0_explore-data.ipynb

{
 "cells": [
  {
   "attachments": {
    "fashion-mnist.png": {
     "image/png": "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"
    }
   },
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "805a97911f01abb36ff8414d8d6aec9293e4fb85"
   },
   "source": [
    " <h1><center><font size=\"6\">CNN with Keras for Fashion MNIST</font></center></h1>\n",
    "\n",
    "![fashion-mnist.png](attachment:fashion-mnist.png)\n",
    "\n",
    "\n",
    "# <a id='0'>Content</a>\n",
    "\n",
    "- <a href='#1'>Introduction</a>  \n",
    "- <a href='#2'>Load packages</a>  \n",
    "- <a href='#3'>Read the data</a>  \n",
    "- <a href='#4'>Data exploration</a>\n",
    "    - <a href='#41'>Class distribution</a>\n",
    "    - <a href='#42'>Images samples</a>\n",
    "- <a href='#5'>Model</a>  \n",
    "    - <a href='#51'>Prepare the model</a>  \n",
    "    - <a href='#52'>Train the model</a>  \n",
    "    - <a href='#53'>Test prediction accuracy</a>   \n",
    "    - <a href='#54'>Validation accuracy and loss</a>   \n",
    "    - <a href='#55'>Add Dropout layers to the model</a>  \n",
    "    - <a href='#56'>Re-train the model</a>   \n",
    "    - <a href='#57'>Check validation accuracy and loss with the new model</a>    \n",
    "    - <a href='#58'>Prediction accuracy with the new model </a>   \n",
    "- <a href='#6'>Visualize the classified images</a>  \n",
    "    - <a href='#61'>Correctly classified images</a>   \n",
    "    - <a href='#62'>Incorrectly classified images</a>   \n",
    "- <a href='#8'>References</a>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "1b1c16628c2f62a18e1dc2068e1d67d7003922b1"
   },
   "source": [
    "# <a id=\"1\">Introduction</a>  \n",
    "\n",
    "\n",
    "## Dataset\n",
    "\n",
    "Fashion-MNIST is a dataset of Zalando's article images—consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. Zalando intends Fashion-MNIST to serve as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms. It shares the same image size and structure of training and testing splits.\n",
    "\n",
    "\n",
    "## Content\n",
    "\n",
    "Each image is 28 pixels in height and 28 pixels in width, for a total of 784 pixels in total.   \n",
    "\n",
    "Each pixel has a single pixel-value associated with it, indicating the lightness or darkness of that pixel, with higher numbers meaning darker. This pixel-value is an integer between 0 and 255.   \n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "5a708dc52b2e5990e2247ed573d50d0f6933b730"
   },
   "source": [
    "# <a id=\"2\">Load packages</a>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "2defa674e4e6d0e7371df92e7d1f388fc5c14bb6"
   },
   "outputs": [],
   "source": [
    "import warnings\n",
    "\n",
    "from IPython.display import SVG\n",
    "from keras.datasets import fashion_mnist\n",
    "from keras.layers import Dense, Flatten, Conv2D, Dropout, MaxPooling2D\n",
    "from keras.models import Sequential\n",
    "from keras.utils import plot_model, to_categorical\n",
    "from keras.utils.vis_utils import model_to_dot\n",
    "from sklearn.metrics import classification_report\n",
    "from sklearn.model_selection import train_test_split\n",
    "import keras\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "import seaborn as sns\n",
    "\n",
    "warnings.filterwarnings('ignore')\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "e17fd2539412442ddb3ecacf84366ddd62faf836"
   },
   "source": [
    "## Parameters"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "d7a44bfc7a28df7026241c4a7e047298446f1554"
   },
   "outputs": [],
   "source": [
    "# Some constants\n",
    "IMG_ROWS = 28\n",
    "IMG_COLS = 28\n",
    "NUM_CLASSES = 10\n",
    "TEST_SIZE = 0.2\n",
    "RANDOM_STATE = 2018\n",
    "#Model\n",
    "NO_EPOCHS = 50\n",
    "BATCH_SIZE = 128"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "f49d65b4af77e87ab34b6854850619911cff1e6d"
   },
   "source": [
    "# <a id=\"3\">Read the data</a>\n",
    "\n",
    "There are 10 different classes of images, as following: \n",
    "\n",
    "* **0**: **T-shirt/top**;   \n",
    "* **1**: **Trouser**;   \n",
    "* **2**: **Pullover**;   \n",
    "* **3**: **Dress**;\n",
    "* **4**: **Coat**;\n",
    "* **5**: **Sandal**;\n",
    "* **6**: **Shirt**;\n",
    "* **7**: **Sneaker**;\n",
    "* **8**: **Bag**;\n",
    "* **9**: **Ankle boot**.\n",
    "\n",
    "Image dimmensions are **28**x**28**.   \n",
    "\n",
    "The train set and test set are given in two separate sets.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "a9c3148fda056ecc88570b302f5185064d5e9fc8"
   },
   "outputs": [],
   "source": [
    "(x_train_data, y_train_data), (x_test_data, y_test_data) = fashion_mnist.load_data()\n",
    "\n",
    "# Create a dictionary for each type of label \n",
    "LABELS = {0 : \"T-shirt/top\", 1: \"Trouser\", 2: \"Pullover\", 3: \"Dress\", 4: \"Coat\",\n",
    "          5: \"Sandal\", 6: \"Shirt\", 7: \"Sneaker\", 8: \"Bag\", 9: \"Ankle Boot\"}"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "290f8f38b2f64dfd1bd3432eb1e9a101dbbaf4e5"
   },
   "source": [
    "# <a id=\"4\">Data exploration</a>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "e8891ae7996aa2ae5bc1c025661a77ae91c2fdfb"
   },
   "source": [
    "![](http://)The dimmension of the original  train,  test set are as following:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "a133496cefc53baa56b9a5eec93bfd064ea6360d"
   },
   "outputs": [],
   "source": [
    "print(\"Fashion MNIST x_train shape: {}\".format(x_train_data.shape))\n",
    "print(\"Fashion MNIST y_train shape: {}\".format(y_train_data.shape))\n",
    "print(\"Fashion MNIST x_test shape: {}\".format(x_test_data.shape))\n",
    "print(\"Fashion MNIST y_test shape: {}\".format(y_test_data.shape))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "dc317fbf40c9b2838f8091e745a64e060a7affc0"
   },
   "source": [
    "## <a id=\"41\">Class distribution</a>\n",
    "\n",
    "Let's see how many number of images are in each class. We start with the train set.\n",
    "\n",
    "### Train set images class distribution"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "da948eb587daf81160a4794a36662f8872c882a4"
   },
   "outputs": [],
   "source": [
    "def get_classes_distribution(y_data):\n",
    "    # Get the count for each label\n",
    "    y = np.bincount(y_data)\n",
    "    ii = np.nonzero(y)[0]\n",
    "    label_counts = zip(ii, y[ii])\n",
    "\n",
    "    # Get total number of samples\n",
    "    total_samples = len(y_data)\n",
    "\n",
    "    # Count the number of items in each class\n",
    "    for label, count in label_counts:\n",
    "        class_name = LABELS[label]\n",
    "        percent = (count / total_samples) * 100\n",
    "        print(\"{:<15s}:  {} or {:.2f}%\".format(class_name, count, percent))\n",
    "        \n",
    "    return label_counts\n",
    "\n",
    "train_label_counts = get_classes_distribution(y_train_data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "fc41d889d3ad6352994cdf8a76a78ed0d22ec141"
   },
   "source": [
    "The classes are equaly distributed in the train set (10% each). Let's check the same for the test set.    \n",
    "Let's also plot the class distribution.\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "7ac8f6fde6bd78553f08469093cebc0e124a5210"
   },
   "outputs": [],
   "source": [
    "def plot_label_per_class(y_data):\n",
    "    \n",
    "    classes = sorted(np.unique(y_data))\n",
    "    f, ax = plt.subplots(1,1, figsize=(12, 4))\n",
    "    g = sns.countplot(y_data, order=classes)\n",
    "    g.set_title(\"Number of labels for each class\")\n",
    "    \n",
    "    for p, label in zip(g.patches, classes):\n",
    "        g.annotate(LABELS[label], (p.get_x(), p.get_height() + 0.2))\n",
    "    \n",
    "    plt.show()\n",
    "    \n",
    "plot_label_per_class(y_train_data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "f7c3328ce859d1809ce916bba73bc57b2d580963"
   },
   "source": [
    "### Test set images class distribution"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "2581093ff1268ae9487fe83b383c2d4be4657000"
   },
   "outputs": [],
   "source": [
    "get_classes_distribution(y_test_data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "fffa6f5248eb037b47a84e1e1ee889ee3470a1d4"
   },
   "source": [
    "Also in the test set the 10 classes are equaly distributed (10% each).  \n",
    "\n",
    "Lets' also plot the class distribution."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "7d770b2cde8f32b83a67da02eeacd17f01904ddd"
   },
   "outputs": [],
   "source": [
    "plot_label_per_class(y_test_data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "588beb3c38b5f5a0b41ca0699426a1477efb8412"
   },
   "source": [
    "## <a id=\"42\">Sample images</a>\n",
    "\n",
    "### Train set images\n",
    "\n",
    "Let's plot some samples for the images.   \n",
    "We add labels to the train set images, with the corresponding fashion item category.  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "62fd41c12740f184ec160d692b3289d4457297d6"
   },
   "outputs": [],
   "source": [
    "def sample_images_data(x_data, y_data):\n",
    "    # An empty list to collect some samples\n",
    "    sample_images = []\n",
    "    sample_labels = []\n",
    "\n",
    "    # Iterate over the keys of the labels dictionary defined in the above cell\n",
    "    for k in LABELS.keys():\n",
    "        # Get four samples for each category\n",
    "        samples = np.where(y_data == k)[0][:4]\n",
    "        # Append the samples to the samples list\n",
    "        for s in samples:\n",
    "            img = x_data[s]\n",
    "            sample_images.append(img)\n",
    "            sample_labels.append(y_data[s])\n",
    "\n",
    "    print(\"Total number of sample images to plot: \", len(sample_images))\n",
    "    return sample_images, sample_labels\n",
    "\n",
    "train_sample_images, train_sample_labels = sample_images_data(x_train_data, y_train_data)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "e5c80cef30985bb5ab0ee48d4fc9891660c2a92f"
   },
   "source": [
    "Let's now plot the images.   \n",
    "The labels are shown above each image."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "4f4f27f5143bdd834302a2c43cdb1d68d2a131a9"
   },
   "outputs": [],
   "source": [
    "def plot_sample_images(data_sample_images, data_sample_labels, cmap=\"gray\"):\n",
    "    # Plot the sample images now\n",
    "    f, ax = plt.subplots(5, 8, figsize=(16, 10))\n",
    "\n",
    "    for i, img in enumerate(data_sample_images):\n",
    "        ax[i//8, i%8].imshow(img, cmap=cmap)\n",
    "        ax[i//8, i%8].axis('off')\n",
    "        ax[i//8, i%8].set_title(LABELS[data_sample_labels[i]])\n",
    "    plt.show()    \n",
    "    \n",
    "plot_sample_images(train_sample_images, train_sample_labels)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "bd30352b2674a20eca5e792d7c186cdcbab25461"
   },
   "source": [
    "### Test set images\n",
    "\n",
    "Let's plot now a selection of the test set images.  \n",
    "Labels are as well added (they are known).  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "616df96bfae0e8065c206e6dfec2847d5d42125d"
   },
   "outputs": [],
   "source": [
    "test_sample_images, test_sample_labels = sample_images_data(x_test_data, y_test_data)\n",
    "plot_sample_images(test_sample_images, test_sample_labels)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "273c79c7470c49fe7a546dd7148e0536be2dee19"
   },
   "source": [
    "# <a id=\"5\">Model</a>\n",
    "\n",
    "We start with preparing the model."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "10b920b2ffc5de4a3ddebe494966d90cca91ee95"
   },
   "source": [
    "## <a id=\"51\">Prepare the model</a>\n",
    "\n",
    "## Data preprocessing\n",
    "\n",
    "First we will do a data preprocessing to prepare for the model.\n",
    "\n",
    "We save label (target) feature as a separate vector."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "db4c308b9fb334a54b4c055b00d2a1fbcf77ab96"
   },
   "outputs": [],
   "source": [
    "def data_preprocessing(x_data, y_data):\n",
    "    out_y = to_categorical(y_data, len(np.unique(y_data)))\n",
    "    num_images = x_data.shape[0]\n",
    "    x_shaped_array = np.expand_dims(x_data, axis=-1)\n",
    "    out_x = x_shaped_array / 255.\n",
    "    \n",
    "    return out_x, out_y\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "2d3cafa55173d40cd9a42df63d4919f03b264c09"
   },
   "source": [
    "We process both the train_data and the test_data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "454d7a8fdca4bdbefc04a9d796de04b6af4a1767"
   },
   "outputs": [],
   "source": [
    "# prepare the data\n",
    "X, y = data_preprocessing(x_train_data, y_train_data)\n",
    "X_test, y_test = data_preprocessing(x_test_data, y_test_data)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "8e25c569f0a0e817ca0462f3e3ea2f50feb480b0"
   },
   "source": [
    "## Split train in train and validation set\n",
    "\n",
    "We further split the train set in train and validation set. The validation set will be 20% from the original train set, therefore the split will be train/validation of 0.8/0.2."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "36eb910dd0868a227a73c9759d6aee0a47ba2b1e"
   },
   "outputs": [],
   "source": [
    "X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=TEST_SIZE, random_state=RANDOM_STATE)\n",
    "\n",
    "print(\"Fashion MNIST X train  -  rows: {}  columns: {}\".format(X_train.shape[0], X_train.shape[1:4]))\n",
    "print(\"Fashion MNIST X valid  -  rows: {}  columns: {}\".format(X_val.shape[0], X_val.shape[1:4]))\n",
    "print(\"Fashion MNIST X test   -  rows: {}  columns: {}\".format(X_test.shape[0], X_test.shape[1:4]))\n",
    "print(\"-----------------------------------------------------------\")\n",
    "print(\"Fashion MNIST y train  -  rows: {}  columns: {}\".format(y_train.shape[0], y_train.shape[1]))\n",
    "print(\"Fashion MNIST y valid  -  rows: {}  columns: {}\".format(y_val.shape[0], y_val.shape[1]))\n",
    "print(\"Fashion MNIST y test   -  rows: {}  columns: {}\".format(y_test.shape[0], y_test.shape[1]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "5bc44b14e07efcd3dda90deaac6884b07285b484"
   },
   "source": [
    "Let's check the class inbalance for the rsulted training set."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "get_classes_distribution(np.argmax(y_train, axis=1))\n",
    "plot_label_per_class(np.argmax(y_train, axis=1))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "07bfe3d5d50b985788a28e546473fd885d96dcef"
   },
   "source": [
    "And, as well, for the validation set."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "af3b3d31c42637235bc07b700db121e37e156afc"
   },
   "outputs": [],
   "source": [
    "get_classes_distribution(np.argmax(y_val, axis=1))\n",
    "plot_label_per_class(np.argmax(y_val, axis=1))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "dd9405b5a3fe0c7bd5bebe0616190a8cf26d2811"
   },
   "source": [
    "## <a id=\"52\">Train the model</a>\n",
    "\n",
    "### Build the model   \n",
    "\n",
    "\n",
    "\n",
    "We will use a **Sequential** model.\n",
    "* The **Sequential** model is a linear stack of layers. It can be first initialized and then we add layers using **add** method or we can add all layers at init stage. The layers added are as follows:\n",
    "\n",
    "* **Conv2D** is a 2D Convolutional layer (i.e. spatial convolution over images). The parameters used are:\n",
    " * filters - the number of filters (Kernels) used with this layer; here filters = 32;\n",
    " * kernel_size - the dimmension of the Kernel: (3 x 3);\n",
    " * activation - is the activation function used, in this case `relu`;\n",
    " * kernel_initializer - the function used for initializing the kernel;\n",
    " * input_shape - is the shape of the image presented to the CNN: in our case is 28 x 28\n",
    " The input and output of the **Conv2D** is a 4D tensor.\n",
    " \n",
    "* **MaxPooling2D** is a Max pooling operation for spatial data. Parameters used here are:\n",
    " * *pool_size*, in this case (2,2), representing the factors by which to downscale in both directions;\n",
    " \n",
    "* **Conv2D** with the following parameters:\n",
    " * filters: 64;\n",
    " * kernel_size : (3 x 3);\n",
    " * activation : `relu`;\n",
    " \n",
    "* **MaxPooling2D** with parameter:\n",
    " * *pool_size* : (2,2);\n",
    "\n",
    "* **Conv2D** with the following parameters:\n",
    " * filters: 128;\n",
    " * kernel_size : (3 x 3);\n",
    " * activation : `relu`;\n",
    " \n",
    "* **Flatten**. This layer Flattens the input. Does not affect the batch size. It is used without parameters;\n",
    "\n",
    "* **Dense**. This layer is a regular fully-connected NN layer. It is used without parameters;\n",
    " * units - this is a positive integer, with the meaning: dimensionality of the output space; in this case is: 128;\n",
    " * activation - activation function : `relu`;\n",
    " \n",
    "* **Dense**. This is the final layer (fully connected). It is used with the parameters:\n",
    " * units: the number of classes (in our case 10);\n",
    " * activation : `softmax`; for this final layer it is used `softmax` activation (standard for multiclass classification)\n",
    " \n",
    "\n",
    "Then we compile the model, specifying as well the following parameters:\n",
    "* *loss*;\n",
    "* *optimizer*;\n",
    "* *metrics*. \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "e31836cd5ec9d86340485404b8f613d1f574aca4"
   },
   "outputs": [],
   "source": [
    "# Model\n",
    "model = Sequential()\n",
    "# Add convolution 2D\n",
    "model.add(Conv2D(16, kernel_size=(3, 3),\n",
    "                 activation='relu',\n",
    "                 kernel_initializer='he_normal',\n",
    "                 input_shape=(IMG_ROWS, IMG_COLS, 1)))\n",
    "model.add(MaxPooling2D((2, 2)))\n",
    "model.add(Conv2D(32, \n",
    "                 kernel_size=(3, 3), \n",
    "                 activation='relu'))\n",
    "model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "model.add(Conv2D(64, (3, 3), activation='relu'))\n",
    "model.add(Flatten())\n",
    "model.add(Dense(64, activation='relu'))\n",
    "model.add(Dense(NUM_CLASSES, activation='softmax'))\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "ecb450d70539a62fb310bb7ed44849d2d01481ee"
   },
   "source": [
    "### Inspect the model\n",
    "\n",
    "Let's check the model we initialized."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "b4b923b11ceaf4a97677f8e24265e3e97ae1653b",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "47062fcbe61fdc131bdc369428415c483167d046"
   },
   "source": [
    "Let's also plot the model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "a9863ccbdc1a3f9d03fc6f3dbbd3f2713995d03e"
   },
   "outputs": [],
   "source": [
    "plot_model(model, to_file='model.png')\n",
    "SVG(model_to_dot(model).create(prog='dot', format='svg'))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "68a2e932241f29e52a3150b3fcf3fe9c21243be2"
   },
   "source": [
    "### Run the model\n",
    "\n",
    "We run the model with the training set. We are also using the validation set (a subset from the orginal training set) for validation."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "400494b7e0525069175625422e8c300bd7b41c51",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "train_model = model.fit(X_train, y_train,\n",
    "                  batch_size=BATCH_SIZE,\n",
    "                  epochs=10,\n",
    "                  verbose=1,\n",
    "                  validation_data=(X_val, y_val))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "8699df44bc0a95f1a43a201d3dd566746d87173f"
   },
   "source": [
    "## <a id=\"53\">Test prediction accuracy</a>\n",
    "\n",
    "We calculate the test loss and accuracy."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "b9e2bb7f25b02d491e34dd0d6d05943e287ae369"
   },
   "outputs": [],
   "source": [
    "score = model.evaluate(X_test, y_test, verbose=0)\n",
    "print('Test loss:', score[0])\n",
    "print('Test accuracy:', score[1])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "4d779b0e47b8263370031c292a231b69decad374"
   },
   "source": [
    "We evaluated the model accuracy based on the predicted values for the test set.  Let's check the validation value during training.\n",
    "\n",
    "## <a id=\"53\">Validation accuracy and loss</a>\n",
    "\n",
    "Let's plot the train and validation accuracy and loss, from the train history."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "6122e4d5c504efcc6f05ca65652cdf307c706049"
   },
   "outputs": [],
   "source": [
    "def plot_accuracy_and_loss(train_model):\n",
    "    hist = train_model.history\n",
    "    acc = hist['accuracy']\n",
    "    val_acc = hist['val_accuracy']\n",
    "    loss = hist['loss']\n",
    "    val_loss = hist['val_loss']\n",
    "    epochs = range(len(acc))\n",
    "    f, ax = plt.subplots(1,2, figsize=(14,6))\n",
    "    ax[0].plot(epochs, acc, 'g', label='Training accuracy')\n",
    "    ax[0].plot(epochs, val_acc, 'r', label='Validation accuracy')\n",
    "    ax[0].set_title('Training and validation accuracy')\n",
    "    ax[0].legend()\n",
    "    ax[1].plot(epochs, loss, 'g', label='Training loss')\n",
    "    ax[1].plot(epochs, val_loss, 'r', label='Validation loss')\n",
    "    ax[1].set_title('Training and validation loss')\n",
    "    ax[1].legend()\n",
    "    plt.show()\n",
    "\n",
    "plot_accuracy_and_loss(train_model)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "678f32a0980699b8aa41a3f98840bcc1eabb2aa8"
   },
   "source": [
    "The validation accuracy does not improve after few epochs and the validation loss is increasing after few epochs. This confirms our assumption that the model is overfitted. We will try to improve the model by adding Dropout layers."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "5c026022d95d3afefc3e2cf47050c3db5f800ee1"
   },
   "source": [
    "## <a id=\"55\">Add Dropout layers to the model</a>\n",
    "\n",
    "We add several Dropout layers to the model, to help avoiding overfitting.    \n",
    "Dropout is helping avoid overfitting in several ways, as explained in <a href='#8'>[6]</a> and <a href='#8'>[7]</a>.  \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "6aff87cd33a993a60d701c35026ce8ea19d8665a"
   },
   "outputs": [],
   "source": [
    "# Model\n",
    "model = Sequential()\n",
    "# Add convolution 2D\n",
    "model.add(Conv2D(16, kernel_size=(3, 3),\n",
    "                 activation='relu',\n",
    "                 kernel_initializer='he_normal',\n",
    "                 input_shape=(IMG_ROWS, IMG_COLS, 1)))\n",
    "model.add(MaxPooling2D((2, 2)))\n",
    "# Add dropouts to the model\n",
    "model.add(Dropout(0.25))\n",
    "model.add(Conv2D(32, \n",
    "                 kernel_size=(3, 3), \n",
    "                 activation='relu'))\n",
    "model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "# Add dropouts to the model\n",
    "model.add(Dropout(0.25))\n",
    "model.add(Conv2D(64, (3, 3), activation='relu'))\n",
    "# Add dropouts to the model\n",
    "model.add(Dropout(0.4))\n",
    "model.add(Flatten())\n",
    "model.add(Dense(64, activation='relu'))\n",
    "# Add dropouts to the model\n",
    "model.add(Dropout(0.3))\n",
    "model.add(Dense(NUM_CLASSES, activation='softmax'))\n",
    "\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "46811947e60224235b1ffaed48e7367b3ba68833"
   },
   "source": [
    "## <a id=\"56\">Re-train the model</a>\n",
    "\n",
    "Let's inspect first the model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "594c8f00280db3fd6d4719791326d73c4769117d"
   },
   "outputs": [],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "ab0eb83773ab435204faef8c88b609a29feb635e"
   },
   "source": [
    "Let's also plot the model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "6abc11f358fb81d2655f80bea7762546b18dee82"
   },
   "outputs": [],
   "source": [
    "plot_model(model, to_file='model.png')\n",
    "SVG(model_to_dot(model).create(prog='dot', format='svg'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "99a964cc1478f1a8583e803ad7eeb35aadd15995"
   },
   "source": [
    "And now let's run the new model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "83029bfb27f99fb6be2ecedc1bd256c0e861456c",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "train_model = model.fit(X_train, y_train,\n",
    "                  batch_size=BATCH_SIZE,\n",
    "                  epochs=10,\n",
    "                  verbose=1,\n",
    "                  validation_data=(X_val, y_val))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "676e1c4a3a64db55d6f0e8498f9c59e0e6182931"
   },
   "source": [
    "## <a id=\"57\">Prediction accuracy with the new model</a>\n",
    "\n",
    "Let's re-evaluate the prediction accuracy with the new model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "6b3940059502bfb6303a85a2cee1095b45304491"
   },
   "outputs": [],
   "source": [
    "plot_accuracy_and_loss(train_model)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "f201d6925f70ad2420b3ee7e3a204cdc77f86723"
   },
   "source": [
    "After adding the Dropout layers, the validation accuracy and validation loss are much better. Let's check now the prediction for the test set.\n",
    "\n",
    "\n",
    "## <a id=\"58\">Prediction accuracy with the new model</a>\n",
    "\n",
    "Let's re-evaluate the test prediction accuracy with the new model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "d8e9293e99e34010bd456af7435c4cc7ccebdefd"
   },
   "outputs": [],
   "source": [
    "score = model.evaluate(X_test, y_test, verbose=0)\n",
    "print('Test loss:', score[0])\n",
    "print('Test accuracy:', score[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "e519f82cc29b0dd8c5145d612da3eb496e2b321d"
   },
   "outputs": [],
   "source": [
    "# get the predictions for the test data\n",
    "predicted_classes = model.predict_classes(X_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "d5989dd811570fe73b1f6b0715b06c428e578669"
   },
   "outputs": [],
   "source": [
    "p = predicted_classes\n",
    "y = y_test_data\n",
    "correct = np.nonzero(p == y)[0]\n",
    "incorrect = np.nonzero(p != y)[0]\n",
    "\n",
    "print(\"Correct predicted classes:\", correct.shape[0])\n",
    "print(\"Incorrect predicted classes:\", incorrect.shape[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "bc3bfd05450aaee55c9e5edc7fb84d884c1d3e0b"
   },
   "outputs": [],
   "source": [
    "target_names = [\"Class {} ({}) :\".format(i, LABELS[i]) for i in range(NUM_CLASSES)]\n",
    "print(classification_report(y_test_data, predicted_classes, target_names=target_names))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "4b0b88588bd3afed148a3c696d997aea4a6c564a"
   },
   "source": [
    "Let's also inspect some of the images. We created two subsets of the predicted images set, correctly and incorrectly classified."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "614afbc7ed760856cf9b47f879334902d1b6545b"
   },
   "source": [
    "# <a id=\"6\">Visualize classified images</a>\n",
    "\n",
    "## <a id=\"61\">Correctly classified images</a>\n",
    "\n",
    "\n",
    "We visualize few images correctly classified."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "e64edf6e7ae3146eafe8da01f4dbc4e05c8bc1ce"
   },
   "outputs": [],
   "source": [
    "def plot_predicted_images(predictions, data_index, x_data, y_data, size=16, cmap=\"gray\"):\n",
    "    # Plot the sample images now\n",
    "    f, ax = plt.subplots(4, 4, figsize=(15, 15))\n",
    "\n",
    "    for i, indx in enumerate(np.random.choice(data_index, size=size, replace=False)):\n",
    "        ax[i//4, i%4].imshow(x_data[indx].reshape(IMG_ROWS,IMG_COLS), cmap=cmap)\n",
    "        ax[i//4, i%4].axis('off')\n",
    "        ax[i//4, i%4].set_title(\"True:{}  Pred:{}\".format(LABELS[y_data[indx]], LABELS[predicted_classes[indx]]))\n",
    "    plt.show()    \n",
    "    \n",
    "plot_predicted_images(predicted_classes, correct, x_test_data, y_test_data, cmap=\"Greens\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "3c7bb1b300a4530d2fa418eb56258a54587e3b90"
   },
   "source": [
    "## <a id=\"62\">Incorrectly classified images</a>\n",
    "\n",
    "Let's see also few images incorrectly classified."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "_uuid": "18b2843f6f8542cd95d0847617e84e959996341b"
   },
   "outputs": [],
   "source": [
    "plot_predicted_images(predicted_classes, incorrect, x_test_data, y_test_data, cmap=\"Reds\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "_uuid": "1b45e3422fb9507da66ce7af67a97b54e7fa7683"
   },
   "source": [
    "# <a id=\"8\">References</a>\n",
    "\n",
    "Original source: https://www.kaggle.com/gpreda/cnn-with-tensorflow-keras-for-fashion-mnist/notebook\n",
    "\n",
    "[1] Fashion MNIST, An MNIST-like dataset of 70,000 28x28 labeled fashion images, https://www.kaggle.com/zalando-research/fashionmnist  \n",
    "[2] DanB, CollinMoris, Deep Learning From Scratch, https://www.kaggle.com/dansbecker/deep-learning-from-scratch  \n",
    "[3] DanB, Dropout and Strides for Larger Models, https://www.kaggle.com/dansbecker/dropout-and-strides-for-larger-models  \n",
    "[4] BGO, CNN with Keras, https://www.kaggle.com/bugraokcu/cnn-with-keras    \n",
    "[5] NAIN, EagerFMINST, https://www.kaggle.com/aakashnain/eagerfmnist  \n",
    "[6] Why Dropounts prevent overfitting in Deep Neural Networks, https://medium.com/@vivek.yadav/why-dropouts-prevent-overfitting-in-deep-neural-networks-937e2543a701  \n",
    "[7] Dropout: A Simple Way to Prevent Neural Networks from Overfitting, https://www.cs.toronto.edu/~hinton/absps/JMLRdropout.pdf"
   ]
  },
  {
   "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": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}