8366 - val_mae: 2.2537\n",
            "Epoch 18/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 13.3380 - mae: 2.4743 - val_loss: 10.2534 - val_mae: 2.3513\n",
            "Epoch 19/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 13.2191 - mae: 2.5240 - val_loss: 9.4234 - val_mae: 2.2435\n",
            "Epoch 20/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 12.6353 - mae: 2.4464 - val_loss: 9.1674 - val_mae: 2.2046\n",
            "Epoch 21/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 12.1251 - mae: 2.3936 - val_loss: 9.3325 - val_mae: 2.2090\n",
            "Epoch 22/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 11.8977 - mae: 2.4042 - val_loss: 9.1391 - val_mae: 2.1904\n",
            "Epoch 23/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 11.4309 - mae: 2.3513 - val_loss: 9.2327 - val_mae: 2.2225\n",
            "Epoch 24/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 11.4186 - mae: 2.3088 - val_loss: 9.4646 - val_mae: 2.2137\n",
            "Epoch 25/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 11.2966 - mae: 2.3455 - val_loss: 9.5623 - val_mae: 2.2133\n",
            "Epoch 26/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 11.3239 - mae: 2.3352 - val_loss: 9.3397 - val_mae: 2.2095\n",
            "Epoch 27/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 10.9282 - mae: 2.2946 - val_loss: 9.6066 - val_mae: 2.2504\n",
            "Epoch 28/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 10.8088 - mae: 2.3125 - val_loss: 10.0747 - val_mae: 2.2888\n",
            "Epoch 29/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 10.6100 - mae: 2.3154 - val_loss: 9.0315 - val_mae: 2.1942\n",
            "Epoch 30/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 10.3206 - mae: 2.2666 - val_loss: 9.1111 - val_mae: 2.2070\n",
            "Epoch 31/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 10.2541 - mae: 2.2378 - val_loss: 9.1771 - val_mae: 2.2261\n",
            "Epoch 32/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 10.2178 - mae: 2.2570 - val_loss: 9.6963 - val_mae: 2.2687\n",
            "Epoch 33/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 10.1990 - mae: 2.2366 - val_loss: 9.2820 - val_mae: 2.2383\n",
            "Epoch 34/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.9224 - mae: 2.2008 - val_loss: 9.8291 - val_mae: 2.2682\n",
            "Epoch 35/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.8418 - mae: 2.2190 - val_loss: 9.8751 - val_mae: 2.2680\n",
            "Epoch 36/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.6956 - mae: 2.2062 - val_loss: 10.0381 - val_mae: 2.3307\n",
            "Epoch 37/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.6691 - mae: 2.2076 - val_loss: 9.8908 - val_mae: 2.2922\n",
            "Epoch 38/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.4417 - mae: 2.1764 - val_loss: 9.8712 - val_mae: 2.3068\n",
            "Epoch 39/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 9.3756 - mae: 2.1847 - val_loss: 9.6599 - val_mae: 2.2768\n",
            "Epoch 40/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.4296 - mae: 2.1896 - val_loss: 9.7394 - val_mae: 2.3023\n",
            "Epoch 41/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 9.2955 - mae: 2.1588 - val_loss: 10.8182 - val_mae: 2.3583\n",
            "Epoch 42/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.0090 - mae: 2.1733 - val_loss: 10.6784 - val_mae: 2.4229\n",
            "Epoch 43/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.1104 - mae: 2.1561 - val_loss: 9.8188 - val_mae: 2.3384\n",
            "Epoch 44/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 9.1659 - mae: 2.1518 - val_loss: 10.1230 - val_mae: 2.3271\n",
            "Epoch 45/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 9.0392 - mae: 2.1388 - val_loss: 10.3550 - val_mae: 2.3834\n",
            "Epoch 46/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.9410 - mae: 2.1004 - val_loss: 10.1056 - val_mae: 2.3424\n",
            "Epoch 47/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.7135 - mae: 2.0781 - val_loss: 11.0393 - val_mae: 2.3935\n",
            "Epoch 48/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 8.9817 - mae: 2.1393 - val_loss: 10.9451 - val_mae: 2.3972\n",
            "Epoch 49/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.5278 - mae: 2.0613 - val_loss: 12.1423 - val_mae: 2.4965\n",
            "Epoch 50/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 8.6911 - mae: 2.1369 - val_loss: 10.6562 - val_mae: 2.3769\n",
            "Epoch 51/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.5663 - mae: 2.1001 - val_loss: 10.3178 - val_mae: 2.4270\n",
            "Epoch 52/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.4884 - mae: 2.0632 - val_loss: 10.8929 - val_mae: 2.4113\n",
            "Epoch 53/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.3182 - mae: 2.0405 - val_loss: 12.8957 - val_mae: 2.5274\n",
            "Epoch 54/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.5717 - mae: 2.1420 - val_loss: 10.5227 - val_mae: 2.4066\n",
            "Epoch 55/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.5186 - mae: 2.0796 - val_loss: 10.8799 - val_mae: 2.4315\n",
            "Epoch 56/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 8.4005 - mae: 2.0631 - val_loss: 11.1523 - val_mae: 2.4575\n",
            "Epoch 57/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 8.2616 - mae: 2.0426 - val_loss: 12.0189 - val_mae: 2.5096\n",
            "Epoch 58/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.2238 - mae: 2.0750 - val_loss: 10.8058 - val_mae: 2.4458\n",
            "Epoch 59/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.2229 - mae: 2.0104 - val_loss: 11.5226 - val_mae: 2.4253\n",
            "Epoch 60/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 8.1652 - mae: 2.0539 - val_loss: 11.1692 - val_mae: 2.4650\n",
            "Epoch 61/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 7.9781 - mae: 2.0157 - val_loss: 11.2512 - val_mae: 2.4410\n",
            "Epoch 62/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.9634 - mae: 2.0326 - val_loss: 11.2064 - val_mae: 2.4765\n",
            "Epoch 63/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 7.9761 - mae: 1.9881 - val_loss: 11.2441 - val_mae: 2.4334\n",
            "Epoch 64/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 7.7827 - mae: 2.0140 - val_loss: 11.0186 - val_mae: 2.5180\n",
            "Epoch 65/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.8843 - mae: 2.0217 - val_loss: 11.0249 - val_mae: 2.4845\n",
            "Epoch 66/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.9058 - mae: 1.9792 - val_loss: 10.7414 - val_mae: 2.4378\n",
            "Epoch 67/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.8468 - mae: 2.0081 - val_loss: 11.2744 - val_mae: 2.4423\n",
            "Epoch 68/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.5752 - mae: 1.9815 - val_loss: 11.0490 - val_mae: 2.4353\n",
            "Epoch 69/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 7.7632 - mae: 1.9845 - val_loss: 11.0473 - val_mae: 2.4575\n",
            "Epoch 70/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.5782 - mae: 1.9467 - val_loss: 10.8195 - val_mae: 2.4443\n",
            "Epoch 71/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.5669 - mae: 1.9522 - val_loss: 11.2569 - val_mae: 2.4526\n",
            "Epoch 72/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.5678 - mae: 1.9449 - val_loss: 11.4452 - val_mae: 2.4871\n",
            "Epoch 73/100\n",
            "10/10 [==============================] - 0s 7ms/step - loss: 7.6307 - mae: 1.9774 - val_loss: 10.9481 - val_mae: 2.4303\n",
            "Epoch 74/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.5582 - mae: 1.9779 - val_loss: 11.2703 - val_mae: 2.4555\n",
            "Epoch 75/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 7.6155 - mae: 1.9944 - val_loss: 11.5177 - val_mae: 2.5154\n",
            "Epoch 76/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.6123 - mae: 1.9508 - val_loss: 10.9443 - val_mae: 2.4666\n",
            "Epoch 77/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.4599 - mae: 1.9450 - val_loss: 12.0903 - val_mae: 2.5149\n",
            "Epoch 78/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.4296 - mae: 1.9632 - val_loss: 11.5154 - val_mae: 2.4665\n",
            "Epoch 79/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 7.3779 - mae: 1.9516 - val_loss: 11.0336 - val_mae: 2.4465\n",
            "Epoch 80/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.3226 - mae: 1.9694 - val_loss: 11.1659 - val_mae: 2.4770\n",
            "Epoch 81/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.4046 - mae: 1.9572 - val_loss: 11.1197 - val_mae: 2.4463\n",
            "Epoch 82/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 7.2160 - mae: 1.9326 - val_loss: 11.2091 - val_mae: 2.4395\n",
            "Epoch 83/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.1788 - mae: 1.9232 - val_loss: 11.3055 - val_mae: 2.4540\n",
            "Epoch 84/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 7.1227 - mae: 1.9188 - val_loss: 12.4297 - val_mae: 2.4729\n",
            "Epoch 85/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.3360 - mae: 1.9370 - val_loss: 11.3292 - val_mae: 2.4649\n",
            "Epoch 86/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 7.1511 - mae: 1.9144 - val_loss: 12.6238 - val_mae: 2.5125\n",
            "Epoch 87/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 7.2362 - mae: 1.9547 - val_loss: 11.6107 - val_mae: 2.4669\n",
            "Epoch 88/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.0221 - mae: 1.9044 - val_loss: 11.3352 - val_mae: 2.4937\n",
            "Epoch 89/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 7.0396 - mae: 1.9170 - val_loss: 11.5825 - val_mae: 2.5152\n",
            "Epoch 90/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.2408 - mae: 1.9192 - val_loss: 11.4595 - val_mae: 2.4605\n",
            "Epoch 91/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 6.9379 - mae: 1.8754 - val_loss: 11.3847 - val_mae: 2.4667\n",
            "Epoch 92/100\n",
            "10/10 [==============================] - 0s 5ms/step - loss: 6.9433 - mae: 1.8777 - val_loss: 11.7018 - val_mae: 2.5207\n",
            "Epoch 93/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 7.0851 - mae: 1.8838 - val_loss: 12.7157 - val_mae: 2.4965\n",
            "Epoch 94/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 6.8661 - mae: 1.9101 - val_loss: 11.5646 - val_mae: 2.4839\n",
            "Epoch 95/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 6.8830 - mae: 1.8803 - val_loss: 11.8982 - val_mae: 2.5192\n",
            "Epoch 96/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 6.6698 - mae: 1.8910 - val_loss: 11.6438 - val_mae: 2.4970\n",
            "Epoch 97/100\n",
            "10/10 [==============================] - 0s 4ms/step - loss: 6.8864 - mae: 1.8903 - val_loss: 12.0042 - val_mae: 2.5136\n",
            "Epoch 98/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 6.8594 - mae: 1.8691 - val_loss: 11.8582 - val_mae: 2.4998\n",
            "Epoch 99/100\n",
            "10/10 [==============================] - 0s 6ms/step - loss: 6.8567 - mae: 1.8991 - val_loss: 11.4934 - val_mae: 2.4822\n",
            "Epoch 100/100\n",
            "10/10 [==============================] - 0s 3ms/step - loss: 6.8060 - mae: 1.8647 - val_loss: 11.3940 - val_mae: 2.5098\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "C9bu7qX8DqRO",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 530
        },
        "outputId": "284ec897-5690-4d82-8e60-593ff5b77c47"
      },
      "source": [
        "# Plotando os resultados obtidos\n",
        "fig, axes = plt.subplots(1, 2, squeeze = False, figsize = (16,8))\n",
        "\n",
        "history_dict = hist.history\n",
        "\n",
        "# Loss - MSE\n",
        "train_loss_values = history_dict[\"loss\"]\n",
        "val_loss_values = history_dict[\"val_loss\"]\n",
        "\n",
        "# Epochs\n",
        "epochs = range(1, len(train_loss_values) + 1)\n",
        "\n",
        "# Metrica - MAE\n",
        "train_mae_values = history_dict[\"mae\"]\n",
        "val_mae_values = history_dict[\"val_mae\"]\n",
        "\n",
        "ax = axes.flat[0]\n",
        "ax.plot(epochs, train_loss_values, \"r\", label = \"Training MSE\")\n",
        "ax.plot(epochs, val_loss_values, \"b\", label = \"Validation MSE\")\n",
        "ax.set_title(\"Training and Validation MSE\")\n",
        "ax.set_xlabel(\"Epochs\")\n",
        "ax.set_ylabel(\"MSE\")\n",
        "ax.legend()\n",
        "\n",
        "ax = axes.flat[1]\n",
        "ax.plot(epochs, train_mae_values, \"r\", label=\"Training MAE\")\n",
        "ax.plot(epochs, val_mae_values, \"b\", label=\"Validation MAE\")\n",
        "ax.set_title(\"Training and Validation MAE\")\n",
        "ax.set_xlabel(\"Epochs\")\n",
        "ax.set_ylabel(\"MSE\")\n",
        "ax.legend()"
      ],
      "id": "C9bu7qX8DqRO",
      "execution_count": 12,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<matplotlib.legend.Legend at 0x7f097d0e1e50>"
            ]
          },
          "metadata": {},
          "execution_count": 12
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 1152x576 with 2 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ruSqlthxDqRP"
      },
      "source": [
        "**g. Comente os resultados obtidos nos gráficos acima. Houve overfitting? Se sim, o que pode ser feito para melhorar a qualidade do modelo?**"
      ],
      "id": "ruSqlthxDqRP"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "q864cYu46wxt"
      },
      "source": [
        "<font color='yellow'>\n",
        "Com base nos gráficos, podemos observer a relação em termos do erro quadrático por epocas, percebe-se que o erro tende a diminuir(não linearmente) com o passar das epocas e ainda podemos comparar os resultados obtidos com os dados de treino e de teste. No caso ideial, desejamos que os gráficos sejam os mais próximos possíveis. Ainda de acordo com o gráfico, próximo as epocas 40 percebe-se um leve incremento das perdas, caracterizando um overfitting e podemos evitar este problema eliminando algumas informações passadas a rede"
      ],
      "id": "q864cYu46wxt"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "JWGU7870DqRP"
      },
      "source": [
        ""
      ],
      "id": "JWGU7870DqRP",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "wyU6dr2BDqRP"
      },
      "source": [
        "**h. Modifique hiperparâmetros do modelo como o número de camadas, as funções de ativação, o número de épocas realizas e o tamanho dos lotes. Comente os resultados obtidos a partir das mudanças realizadas.**"
      ],
      "id": "wyU6dr2BDqRP"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "vbZQ9Allgirj"
      },
      "source": [
        "<font color='yellow'>\n",
        "Observe que modifiquei os hiperâmetros nos blocos anteriores\n",
        "</font>"
      ],
      "id": "vbZQ9Allgirj"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "QwfW18QyDqRP"
      },
      "source": [
        "### 1.4 Teste do modelo\n",
        "\n",
        "O teste do modelo pode ser realizado a partir da função **evaluate**, que recebe os dados de treino e retorna o valor de loss calculado para esse conjunto e os valores de cada métrica da lista fornecida durante a compilação do modelo. \n",
        "\n",
        "* É uma prática comum realizar ajustes no modelo com base no conjunto de validação e só utilizar o conjunto de testes após a definição dos hiperparâmetros definitivos.\n",
        "* Como os hiperparâmetros são ajustados a partir dos resultados obtidos para o conjunto de validação, o modelo pode acabar. sobreajustando aos dados de validação, então é interessante mudar os dados desse conjunto com frequência.\n",
        "* Para mudar os dados de validação basta alterar a semente na função train_test_split.\n",
        "* Crie um novo modelo do zero após a realização de mudanças nos conjuntos de treino/validação."
      ],
      "id": "QwfW18QyDqRP"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ddBXpOLvDqRQ",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "864c776d-ba48-443c-d180-ccc6c6475a13"
      },
      "source": [
        "test_mse, test_mae = model.evaluate( test_data, test_targets)\n",
        "\n",
        "# Observa-se que os dados de saída são normalizados em milhares de dólares, \n",
        "# então é necessário multiplicar por 1000 para obter os valores absolutos\n",
        "print(\"Erro médio absoluto de teste: ${:.2f}\".format(1000*test_mae))"
      ],
      "id": "ddBXpOLvDqRQ",
      "execution_count": 15,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "4/4 [==============================] - 0s 3ms/step - loss: 20.1080 - mae: 2.7468\n",
            "Erro médio absoluto de teste: $2746.80\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "VF5mKk5tDqRQ"
      },
      "source": [
        "**i. Verifique novamente a faixa de valores dos gabaritos no banco de dados. Considerando a extensão dessa faixa de valores, pode-se dizer que as predições do modelo são significativas?**"
      ],
      "id": "VF5mKk5tDqRQ"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Sl7rvC9LDqRQ"
      },
      "source": [
        ""
      ],
      "id": "Sl7rvC9LDqRQ",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "KGi9MobjDqRQ"
      },
      "source": [
        "## Referências\n",
        "* Chollet, Francois. Deep learning with Python. Simon and Schuster, 2017."
      ],
      "id": "KGi9MobjDqRQ"
    }
  ]
}