handson-ml/11_training_deep_neural_networks.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**Chapter 11 – Training Deep Neural Networks**"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "_This notebook contains all the sample code and solutions to the exercises in chapter 11._"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<table align=\"left\">\n",
    "  <td>\n",
    "    <a target=\"_blank\" href=\"https://colab.research.google.com/github/ageron/handson-ml2/blob/master/11_training_deep_neural_networks.ipynb\"><img src=\"https://www.tensorflow.org/images/colab_logo_32px.png\" />Run in Google Colab</a>\n",
    "  </td>\n",
    "</table>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Setup"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "First, let's import a few common modules, ensure MatplotLib plots figures inline and prepare a function to save the figures. We also check that Python 3.5 or later is installed (although Python 2.x may work, it is deprecated so we strongly recommend you use Python 3 instead), as well as Scikit-Learn ≥0.20 and TensorFlow ≥2.0."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Python ≥3.5 is required\n",
    "import sys\n",
    "assert sys.version_info >= (3, 5)\n",
    "\n",
    "# Scikit-Learn ≥0.20 is required\n",
    "import sklearn\n",
    "assert sklearn.__version__ >= \"0.20\"\n",
    "\n",
    "try:\n",
    "    # %tensorflow_version only exists in Colab.\n",
    "    %tensorflow_version 2.x\n",
    "except Exception:\n",
    "    pass\n",
    "\n",
    "# TensorFlow ≥2.0 is required\n",
    "import tensorflow as tf\n",
    "from tensorflow import keras\n",
    "assert tf.__version__ >= \"2.0\"\n",
    "\n",
    "%load_ext tensorboard\n",
    "\n",
    "# Common imports\n",
    "import numpy as np\n",
    "import os\n",
    "\n",
    "# to make this notebook's output stable across runs\n",
    "np.random.seed(42)\n",
    "\n",
    "# To plot pretty figures\n",
    "%matplotlib inline\n",
    "import matplotlib as mpl\n",
    "import matplotlib.pyplot as plt\n",
    "mpl.rc('axes', labelsize=14)\n",
    "mpl.rc('xtick', labelsize=12)\n",
    "mpl.rc('ytick', labelsize=12)\n",
    "\n",
    "# Where to save the figures\n",
    "PROJECT_ROOT_DIR = \".\"\n",
    "CHAPTER_ID = \"deep\"\n",
    "IMAGES_PATH = os.path.join(PROJECT_ROOT_DIR, \"images\", CHAPTER_ID)\n",
    "os.makedirs(IMAGES_PATH, exist_ok=True)\n",
    "\n",
    "def save_fig(fig_id, tight_layout=True, fig_extension=\"png\", resolution=300):\n",
    "    path = os.path.join(IMAGES_PATH, fig_id + \".\" + fig_extension)\n",
    "    print(\"Saving figure\", fig_id)\n",
    "    if tight_layout:\n",
    "        plt.tight_layout()\n",
    "    plt.savefig(path, format=fig_extension, dpi=resolution)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Vanishing/Exploding Gradients Problem"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "def logit(z):\n",
    "    return 1 / (1 + np.exp(-z))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "z = np.linspace(-5, 5, 200)\n",
    "\n",
    "plt.plot([-5, 5], [0, 0], 'k-')\n",
    "plt.plot([-5, 5], [1, 1], 'k--')\n",
    "plt.plot([0, 0], [-0.2, 1.2], 'k-')\n",
    "plt.plot([-5, 5], [-3/4, 7/4], 'g--')\n",
    "plt.plot(z, logit(z), \"b-\", linewidth=2)\n",
    "props = dict(facecolor='black', shrink=0.1)\n",
    "plt.annotate('Saturating', xytext=(3.5, 0.7), xy=(5, 1), arrowprops=props, fontsize=14, ha=\"center\")\n",
    "plt.annotate('Saturating', xytext=(-3.5, 0.3), xy=(-5, 0), arrowprops=props, fontsize=14, ha=\"center\")\n",
    "plt.annotate('Linear', xytext=(2, 0.2), xy=(0, 0.5), arrowprops=props, fontsize=14, ha=\"center\")\n",
    "plt.grid(True)\n",
    "plt.title(\"Sigmoid activation function\", fontsize=14)\n",
    "plt.axis([-5, 5, -0.2, 1.2])\n",
    "\n",
    "save_fig(\"sigmoid_saturation_plot\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Xavier and He Initialization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "[name for name in dir(keras.initializers) if not name.startswith(\"_\")]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.layers.Dense(10, activation=\"relu\", kernel_initializer=\"he_normal\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "init = keras.initializers.VarianceScaling(scale=2., mode='fan_avg',\n",
    "                                          distribution='uniform')\n",
    "keras.layers.Dense(10, activation=\"relu\", kernel_initializer=init)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Nonsaturating Activation Functions"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Leaky ReLU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def leaky_relu(z, alpha=0.01):\n",
    "    return np.maximum(alpha*z, z)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.plot(z, leaky_relu(z, 0.05), \"b-\", linewidth=2)\n",
    "plt.plot([-5, 5], [0, 0], 'k-')\n",
    "plt.plot([0, 0], [-0.5, 4.2], 'k-')\n",
    "plt.grid(True)\n",
    "props = dict(facecolor='black', shrink=0.1)\n",
    "plt.annotate('Leak', xytext=(-3.5, 0.5), xy=(-5, -0.2), arrowprops=props, fontsize=14, ha=\"center\")\n",
    "plt.title(\"Leaky ReLU activation function\", fontsize=14)\n",
    "plt.axis([-5, 5, -0.5, 4.2])\n",
    "\n",
    "save_fig(\"leaky_relu_plot\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "[m for m in dir(keras.activations) if not m.startswith(\"_\")]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "[m for m in dir(keras.layers) if \"relu\" in m.lower()]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's train a neural network on Fashion MNIST using the Leaky ReLU:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "(X_train_full, y_train_full), (X_test, y_test) = keras.datasets.fashion_mnist.load_data()\n",
    "X_train_full = X_train_full / 255.0\n",
    "X_test = X_test / 255.0\n",
    "X_valid, X_train = X_train_full[:5000], X_train_full[5000:]\n",
    "y_valid, y_train = y_train_full[:5000], y_train_full[5000:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, kernel_initializer=\"he_normal\"),\n",
    "    keras.layers.LeakyReLU(),\n",
    "    keras.layers.Dense(100, kernel_initializer=\"he_normal\"),\n",
    "    keras.layers.LeakyReLU(),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "history = model.fit(X_train, y_train, epochs=10,\n",
    "                    validation_data=(X_valid, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's try PReLU:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, kernel_initializer=\"he_normal\"),\n",
    "    keras.layers.PReLU(),\n",
    "    keras.layers.Dense(100, kernel_initializer=\"he_normal\"),\n",
    "    keras.layers.PReLU(),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model.fit(X_train, y_train, epochs=10,\n",
    "                    validation_data=(X_valid, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### ELU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def elu(z, alpha=1):\n",
    "    return np.where(z < 0, alpha * (np.exp(z) - 1), z)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.plot(z, elu(z), \"b-\", linewidth=2)\n",
    "plt.plot([-5, 5], [0, 0], 'k-')\n",
    "plt.plot([-5, 5], [-1, -1], 'k--')\n",
    "plt.plot([0, 0], [-2.2, 3.2], 'k-')\n",
    "plt.grid(True)\n",
    "plt.title(r\"ELU activation function ($\\alpha=1$)\", fontsize=14)\n",
    "plt.axis([-5, 5, -2.2, 3.2])\n",
    "\n",
    "save_fig(\"elu_plot\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Implementing ELU in TensorFlow is trivial, just specify the activation function when building each layer:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.layers.Dense(10, activation=\"elu\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### SELU"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This activation function was proposed in this [great paper](https://arxiv.org/pdf/1706.02515.pdf) by Günter Klambauer, Thomas Unterthiner and Andreas Mayr, published in June 2017. During training, a neural network composed exclusively of a stack of dense layers using the SELU activation function and LeCun initialization will self-normalize: the output of each layer will tend to preserve the same mean and variance during training, which solves the vanishing/exploding gradients problem. As a result, this activation function outperforms the other activation functions very significantly for such neural nets, so you should really try it out. Unfortunately, the self-normalizing property of the SELU activation function is easily broken: you cannot use ℓ<sub>1</sub> or ℓ<sub>2</sub> regularization, regular dropout, max-norm, skip connections or other non-sequential topologies (so recurrent neural networks won't self-normalize). However, in practice it works quite well with sequential CNNs. If you break self-normalization, SELU will not necessarily outperform other activation functions."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "from scipy.special import erfc\n",
    "\n",
    "# alpha and scale to self normalize with mean 0 and standard deviation 1\n",
    "# (see equation 14 in the paper):\n",
    "alpha_0_1 = -np.sqrt(2 / np.pi) / (erfc(1/np.sqrt(2)) * np.exp(1/2) - 1)\n",
    "scale_0_1 = (1 - erfc(1 / np.sqrt(2)) * np.sqrt(np.e)) * np.sqrt(2 * np.pi) * (2 * erfc(np.sqrt(2))*np.e**2 + np.pi*erfc(1/np.sqrt(2))**2*np.e - 2*(2+np.pi)*erfc(1/np.sqrt(2))*np.sqrt(np.e)+np.pi+2)**(-1/2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "def selu(z, scale=scale_0_1, alpha=alpha_0_1):\n",
    "    return scale * elu(z, alpha)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.plot(z, selu(z), \"b-\", linewidth=2)\n",
    "plt.plot([-5, 5], [0, 0], 'k-')\n",
    "plt.plot([-5, 5], [-1.758, -1.758], 'k--')\n",
    "plt.plot([0, 0], [-2.2, 3.2], 'k-')\n",
    "plt.grid(True)\n",
    "plt.title(\"SELU activation function\", fontsize=14)\n",
    "plt.axis([-5, 5, -2.2, 3.2])\n",
    "\n",
    "save_fig(\"selu_plot\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "By default, the SELU hyperparameters (`scale` and `alpha`) are tuned in such a way that the mean output of each neuron remains close to 0, and the standard deviation remains close to 1 (assuming the inputs are standardized with mean 0 and standard deviation 1 too). Using this activation function, even a 1,000 layer deep neural network preserves roughly mean 0 and standard deviation 1 across all layers, avoiding the exploding/vanishing gradients problem:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.random.seed(42)\n",
    "Z = np.random.normal(size=(500, 100)) # standardized inputs\n",
    "for layer in range(1000):\n",
    "    W = np.random.normal(size=(100, 100), scale=np.sqrt(1 / 100)) # LeCun initialization\n",
    "    Z = selu(np.dot(Z, W))\n",
    "    means = np.mean(Z, axis=0).mean()\n",
    "    stds = np.std(Z, axis=0).mean()\n",
    "    if layer % 100 == 0:\n",
    "        print(\"Layer {}: mean {:.2f}, std deviation {:.2f}\".format(layer, means, stds))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Using SELU is easy:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.layers.Dense(10, activation=\"selu\",\n",
    "                   kernel_initializer=\"lecun_normal\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's create a neural net for Fashion MNIST with 100 hidden layers, using the SELU activation function:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.random.seed(42)\n",
    "tf.random.set_seed(42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential()\n",
    "model.add(keras.layers.Flatten(input_shape=[28, 28]))\n",
    "model.add(keras.layers.Dense(300, activation=\"selu\",\n",
    "                             kernel_initializer=\"lecun_normal\"))\n",
    "for layer in range(99):\n",
    "    model.add(keras.layers.Dense(100, activation=\"selu\",\n",
    "                                 kernel_initializer=\"lecun_normal\"))\n",
    "model.add(keras.layers.Dense(10, activation=\"softmax\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's train it. Do not forget to scale the inputs to mean 0 and standard deviation 1:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "pixel_means = X_train.mean(axis=0, keepdims=True)\n",
    "pixel_stds = X_train.std(axis=0, keepdims=True)\n",
    "X_train_scaled = (X_train - pixel_means) / pixel_stds\n",
    "X_valid_scaled = (X_valid - pixel_means) / pixel_stds\n",
    "X_test_scaled = (X_test - pixel_means) / pixel_stds"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model.fit(X_train_scaled, y_train, epochs=5,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now look at what happens if we try to use the ReLU activation function instead:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.random.seed(42)\n",
    "tf.random.set_seed(42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential()\n",
    "model.add(keras.layers.Flatten(input_shape=[28, 28]))\n",
    "model.add(keras.layers.Dense(300, activation=\"relu\", kernel_initializer=\"he_normal\"))\n",
    "for layer in range(99):\n",
    "    model.add(keras.layers.Dense(100, activation=\"relu\", kernel_initializer=\"he_normal\"))\n",
    "model.add(keras.layers.Dense(10, activation=\"softmax\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model.fit(X_train_scaled, y_train, epochs=5,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Not great at all, we suffered from the vanishing/exploding gradients problem."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Batch Normalization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.BatchNormalization(),\n",
    "    keras.layers.Dense(300, activation=\"relu\"),\n",
    "    keras.layers.BatchNormalization(),\n",
    "    keras.layers.Dense(100, activation=\"relu\"),\n",
    "    keras.layers.BatchNormalization(),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [],
   "source": [
    "bn1 = model.layers[1]\n",
    "[(var.name, var.trainable) for var in bn1.variables]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [],
   "source": [
    "#bn1.updates #deprecated"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model.fit(X_train, y_train, epochs=10,\n",
    "                    validation_data=(X_valid, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Sometimes applying BN before the activation function works better (there's a debate on this topic). Moreover, the layer before a `BatchNormalization` layer does not need to have bias terms, since the `BatchNormalization` layer some as well, it would be a waste of parameters, so you can set `use_bias=False` when creating those layers:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.BatchNormalization(),\n",
    "    keras.layers.Dense(300, use_bias=False),\n",
    "    keras.layers.BatchNormalization(),\n",
    "    keras.layers.Activation(\"relu\"),\n",
    "    keras.layers.Dense(100, use_bias=False),\n",
    "    keras.layers.BatchNormalization(),\n",
    "    keras.layers.Activation(\"relu\"),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model.fit(X_train, y_train, epochs=10,\n",
    "                    validation_data=(X_valid, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Gradient Clipping"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "All Keras optimizers accept `clipnorm` or `clipvalue` arguments:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.SGD(clipvalue=1.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.SGD(clipnorm=1.0)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Reusing Pretrained Layers"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Reusing a Keras model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's split the fashion MNIST training set in two:\n",
    "* `X_train_A`: all images of all items except for sandals and shirts (classes 5 and 6).\n",
    "* `X_train_B`: a much smaller training set of just the first 200 images of sandals or shirts.\n",
    "\n",
    "The validation set and the test set are also split this way, but without restricting the number of images.\n",
    "\n",
    "We will train a model on set A (classification task with 8 classes), and try to reuse it to tackle set B (binary classification). We hope to transfer a little bit of knowledge from task A to task B, since classes in set A (sneakers, ankle boots, coats, t-shirts, etc.) are somewhat similar to classes in set B (sandals and shirts). However, since we are using `Dense` layers, only patterns that occur at the same location can be reused (in contrast, convolutional layers will transfer much better, since learned patterns can be detected anywhere on the image, as we will see in the CNN chapter)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [],
   "source": [
    "def split_dataset(X, y):\n",
    "    y_5_or_6 = (y == 5) | (y == 6) # sandals or shirts\n",
    "    y_A = y[~y_5_or_6]\n",
    "    y_A[y_A > 6] -= 2 # class indices 7, 8, 9 should be moved to 5, 6, 7\n",
    "    y_B = (y[y_5_or_6] == 6).astype(np.float32) # binary classification task: is it a shirt (class 6)?\n",
    "    return ((X[~y_5_or_6], y_A),\n",
    "            (X[y_5_or_6], y_B))\n",
    "\n",
    "(X_train_A, y_train_A), (X_train_B, y_train_B) = split_dataset(X_train, y_train)\n",
    "(X_valid_A, y_valid_A), (X_valid_B, y_valid_B) = split_dataset(X_valid, y_valid)\n",
    "(X_test_A, y_test_A), (X_test_B, y_test_B) = split_dataset(X_test, y_test)\n",
    "X_train_B = X_train_B[:200]\n",
    "y_train_B = y_train_B[:200]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [],
   "source": [
    "X_train_A.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [],
   "source": [
    "X_train_B.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_train_A[:30]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_train_B[:30]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_A = keras.models.Sequential()\n",
    "model_A.add(keras.layers.Flatten(input_shape=[28, 28]))\n",
    "for n_hidden in (300, 100, 50, 50, 50):\n",
    "    model_A.add(keras.layers.Dense(n_hidden, activation=\"selu\"))\n",
    "model_A.add(keras.layers.Dense(8, activation=\"softmax\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_A.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "                optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "                metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model_A.fit(X_train_A, y_train_A, epochs=20,\n",
    "                    validation_data=(X_valid_A, y_valid_A))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_A.save(\"my_model_A.h5\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_B = keras.models.Sequential()\n",
    "model_B.add(keras.layers.Flatten(input_shape=[28, 28]))\n",
    "for n_hidden in (300, 100, 50, 50, 50):\n",
    "    model_B.add(keras.layers.Dense(n_hidden, activation=\"selu\"))\n",
    "model_B.add(keras.layers.Dense(1, activation=\"sigmoid\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_B.compile(loss=\"binary_crossentropy\",\n",
    "                optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "                metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model_B.fit(X_train_B, y_train_B, epochs=20,\n",
    "                      validation_data=(X_valid_B, y_valid_B))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_B.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_A = keras.models.load_model(\"my_model_A.h5\")\n",
    "model_B_on_A = keras.models.Sequential(model_A.layers[:-1])\n",
    "model_B_on_A.add(keras.layers.Dense(1, activation=\"sigmoid\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 61,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_A_clone = keras.models.clone_model(model_A)\n",
    "model_A_clone.set_weights(model_A.get_weights())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 62,
   "metadata": {},
   "outputs": [],
   "source": [
    "for layer in model_B_on_A.layers[:-1]:\n",
    "    layer.trainable = False\n",
    "\n",
    "model_B_on_A.compile(loss=\"binary_crossentropy\",\n",
    "                     optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "                     metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 63,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model_B_on_A.fit(X_train_B, y_train_B, epochs=4,\n",
    "                           validation_data=(X_valid_B, y_valid_B))\n",
    "\n",
    "for layer in model_B_on_A.layers[:-1]:\n",
    "    layer.trainable = True\n",
    "\n",
    "model_B_on_A.compile(loss=\"binary_crossentropy\",\n",
    "                     optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "                     metrics=[\"accuracy\"])\n",
    "history = model_B_on_A.fit(X_train_B, y_train_B, epochs=16,\n",
    "                           validation_data=(X_valid_B, y_valid_B))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "So, what's the final verdict?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 64,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_B.evaluate(X_test_B, y_test_B)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_B_on_A.evaluate(X_test_B, y_test_B)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Great! We got quite a bit of transfer: the error rate dropped by a factor of 4.5!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 66,
   "metadata": {},
   "outputs": [],
   "source": [
    "(100 - 97.05) / (100 - 99.35)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Faster Optimizers"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Momentum optimization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.SGD(lr=0.001, momentum=0.9)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Nesterov Accelerated Gradient"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 68,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.SGD(lr=0.001, momentum=0.9, nesterov=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## AdaGrad"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 69,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.Adagrad(lr=0.001)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## RMSProp"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 70,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.RMSprop(lr=0.001, rho=0.9)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Adam Optimization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 71,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Adamax Optimization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 72,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.Adamax(lr=0.001, beta_1=0.9, beta_2=0.999)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Nadam Optimization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.Nadam(lr=0.001, beta_1=0.9, beta_2=0.999)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Learning Rate Scheduling"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Power Scheduling"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```lr = lr0 / (1 + steps / s)**c```\n",
    "* Keras uses `c=1` and `s = 1 / decay`"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 74,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.SGD(lr=0.01, decay=1e-4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 75,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 76,
   "metadata": {},
   "outputs": [],
   "source": [
    "n_epochs = 25\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 77,
   "metadata": {},
   "outputs": [],
   "source": [
    "import math\n",
    "\n",
    "learning_rate = 0.01\n",
    "decay = 1e-4\n",
    "batch_size = 32\n",
    "n_steps_per_epoch = math.ceil(len(X_train) / batch_size)\n",
    "epochs = np.arange(n_epochs)\n",
    "lrs = learning_rate / (1 + decay * epochs * n_steps_per_epoch)\n",
    "\n",
    "plt.plot(epochs, lrs,  \"o-\")\n",
    "plt.axis([0, n_epochs - 1, 0, 0.01])\n",
    "plt.xlabel(\"Epoch\")\n",
    "plt.ylabel(\"Learning Rate\")\n",
    "plt.title(\"Power Scheduling\", fontsize=14)\n",
    "plt.grid(True)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Exponential Scheduling"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```lr = lr0 * 0.1**(epoch / s)```"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "metadata": {},
   "outputs": [],
   "source": [
    "def exponential_decay_fn(epoch):\n",
    "    return 0.01 * 0.1**(epoch / 20)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 79,
   "metadata": {},
   "outputs": [],
   "source": [
    "def exponential_decay(lr0, s):\n",
    "    def exponential_decay_fn(epoch):\n",
    "        return lr0 * 0.1**(epoch / s)\n",
    "    return exponential_decay_fn\n",
    "\n",
    "exponential_decay_fn = exponential_decay(lr0=0.01, s=20)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 80,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=\"nadam\", metrics=[\"accuracy\"])\n",
    "n_epochs = 25"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 81,
   "metadata": {},
   "outputs": [],
   "source": [
    "lr_scheduler = keras.callbacks.LearningRateScheduler(exponential_decay_fn)\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid),\n",
    "                    callbacks=[lr_scheduler])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 82,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.plot(history.epoch, history.history[\"lr\"], \"o-\")\n",
    "plt.axis([0, n_epochs - 1, 0, 0.011])\n",
    "plt.xlabel(\"Epoch\")\n",
    "plt.ylabel(\"Learning Rate\")\n",
    "plt.title(\"Exponential Scheduling\", fontsize=14)\n",
    "plt.grid(True)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The schedule function can take the current learning rate as a second argument:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 83,
   "metadata": {},
   "outputs": [],
   "source": [
    "def exponential_decay_fn(epoch, lr):\n",
    "    return lr * 0.1**(1 / 20)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "If you want to update the learning rate at each iteration rather than at each epoch, you must write your own callback class:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 84,
   "metadata": {},
   "outputs": [],
   "source": [
    "K = keras.backend\n",
    "\n",
    "class ExponentialDecay(keras.callbacks.Callback):\n",
    "    def __init__(self, s=40000):\n",
    "        super().__init__()\n",
    "        self.s = s\n",
    "\n",
    "    def on_batch_begin(self, batch, logs=None):\n",
    "        # Note: the `batch` argument is reset at each epoch\n",
    "        lr = K.get_value(self.model.optimizer.lr)\n",
    "        K.set_value(self.model.optimizer.lr, lr * 0.1**(1 / s))\n",
    "\n",
    "    def on_epoch_end(self, epoch, logs=None):\n",
    "        logs = logs or {}\n",
    "        logs['lr'] = K.get_value(self.model.optimizer.lr)\n",
    "\n",
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "lr0 = 0.01\n",
    "optimizer = keras.optimizers.Nadam(lr=lr0)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])\n",
    "n_epochs = 25\n",
    "\n",
    "s = 20 * len(X_train) // 32 # number of steps in 20 epochs (batch size = 32)\n",
    "exp_decay = ExponentialDecay(s)\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid),\n",
    "                    callbacks=[exp_decay])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 85,
   "metadata": {},
   "outputs": [],
   "source": [
    "n_steps = n_epochs * len(X_train) // 32\n",
    "steps = np.arange(n_steps)\n",
    "lrs = lr0 * 0.1**(steps / s)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 86,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "plt.plot(steps, lrs, \"-\", linewidth=2)\n",
    "plt.axis([0, n_steps - 1, 0, lr0 * 1.1])\n",
    "plt.xlabel(\"Batch\")\n",
    "plt.ylabel(\"Learning Rate\")\n",
    "plt.title(\"Exponential Scheduling (per batch)\", fontsize=14)\n",
    "plt.grid(True)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Piecewise Constant Scheduling"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 87,
   "metadata": {},
   "outputs": [],
   "source": [
    "def piecewise_constant_fn(epoch):\n",
    "    if epoch < 5:\n",
    "        return 0.01\n",
    "    elif epoch < 15:\n",
    "        return 0.005\n",
    "    else:\n",
    "        return 0.001"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 88,
   "metadata": {},
   "outputs": [],
   "source": [
    "def piecewise_constant(boundaries, values):\n",
    "    boundaries = np.array([0] + boundaries)\n",
    "    values = np.array(values)\n",
    "    def piecewise_constant_fn(epoch):\n",
    "        return values[np.argmax(boundaries > epoch) - 1]\n",
    "    return piecewise_constant_fn\n",
    "\n",
    "piecewise_constant_fn = piecewise_constant([5, 15], [0.01, 0.005, 0.001])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 89,
   "metadata": {},
   "outputs": [],
   "source": [
    "lr_scheduler = keras.callbacks.LearningRateScheduler(piecewise_constant_fn)\n",
    "\n",
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=\"nadam\", metrics=[\"accuracy\"])\n",
    "n_epochs = 25\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid),\n",
    "                    callbacks=[lr_scheduler])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 90,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.plot(history.epoch, [piecewise_constant_fn(epoch) for epoch in history.epoch], \"o-\")\n",
    "plt.axis([0, n_epochs - 1, 0, 0.011])\n",
    "plt.xlabel(\"Epoch\")\n",
    "plt.ylabel(\"Learning Rate\")\n",
    "plt.title(\"Piecewise Constant Scheduling\", fontsize=14)\n",
    "plt.grid(True)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Performance Scheduling"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 91,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 92,
   "metadata": {},
   "outputs": [],
   "source": [
    "lr_scheduler = keras.callbacks.ReduceLROnPlateau(factor=0.5, patience=5)\n",
    "\n",
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "optimizer = keras.optimizers.SGD(lr=0.02, momentum=0.9)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])\n",
    "n_epochs = 25\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid),\n",
    "                    callbacks=[lr_scheduler])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 93,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.plot(history.epoch, history.history[\"lr\"], \"bo-\")\n",
    "plt.xlabel(\"Epoch\")\n",
    "plt.ylabel(\"Learning Rate\", color='b')\n",
    "plt.tick_params('y', colors='b')\n",
    "plt.gca().set_xlim(0, n_epochs - 1)\n",
    "plt.grid(True)\n",
    "\n",
    "ax2 = plt.gca().twinx()\n",
    "ax2.plot(history.epoch, history.history[\"val_loss\"], \"r^-\")\n",
    "ax2.set_ylabel('Validation Loss', color='r')\n",
    "ax2.tick_params('y', colors='r')\n",
    "\n",
    "plt.title(\"Reduce LR on Plateau\", fontsize=14)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### tf.keras schedulers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 94,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "s = 20 * len(X_train) // 32 # number of steps in 20 epochs (batch size = 32)\n",
    "learning_rate = keras.optimizers.schedules.ExponentialDecay(0.01, s, 0.1)\n",
    "optimizer = keras.optimizers.SGD(learning_rate)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])\n",
    "n_epochs = 25\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "For piecewise constant scheduling, try this:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 95,
   "metadata": {},
   "outputs": [],
   "source": [
    "learning_rate = keras.optimizers.schedules.PiecewiseConstantDecay(\n",
    "    boundaries=[5. * n_steps_per_epoch, 15. * n_steps_per_epoch],\n",
    "    values=[0.01, 0.005, 0.001])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1Cycle scheduling"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 96,
   "metadata": {},
   "outputs": [],
   "source": [
    "K = keras.backend\n",
    "\n",
    "class ExponentialLearningRate(keras.callbacks.Callback):\n",
    "    def __init__(self, factor):\n",
    "        self.factor = factor\n",
    "        self.rates = []\n",
    "        self.losses = []\n",
    "    def on_batch_end(self, batch, logs):\n",
    "        self.rates.append(K.get_value(self.model.optimizer.lr))\n",
    "        self.losses.append(logs[\"loss\"])\n",
    "        K.set_value(self.model.optimizer.lr, self.model.optimizer.lr * self.factor)\n",
    "\n",
    "def find_learning_rate(model, X, y, epochs=1, batch_size=32, min_rate=10**-5, max_rate=10):\n",
    "    init_weights = model.get_weights()\n",
    "    iterations = math.ceil(len(X) / batch_size) * epochs\n",
    "    factor = np.exp(np.log(max_rate / min_rate) / iterations)\n",
    "    init_lr = K.get_value(model.optimizer.lr)\n",
    "    K.set_value(model.optimizer.lr, min_rate)\n",
    "    exp_lr = ExponentialLearningRate(factor)\n",
    "    history = model.fit(X, y, epochs=epochs, batch_size=batch_size,\n",
    "                        callbacks=[exp_lr])\n",
    "    K.set_value(model.optimizer.lr, init_lr)\n",
    "    model.set_weights(init_weights)\n",
    "    return exp_lr.rates, exp_lr.losses\n",
    "\n",
    "def plot_lr_vs_loss(rates, losses):\n",
    "    plt.plot(rates, losses)\n",
    "    plt.gca().set_xscale('log')\n",
    "    plt.hlines(min(losses), min(rates), max(rates))\n",
    "    plt.axis([min(rates), max(rates), min(losses), (losses[0] + min(losses)) / 2])\n",
    "    plt.xlabel(\"Learning rate\")\n",
    "    plt.ylabel(\"Loss\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 97,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=keras.optimizers.SGD(lr=1e-3),\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 98,
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 128\n",
    "rates, losses = find_learning_rate(model, X_train_scaled, y_train, epochs=1, batch_size=batch_size)\n",
    "plot_lr_vs_loss(rates, losses)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 99,
   "metadata": {},
   "outputs": [],
   "source": [
    "class OneCycleScheduler(keras.callbacks.Callback):\n",
    "    def __init__(self, iterations, max_rate, start_rate=None,\n",
    "                 last_iterations=None, last_rate=None):\n",
    "        self.iterations = iterations\n",
    "        self.max_rate = max_rate\n",
    "        self.start_rate = start_rate or max_rate / 10\n",
    "        self.last_iterations = last_iterations or iterations // 10 + 1\n",
    "        self.half_iteration = (iterations - self.last_iterations) // 2\n",
    "        self.last_rate = last_rate or self.start_rate / 1000\n",
    "        self.iteration = 0\n",
    "    def _interpolate(self, iter1, iter2, rate1, rate2):\n",
    "        return ((rate2 - rate1) * (self.iteration - iter1)\n",
    "                / (iter2 - iter1) + rate1)\n",
    "    def on_batch_begin(self, batch, logs):\n",
    "        if self.iteration < self.half_iteration:\n",
    "            rate = self._interpolate(0, self.half_iteration, self.start_rate, self.max_rate)\n",
    "        elif self.iteration < 2 * self.half_iteration:\n",
    "            rate = self._interpolate(self.half_iteration, 2 * self.half_iteration,\n",
    "                                     self.max_rate, self.start_rate)\n",
    "        else:\n",
    "            rate = self._interpolate(2 * self.half_iteration, self.iterations,\n",
    "                                     self.start_rate, self.last_rate)\n",
    "        self.iteration += 1\n",
    "        K.set_value(self.model.optimizer.lr, rate)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 100,
   "metadata": {},
   "outputs": [],
   "source": [
    "n_epochs = 25\n",
    "onecycle = OneCycleScheduler(math.ceil(len(X_train) / batch_size) * n_epochs, max_rate=0.05)\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs, batch_size=batch_size,\n",
    "                    validation_data=(X_valid_scaled, y_valid),\n",
    "                    callbacks=[onecycle])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Avoiding Overfitting Through Regularization"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## $\\ell_1$ and $\\ell_2$ regularization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 101,
   "metadata": {},
   "outputs": [],
   "source": [
    "layer = keras.layers.Dense(100, activation=\"elu\",\n",
    "                           kernel_initializer=\"he_normal\",\n",
    "                           kernel_regularizer=keras.regularizers.l2(0.01))\n",
    "# or l1(0.1) for ℓ1 regularization with a factor or 0.1\n",
    "# or l1_l2(0.1, 0.01) for both ℓ1 and ℓ2 regularization, with factors 0.1 and 0.01 respectively"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 102,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dense(300, activation=\"elu\",\n",
    "                       kernel_initializer=\"he_normal\",\n",
    "                       kernel_regularizer=keras.regularizers.l2(0.01)),\n",
    "    keras.layers.Dense(100, activation=\"elu\",\n",
    "                       kernel_initializer=\"he_normal\",\n",
    "                       kernel_regularizer=keras.regularizers.l2(0.01)),\n",
    "    keras.layers.Dense(10, activation=\"softmax\",\n",
    "                       kernel_regularizer=keras.regularizers.l2(0.01))\n",
    "])\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=\"nadam\", metrics=[\"accuracy\"])\n",
    "n_epochs = 2\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 103,
   "metadata": {},
   "outputs": [],
   "source": [
    "from functools import partial\n",
    "\n",
    "RegularizedDense = partial(keras.layers.Dense,\n",
    "                           activation=\"elu\",\n",
    "                           kernel_initializer=\"he_normal\",\n",
    "                           kernel_regularizer=keras.regularizers.l2(0.01))\n",
    "\n",
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    RegularizedDense(300),\n",
    "    RegularizedDense(100),\n",
    "    RegularizedDense(10, activation=\"softmax\")\n",
    "])\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=\"nadam\", metrics=[\"accuracy\"])\n",
    "n_epochs = 2\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Dropout"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 104,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.Dropout(rate=0.2),\n",
    "    keras.layers.Dense(300, activation=\"elu\", kernel_initializer=\"he_normal\"),\n",
    "    keras.layers.Dropout(rate=0.2),\n",
    "    keras.layers.Dense(100, activation=\"elu\", kernel_initializer=\"he_normal\"),\n",
    "    keras.layers.Dropout(rate=0.2),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=\"nadam\", metrics=[\"accuracy\"])\n",
    "n_epochs = 2\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Alpha Dropout"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 105,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 106,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    keras.layers.AlphaDropout(rate=0.2),\n",
    "    keras.layers.Dense(300, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.AlphaDropout(rate=0.2),\n",
    "    keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\"),\n",
    "    keras.layers.AlphaDropout(rate=0.2),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "optimizer = keras.optimizers.SGD(lr=0.01, momentum=0.9, nesterov=True)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])\n",
    "n_epochs = 20\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 107,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.evaluate(X_test_scaled, y_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 108,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.evaluate(X_train_scaled, y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 109,
   "metadata": {},
   "outputs": [],
   "source": [
    "history = model.fit(X_train_scaled, y_train)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## MC Dropout"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 110,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 111,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_probas = np.stack([model(X_test_scaled, training=True)\n",
    "                     for sample in range(100)])\n",
    "y_proba = y_probas.mean(axis=0)\n",
    "y_std = y_probas.std(axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 112,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.round(model.predict(X_test_scaled[:1]), 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 113,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.round(y_probas[:, :1], 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 114,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.round(y_proba[:1], 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 115,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_std = y_probas.std(axis=0)\n",
    "np.round(y_std[:1], 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 116,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_pred = np.argmax(y_proba, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 117,
   "metadata": {},
   "outputs": [],
   "source": [
    "accuracy = np.sum(y_pred == y_test) / len(y_test)\n",
    "accuracy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 118,
   "metadata": {},
   "outputs": [],
   "source": [
    "class MCDropout(keras.layers.Dropout):\n",
    "    def call(self, inputs):\n",
    "        return super().call(inputs, training=True)\n",
    "\n",
    "class MCAlphaDropout(keras.layers.AlphaDropout):\n",
    "    def call(self, inputs):\n",
    "        return super().call(inputs, training=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 119,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 120,
   "metadata": {},
   "outputs": [],
   "source": [
    "mc_model = keras.models.Sequential([\n",
    "    MCAlphaDropout(layer.rate) if isinstance(layer, keras.layers.AlphaDropout) else layer\n",
    "    for layer in model.layers\n",
    "])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 121,
   "metadata": {},
   "outputs": [],
   "source": [
    "mc_model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 122,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.SGD(lr=0.01, momentum=0.9, nesterov=True)\n",
    "mc_model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 123,
   "metadata": {},
   "outputs": [],
   "source": [
    "mc_model.set_weights(model.get_weights())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now we can use the model with MC Dropout:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 124,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.round(np.mean([mc_model.predict(X_test_scaled[:1]) for sample in range(100)], axis=0), 2)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Max norm"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 125,
   "metadata": {},
   "outputs": [],
   "source": [
    "layer = keras.layers.Dense(100, activation=\"selu\", kernel_initializer=\"lecun_normal\",\n",
    "                           kernel_constraint=keras.constraints.max_norm(1.))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 126,
   "metadata": {},
   "outputs": [],
   "source": [
    "MaxNormDense = partial(keras.layers.Dense,\n",
    "                       activation=\"selu\", kernel_initializer=\"lecun_normal\",\n",
    "                       kernel_constraint=keras.constraints.max_norm(1.))\n",
    "\n",
    "model = keras.models.Sequential([\n",
    "    keras.layers.Flatten(input_shape=[28, 28]),\n",
    "    MaxNormDense(300),\n",
    "    MaxNormDense(100),\n",
    "    keras.layers.Dense(10, activation=\"softmax\")\n",
    "])\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\", optimizer=\"nadam\", metrics=[\"accuracy\"])\n",
    "n_epochs = 2\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs,\n",
    "                    validation_data=(X_valid_scaled, y_valid))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Exercises"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. to 7."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "See appendix A."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 8. Deep Learning on CIFAR10"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### a.\n",
    "*Exercise: Build a DNN with 20 hidden layers of 100 neurons each (that's too many, but it's the point of this exercise). Use He initialization and the ELU activation function.*"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 127,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.backend.clear_session()\n",
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential()\n",
    "model.add(keras.layers.Flatten(input_shape=[32, 32, 3]))\n",
    "for _ in range(20):\n",
    "    model.add(keras.layers.Dense(100,\n",
    "                                 activation=\"elu\",\n",
    "                                 kernel_initializer=\"he_normal\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### b.\n",
    "*Exercise: Using Nadam optimization and early stopping, train the network on the CIFAR10 dataset. You can load it with `keras.datasets.cifar10.load_data()`. The dataset is composed of 60,000 32 × 32–pixel color images (50,000 for training, 10,000 for testing) with 10 classes, so you'll need a softmax output layer with 10 neurons. Remember to search for the right learning rate each time you change the model's architecture or hyperparameters.*"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's add the output layer to the model:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 128,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.add(keras.layers.Dense(10, activation=\"softmax\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's use a Nadam optimizer with a learning rate of 5e-5. I tried learning rates 1e-5, 3e-5, 1e-4, 3e-4, 1e-3, 3e-3 and 1e-2, and I compared their learning curves for 10 epochs each (using the TensorBoard callback, below). The learning rates 3e-5 and 1e-4 were pretty good, so I tried 5e-5, which turned out to be slightly better."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 129,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = keras.optimizers.Nadam(lr=5e-5)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=optimizer,\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's load the CIFAR10 dataset. We also want to use early stopping, so we need a validation set. Let's use the first 5,000 images of the original training set as the validation set:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 130,
   "metadata": {},
   "outputs": [],
   "source": [
    "(X_train_full, y_train_full), (X_test, y_test) = keras.datasets.cifar10.load_data()\n",
    "\n",
    "X_train = X_train_full[5000:]\n",
    "y_train = y_train_full[5000:]\n",
    "X_valid = X_train_full[:5000]\n",
    "y_valid = y_train_full[:5000]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now we can create the callbacks we need and train the model:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 131,
   "metadata": {},
   "outputs": [],
   "source": [
    "early_stopping_cb = keras.callbacks.EarlyStopping(patience=20)\n",
    "model_checkpoint_cb = keras.callbacks.ModelCheckpoint(\"my_cifar10_model.h5\", save_best_only=True)\n",
    "run_index = 1 # increment every time you train the model\n",
    "run_logdir = os.path.join(os.curdir, \"my_cifar10_logs\", \"run_{:03d}\".format(run_index))\n",
    "tensorboard_cb = keras.callbacks.TensorBoard(run_logdir)\n",
    "callbacks = [early_stopping_cb, model_checkpoint_cb, tensorboard_cb]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 132,
   "metadata": {},
   "outputs": [],
   "source": [
    "%tensorboard --logdir=./my_cifar10_logs --port=6006"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 133,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.fit(X_train, y_train, epochs=100,\n",
    "          validation_data=(X_valid, y_valid),\n",
    "          callbacks=callbacks)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 134,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.load_model(\"my_cifar10_model.h5\")\n",
    "model.evaluate(X_valid, y_valid)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The model with the lowest validation loss gets about 47.6% accuracy on the validation set. It took 27 epochs to reach the lowest validation loss, with roughly 8 seconds per epoch on my laptop (without a GPU). Let's see if we can improve performance using Batch Normalization."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### c.\n",
    "*Exercise: Now try adding Batch Normalization and compare the learning curves: Is it converging faster than before? Does it produce a better model? How does it affect training speed?*"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The code below is very similar to the code above, with a few changes:\n",
    "\n",
    "* I added a BN layer after every Dense layer (before the activation function), except for the output layer. I also added a BN layer before the first hidden layer.\n",
    "* I changed the learning rate to 5e-4. I experimented with 1e-5, 3e-5, 5e-5, 1e-4, 3e-4, 5e-4, 1e-3 and 3e-3, and I chose the one with the best validation performance after 20 epochs.\n",
    "* I renamed the run directories to run_bn_* and the model file name to my_cifar10_bn_model.h5."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 135,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.backend.clear_session()\n",
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential()\n",
    "model.add(keras.layers.Flatten(input_shape=[32, 32, 3]))\n",
    "model.add(keras.layers.BatchNormalization())\n",
    "for _ in range(20):\n",
    "    model.add(keras.layers.Dense(100, kernel_initializer=\"he_normal\"))\n",
    "    model.add(keras.layers.BatchNormalization())\n",
    "    model.add(keras.layers.Activation(\"elu\"))\n",
    "model.add(keras.layers.Dense(10, activation=\"softmax\"))\n",
    "\n",
    "optimizer = keras.optimizers.Nadam(lr=5e-4)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=optimizer,\n",
    "              metrics=[\"accuracy\"])\n",
    "\n",
    "early_stopping_cb = keras.callbacks.EarlyStopping(patience=20)\n",
    "model_checkpoint_cb = keras.callbacks.ModelCheckpoint(\"my_cifar10_bn_model.h5\", save_best_only=True)\n",
    "run_index = 1 # increment every time you train the model\n",
    "run_logdir = os.path.join(os.curdir, \"my_cifar10_logs\", \"run_bn_{:03d}\".format(run_index))\n",
    "tensorboard_cb = keras.callbacks.TensorBoard(run_logdir)\n",
    "callbacks = [early_stopping_cb, model_checkpoint_cb, tensorboard_cb]\n",
    "\n",
    "model.fit(X_train, y_train, epochs=100,\n",
    "          validation_data=(X_valid, y_valid),\n",
    "          callbacks=callbacks)\n",
    "\n",
    "model = keras.models.load_model(\"my_cifar10_bn_model.h5\")\n",
    "model.evaluate(X_valid, y_valid)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* *Is the model converging faster than before?* Much faster! The previous model took 27 epochs to reach the lowest validation loss, while the new model achieved that same loss in just 5 epochs and continued to make progress until the 16th epoch. The BN layers stabilized training and allowed us to use a much larger learning rate, so convergence was faster.\n",
    "* *Does BN produce a better model?* Yes! The final model is also much better, with 54.0% accuracy instead of 47.6%. It's still not a very good model, but at least it's much better than before (a Convolutional Neural Network would do much better, but that's a different topic, see chapter 14).\n",
    "* *How does BN affect training speed?* Although the model converged much faster, each epoch took about 12s instead of 8s, because of the extra computations required by the BN layers. But overall the training time (wall time) was shortened significantly!"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### d.\n",
    "*Exercise: Try replacing Batch Normalization with SELU, and make the necessary adjustements to ensure the network self-normalizes (i.e., standardize the input features, use LeCun normal initialization, make sure the DNN contains only a sequence of dense layers, etc.).*"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 136,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "keras.backend.clear_session()\n",
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential()\n",
    "model.add(keras.layers.Flatten(input_shape=[32, 32, 3]))\n",
    "for _ in range(20):\n",
    "    model.add(keras.layers.Dense(100,\n",
    "                                 kernel_initializer=\"lecun_normal\",\n",
    "                                 activation=\"selu\"))\n",
    "model.add(keras.layers.Dense(10, activation=\"softmax\"))\n",
    "\n",
    "optimizer = keras.optimizers.Nadam(lr=7e-4)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=optimizer,\n",
    "              metrics=[\"accuracy\"])\n",
    "\n",
    "early_stopping_cb = keras.callbacks.EarlyStopping(patience=20)\n",
    "model_checkpoint_cb = keras.callbacks.ModelCheckpoint(\"my_cifar10_selu_model.h5\", save_best_only=True)\n",
    "run_index = 1 # increment every time you train the model\n",
    "run_logdir = os.path.join(os.curdir, \"my_cifar10_logs\", \"run_selu_{:03d}\".format(run_index))\n",
    "tensorboard_cb = keras.callbacks.TensorBoard(run_logdir)\n",
    "callbacks = [early_stopping_cb, model_checkpoint_cb, tensorboard_cb]\n",
    "\n",
    "X_means = X_train.mean(axis=0)\n",
    "X_stds = X_train.std(axis=0)\n",
    "X_train_scaled = (X_train - X_means) / X_stds\n",
    "X_valid_scaled = (X_valid - X_means) / X_stds\n",
    "X_test_scaled = (X_test - X_means) / X_stds\n",
    "\n",
    "model.fit(X_train_scaled, y_train, epochs=100,\n",
    "          validation_data=(X_valid_scaled, y_valid),\n",
    "          callbacks=callbacks)\n",
    "\n",
    "model = keras.models.load_model(\"my_cifar10_selu_model.h5\")\n",
    "model.evaluate(X_valid_scaled, y_valid)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 137,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = keras.models.load_model(\"my_cifar10_selu_model.h5\")\n",
    "model.evaluate(X_valid_scaled, y_valid)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We get 47.9% accuracy, which is not much better than the original model (47.6%), and not as good as the model using batch normalization (54.0%). However, convergence was almost as fast as with the BN model, plus each epoch took only 7 seconds. So it's by far the fastest model to train so far."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### e.\n",
    "*Exercise: Try regularizing the model with alpha dropout. Then, without retraining your model, see if you can achieve better accuracy using MC Dropout.*"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 138,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.backend.clear_session()\n",
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential()\n",
    "model.add(keras.layers.Flatten(input_shape=[32, 32, 3]))\n",
    "for _ in range(20):\n",
    "    model.add(keras.layers.Dense(100,\n",
    "                                 kernel_initializer=\"lecun_normal\",\n",
    "                                 activation=\"selu\"))\n",
    "\n",
    "model.add(keras.layers.AlphaDropout(rate=0.1))\n",
    "model.add(keras.layers.Dense(10, activation=\"softmax\"))\n",
    "\n",
    "optimizer = keras.optimizers.Nadam(lr=5e-4)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=optimizer,\n",
    "              metrics=[\"accuracy\"])\n",
    "\n",
    "early_stopping_cb = keras.callbacks.EarlyStopping(patience=20)\n",
    "model_checkpoint_cb = keras.callbacks.ModelCheckpoint(\"my_cifar10_alpha_dropout_model.h5\", save_best_only=True)\n",
    "run_index = 1 # increment every time you train the model\n",
    "run_logdir = os.path.join(os.curdir, \"my_cifar10_logs\", \"run_alpha_dropout_{:03d}\".format(run_index))\n",
    "tensorboard_cb = keras.callbacks.TensorBoard(run_logdir)\n",
    "callbacks = [early_stopping_cb, model_checkpoint_cb, tensorboard_cb]\n",
    "\n",
    "X_means = X_train.mean(axis=0)\n",
    "X_stds = X_train.std(axis=0)\n",
    "X_train_scaled = (X_train - X_means) / X_stds\n",
    "X_valid_scaled = (X_valid - X_means) / X_stds\n",
    "X_test_scaled = (X_test - X_means) / X_stds\n",
    "\n",
    "model.fit(X_train_scaled, y_train, epochs=100,\n",
    "          validation_data=(X_valid_scaled, y_valid),\n",
    "          callbacks=callbacks)\n",
    "\n",
    "model = keras.models.load_model(\"my_cifar10_alpha_dropout_model.h5\")\n",
    "model.evaluate(X_valid_scaled, y_valid)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The model reaches 48.9% accuracy on the validation set. That's very slightly better than without dropout (47.6%). With an extensive hyperparameter search, it might be possible to do better (I tried dropout rates of 5%, 10%, 20% and 40%, and learning rates 1e-4, 3e-4, 5e-4, and 1e-3), but probably not much better in this case."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's use MC Dropout now. We will need the `MCAlphaDropout` class we used earlier, so let's just copy it here for convenience:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 139,
   "metadata": {},
   "outputs": [],
   "source": [
    "class MCAlphaDropout(keras.layers.AlphaDropout):\n",
    "    def call(self, inputs):\n",
    "        return super().call(inputs, training=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's create a new model, identical to the one we just trained (with the same weights), but with `MCAlphaDropout` dropout layers instead of `AlphaDropout` layers:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 140,
   "metadata": {},
   "outputs": [],
   "source": [
    "mc_model = keras.models.Sequential([\n",
    "    MCAlphaDropout(layer.rate) if isinstance(layer, keras.layers.AlphaDropout) else layer\n",
    "    for layer in model.layers\n",
    "])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Then let's add a couple utility functions. The first will run the model many times (10 by default) and it will return the mean predicted class probabilities. The second will use these mean probabilities to predict the most likely class for each instance:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 141,
   "metadata": {},
   "outputs": [],
   "source": [
    "def mc_dropout_predict_probas(mc_model, X, n_samples=10):\n",
    "    Y_probas = [mc_model.predict(X) for sample in range(n_samples)]\n",
    "    return np.mean(Y_probas, axis=0)\n",
    "\n",
    "def mc_dropout_predict_classes(mc_model, X, n_samples=10):\n",
    "    Y_probas = mc_dropout_predict_probas(mc_model, X, n_samples)\n",
    "    return np.argmax(Y_probas, axis=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's make predictions for all the instances in the validation set, and compute the accuracy:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 142,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.backend.clear_session()\n",
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "y_pred = mc_dropout_predict_classes(mc_model, X_valid_scaled)\n",
    "accuracy = np.mean(y_pred == y_valid[:, 0])\n",
    "accuracy"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We get no accuracy improvement in this case (we're still at 48.9% accuracy).\n",
    "\n",
    "So the best model we got in this exercise is the Batch Normalization model."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### f.\n",
    "*Exercise: Retrain your model using 1cycle scheduling and see if it improves training speed and model accuracy.*"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 143,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.backend.clear_session()\n",
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential()\n",
    "model.add(keras.layers.Flatten(input_shape=[32, 32, 3]))\n",
    "for _ in range(20):\n",
    "    model.add(keras.layers.Dense(100,\n",
    "                                 kernel_initializer=\"lecun_normal\",\n",
    "                                 activation=\"selu\"))\n",
    "\n",
    "model.add(keras.layers.AlphaDropout(rate=0.1))\n",
    "model.add(keras.layers.Dense(10, activation=\"softmax\"))\n",
    "\n",
    "optimizer = keras.optimizers.SGD(lr=1e-3)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=optimizer,\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 144,
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 128\n",
    "rates, losses = find_learning_rate(model, X_train_scaled, y_train, epochs=1, batch_size=batch_size)\n",
    "plot_lr_vs_loss(rates, losses)\n",
    "plt.axis([min(rates), max(rates), min(losses), (losses[0] + min(losses)) / 1.4])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 145,
   "metadata": {},
   "outputs": [],
   "source": [
    "keras.backend.clear_session()\n",
    "tf.random.set_seed(42)\n",
    "np.random.seed(42)\n",
    "\n",
    "model = keras.models.Sequential()\n",
    "model.add(keras.layers.Flatten(input_shape=[32, 32, 3]))\n",
    "for _ in range(20):\n",
    "    model.add(keras.layers.Dense(100,\n",
    "                                 kernel_initializer=\"lecun_normal\",\n",
    "                                 activation=\"selu\"))\n",
    "\n",
    "model.add(keras.layers.AlphaDropout(rate=0.1))\n",
    "model.add(keras.layers.Dense(10, activation=\"softmax\"))\n",
    "\n",
    "optimizer = keras.optimizers.SGD(lr=1e-2)\n",
    "model.compile(loss=\"sparse_categorical_crossentropy\",\n",
    "              optimizer=optimizer,\n",
    "              metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 146,
   "metadata": {},
   "outputs": [],
   "source": [
    "n_epochs = 15\n",
    "onecycle = OneCycleScheduler(math.ceil(len(X_train_scaled) / batch_size) * n_epochs, max_rate=0.05)\n",
    "history = model.fit(X_train_scaled, y_train, epochs=n_epochs, batch_size=batch_size,\n",
    "                    validation_data=(X_valid_scaled, y_valid),\n",
    "                    callbacks=[onecycle])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "One cycle allowed us to train the model in just 15 epochs, each taking only 2 seconds (thanks to the larger batch size). This is several times faster than the fastest model we trained so far. Moreover, we improved the model's performance (from 47.6% to 52.0%). The batch normalized model reaches a slightly better performance (54%), but it's much slower to train."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.7.9"
  },
  "nav_menu": {
   "height": "360px",
   "width": "416px"
  },
  "toc": {
   "navigate_menu": true,
   "number_sections": true,
   "sideBar": true,
   "threshold": 6,
   "toc_cell": false,
   "toc_section_display": "block",
   "toc_window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}