tp1.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4",
      "toc_visible": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "gpuClass": "standard"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "# **Redes Neuronales Artificiales**: Trabajo Práctico 1\n",
        "\n",
        "#### Integrantes:\n",
        "\n",
        "- Maximiliano Dacko, LU  284/21\n",
        "- María Marino, LU 450/21\n",
        "- Giovanni Franco Gabriel Marraffini, LU 292/21"
      ],
      "metadata": {
        "id": "uh0IJK1M9zHr"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Clase `MLP`: Multi Level Perceptron"
      ],
      "metadata": {
        "id": "RTv4jpg5_UdK"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "En primer lugar, se crea una clase para facilitar las acciones más utilizadas por la red neuronal. Esta consta de un constructor donde se definen la cantidad de neuronas de cada capa y la función de activación. Asimismo se crean las matrices pseudoaleatorias con una distribución normal.\n",
        "\n",
        "Además, se utilizan como métodos públicos una función de activación de la red, una función de corrección de los pesos que utiliza el algortimo de *backpropagation* del error y una función para obtener los valores de salida dados nuevos datos de entrada.\n",
        "\n",
        "Se decidió utilizar la misma función de activación para todas las capas. "
      ],
      "metadata": {
        "id": "1eJxl6z1FTPG"
      }
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "WtWGriXIPiva"
      },
      "outputs": [],
      "source": [
        "import numpy as np\n",
        "\n",
        "def sigmoid(x):\n",
        "  return 1 / (1 + np.exp(-x))\n",
        "\n",
        "class MLP:\n",
        "    def __init__(self, sizes: list[int], batchSize: int, f: str = 'tanh'):\n",
        "        self.S = sizes\n",
        "        self.L = len(self.S)\n",
        "        self.N = batchSize\n",
        "        self.Y = [np.empty(shape=(self.N, self.S[i - 1] + 1)) for i in range(self.L)]\n",
        "        self.W = [None] + [np.random.normal(0, 0.5, (self.S[i - 1] + 1, self.S[i])) for i in range(1, self.L)]\n",
        "        self._f = f\n",
        "\n",
        "    \n",
        "    def _add_bias(self, V):\n",
        "        bias = np.ones((len(V),1))\n",
        "        return np.hstack([V, bias])\n",
        "\n",
        "    def _sub_bias(self, V):\n",
        "        return V[:, :-1]\n",
        "    \n",
        "    @property\n",
        "    def f(self):\n",
        "        return self._f\n",
        "\n",
        "    @f.setter\n",
        "    def f(self, fun):\n",
        "        if fun != 'tanh' and fun != 'sigmoid': \n",
        "            raise Exception('Activation function should be  \\'tanh\\' or \\'sigmoid\\'')\n",
        "    \n",
        "    def activation(self, Xh):\n",
        "        self.Y = [np.empty(shape=(self.N, self.S[i - 1] + 1)) for i in range(self.L)]\n",
        "        Y_b = Xh\n",
        "        for k in range(1, self.L):\n",
        "            self.Y[k - 1] = self._add_bias(Y_b)\n",
        "            if self.f == 'sigmoid':\n",
        "                Y_b = sigmoid(self.Y[k - 1] @ self.W[k])\n",
        "            else:\n",
        "                Y_b = np.tanh(self.Y[k - 1] @ self.W[k])\n",
        "        self.Y[self.L - 1] = Y_b\n",
        "        return self.Y\n",
        "    \n",
        "    def correction(self, Yh, Zh):\n",
        "        D = [np.empty(shape=(self.N, self.S[i - 1] + 1)) for i in range(self.L)]\n",
        "        deltaW = [None] + [np.empty(shape=(self.S[i - 1] + 1, self.S[i])) for i in range(1, self.L)]\n",
        "        E = Zh - Yh[self.L - 1]\n",
        "        if self.f == 'sigmoid':\n",
        "            dY = Yh[self.L - 1] * (1 - Yh[self.L - 1]) # Derivative of sigmoid\n",
        "        else:\n",
        "            dY = 1 - np.square(Yh[self.L - 1]) # Derivative of tanh\n",
        "        D[self.L - 1] = E * dY\n",
        "        for k in reversed(range(1, self.L)):\n",
        "            deltaW[k] = Yh[k - 1].T @ D[k]\n",
        "            E = D[k] @ self.W[k].T\n",
        "            if self.f == 'sigmoid':\n",
        "                dY = Yh[k - 1] * (1 - Yh[k - 1]) # Derivative of sigmoid\n",
        "            else:\n",
        "                dY = 1 - np.square(Yh[k - 1]) # Derivative of tanh\n",
        "            D[k - 1] = self._sub_bias(E * dY)\n",
        "        return deltaW\n",
        "\n",
        "    def predict(self, x):\n",
        "      return self.activation(x)[-1]"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# **Ejercicio 1**"
      ],
      "metadata": {
        "id": "a65siJFgTxiE"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Se subieron los datos a un repositorio público para facilitar su importación.\n",
        "\n",
        "Por el enunciado, sabemos que el conjunto de datos consta de las siguientes columnas:\n",
        "1. *Diagnóstico, dónde 1 = maligno y 0 = benigno*. (**Nuestra variable objetivo**).\n",
        "2. *Radio: media de la distancia desde el centro a los puntos deperímetro*\n",
        "3. *Textura: desviación estándar de los valores en escala de gris*\n",
        "4. *Perímetro*\n",
        "5. *Área*\n",
        "6. *Suavidad: variaciones locales en la longitud del radio*\n",
        "7. *Compacidad: perímetro^2 / área - 1*\n",
        "8. *Concavidad: severidad de las porciones cóncavas del contorno*\n",
        "9. *Puntos cóncavos: proporción de porciones cóncavas del contorno*\n",
        "10. *Simetría*\n",
        "11. *Dimensión fractal*"
      ],
      "metadata": {
        "id": "LHVE0m4wQFEG"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "data = np.loadtxt('https://raw.githubusercontent.com/marinomaria/UBA-redes_neuronales/main/dataset/tp1-ej1.txt', delimiter = ',')\n",
        "\n",
        "datasetSize = data.shape[0]\n",
        "\n",
        "inputSize = 10\n",
        "outputSize = 1"
      ],
      "metadata": {
        "id": "lH9SwGEzRzFv"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "Se decide separar el 10% de los datos para luego evaluar con estos al final del entrenamiento y obtener un estimado del error real que podría llegar a cometer el modelo."
      ],
      "metadata": {
        "id": "aTRuPwKqRTHK"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "np.random.shuffle(data)\n",
        "\n",
        "evalSetSize = int(datasetSize * 0.1)\n",
        "trainSet = data[: - evalSetSize, :]\n",
        "evalSet = data[- evalSetSize:, :]\n",
        "\n",
        "trainSet.shape, evalSet.shape"
      ],
      "metadata": {
        "id": "tgvFggGEgBc2",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "e24a976f-2dd3-46cb-9692-750b74653b9a"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "((369, 11), (41, 11))"
            ]
          },
          "metadata": {},
          "execution_count": 64
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "### Estandarización"
      ],
      "metadata": {
        "id": "1-iafbtNaHAJ"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Para cada variable miramos su rango y promedio, y luego procedemos a estandarizar todo el dataset de entrenamiento **sin** la variable objetivo:"
      ],
      "metadata": {
        "id": "TprcqetEAwIS"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "for j in range(trainSet.shape[1]):\n",
        "  print(f'Columna {j + 1}: \\n Rango: {trainSet[:, j].min()} - {trainSet[:, j].max()} \\n Promedio: {trainSet[:, j].mean()}')"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "qRB7oMdqBDvl",
        "outputId": "d34adef2-994e-4f7e-8f71-443dbb84a829"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Columna 1: \n",
            " Rango: 0.0 - 1.0 \n",
            " Promedio: 0.4878048780487805\n",
            "Columna 2: \n",
            " Rango: 6.981 - 28.11 \n",
            " Promedio: 18.601371273712733\n",
            "Columna 3: \n",
            " Rango: 10.034 - 33.44 \n",
            " Promedio: 21.19330081300813\n",
            "Columna 4: \n",
            " Rango: 43.79 - 188.5 \n",
            " Promedio: 113.41271544715447\n",
            "Columna 5: \n",
            " Rango: 143.5 - 2501.0 \n",
            " Promedio: 850.2510623306233\n",
            "Columna 6: \n",
            " Rango: 0.053 - 4.099 \n",
            " Promedio: 1.788737127371274\n",
            "Columna 7: \n",
            " Rango: 0.387 - 3.345 \n",
            " Promedio: 1.885661246612466\n",
            "Columna 8: \n",
            " Rango: 0.427 - 7.283 \n",
            " Promedio: 3.6408428184281845\n",
            "Columna 9: \n",
            " Rango: 0.201 - 9.013 \n",
            " Promedio: 4.536311653116531\n",
            "Columna 10: \n",
            " Rango: 0.106 - 1.304 \n",
            " Promedio: 0.7225853658536586\n",
            "Columna 11: \n",
            " Rango: 0.05 - 2.974 \n",
            " Promedio: 1.5473387533875338\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "z = trainSet[:, 0].reshape(trainSet.shape[0], 1)\n",
        "x = trainSet[:, 1:trainSet.shape[1]]\n",
        "\n",
        "z.shape, x.shape"
      ],
      "metadata": {
        "id": "WL3PdKypUsBH",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "d9550a35-181f-4dae-d84e-a33e6511c96c"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "((369, 1), (369, 10))"
            ]
          },
          "metadata": {},
          "execution_count": 66
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "La estandarización utilizada es:\n",
        "\n",
        "$z = \\dfrac{x_{ij} - \\mu_{j}}{\\sigma_{j}}$\n",
        "\n",
        "donde $x_{ij}$ es el valor original para la instancia $i$ en su variable $j$, $\\mu_j$ y $\\sigma_j$ son respectivamente la media y la desviación estándar de la variable $j$."
      ],
      "metadata": {
        "id": "ij3LqZ9nGA29"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "x = (x - x.mean(0)) / x.std(0)"
      ],
      "metadata": {
        "id": "UAeQFcJ4bTno"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "### Entrenamiento"
      ],
      "metadata": {
        "id": "sutyEabMglU8"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Definimos el *batch size* (tamaño del lote) como un 10% del total de nuestro dataset. Teniendo en cuenta que ya quitamos un 10% para validación, nos quedan 9 lotes para entrenamiento.\n",
        "\n",
        "Asimismo inicializamos una instancia de la clase `MLP`, definiendo aquí parte de su arquitectura. En este caso, nuestra red neuronal tendrá una única capa oculta cuyo tamaño será el doble del tamaño de entrada, y todas las capas utilizarán la función sigmoide como función de activación."
      ],
      "metadata": {
        "id": "gNA-WMvcG6S5"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "bSize = int(datasetSize * 0.1)\n",
        "M = MLP([inputSize, inputSize * 2, outputSize], bSize, 'sigmoid')"
      ],
      "metadata": {
        "id": "Wz7ycCxVUmiI"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "Entrenamos el modelo de a lotes. Elegimos entrenar por 40000 épocas para cumplir con lo pedido en el enunciado y que el modelo no tarde más de un minuto en entrenarse.\n",
        "\n",
        "Cada ciclo del `while loop` exterior representa una época (*epoch*) y en él permutamos el orden de los datos para intentar evitar estancarnos en un mínimo local de nuestra función de error.\n",
        "\n",
        "En cada iteración del `for loop` ingresamos un lote al modelo, activando la red y luego corrigiendo sus pesos con los métodos ya mencionados de la clase `MLP`.\n",
        "\n",
        "Para calcular el error, sumamos el cuadrado de la diferencia entre el valor esperado (variable objetivo) y el valor predicho por el modelo, y luego dividimos por la cantidad de términos de la suma (error cuadrático medio).\n",
        "\n",
        "Por último, imprimimos para algunas épocas el error estimado para poder observar su evolución."
      ],
      "metadata": {
        "id": "Rhg9yQnZJCUk"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "minError = 0.0001\n",
        "maxEpoch = 40000\n",
        "e = 1.0\n",
        "eta = 0.001\n",
        "epoch = 0\n",
        "p = x.shape[0]\n",
        "\n",
        "while e > minError and epoch < maxEpoch:\n",
        "    stochastic = np.random.permutation(p)\n",
        "    e = 0.0\n",
        "    for i in range(0, p, bSize):\n",
        "        batch = stochastic[i : i + bSize]\n",
        "        xi = x[batch].reshape(bSize, 10)\n",
        "        zi = z[batch].reshape(bSize, 1)\n",
        "        yi = M.activation(xi)\n",
        "        e += np.sum(np.square(zi - yi[-1]))\n",
        "        dW = M.correction(yi, zi)\n",
        "        for j in range(1, M.L):   \n",
        "          M.W[j] += eta * dW[j]\n",
        "    e /= p\n",
        "    if (epoch % 10000 == 0 or epoch == maxEpoch - 1):\n",
        "      print(f'Epoch {epoch}, error: {e}')\n",
        "    epoch += 1 "
      ],
      "metadata": {
        "id": "ddVTnDwob6IJ",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "7bc53e44-f7ad-41bd-858e-abfa691e22a2"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch 0, error: 0.3196186232210721\n",
            "Epoch 10000, error: 0.03849632065520656\n",
            "Epoch 20000, error: 0.013449500631073806\n",
            "Epoch 30000, error: 0.008579452637621263\n",
            "Epoch 39999, error: 0.006541638692360828\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "### Validación"
      ],
      "metadata": {
        "id": "ssUQ5dHLLH1X"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Para validar el desempeño del modelo lo utilizamos para predecir el resultado del 10% del dataset que **no fue utilizado en su entrenamiento**.\n",
        "\n",
        "Primero estandarizamos este dataset de validación como hicimos con los datos de entrenamiento, y luego comparamos los resultados obtenidos con los esperados."
      ],
      "metadata": {
        "id": "25SPnj45LcfJ"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "zEval = evalSet[:, 0:1].reshape(evalSet.shape[0], 1)\n",
        "xEval = evalSet[:, 1:11]\n",
        "xEval = (xEval - xEval.mean(0)) / xEval.std(0)\n",
        "\n",
        "pred = np.zeros_like(zEval)\n",
        "\n",
        "for i in range(len(evalSet)):\n",
        "  pred[i] = M.predict(xEval[i].reshape(1, 10))\n",
        "\n",
        "print(f'El modelo acertó en {np.sum(zEval == np.round(pred))} de {len(evalSet)} instancias ({100 * np.round(np.sum(zEval == np.round(pred)) / len(evalSet), 2)} %).\\n\\n')\n",
        "\n",
        "for i in range(len(zEval)):\n",
        "  print(f'Esperado: {zEval[i]}  Predicho: {pred[i]}   Redondeado: {np.round(pred[i])}')"
      ],
      "metadata": {
        "id": "s2s1WDV_g33F",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "c7b4ed88-34c8-4aa2-d9d7-beb484d3739d"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "El modelo acertó en 33 de 41 instancias (80.0 %).\n",
            "\n",
            "\n",
            "Esperado: [0.]  Predicho: [0.00345947]   Redondeado: [0.]\n",
            "Esperado: [1.]  Predicho: [0.99742525]   Redondeado: [1.]\n",
            "Esperado: [1.]  Predicho: [0.90111181]   Redondeado: [1.]\n",
            "Esperado: [1.]  Predicho: [0.99930487]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.60292348]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.08984131]   Redondeado: [0.]\n",
            "Esperado: [1.]  Predicho: [0.49259027]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.2505731]   Redondeado: [0.]\n",
            "Esperado: [1.]  Predicho: [0.96465163]   Redondeado: [1.]\n",
            "Esperado: [1.]  Predicho: [0.97147201]   Redondeado: [1.]\n",
            "Esperado: [1.]  Predicho: [0.9192532]   Redondeado: [1.]\n",
            "Esperado: [1.]  Predicho: [0.98344597]   Redondeado: [1.]\n",
            "Esperado: [1.]  Predicho: [0.96642993]   Redondeado: [1.]\n",
            "Esperado: [1.]  Predicho: [0.99840584]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.5004319]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.5106831]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.19490492]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.0004165]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.02409314]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.06968876]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.54383613]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.31008608]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.00089046]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.97226675]   Redondeado: [1.]\n",
            "Esperado: [1.]  Predicho: [0.98176726]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.40819013]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.06184021]   Redondeado: [0.]\n",
            "Esperado: [1.]  Predicho: [0.97813642]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.0091476]   Redondeado: [0.]\n",
            "Esperado: [1.]  Predicho: [0.99944124]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.00089525]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.91807712]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.05919815]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.00015688]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.01034764]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.00117699]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.02427213]   Redondeado: [0.]\n",
            "Esperado: [1.]  Predicho: [0.27870447]   Redondeado: [0.]\n",
            "Esperado: [0.]  Predicho: [0.02916192]   Redondeado: [0.]\n",
            "Esperado: [1.]  Predicho: [0.64081789]   Redondeado: [1.]\n",
            "Esperado: [0.]  Predicho: [0.05629045]   Redondeado: [0.]\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "#### Validación cruzada con *K-folds*"
      ],
      "metadata": {
        "id": "U8jIFfISeUnw"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Por último, para tener un mayor entendimiento de la performance del modelo en el \"caso real\" utilizamos el método *K-folds*.\n",
        "\n",
        "La siguiente celda de código demora más en correr ya que entrena múltiples modelos con la misma arquitectura que el original pero variando los conjuntos de datos de entrenamiento y validación."
      ],
      "metadata": {
        "id": "8snlgstqNXeQ"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "s = np.random.permutation(datasetSize)\n",
        "\n",
        "pred = np.empty(shape=(datasetSize, outputSize))\n",
        "\n",
        "print(f'Cantidad de folds: {int(datasetSize / bSize)}. Tamaño de cada fold: {bSize} instancias.')\n",
        "\n",
        "for i in range(0, datasetSize, bSize):\n",
        "    fold = s[i : i + 82]\n",
        "    evalSet = data[fold]\n",
        "    trainSet = np.delete(data, fold, 0)\n",
        "\n",
        "    x = trainSet[:, 1:11]\n",
        "    z = trainSet[:, 0]\n",
        "    \n",
        "    x = (x - x.mean(0)) / x.std(0)\n",
        "\n",
        "    M = MLP([inputSize, inputSize * 2, outputSize], bSize, 'sigmoid')\n",
        "\n",
        "    e = 1.0\n",
        "    epoch = 0\n",
        "    p = len(trainSet)\n",
        "\n",
        "    while e > minError and epoch < maxEpoch:\n",
        "        stochastic = np.random.permutation(p)\n",
        "        e = 0.0\n",
        "        for j in range(0, p, bSize):\n",
        "            batch = stochastic[j : j + bSize]\n",
        "            xj = x[batch].reshape(bSize, M.S[0])\n",
        "            zj = z[batch].reshape(bSize, M.S[-1])\n",
        "            y = M.activation(xj)\n",
        "            e += np.sum(np.square(zj - y[-1]))\n",
        "            dW = M.correction(y, zj)\n",
        "            for k in range(1, M.L):\n",
        "                M.W[k] += eta * dW[k]\n",
        "        e /= (p / bSize)\n",
        "        epoch += 1 \n",
        "\n",
        "    xEval = evalSet[:, 1:11]\n",
        "    zEval = evalSet[:, 0]\n",
        "\n",
        "    xEval = (xEval - xEval.mean(0)) / xEval.std(0)\n",
        "\n",
        "    pred[fold] = M.predict(xEval)\n",
        "    print(f'Modelo entrenado sin el fold {int(i / bSize)}. Error en la última época: {e}')\n"
      ],
      "metadata": {
        "id": "H7GuSlbteTSh",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "9f224d18-6f84-419e-8a44-b4e4fd2f23e2"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Cantidad de folds: 10. Tamaño de cada fold: 41 instancias.\n",
            "Modelo entrenado sin el fold 0. Error en la última época: 0.36800509811383275\n",
            "Modelo entrenado sin el fold 1. Error en la última época: 0.1346939025501628\n",
            "Modelo entrenado sin el fold 2. Error en la última época: 0.13720234639960358\n",
            "Modelo entrenado sin el fold 3. Error en la última época: 0.25093974040198885\n",
            "Modelo entrenado sin el fold 4. Error en la última época: 0.1825338687636233\n",
            "Modelo entrenado sin el fold 5. Error en la última época: 0.15329515656697157\n",
            "Modelo entrenado sin el fold 6. Error en la última época: 0.1904755478063563\n",
            "Modelo entrenado sin el fold 7. Error en la última época: 0.18781691899783984\n",
            "Modelo entrenado sin el fold 8. Error en la última época: 0.21331884453630232\n",
            "Modelo entrenado sin el fold 9. Error en la última época: 0.19838848849628624\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "Los *K-folds* se organizan de forma tal que para cada instancia se obtiene un resultado con alguno de los $K$ modelos entrenados. Guardamos esos resultados en el arreglo `pred` y los comparamos con los valores esperados para construir las métricas de *Recall* y Precisión.\n",
        "\n",
        "$\\text{Recall} = \\dfrac{\\text{Verdaderos Positivos}}{\\text{Verdaderos Positivos} + \\text{Falsos Negativos}}$\n",
        "\n",
        "$\\text{Precisión} = \\dfrac{\\text{Verdaderos Positivos}}{\\text{Verdaderos Positivos} + \\text{Falsos Positivos}}$\n"
      ],
      "metadata": {
        "id": "sjRnewup1h9r"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "expected = data[:, 0].reshape(datasetSize, outputSize)\n",
        "\n",
        "tp = fn = tn = fp = 0\n",
        "\n",
        "for i in range(len(pred)):\n",
        "    if expected[i] == 1:\n",
        "        if np.round(pred[i]) == 1:\n",
        "            tp += 1\n",
        "        else:\n",
        "            fn += 1\n",
        "    else:\n",
        "        if np.round(pred[i]) == 1:\n",
        "            fp += 1\n",
        "        else:\n",
        "            tn += 1\n",
        "\n",
        "print('Recall:', tp / (tp + fn))\n",
        "print('Precisión:', tp / (tp + fp))\n",
        "\n",
        "print(f'\\n\\nSe cometieron {fp + fn} errores en {datasetSize} diagnósticos.')\n",
        "print(f'Diagnosticamos como benignos a {fn} tumores malignos.')\n",
        "print(f'Diagnosticamos como malignos a {fp} tumores benignos.')"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "djz3Gfn5e8Qg",
        "outputId": "d8b81977-3207-422d-db69-3903bbbb9943"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Recall: 0.9076923076923077\n",
            "Precisión: 0.8984771573604061\n",
            "\n",
            "\n",
            "Se cometieron 38 errores en 410 diagnósticos.\n",
            "Diagnosticamos como benignos a 18 tumores malignos.\n",
            "Diagnosticamos como malignos a 20 tumores benignos.\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "En este caso la medida principal que queremos optimizar es el *recall*, ya que al tratarse de un problema de diagnóstico de cáncer de mama es preferible sobreestimar (dar un resultado positivo para tumores benignos) que subestimar, por el impacto que puede tener en las posibilidades de tratamiento y en la calidad de vida de un paciente la detección tardía de un tumor maligno.\n",
        "\n",
        "Como alternativa para mejorar nuestro modelo podríamos agregar un *threshold* $\\epsilon$, y sólo diagnosticar como benignos aquellos casos donde el modelo devuelva resultados menores a $0,5 + \\epsilon$."
      ],
      "metadata": {
        "id": "Ldr9GHvr2O0a"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "# **Ejercicio 2**"
      ],
      "metadata": {
        "id": "AAVnjPi0CDl5"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Para resolver este ejericio se utiliza la misma clase que en el ejericio anterior."
      ],
      "metadata": {
        "id": "d5Z9E_7nxnHV"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Al igual que antes se subieron los datos a un repositorio público para facilitar su lectura.\n",
        "\n",
        "Sabemos por la consigna de este ejercicio que las columnas del dataset representan:\n",
        "\n",
        "1. Compacidad Relativa\n",
        "2. Área de la Superficie Total\n",
        "3. Área de las Paredes\n",
        "4. Área del Techo\n",
        "5. Altura Total\n",
        "6. Orientación\n",
        "7. Área de Reflejo Total \n",
        "8. Distribución del Área de Reflejo\n",
        "9. Carga de Calefacción\n",
        "10. Carga de Refrigeración\n",
        "\n",
        "Las últimas dos columnas serán las variables objetivo en este ejercicio."
      ],
      "metadata": {
        "id": "RvzBggueyWyI"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "data = np.loadtxt('https://raw.githubusercontent.com/marinomaria/UBA-redes_neuronales/main/dataset/tp1-ej2.txt', delimiter = ',')\n",
        "\n",
        "datasetSize = data.shape[0]\n",
        "inputSize = 8\n",
        "outputSize = 2"
      ],
      "metadata": {
        "id": "4QUOqITmCAXM"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "Separamos un 10% del dataset para validar posteriormente el desempeño del modelo en datos que no hayan sido utilizado en su entrenamiento."
      ],
      "metadata": {
        "id": "QK8O8bLF1nxi"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "np.random.shuffle(data)\n",
        "\n",
        "evalSetSize = int(datasetSize * 0.1)\n",
        "trainSet = data[: - evalSetSize, :]\n",
        "evalSet = data[- evalSetSize:, :]\n",
        "\n",
        "trainSetSize = datasetSize - evalSetSize\n",
        "\n",
        "trainSet.shape, evalSet.shape"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "hgaWNTOkKSxJ",
        "outputId": "a0616589-f354-4899-dab3-79ef888d3322"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "((450, 10), (50, 10))"
            ]
          },
          "metadata": {},
          "execution_count": 72
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "### Estandarización"
      ],
      "metadata": {
        "id": "mgXFtAtT1gd-"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "for j in range(inputSize + outputSize):\n",
        "  print(f'Columna {j + 1}: \\n Rango: {trainSet[:, j].min()} - {trainSet[:, j].max()} \\n Promedio: {trainSet[:, j].mean()}')"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "AODWwAzU2Eiy",
        "outputId": "84c16bc5-19d0-4783-e4e2-45eeedb50390"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Columna 1: \n",
            " Rango: 0.62 - 0.98 \n",
            " Promedio: 0.7616444444444445\n",
            "Columna 2: \n",
            " Rango: 514.5 - 808.5 \n",
            " Promedio: 673.4233333333333\n",
            "Columna 3: \n",
            " Rango: 245.0 - 416.5 \n",
            " Promedio: 318.44555555555553\n",
            "Columna 4: \n",
            " Rango: 110.25 - 220.5 \n",
            " Promedio: 177.48888888888888\n",
            "Columna 5: \n",
            " Rango: 3.5 - 7.0 \n",
            " Promedio: 5.234444444444445\n",
            "Columna 6: \n",
            " Rango: 2.0 - 5.0 \n",
            " Promedio: 3.397777777777778\n",
            "Columna 7: \n",
            " Rango: 0.0 - 0.4 \n",
            " Promedio: 0.23077777777777778\n",
            "Columna 8: \n",
            " Rango: 0.0 - 5.0 \n",
            " Promedio: 2.7777777777777777\n",
            "Columna 9: \n",
            " Rango: 6.04 - 42.96 \n",
            " Promedio: 22.15886666666666\n",
            "Columna 10: \n",
            " Rango: 10.9 - 47.59 \n",
            " Promedio: 24.457\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "Observamos que los rangos de las columnas tienen distintos órdenes de magnitud. \n",
        "\n",
        "Estandarizamos utilizando una fórmula similar a la que empleamos en el Ejercicio 1, sólo que \n",
        "- Ahora dividiremos por la desviación estándar de **todo el conjunto de datos de entreanmiento**, no sólo de la variable $j$\n",
        "- También estandarizaremos las variables objetivo.\n",
        "\n",
        "$z = \\dfrac{x_{ij} - \\mu_{j}}{\\sigma}$\n",
        "\n",
        "donde $x_{ij}$ es el valor original para la instancia $i$ en su variable $j$, $\\mu_j$ es la media de la variable $j$ y $\\sigma$ es la desviación estándar del conjunto de datos de entrenamiento."
      ],
      "metadata": {
        "id": "LVOg6kh_120m"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "trainSet = (trainSet - trainSet.mean(0)) / trainSet.std()"
      ],
      "metadata": {
        "id": "KdGRKLH8BM-w"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "### Entrenamiento"
      ],
      "metadata": {
        "id": "anYFxGxA5O3Z"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Separamos en variables de entrada y variables objetivo."
      ],
      "metadata": {
        "id": "ws1__9RM5Nsb"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "z = trainSet[:, inputSize:].reshape(trainSet.shape[0], 2)\n",
        "x = trainSet[:, 0:inputSize]\n",
        "\n",
        "x.shape, z.shape"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "ajzmzKRnDDU5",
        "outputId": "aa7e3e82-f133-49b6-9018-46a6c6eab14c"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "((450, 8), (450, 2))"
            ]
          },
          "metadata": {},
          "execution_count": 75
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "Definimos el *batch size* (tamaño del lote) como un 10% del conjunto total de datos. Teniendo en cuenta que ya quitamos un 10% para validación, nos quedan 9 lotes para entrenamiento.\n",
        "\n",
        "\n",
        "Asimismo inicializamos una instancia de la clase `MLP`, definiendo aquí parte de su arquitectura. En este caso, nuestra red neuronal posee una capa oculta, cuyo tamaño será el doble del tamaño de entrada. Todas las capas utilizarán la función tangente hiperbólica como función de activación."
      ],
      "metadata": {
        "id": "1fJ6mqBV5KBT"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "bSize = int(datasetSize * 0.1)\n",
        "M = MLP([inputSize, inputSize * 2, outputSize], bSize, 'tanh')"
      ],
      "metadata": {
        "id": "j_nWpBq6FZh7"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "Entrenamos el modelo de a lotes. Aquí también elegimos entrenar por 40000 épocas para cumplir con la cota temporal del enunciado (un minuto).\n",
        "\n",
        "Cada ciclo del `while loop` exterior representa una época (*epoch*) y en él permutamos el orden de los datos para intentar evitar estancarnos en un mínimo local de nuestra función de error.\n",
        "\n",
        "En cada iteración del `for loop` ingresamos un lote al modelo, activando la red y luego corrigiendo sus pesos con los métodos ya mencionados de la clase `MLP`.\n",
        "\n",
        "Para calcular el error, sumamos el cuadrado de la diferencia entre el valor esperado (variable objetivo) y el valor predicho por el modelo, y luego dividimos por la cantidad de términos de la suma (error cuadrático medio).\n",
        "\n",
        "Por último, imprimimos para algunas épocas el error estimado para poder observar su evolución."
      ],
      "metadata": {
        "id": "MauwOnxs57qQ"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "minError = 0.000001\n",
        "maxEpoch = 40000\n",
        "e = 1.0\n",
        "epoch = 0\n",
        "eta = 0.001\n",
        "p = len(trainSet)\n",
        "\n",
        "while e > minError and epoch < maxEpoch:\n",
        "    stochastic = np.random.permutation(p)\n",
        "    e = 0.0\n",
        "    for i in range(0, p, bSize):\n",
        "        batch = stochastic[i : i + bSize]\n",
        "        xi = x[batch].reshape(bSize, M.S[0])\n",
        "        zi = z[batch].reshape(bSize, M.S[-1])\n",
        "        y = M.activation(xi)\n",
        "        e += np.sum(np.square(zi - y[-1]))\n",
        "        dW = M.correction(y, zi)\n",
        "        for i in range(1, M.L):\n",
        "          M.W[i] += eta * dW[i]\n",
        "    e /= p\n",
        "    if epoch % 10000 == 0 or epoch == maxEpoch - 1:\n",
        "      print(f'Epoch {epoch}, error: {e}')\n",
        "    epoch += 1 "
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "W-VB9nZaE46A",
        "outputId": "0f6b4b12-f600-4fa5-bc36-b80f3fc423b0"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch 0, error: 0.5581709574724008\n",
            "Epoch 10000, error: 0.0006484536721248918\n",
            "Epoch 20000, error: 0.0006084436133713308\n",
            "Epoch 30000, error: 0.0005776930113166594\n",
            "Epoch 39999, error: 0.0005624972118241038\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "### Validación"
      ],
      "metadata": {
        "id": "xxtji7Vc6VN4"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Para validar el desempeño del modelo lo utilizamos para predecir el resultado del 10% del dataset que **no fue utilizado en su entrenamiento**.\n",
        "\n",
        "Primero estandarizamos este dataset de validación como hicimos con los datos de entrenamiento, y luego comparamos los resultados obtenidos con los esperados."
      ],
      "metadata": {
        "id": "n5XPr7UR7grs"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "evalSet = data[- evalSetSize:, :]\n",
        "evalSet = (evalSet - evalSet.mean(0)) / evalSet.std()\n",
        "\n",
        "zEval = evalSet[:, inputSize:].reshape(evalSetSize, 2)\n",
        "xEval = evalSet[:, 0:inputSize]\n",
        "\n",
        "pred = np.zeros_like(zEval)\n",
        "\n",
        "for i in range(len(evalSet)):\n",
        "  pred[i] = M.predict(xEval[i].reshape(1, 8))\n",
        "  print(f'Esperado: {zEval[i]}  Predicho: {pred[i]}   Diferencia: {np.linalg.norm(zEval[i] - pred[i], 1)}')\n",
        "\n",
        "print('\\n\\nError cuadrático medio por variable:', np.mean(np.square(zEval - pred), axis = 0))\n",
        "print('Error absoluto medio:', np.mean(np.linalg.norm(zEval - pred, 1, axis=1)))"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "DsI7JtNgIP2z",
        "outputId": "6a4187ac-57ce-40f6-a553-fd8acd0a4b2f"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Esperado: [-0.0356849  -0.04237446]  Predicho: [-0.03589265 -0.0391285 ]   Diferencia: 0.0034537121031655016\n",
            "Esperado: [-0.0501041  -0.05093143]  Predicho: [-0.0397359  -0.03701973]   Diferencia: 0.02427988712121181\n",
            "Esperado: [-0.05067141 -0.05244426]  Predicho: [-0.04595925 -0.04136567]   Diferencia: 0.0157907556092935\n",
            "Esperado: [-0.03317927 -0.02530781]  Predicho: [-0.03556766 -0.03616525]   Diferencia: 0.01324583180345762\n",
            "Esperado: [-0.04528194 -0.04904038]  Predicho: [-0.03422101 -0.03784294]   Diferencia: 0.022258372104101427\n",
            "Esperado: [-0.03729229 -0.03641768]  Predicho: [-0.04684053 -0.04504546]   Diferencia: 0.018176020349276498\n",
            "Esperado: [-0.02396044 -0.01954013]  Predicho: [-0.04083559 -0.0401978 ]   Diferencia: 0.03753281181394267\n",
            "Esperado: [0.03234532 0.02395383]  Predicho: [0.05113936 0.04598008]   Diferencia: 0.04082028896084869\n",
            "Esperado: [0.01144931 0.00868366]  Predicho: [0.01780358 0.02177446]   Diferencia: 0.019445064345141216\n",
            "Esperado: [0.03504006 0.02589214]  Predicho: [0.05109599 0.04588897]   Diferencia: 0.03605275573116643\n",
            "Esperado: [-0.0727966  -0.05991387]  Predicho: [-0.0517471  -0.04391035]   Diferencia: 0.03705301336358739\n",
            "Esperado: [0.09186586 0.09921729]  Predicho: [0.05399047 0.0592089 ]   Diferencia: 0.07788377706127836\n",
            "Esperado: [-0.03790688 -0.03774141]  Predicho: [-0.04452614 -0.04456048]   Diferencia: 0.013438330222573583\n",
            "Esperado: [0.02818503 0.0259867 ]  Predicho: [0.0266321  0.02578409]   Diferencia: 0.0017555329140289191\n",
            "Esperado: [-0.02225851 -0.01641992]  Predicho: [-0.04331257 -0.0417153 ]   Diferencia: 0.046349455108964405\n",
            "Esperado: [-0.04835488 -0.04672386]  Predicho: [-0.05036029 -0.04624647]   Diferencia: 0.00248279911107889\n",
            "Esperado: [0.06789691 0.05099572]  Predicho: [0.04887427 0.04471999]   Diferencia: 0.025298369267262734\n",
            "Esperado: [-0.03322655 -0.02488233]  Predicho: [-0.04052097 -0.03920337]   Diferencia: 0.02161546706788778\n",
            "Esperado: [-0.04632201 -0.0440764 ]  Predicho: [-0.05578456 -0.04778409]   Diferencia: 0.013170238577296788\n",
            "Esperado: [0.0470009  0.05487236]  Predicho: [0.02627346 0.01853716]   Diferencia: 0.05706263777554173\n",
            "Esperado: [-0.03365203 -0.0349994 ]  Predicho: [-0.03878375 -0.042079  ]   Diferencia: 0.012211318963310562\n",
            "Esperado: [0.0323926  0.02239372]  Predicho: [0.05484711 0.04766086]   Diferencia: 0.047721655588349704\n",
            "Esperado: [-0.04972589 -0.04653475]  Predicho: [-0.05003099 -0.04595729]   Diferencia: 0.0008825663265307551\n",
            "Esperado: [-0.05591905 -0.05253881]  Predicho: [-0.05085494 -0.04638824]   Diferencia: 0.011214675809444234\n",
            "Esperado: [0.04213146 0.05510874]  Predicho: [0.02982344 0.02008105]   Diferencia: 0.0473357172328006\n",
            "Esperado: [-0.02216395 -0.01660902]  Predicho: [-0.0411688  -0.03887112]   Diferencia: 0.041266946198689425\n",
            "Esperado: [-0.04603836 -0.04473826]  Predicho: [-0.03973465 -0.04006793]   Diferencia: 0.010974042376340996\n",
            "Esperado: [0.09588432 0.06716413]  Predicho: [0.058487   0.06219375]   Diferencia: 0.04236770827808907\n",
            "Esperado: [-0.05119144 -0.04757483]  Predicho: [-0.03754947 -0.03776179]   Diferencia: 0.023455007183772462\n",
            "Esperado: [0.01130748 0.02225189]  Predicho: [0.05063951 0.04623562]   Diferencia: 0.06331576519410462\n",
            "Esperado: [-0.03365203 -0.02951538]  Predicho: [-0.03653594 -0.03597849]   Diferencia: 0.009347020248724337\n",
            "Esperado: [0.0323926  0.04366794]  Predicho: [0.05148596 0.04666687]   Diferencia: 0.022092291502616164\n",
            "Esperado: [0.00322328 0.02021902]  Predicho: [0.04979292 0.04580422]   Diferencia: 0.07215484937063318\n",
            "Esperado: [-0.03256468 -0.02450412]  Predicho: [-0.0344637  -0.03475853]   Diferencia: 0.012153422824922655\n",
            "Esperado: [0.04671724 0.03355086]  Predicho: [0.01558274 0.01390127]   Diferencia: 0.050784092575445174\n",
            "Esperado: [0.01967534 0.01147295]  Predicho: [0.01384358 0.01297832]   Diferencia: 0.0073371382635202265\n",
            "Esperado: [-0.0323283  -0.02563874]  Predicho: [-0.03310641 -0.03464071]   Diferencia: 0.00978006828842972\n",
            "Esperado: [-0.02296765 -0.01613626]  Predicho: [-0.04223115 -0.04029127]   Diferencia: 0.04341851088156044\n",
            "Esperado: [0.05177578 0.04399887]  Predicho: [0.05864038 0.05156853]   Diferencia: 0.014434263313690618\n",
            "Esperado: [0.08397076 0.05751982]  Predicho: [0.05407812 0.05975887]   Diferencia: 0.032131695659426716\n",
            "Esperado: [-0.00646831  0.01738246]  Predicho: [0.03874217 0.03954431]   Diferencia: 0.06737233028647767\n",
            "Esperado: [0.02279556 0.00570527]  Predicho: [0.02459191 0.01772179]   Diferencia: 0.013812867179887974\n",
            "Esperado: [0.01211118 0.02452114]  Predicho: [0.05172117 0.04738553]   Diferencia: 0.06247438249585325\n",
            "Esperado: [0.00175772 0.0184698 ]  Predicho: [0.03574049 0.03892056]   Diferencia: 0.05443351471517254\n",
            "Esperado: [0.01178024 0.01969898]  Predicho: [0.04295124 0.04210481]   Diferencia: 0.05357681664319843\n",
            "Esperado: [-0.03795415 -0.0359922 ]  Predicho: [-0.05224295 -0.04658246]   Diferencia: 0.02487905910024065\n",
            "Esperado: [0.0959316  0.06498943]  Predicho: [0.05132041 0.05814304]   Diferencia: 0.05145758447258938\n",
            "Esperado: [0.06373662 0.05827623]  Predicho: [0.07099414 0.06044838]   Diferencia: 0.009429668269985259\n",
            "Esperado: [0.05711797 0.0411623 ]  Predicho: [0.05285393 0.04685214]   Diferencia: 0.009953875676563056\n",
            "Esperado: [-0.03681953 -0.03655951]  Predicho: [-0.04768669 -0.04249689]   Diferencia: 0.016804539592635406\n",
            "\n",
            "\n",
            "Error cuadrático medio por variable: [0.00043635 0.00027305]\n",
            "Error absoluto medio: 0.029314770298482225\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "import matplotlib.pyplot as plt\n",
        "\n",
        "\n",
        "plt.scatter(range(evalSetSize), zEval[:, 0], c ='red', alpha=0.5)\n",
        "plt.scatter(range(evalSetSize), pred[:, 0], c ='blue', alpha=0.5)\n",
        "plt.ylabel('Carga de Calefacción (estandarizada)')\n",
        "plt.xlabel('Índice')\n",
        "plt.legend(['Esperados', 'Predichos'])\n",
        "plt.show()\n",
        "\n",
        "plt.scatter(range(evalSetSize), zEval[:, 1], c ='red', alpha=0.5)\n",
        "plt.scatter(range(evalSetSize), pred[:, 1], c ='blue', alpha=0.5)\n",
        "plt.ylabel('Carga de Refrigeración (estandarizada)')\n",
        "plt.xlabel('Índice')\n",
        "plt.legend(['Esperados', 'Predichos'])\n",
        "plt.show()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 881
        },
        "id": "6wZ9L1SQoZPy",
        "outputId": "90a654c3-81d5-4e57-9405-91e01cc68b2b"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 640x480 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {}
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 640x480 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "El error cuadrático medio para el conjunto de datos de evaluación es cercano al obtenido en la última época de entrenamiento del modelo. Con esto entendemos que no estamos sobre-ajustando el modelo a los datos de entrenamiento, sino que el ajuste es adecuado.\n",
        "\n",
        "Para analizar la magnitud del error preferimos utilizar el error absoluto medio, que se encuentra en la misma escala que las variables objetivo estandarizadas. El rango de las mismas es [-1, 1]. La media del error absoluto nos indica que tomando el modulo del \"vector error\", es decir, la diferencia entre los valores reales  y los predichos para ambas variables, nos daría en promedio 0.03 aproximadamente. Obviamente esto toma en cuenta ambas variables, pero si queremos analizar cada una por separado, podemos tomar la raiz del error cuadratico medio calculado para cada variable objetivo. En este caso para la primera variable nos quedaría una diferencia estimada en promedio de 0,02 y para la segunda variable, una diferencia estimada de 0.018 aunque obviamente estos valores pueden variar ligeramente en diferentes corridas del algoritmo."
      ],
      "metadata": {
        "id": "Pl3EgYpWRx4w"
      }
    }
  ]
}