chapter03_introduction-to-keras-and-tf.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "1yX7I6IkPe-N"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/ayrna/deep-learning-with-python-notebooks/blob/master/chapter03_introduction-to-keras-and-tf.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "vLlJ5DqIKTIS"
      },
      "source": [
        "Este cuaderno está basado en los cuadernos del libro [Deep Learning with Python, Second Edition](https://www.manning.com/books/deep-learning-with-python-second-edition?a_aid=keras&a_bid=76564dff).\n",
        "\n",
        "> This is a companion notebook for the book [Deep Learning with Python, Second Edition](https://www.manning.com/books/deep-learning-with-python-second-edition?a_aid=keras&a_bid=76564dff). For readability, it only contains runnable code blocks and section titles, and omits everything else in the book: text paragraphs, figures, and pseudocode.\n",
        "> **If you want to be able to follow what's going on, I recommend reading the notebook side by side with your copy of the book.**\n",
        "> This notebook was generated for TensorFlow 2.6."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Lzy5cZNbKTIX"
      },
      "source": [
        "# Introducción a Keras y TensorFlow\n",
        "\n",
        "![imagen.png](data:image/png;base64,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)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "pKyiVRJBKTIn"
      },
      "source": [
        "## Primeros pasos con TensorFlow\n",
        "\n",
        "Entrenar redes neuronales gira en torno a varios conceptos:\n",
        "\n",
        "1. La manipulación de tensores, la *infraestructura* que subyace a todo el aprendizaje automático moderno. Esto se traduce en las APIs de **TensorFlow**:\n",
        " * *Tensores*, incluyendo tensores especiales que almacenan el estado de la red (variables)\n",
        " * *Operaciones de tensor* como la `suma`, `relu`, `matmul`\n",
        " * *Retropropagación (Backpropagation)*, una forma de calcular el gradiente de expresiones matemáticas (usado TensorFlow a través del objeto `GradientTape`)\n",
        "2. Conceptos de aprendizaje profundo de alto nivel. Esto se traduce en las APIs de **Keras**:\n",
        " * *Capas*, que se combinan en un modelo\n",
        " * *Función de pérdida*, que define la señal de retroalimentación utilizada para el aprendizaje\n",
        " * *Optimizador*, que determina cómo se realiza el ajuste/aprendizaje a los datos.\n",
        " * *Métricas* para evaluar el rendimiento del modelo, como la precisión\n",
        " * Un *bucle de entrenamiento* que realiza el descenso de gradiente estocástico en mini lotes"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "jYFtk25gKTIo"
      },
      "source": [
        "#### Tensores constantes y variables"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "uVD4LEMgOPRC"
      },
      "source": [
        "Los tensores son la estructura de datos básica de TensorFlow y Keras. Deben crearse con un contenido y no se les puede asignar un valor.\n",
        "\n",
        "Cuando es necesario cambiar el estado de una variable necesitaremos el tipo `tf.Variable`, que deben crearse con un valor inicial también."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "NjdQ6GXzKTIo"
      },
      "source": [
        "**All-ones o all-zeros tensors**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "6DhYi-8CKTIp"
      },
      "outputs": [],
      "source": [
        "import tensorflow as tf\n",
        "x = tf.ones(shape=(2, 1))\n",
        "print(x)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "43DzLxyvKTIr"
      },
      "outputs": [],
      "source": [
        "x = tf.zeros(shape=(2, 1))\n",
        "print(x)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "gSA4b6GHKTIr"
      },
      "source": [
        "**Tensores aleatorios**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "wEmRkf-HKTIr"
      },
      "outputs": [],
      "source": [
        "x = tf.random.normal(shape=(3, 1), mean=0., stddev=1.)\n",
        "print(x)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "05CGedl5KTIs"
      },
      "outputs": [],
      "source": [
        "x = tf.random.uniform(shape=(3, 1), minval=0., maxval=1.)\n",
        "print(x)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "fhiR3zqfKTIt"
      },
      "source": [
        "**Los arrays de NumPy son asignables pero los tensores no**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "TdnfCFsUKTIt"
      },
      "outputs": [],
      "source": [
        "import numpy as np\n",
        "x = np.ones(shape=(2, 2))\n",
        "x[0, 0] = 0."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "hMELAOHUO9WI"
      },
      "outputs": [],
      "source": [
        "try:\n",
        "  x = tf.ones(shape=(2, 2))\n",
        "  x[0, 0] = 0.\n",
        "except Exception as e:\n",
        "  print(e)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "nNU6758FKTIt"
      },
      "source": [
        "**Crear un TensorFlow variable**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "R5Uubtf1KTIu"
      },
      "outputs": [],
      "source": [
        "v = tf.Variable(initial_value=tf.random.normal(shape=(3, 1)))\n",
        "print(v)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "tgiNms9hKTIv"
      },
      "source": [
        "**Asignar un valor a un TensorFlow variable**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "m8pAzAXbKTIv"
      },
      "outputs": [],
      "source": [
        "v.assign(tf.ones((3, 1)))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "v1iz3U4bKTIw"
      },
      "source": [
        "**Asignar un valor a un subconjunto de un TensorFlow variable**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "85C5QwR-KTIw"
      },
      "outputs": [],
      "source": [
        "v[0, 0].assign(3.)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "rDjhxQ-aKTIw"
      },
      "source": [
        "**Suma con `assign_add`**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "eStLXrkJKTIx"
      },
      "outputs": [],
      "source": [
        "v.assign_add(tf.ones((3, 1)))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "0bFwaV0LKTIy"
      },
      "source": [
        "#### Operaciones matemáticas con Tensores"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "rZ0pHWDxKTIy"
      },
      "source": [
        "**Operaciones básicas**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "WvFbtDZTKTIy"
      },
      "outputs": [],
      "source": [
        "a = tf.ones((2, 2))\n",
        "b = tf.square(a)\n",
        "c = tf.sqrt(a)\n",
        "d = b + c\n",
        "e = tf.matmul(a, b)\n",
        "e *= d"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "J0ISrkpSKTIy"
      },
      "source": [
        "#### GradientTape API"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "7vRfqlJQKTIy"
      },
      "source": [
        "**Cómo usar `GradientTape`**\n",
        "\n",
        "Basta con abrir un *ámbito* de `GradientTape`, aplicar algún cálculo a uno o varios tensores de entrada, y recuperar el gradiente del resultado con respecto a las entradas.\n",
        "\n",
        "$f(x)=x^2$\n",
        "\n",
        "$\n",
        "\\frac{\\partial f}{\\partial x} = 2x\n",
        "$\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "7u852m1xKTI0"
      },
      "outputs": [],
      "source": [
        "input_var = tf.Variable(initial_value=3.)\n",
        "with tf.GradientTape() as tape:\n",
        "   result = tf.square(input_var)\n",
        "gradient = tape.gradient(result, input_var)\n",
        "# la función sería f(x)=x^2; f'(x)=2x, siendo x = 3 f'(x)=6\n",
        "print(input_var)\n",
        "print(gradient)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "SmgIjR9aKTI0"
      },
      "source": [
        "**`GradientTape` con tensores constantes como entrada**"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "YRcTyxsiRvFN"
      },
      "source": [
        "Hasta ahora, sólo hemos visto el caso en que los tensores de entrada en `tape.gradient()` eran variables TensorFlow. En realidad es posible que estas entradas sean cualquier tensor arbitrario. Sin embargo, sólo las variables entrenables son *rastreadas* por defecto. Con un tensor constante, hay que marcarlo manualmente como rastreado llamando a `tape.watch()` sobre él. Esto se hace por motivos de eficiencia computacional."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "YTvvjZbKKTI1"
      },
      "outputs": [],
      "source": [
        "input_const = tf.constant(3.)\n",
        "with tf.GradientTape() as tape:\n",
        "   tape.watch(input_const)\n",
        "   result = tf.square(input_const)\n",
        "gradient = tape.gradient(result, input_const)\n",
        "print(result)\n",
        "print(gradient)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "UQH2QkHjKTI2"
      },
      "source": [
        "#### Un clasificador lineal completo en TensorFlow"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "h_CaD-nmKTI2"
      },
      "source": [
        "**Generación del conjunto de datos**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "AaRLZE7MKTI2"
      },
      "outputs": [],
      "source": [
        "num_samples_per_class = 1000\n",
        "negative_samples = np.random.multivariate_normal(\n",
        "    mean=[0, 3],\n",
        "    cov=[[1, 0.5],[0.5, 1]],\n",
        "    size=num_samples_per_class)\n",
        "positive_samples = np.random.multivariate_normal(\n",
        "    mean=[3, 0],\n",
        "    cov=[[1, 0.5],[0.5, 1]],\n",
        "    size=num_samples_per_class)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "N1wa_98KKTI3"
      },
      "outputs": [],
      "source": [
        "inputs = np.vstack((negative_samples, positive_samples)).astype(np.float32)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "IbNx-pwBKTI3"
      },
      "outputs": [],
      "source": [
        "targets = np.vstack((np.zeros((num_samples_per_class, 1), dtype=\"float32\"),\n",
        "                     np.ones((num_samples_per_class, 1), dtype=\"float32\")))"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "AZTuA8VWKTI4"
      },
      "outputs": [],
      "source": [
        "import matplotlib.pyplot as plt\n",
        "plt.scatter(inputs[:, 0], inputs[:, 1], c=targets[:, 0])\n",
        "plt.show()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "s-hKLGpdKTI4"
      },
      "source": [
        "**Creación de la variables del modelo lineal**\n",
        "\n",
        "Inicializamos `W` con valores aleatorios y `b` a cero.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "nqadhSZTKTI4"
      },
      "outputs": [],
      "source": [
        "input_dim = 2\n",
        "output_dim = 1\n",
        "W = tf.Variable(initial_value=tf.random.uniform(shape=(input_dim, output_dim)))\n",
        "b = tf.Variable(initial_value=tf.zeros(shape=(output_dim,)))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9Tj-9u9wKTI6"
      },
      "source": [
        "**Propagación hacia delante**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "ntvoFKavKTI6"
      },
      "outputs": [],
      "source": [
        "def model(inputs):\n",
        "    return tf.matmul(inputs, W) + b"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "7-_dDjgJKTI6"
      },
      "source": [
        "**Función de pérdia de error cuadrático medio (MSE)**"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "BWOFtzq9iuld"
      },
      "source": [
        "* `persample_loss` es un tensor con la misma forma que targets y predictions y almacenará la pérdida correspondiente a cada patrón.\n",
        "* Devolvemos la contribución media al error del lote. `reduce_mean` por defecto reduce todas las dimensiones a 1, pero podría utilizarse para calcular medias por ejes del tensor."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "n9gDDk0eKTI6"
      },
      "outputs": [],
      "source": [
        "def square_loss(targets, predictions):\n",
        "    per_sample_losses = tf.square(targets - predictions)\n",
        "    return tf.reduce_mean(per_sample_losses)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Q4NQxVCqKTI7"
      },
      "source": [
        "**El paso de entrenamiento**"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "R0-GJqnlhAR3"
      },
      "source": [
        "Recordad que `GradientTape` nos permite derivar respecto a listas de parámetros. En este caso nos devuelve el gradiente parcial respecto a cada parámetro de la lista."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "5YfXjhpkKTI7"
      },
      "outputs": [],
      "source": [
        "learning_rate = 0.1\n",
        "\n",
        "def training_step(inputs, targets):\n",
        "    with tf.GradientTape() as tape:\n",
        "        predictions = model(inputs)\n",
        "        loss = square_loss(targets, predictions)\n",
        "    grad_loss_wrt_W, grad_loss_wrt_b = tape.gradient(loss, [W, b])\n",
        "    # Actualizamos los pesos restando el gradiente (dirección contraria)\n",
        "    W.assign_sub(grad_loss_wrt_W * learning_rate)\n",
        "    b.assign_sub(grad_loss_wrt_b * learning_rate)\n",
        "    return loss"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "dBoJ6K_vKTI8"
      },
      "source": [
        "**Bucle de aprendizaje con el conjunto completo (batch training)**\n",
        "\n",
        "Para simplificar, haremos un entrenamiento por lotes en lugar de un mini-entrenamiento por lotes: ejecutaremos cada paso de entrenamiento (cálculo del gradiente y actualización de pesos) para todos los datos, en lugar de iterar sobre los datos en pequeños lotes.\n",
        "\n",
        "Esto significa que cada paso de entrenamiento tardará mucho más tiempo en ejecutarse pero cada actualización del gradiente será mucho más eficaz para reducir la función de pérdida al incluir todos los datos de entrenamiento. Esto significa que harán falta menos pasos de entrenamiento y que la tasa de aprendizaje debería ser mayor que al utilizar lotes pequeños (`learning_rate = 0.1`).\n",
        "\n",
        "❓❓❓ **EJERCICIO**\n",
        "* Prueba a utilizar distintas tasas de aprendizaje en este ejemplo y observa el resultado\n",
        "* ¿[Tiene sentido que la tasa de aprendizaje sea mayor que 1](https://ai.stackexchange.com/questions/23740/why-is-the-learning-rate-generally-beneath-1)?\n",
        "* Intenta programar otra función de pérdida: el RMSE.\n",
        "\n",
        "❓❓❓"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "j5y2nK1KKTI8"
      },
      "outputs": [],
      "source": [
        "for step in range(40):\n",
        "    loss = training_step(inputs, targets)\n",
        "    print(f\"Loss at step {step}: {loss:.4f}\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "KJjDn_v6p0b0"
      },
      "source": [
        "Como vemos la función de pérdida más o menos se estabiliza.\n",
        "\n",
        "Ahora vamos a predecir la clase de los patrones. En este ejemplo de clasificación binaria el umbral de decisión entre las clases es `0.5` donde se asigna la clase \"1\" para los valores mayores que el umbral."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "MLuywB26KTI8"
      },
      "outputs": [],
      "source": [
        "predictions = model(inputs)\n",
        "plt.scatter(inputs[:, 0], inputs[:, 1], c=predictions[:, 0] > 0.5)\n",
        "plt.show()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "OjvzPWm2KTI9"
      },
      "outputs": [],
      "source": [
        "x = np.linspace(-1, 4, 100)\n",
        "y = - W[0] /  W[1] * x + (0.5 - b) / W[1]\n",
        "plt.plot(x, y, \"-r\")\n",
        "plt.scatter(inputs[:, 0], inputs[:, 1], c=predictions[:, 0] > 0.5)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "GsbXu-6hKTI9"
      },
      "source": [
        "## Anatomía de una red neuronal para entender conceptos de Keras"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "atUcKC3MKTI-"
      },
      "source": [
        "### Capas (Layers)\n",
        "\n",
        "La estructura de datos fundamental en las redes neuronales es la *capa*. Una capa es un módulo de procesamiento de datos que toma como entrada uno o más tensores y que da como salida uno o más tensores. Algunas capas no tienen estado, pero lo más frecuente es que las capas tengan un estado: los *pesos* de la capa, uno o varios tensores aprendidos con el descenso por gradiente, que en conjunto almacenan el *conocimiento* de la red.\n",
        "\n",
        "Según el tipo de dato con el que trabajemos habrá tipos de capas más apropiados para trabajar con estos.\n",
        "* Datos vectoriales simples, también llamados tabulares (\"hoja de cálculo\"), almacenados en tensores de rango 2 `(muestra, características)`, suelen ser procesados por capas densamente conectadas, también llamadas capas totalmente conectadas (la clase `Dense` en Keras).\n",
        "* Los datos con disposición temporal o de secuencia, tensores de rango 3 `(muestras, pasos de tiempo, características)`, se procesan normalmente por capas recurrentes, como una capa `LSTM`, o capas de convolución 1D (`Conv1D`).\n",
        "* Los datos de imágenes, almacenados en tensores de rango 4, suelen ser procesados por capas de convolución 2D (`Conv2D`)."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "D21FZwD4KTI-"
      },
      "source": [
        "#### La clase `Layer` en Keras\n",
        "\n",
        "\n",
        "Casi todo en Keras gira en torno a un objeto de tipo Layer, que encapsula un estado de los pesos y un cálculo sobre la red (*forward pass*). Los pesos se definen en `build()` y el procesamiento en `call()`.\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "HgMFet7GKTI-"
      },
      "source": [
        "**Una capa `Dense` implementada como subclase de `Layer`**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "nKFXB9GSKTI-"
      },
      "outputs": [],
      "source": [
        "from tensorflow import keras\n",
        "\n",
        "class SimpleDense(keras.layers.Layer):\n",
        "\n",
        "    def __init__(self, units, activation=None):\n",
        "        super().__init__()\n",
        "        self.units = units\n",
        "        self.activation = activation\n",
        "\n",
        "    def build(self, input_shape):\n",
        "        input_dim = input_shape[-1]\n",
        "        self.W = self.add_weight(shape=(input_dim, self.units),\n",
        "                                 initializer=\"random_normal\")\n",
        "        self.b = self.add_weight(shape=(self.units,),\n",
        "                                 initializer=\"zeros\")\n",
        "\n",
        "    def call(self, inputs):\n",
        "        y = tf.matmul(inputs, self.W) + self.b\n",
        "        if self.activation is not None:\n",
        "            y = self.activation(y)\n",
        "        return y"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_Lp8AkEyu02i"
      },
      "source": [
        "Esta capa que hemos creado se puede instanciar y usarse como una función:"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "eZmz_BQRKTI_"
      },
      "outputs": [],
      "source": [
        "my_dense = SimpleDense(units=32, activation=tf.nn.relu)\n",
        "input_tensor = tf.ones(shape=(2, 784))\n",
        "output_tensor = my_dense(input_tensor)\n",
        "print(output_tensor.shape)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "l-3t_thIu-1s"
      },
      "outputs": [],
      "source": [
        "my_dense = SimpleDense(units=512, activation=tf.nn.tanh)\n",
        "input_tensor = tf.ones(shape=(2, 784))\n",
        "output_tensor = my_dense(input_tensor)\n",
        "print(output_tensor.shape)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "tYqRul6oKTI_"
      },
      "source": [
        "#### Deducción de la forma de la capa automática: construcción de capas sobre la marcha"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "4r32G-GMvemc"
      },
      "source": [
        "En Keras en general sólo necesitaremos especificar la forma de los tensores de salida de la capa, ya que las capas que se añaden al modelo se construyen dinámicamente para adaptar cada capa a la forma de la capa de entrada (la salida de la capa anterior)."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "aeaWJ7kvv9vM"
      },
      "source": [
        "La siguiente capa sólo puede conectarse a una capa que espere vectores de 32 dimensiones como entrada."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "SG87ONtCKTI_"
      },
      "outputs": [],
      "source": [
        "from tensorflow.keras import layers\n",
        "layer = layers.Dense(32, activation=\"relu\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "q31sRXyVwE-_"
      },
      "source": [
        "En Keras, no se implementa la construcción de la capa del todo en el constructor `__build__()` en espera de conocer cuál será la forma de la entrada desde la capa anterior."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "i0iPGmzDKTI_"
      },
      "outputs": [],
      "source": [
        "from tensorflow.keras import models\n",
        "from tensorflow.keras import layers\n",
        "model = models.Sequential([\n",
        "    layers.Dense(32, activation=\"relu\"),\n",
        "    layers.Dense(32)\n",
        "])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "E6YKkaxtwcQx"
      },
      "source": [
        "En el ejemplo de red que hicimos desde cero necesitábamos especificar perfectamente las dimensiones de entrada y salida de la secuencia de capas:\n",
        "\n",
        "```\n",
        "model = NaiveSequential([\n",
        "    NaiveDense(input_size=784, output_size=32, activation=\"relu\"),\n",
        "    NaiveDense(input_size=32, output_size=64, activation=\"relu\"),\n",
        "    NaiveDense(input_size=64, output_size=32, activation=\"relu\"),\n",
        "    NaiveDense(input_size=32, output_size=10, activation=\"softmax\")\n",
        "])\n",
        "```\n",
        "\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "GOwolWkWwnoI"
      },
      "source": [
        "En Keras basta con lo siguiente:"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "jZgfKF5pKTJA"
      },
      "outputs": [],
      "source": [
        "model = keras.Sequential([\n",
        "    SimpleDense(32, activation=\"relu\"),\n",
        "    SimpleDense(64, activation=\"relu\"),\n",
        "    SimpleDense(32, activation=\"relu\"),\n",
        "    SimpleDense(10, activation=\"softmax\")\n",
        "])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "AqGUwElFKTJA"
      },
      "source": [
        "### De las capas a los modelos\n",
        "\n",
        "Un modelo de aprendizaje profundo consiste en un grafo de capas, clase `Model` en Keras. Hasta ahora sólo hemos visto el modelo `Sequential`, que es una subclase de `Model` pero existen muchas otras topologías de red más avanzadas (Transformers, Residual...)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "vfY2dELVKTJA"
      },
      "source": [
        "### El paso \"compile\" step para configurar el proceso de aprendizaje"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "YD_WVkPnxwC4"
      },
      "source": [
        "Aquí vamos a configurar:\n",
        "* Función de pérdida (o función objetivo).\n",
        "* El optimizador.\n",
        "* Métricas de evaluación sobre el conjunto de train y opcionalmente validación.\n",
        "\n",
        "Hay dos alternativas para especificar:"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "--uxOt36KTJB"
      },
      "outputs": [],
      "source": [
        "model = keras.Sequential([keras.layers.Dense(1)])\n",
        "model.compile(optimizer=\"rmsprop\",\n",
        "              loss=\"mean_squared_error\",\n",
        "              metrics=[\"accuracy\"])"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "7UONW9yKKTJB"
      },
      "outputs": [],
      "source": [
        "model.compile(optimizer=keras.optimizers.RMSprop(),\n",
        "              loss=keras.losses.MeanSquaredError(),\n",
        "              metrics=[keras.metrics.BinaryAccuracy()])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "VFumPtBWKTJB"
      },
      "source": [
        "### La importancia de la función de pérdida ❗❗❗❗\n",
        "\n",
        "La elección de la función de pérdida correcta para el problema adecuado es extremadamente importante: la red tomará cualquier atajo que pueda para minimizar la pérdida, por lo que si el objetivo no está totalmente correlacionado con la tarea en cuestión, la red terminará haciendo cosas que probablemente no queramos.\n",
        "\n",
        "Ejemplos:\n",
        "* ¿detector de emociones?\n",
        "* ¿maximizar el bienestar promedio de los seres humanos?"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "RNE9EC25KTJC"
      },
      "source": [
        "### El método fit()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Fa6z8y8lKTJC"
      },
      "source": [
        "**Calling `fit()` with NumPy data**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "1AQi7fsRKTJC"
      },
      "outputs": [],
      "source": [
        "history = model.fit(\n",
        "    inputs,\n",
        "    targets,\n",
        "    epochs=5,\n",
        "    batch_size=128\n",
        ")"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "uQKA8RTNKTJD"
      },
      "outputs": [],
      "source": [
        "history.history"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "fWOpQ7csKTJD"
      },
      "source": [
        "### Monitorización de la pérdida y las métricas sobre datos de validación"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "fXJb3QNdKTJD"
      },
      "source": [
        "**Parámetro `validation_data`**\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "tEn2yZFyKTJD"
      },
      "outputs": [],
      "source": [
        "model = keras.Sequential([keras.layers.Dense(1)])\n",
        "model.compile(optimizer=keras.optimizers.RMSprop(learning_rate=0.1),\n",
        "              loss=keras.losses.MeanSquaredError(),\n",
        "              metrics=[keras.metrics.BinaryAccuracy()])\n",
        "\n",
        "indices_permutation = np.random.permutation(len(inputs))\n",
        "shuffled_inputs = inputs[indices_permutation]\n",
        "shuffled_targets = targets[indices_permutation]\n",
        "\n",
        "num_validation_samples = int(0.3 * len(inputs))\n",
        "val_inputs = shuffled_inputs[:num_validation_samples]\n",
        "val_targets = shuffled_targets[:num_validation_samples]\n",
        "training_inputs = shuffled_inputs[num_validation_samples:]\n",
        "training_targets = shuffled_targets[num_validation_samples:]\n",
        "model.fit(\n",
        "    training_inputs,\n",
        "    training_targets,\n",
        "    epochs=5,\n",
        "    batch_size=16,\n",
        "    validation_data=(val_inputs, val_targets)\n",
        ")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "lmVx8Z-qKTJE"
      },
      "source": [
        "### Cómo usar el modelo entrenado"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "jffLo6HvzF4S"
      },
      "source": [
        "Se puede llamar al modelo como una función, pero esto evaluaría toda la base de datos de test directamente en lugar de por lotes.\n",
        "\n",
        "```\n",
        "predictions = model(new_inputs)\n",
        "```\n",
        "\n",
        "Habitualmente será más adecuado usar la función `predict`"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "__d2SikiKTJE"
      },
      "outputs": [],
      "source": [
        "predictions = model.predict(val_inputs, batch_size=128)\n",
        "print(predictions[:10])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "rIK1RSHxzi7X"
      },
      "source": [
        "❓❓❓\n",
        "**EJERCICIO**\n",
        "\n",
        "* Prueba a utilizar el modelo anterior de 4 capas con la base de datos MNIST\n",
        "  * No bastará con copiar y pegar. Presta atención a: variables de entrada, función de pérdida adecuada, neuronas por capa... Deberás calcular la precisión media en test.\n",
        "* Prueba a visualizar el historial de ajuste del modelo\n",
        "\n",
        "❓❓❓"
      ]
    },
    {
      "cell_type": "code",
      "source": [],
      "metadata": {
        "id": "bxhnOcIqftgz"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Puedes usar este código para visualizar el histórico de entrenamiento\n",
        "import matplotlib.pyplot as plt\n",
        "history_dict = history.history\n",
        "loss_values = history_dict[\"loss\"]\n",
        "acc_values = history_dict[\"accuracy\"]\n",
        "plt.plot(loss_values, \"b-\", label=\"Training loss\")\n",
        "plt.plot(acc_values, \"bo\", label=\"Training accuracy\")\n",
        "plt.xlabel(\"Epochs\")\n",
        "plt.ylabel(\"Loss/Accuracy\")\n",
        "plt.legend()"
      ],
      "metadata": {
        "id": "W4-ZgRhDhSEB"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "qT_q7bpOKTJE"
      },
      "source": [
        "## Resumen\n",
        "* TensorFlow es un *framework* que puede ejecutarse en CPU, GPU y TPU.\n",
        "* TensorFlow puede calcular el gradiente de cualquier expresión diferenciable.\n",
        "* Keras es un API para hacer modelos de aprendizaje profundo con TF.\n",
        "* Objetos claves de TF: tensores, variables, operaciones con tensores y *GradientTape*.\n",
        "* La clase central de Keras es `Layer`, que encapsula algunos pesos y cálculos. Las capas se unen o ensamblan en **modelos**.\n",
        "* Antes de entrenar un modelo es necesario elegir un **optimizador**, una **función de pérdida** y una o varias **métricas**: `model.compile()`\n",
        "* El método `model.fit()` ejecuta el descenso por gradiente por mini-lotes. Permite monitorizar la pérdida y métricas y incluir datos de **validación**.\n",
        "* Con el modelo entrenado, se pueden hacer predicciones con `model.predict()`."
      ]
    }
  ],
  "metadata": {
    "colab": {
      "name": "chapter03_introduction-to-keras-and-tf.i",
      "provenance": [],
      "toc_visible": true
    },
    "kernelspec": {
      "display_name": "Python 3 (ipykernel)",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.9.16"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}