tutorials/Part 01 - Building a 2 target features dataset.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Functional Encryption - Classification and information leakage\n",
    "\n",
    "Our start point is the work on encrypted classification using Function Encryption of the paper [Reading in the Dark: Classifying Encrypted Digits with Functional Encryption](https://eprint.iacr.org/2018/206), and the associated [GitHub repository](https://github.com/edufoursans/reading-in-the-dark).\n",
    "\n",
    "More specifically, the paper provides a new Functional Encryption scheme for quadratic multi-variate polynomials, which can under some hypothesis be seen as a single hidden layer neural network with a quadratic activation.\n",
    "In the paper, the output corresponds to 1 element per class, so 10 on their example using MNIST, and it is made in clear. The argmax is then taken in clear to determine the class. This approach is standard and was used in [CryptoNets](https://www.microsoft.com/en-us/research/wp-content/uploads/2016/04/CryptonetsTechReport.pdf), [MiniONN](https://eprint.iacr.org/2017/452.pdf), [CryptoDL](https://arxiv.org/pdf/1711.05189.pdf) and other recent papers on Homomorphic Encryption (HE) (like [this paper of 2018](https://arxiv.org/pdf/1807.08459.pdf)).\n",
    "\n",
    "\n",
    "However, there has been very little discussion about the sensitivity of this output, because in the case of MNIST 10 real values are output, which in term of bits of information is huge. Our starting point is therefore: can this output disclose extra information about the initial input or about charasteristics of this input, apart from the label we want to classify?\n",
    "\n",
    "To this aim, we have built a specific dataset which is very similar to MNIST, used in the original paper but which is composed of 10 letter characters of 2 differents fonts. Our goal is two-fold:\n",
    "\n",
    " - Evaluate how the output in clear can be leveraged with a public neural network to make better predictions than a simple `argmax` function on the character recognition task.\n",
    " - Analyse to what extent the output in clear of the model can reveal information about the font used, using an \"collateral\" network.\n",
    " \n",
    "**But first, let's build this dataset!**\n",
    "\n",
    "\n",
    "_Disclaimer about existing datasets:_\n",
    " > Writer identification is not a novel task, several datasets have proposed among which the CDAR 2013 Competition  on  Writer  Identification, ICDAR  2011  Writer  Identification  Contest, and the CVL-Database datasets. However these datasets don't consider characters but complete pages and text extracts which is not suitable for our analysis.\n",
    " \n",
    " > EMNIST however could be an interesting dataset but they don't provide the two labels (value, authors) together. and there is at most 150 characters per authors which is not enough to correctly perform learning.\n",
    " \n",
    " > There is also this excellent work @TalwalkarLab to create [FEMNIST](https://github.com/TalwalkarLab/leaf) which could be used."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 1. Build the dataset\n",
    "\n",
    "But first, let's build the dataset!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "from string import ascii_lowercase\n",
    "from scipy.ndimage.interpolation import map_coordinates\n",
    "from scipy.ndimage.filters import gaussian_filter\n",
    "import torch.utils.data as utils\n",
    "import pickle"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1.1 Fonts"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We have done previous work on font distinguishability in a classification task and we've select the two fonts *cursive, Georgia* that are of medium difficulty to distinguish. We will extend our analysis two more couples but this is a fair start."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "families = ['cursive', 'Georgia']"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Here what they look like!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x144 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(12,2))\n",
    "for i, family in enumerate(families):\n",
    "    ax = plt.subplot(1, 5, i+1)\n",
    "    ax.set_title(family)\n",
    "    ax.text(0.3, 0.4, '123', size=50, family=family)\n",
    "\n",
    "    # Rm axes, draw and get the rgba shape of the digits\n",
    "    plt.axis('off')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1.2 Noising process\n",
    "Our dataset is artificial. Hence we need to add some noise to make it more interesting for a classification task.\n",
    "Currently the noise boils down to:\n",
    "\n",
    " - A random but moderate rotation\n",
    " - A random but moderation size variation\n",
    " - A deformation using a Gaussian filter\n",
    "\n",
    "When we are done, we recenter the data (as it is done for MNIST).\n",
    "\n",
    "Let's do it for one character to see how it looks like!"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Utils functions to handle rgb / rgba conversion"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "def rgba_to_rgb(rgba):\n",
    "        if rgba[0] > 0:\n",
    "            return np.ones(3) * (256 - rgba[0])\n",
    "        return rgba[1:]\n",
    "    \n",
    "def convert_to_rgb(data):\n",
    "    return np.apply_along_axis(rgba_to_rgb, 2, data) "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Deformation function"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "def elastic_transform(image, alpha, sigma, random_state=None):\n",
    "    \"\"\"Elastic deformation of images as described in [Simard2003]_.\n",
    "    .. [Simard2003] Simard, Steinkraus and Platt, \"Best Practices for\n",
    "       Convolutional Neural Networks applied to Visual Document Analysis\", in\n",
    "       Proc. of the International Conference on Document Analysis and\n",
    "       Recognition, 2003.\n",
    "    \"\"\"\n",
    "    if random_state is None:\n",
    "        random_state = np.random.RandomState(None)\n",
    "\n",
    "    shape = image.shape\n",
    "    dx = gaussian_filter((random_state.rand(*shape) * 2 - 1), sigma, mode=\"constant\", cval=0) * alpha\n",
    "    dy = gaussian_filter((random_state.rand(*shape) * 2 - 1), sigma, mode=\"constant\", cval=0) * alpha\n",
    "    dz = np.zeros_like(dx)\n",
    "\n",
    "    x, y, z = np.meshgrid(np.arange(shape[0]), np.arange(shape[1]), np.arange(shape[2]))\n",
    "    indices = np.reshape(y+dy, (-1, 1)), np.reshape(x+dx, (-1, 1)), np.reshape(z, (-1, 1))\n",
    "\n",
    "    distored_image = map_coordinates(image, indices, order=1, mode='reflect')\n",
    "    return distored_image.reshape(image.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Recentering function"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def center(data):\n",
    "    # Inverse black and white\n",
    "    wb_data = np.ones(data.shape) * 255 - data\n",
    "    \n",
    "    # normalize\n",
    "    prob_data = wb_data / np.sum(wb_data)\n",
    "    \n",
    "    # marginal distributions\n",
    "    dx = np.sum(prob_data, (1, 2))\n",
    "    dy = np.sum(prob_data, (0, 2))\n",
    "\n",
    "    # expected values\n",
    "    (X, Y, Z) = prob_data.shape\n",
    "    cx = np.sum(dx * np.arange(X))\n",
    "    cy = np.sum(dy * np.arange(Y))\n",
    "    \n",
    "    # Check bounds\n",
    "    assert cx > X/4 and cx < 3 * X/4, f\"ERROR: {cx} > {X/4} and {cx} < {3 * X/4}\"\n",
    "    assert cy > Y/4 and cy < 3 * Y/4, f\"ERROR: {cy} > {Y/4} and {cy} < {3 * Y/4}\"\n",
    "    \n",
    "    # print('Center', cx, cy)\n",
    "    \n",
    "    x_min = int(round(cx - X/4))\n",
    "    x_max = int(round(cx + X/4))\n",
    "    y_min = int(round(cy - Y/4))\n",
    "    y_max = int(round(cy + Y/4))\n",
    "    \n",
    "    return data[x_min:x_max, y_min:y_max, :]\n",
    "    "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### One first example!\n",
    "Ok let's how it looks like!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 200x200 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Sample a rotation angle\n",
    "rotation = np.random.normal(0, 10) # 10 degrees (out of 180)\n",
    "# Sample a text size +/- 5% in std\n",
    "size = 100 + np.random.normal(0, 5) \n",
    "# Sample a family\n",
    "family_idx = np.random.randint(len(families))\n",
    "family = families[family_idx]\n",
    "# Sample a char\n",
    "digit = np.random.randint(10)\n",
    "letter = str(digit)\n",
    "\n",
    "# Show rotated letter\n",
    "fig = plt.figure(figsize=(2,2), dpi=100)\n",
    "fig.text(0.4, 0.4, letter, size=size, rotation=rotation, family=family)\n",
    "plt.axis('off')\n",
    "plt.show()\n",
    "plt.close(fig)\n",
    "\n",
    "# Load the letter in smaller size\n",
    "fig = plt.figure(figsize=(2,2), dpi=100)\n",
    "fig.text(0.4, 0.4, letter, size=50, rotation=rotation, family=family)\n",
    "\n",
    "# Rm axes, draw and get the rgba shape of the letter\n",
    "plt.axis('off')\n",
    "fig.canvas.draw()\n",
    "data = np.frombuffer(fig.canvas.tostring_argb(), dtype=np.uint8)\n",
    "data = data.reshape(fig.canvas.get_width_height()[::-1] + (4,))\n",
    "plt.close(fig)\n",
    "# Convert to rgb\n",
    "data = convert_to_rgb(data)\n",
    "\n",
    "# Re-center the data\n",
    "data = center(data)\n",
    "\n",
    "plt.show()\n",
    "plt.axis('off')\n",
    "plt.imshow(data)\n",
    "plt.show()\n",
    "\n",
    "# Apply an elastic deformation\n",
    "data = elastic_transform(data, alpha=1100, sigma=7) # 991; 9\n",
    "\n",
    "plt.axis('off')\n",
    "plt.imshow(data)\n",
    "plt.show()\n",
    "\n",
    "# Free memory space\n",
    "plt.close(fig)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Notice that the images here are provided with high resolution, they will then be converted to 28x28 pixels and will becomes way harder to analyse.\n",
    "\n",
    "Feel free to change the rotation std, the text size std, and the `alpha` and `sigma` parameters in the deformation. You'll see that the task can be made incredibly harder!"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Building the dataset\n",
    "Ok now we are equipped to build a complete dataset!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "# We define a variance parameter for rotation and one for text size\n",
    "std_rotation = 10\n",
    "std_size = 2.5 # 5% of reference size, here 50"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train0\n",
      "0 %\n",
      "1 %\n",
      "2 %\n",
      "3 %\n",
      "4 %\n",
      "5 %\n",
      "6 %\n",
      "7 %\n",
      "8 %\n",
      "9 %\n",
      "10 %\n",
      "11 %\n",
      "12 %\n",
      "13 %\n",
      "14 %\n",
      "15 %\n",
      "16 %\n",
      "17 %\n",
      "18 %\n",
      "19 %\n",
      "20 %\n",
      "21 %\n",
      "22 %\n",
      "23 %\n",
      "24 %\n",
      "25 %\n",
      "26 %\n",
      "27 %\n",
      "28 %\n",
      "29 %\n",
      "30 %\n",
      "31 %\n",
      "32 %\n",
      "33 %\n",
      "34 %\n",
      "35 %\n",
      "36 %\n",
      "37 %\n",
      "38 %\n",
      "39 %\n",
      "40 %\n",
      "41 %\n",
      "42 %\n",
      "43 %\n",
      "44 %\n",
      "45 %\n",
      "46 %\n",
      "47 %\n",
      "48 %\n",
      "49 %\n",
      "50 %\n",
      "51 %\n",
      "52 %\n",
      "53 %\n",
      "54 %\n",
      "55 %\n",
      "56 %\n",
      "57 %\n",
      "58 %\n",
      "59 %\n",
      "60 %\n",
      "61 %\n",
      "62 %\n",
      "63 %\n",
      "64 %\n",
      "65 %\n",
      "66 %\n",
      "67 %\n",
      "68 %\n",
      "69 %\n",
      "70 %\n",
      "71 %\n",
      "72 %\n",
      "73 %\n",
      "74 %\n",
      "75 %\n",
      "76 %\n",
      "77 %\n",
      "78 %\n",
      "79 %\n",
      "80 %\n",
      "81 %\n",
      "82 %\n",
      "83 %\n",
      "84 %\n",
      "85 %\n",
      "86 %\n",
      "87 %\n",
      "88 %\n",
      "89 %\n",
      "90 %\n",
      "91 %\n",
      "92 %\n",
      "93 %\n",
      "94 %\n",
      "95 %\n",
      "96 %\n",
      "97 %\n",
      "98 %\n",
      "99 %\n",
      "Train1\n",
      "0 %\n",
      "1 %\n",
      "2 %\n",
      "3 %\n",
      "4 %\n",
      "5 %\n",
      "6 %\n",
      "7 %\n",
      "8 %\n",
      "9 %\n",
      "10 %\n",
      "11 %\n",
      "12 %\n",
      "13 %\n",
      "14 %\n",
      "15 %\n",
      "16 %\n",
      "17 %\n",
      "18 %\n",
      "19 %\n",
      "20 %\n",
      "21 %\n",
      "22 %\n",
      "23 %\n",
      "24 %\n",
      "25 %\n",
      "26 %\n",
      "27 %\n",
      "28 %\n",
      "29 %\n",
      "30 %\n",
      "31 %\n",
      "32 %\n",
      "33 %\n",
      "34 %\n",
      "35 %\n",
      "36 %\n",
      "37 %\n",
      "38 %\n",
      "39 %\n",
      "40 %\n",
      "41 %\n",
      "42 %\n",
      "43 %\n",
      "44 %\n",
      "45 %\n",
      "46 %\n",
      "47 %\n",
      "48 %\n",
      "49 %\n",
      "50 %\n",
      "51 %\n",
      "52 %\n",
      "53 %\n",
      "54 %\n",
      "55 %\n",
      "56 %\n",
      "57 %\n",
      "58 %\n",
      "59 %\n",
      "60 %\n",
      "61 %\n",
      "62 %\n",
      "63 %\n",
      "64 %\n",
      "65 %\n",
      "66 %\n",
      "67 %\n",
      "68 %\n",
      "69 %\n",
      "70 %\n",
      "71 %\n",
      "72 %\n",
      "73 %\n",
      "74 %\n",
      "75 %\n",
      "76 %\n",
      "77 %\n",
      "78 %\n",
      "79 %\n",
      "80 %\n",
      "81 %\n",
      "82 %\n",
      "83 %\n",
      "84 %\n",
      "85 %\n",
      "86 %\n",
      "87 %\n",
      "88 %\n",
      "89 %\n",
      "90 %\n",
      "91 %\n",
      "92 %\n",
      "93 %\n",
      "94 %\n",
      "95 %\n",
      "96 %\n",
      "97 %\n",
      "98 %\n",
      "99 %\n",
      "Train2\n",
      "0 %\n",
      "1 %\n",
      "2 %\n",
      "3 %\n",
      "4 %\n",
      "5 %\n",
      "6 %\n",
      "7 %\n",
      "8 %\n",
      "9 %\n",
      "10 %\n",
      "11 %\n",
      "12 %\n",
      "13 %\n",
      "14 %\n",
      "15 %\n",
      "16 %\n",
      "17 %\n",
      "18 %\n",
      "19 %\n",
      "20 %\n",
      "21 %\n",
      "22 %\n",
      "23 %\n",
      "24 %\n",
      "25 %\n",
      "26 %\n",
      "27 %\n",
      "28 %\n",
      "29 %\n",
      "30 %\n",
      "31 %\n",
      "32 %\n",
      "33 %\n",
      "34 %\n",
      "35 %\n",
      "36 %\n",
      "37 %\n",
      "38 %\n",
      "39 %\n",
      "40 %\n",
      "41 %\n",
      "42 %\n",
      "43 %\n",
      "44 %\n",
      "45 %\n",
      "46 %\n",
      "47 %\n",
      "48 %\n",
      "49 %\n",
      "50 %\n",
      "51 %\n",
      "52 %\n",
      "53 %\n",
      "54 %\n",
      "55 %\n",
      "56 %\n",
      "57 %\n",
      "58 %\n",
      "59 %\n",
      "60 %\n",
      "61 %\n",
      "62 %\n",
      "63 %\n",
      "64 %\n",
      "65 %\n",
      "66 %\n",
      "67 %\n",
      "68 %\n",
      "69 %\n",
      "70 %\n",
      "71 %\n",
      "72 %\n",
      "73 %\n",
      "74 %\n",
      "75 %\n",
      "76 %\n",
      "77 %\n",
      "78 %\n",
      "79 %\n",
      "80 %\n",
      "81 %\n",
      "82 %\n",
      "83 %\n",
      "84 %\n",
      "85 %\n",
      "86 %\n",
      "87 %\n",
      "88 %\n",
      "89 %\n",
      "90 %\n",
      "91 %\n",
      "92 %\n",
      "93 %\n",
      "94 %\n",
      "95 %\n",
      "96 %\n",
      "97 %\n",
      "98 %\n",
      "99 %\n",
      "Train3\n",
      "0 %\n",
      "1 %\n",
      "2 %\n",
      "3 %\n",
      "4 %\n",
      "5 %\n",
      "6 %\n",
      "7 %\n",
      "8 %\n",
      "9 %\n",
      "10 %\n",
      "11 %\n",
      "12 %\n",
      "13 %\n",
      "14 %\n",
      "15 %\n",
      "16 %\n",
      "17 %\n",
      "18 %\n",
      "19 %\n",
      "20 %\n",
      "21 %\n",
      "22 %\n",
      "23 %\n",
      "24 %\n",
      "25 %\n",
      "26 %\n",
      "27 %\n",
      "28 %\n",
      "29 %\n",
      "30 %\n",
      "31 %\n",
      "32 %\n",
      "33 %\n",
      "34 %\n",
      "35 %\n",
      "36 %\n",
      "37 %\n",
      "38 %\n",
      "39 %\n",
      "40 %\n",
      "41 %\n",
      "42 %\n",
      "43 %\n",
      "44 %\n",
      "45 %\n",
      "46 %\n",
      "47 %\n",
      "48 %\n",
      "49 %\n",
      "50 %\n",
      "51 %\n",
      "52 %\n",
      "53 %\n",
      "54 %\n",
      "55 %\n",
      "56 %\n",
      "57 %\n",
      "58 %\n",
      "59 %\n",
      "60 %\n",
      "61 %\n",
      "62 %\n",
      "63 %\n",
      "64 %\n",
      "65 %\n",
      "66 %\n",
      "67 %\n",
      "68 %\n",
      "69 %\n",
      "70 %\n",
      "71 %\n",
      "72 %\n",
      "73 %\n",
      "74 %\n",
      "75 %\n",
      "76 %\n",
      "77 %\n",
      "78 %\n",
      "79 %\n",
      "80 %\n",
      "81 %\n",
      "82 %\n",
      "83 %\n",
      "84 %\n",
      "85 %\n",
      "86 %\n",
      "87 %\n",
      "88 %\n",
      "89 %\n",
      "90 %\n",
      "91 %\n",
      "92 %\n",
      "93 %\n",
      "94 %\n",
      "95 %\n",
      "96 %\n",
      "97 %\n",
      "98 %\n",
      "99 %\n",
      "Test\n",
      "0 %\n",
      "1 %\n",
      "2 %\n",
      "3 %\n",
      "4 %\n",
      "5 %\n",
      "6 %\n",
      "7 %\n",
      "8 %\n",
      "9 %\n",
      "10 %\n",
      "11 %\n",
      "12 %\n",
      "13 %\n",
      "14 %\n",
      "15 %\n",
      "16 %\n",
      "17 %\n",
      "18 %\n",
      "19 %\n",
      "20 %\n",
      "21 %\n",
      "22 %\n",
      "23 %\n",
      "24 %\n",
      "25 %\n",
      "26 %\n",
      "27 %\n",
      "28 %\n",
      "29 %\n",
      "30 %\n",
      "31 %\n",
      "32 %\n",
      "33 %\n",
      "34 %\n",
      "35 %\n",
      "36 %\n",
      "37 %\n",
      "38 %\n",
      "39 %\n",
      "40 %\n",
      "41 %\n",
      "42 %\n",
      "43 %\n",
      "44 %\n",
      "45 %\n",
      "46 %\n",
      "47 %\n",
      "48 %\n",
      "49 %\n",
      "50 %\n",
      "51 %\n",
      "52 %\n",
      "53 %\n",
      "54 %\n",
      "55 %\n",
      "56 %\n",
      "57 %\n",
      "58 %\n",
      "59 %\n",
      "60 %\n",
      "61 %\n",
      "62 %\n",
      "63 %\n",
      "64 %\n",
      "65 %\n",
      "66 %\n",
      "67 %\n",
      "68 %\n",
      "69 %\n",
      "70 %\n",
      "71 %\n",
      "72 %\n",
      "73 %\n",
      "74 %\n",
      "75 %\n",
      "76 %\n",
      "77 %\n",
      "78 %\n",
      "79 %\n",
      "80 %\n",
      "81 %\n",
      "82 %\n",
      "83 %\n",
      "84 %\n",
      "85 %\n",
      "86 %\n",
      "87 %\n",
      "88 %\n",
      "89 %\n",
      "90 %\n",
      "91 %\n",
      "92 %\n",
      "93 %\n",
      "94 %\n",
      "95 %\n",
      "96 %\n",
      "97 %\n",
      "98 %\n",
      "99 %\n",
      "CPU times: user 21min 58s, sys: 3min 32s, total: 25min 30s\n",
      "Wall time: 26min 33s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "def build_dataset(dataset_size, verbose=False):\n",
    "    dataset_data = []\n",
    "    dataset_target_char = []\n",
    "    dataset_target_family = []\n",
    "    for i in range(dataset_size):\n",
    "        if i % int(dataset_size/100) == 0:\n",
    "            print(round(i / dataset_size * 100), '%')\n",
    "        rotation = np.random.normal(0, std_rotation)\n",
    "        size = 50 + np.random.normal(0, std_size) \n",
    "        family_idx = np.random.randint(len(families))\n",
    "        family = families[family_idx]\n",
    "        digit = np.random.randint(10)\n",
    "        letter = str(digit)\n",
    "\n",
    "        fig = plt.figure(figsize=(2,2), dpi=28)\n",
    "        fig.text(0.4, 0.4, letter, size=size, rotation=rotation, family=family)\n",
    "\n",
    "        # Rm axes, draw and get the rgba shape of the letter\n",
    "        plt.axis('off')\n",
    "        fig.canvas.draw()\n",
    "        data = np.frombuffer(fig.canvas.tostring_argb(), dtype=np.uint8)\n",
    "        data = data.reshape(fig.canvas.get_width_height()[::-1] + (4,))\n",
    "\n",
    "        # Convert to rgb\n",
    "        data = convert_to_rgb(data)\n",
    "\n",
    "        # Center the data\n",
    "        data = center(data)\n",
    "\n",
    "        if verbose:\n",
    "            plt.show()\n",
    "            plt.axis('off')\n",
    "            plt.imshow(data)\n",
    "            plt.show()\n",
    "            \n",
    "        # Apply an elastic deformation\n",
    "        data = elastic_transform(data, alpha=991, sigma=9)\n",
    "\n",
    "        if verbose:\n",
    "            plt.axis('off')\n",
    "            plt.imshow(data)\n",
    "            plt.show()\n",
    "\n",
    "        # Free memory space\n",
    "        plt.close(fig)\n",
    "\n",
    "        # Append data to the datasets\n",
    "        target_char = np.array([digit])\n",
    "        target_family = np.array([family_idx])\n",
    "        dataset_data.append(data[:,:,0])\n",
    "        dataset_target_char.append(target_char)\n",
    "        dataset_target_family.append(target_family)\n",
    "        \n",
    "    return dataset_data, dataset_target_char, dataset_target_family\n",
    "\n",
    "\n",
    "def save_dataset(info, data, target_char, target_family):\n",
    "    with open('dataset/character_dataset_{}.pkl'.format(info), 'wb') as output:\n",
    "        dataset = data, target_char, target_family\n",
    "        pickle.dump(dataset, output, pickle.HIGHEST_PROTOCOL)\n",
    "\n",
    "# WE BUILD 10000 PER 10000 FOR MEMORY PURPOSE\n",
    "# WARNING: This step takes time, and you may need to restart your notebook for memory purposes.\n",
    "# If so, change the following range to avoid the loops already processed\n",
    "for i in range(6):\n",
    "    print(f\"Train{i}\")\n",
    "    train_data, train_target_char, train_target_family = build_dataset(10000)\n",
    "    save_dataset(f\"train{i}\", train_data, train_target_char, train_target_family)\n",
    "    \n",
    "print(\"Test\")\n",
    "test_data, test_target_char, test_target_family = build_dataset(10000)\n",
    "save_dataset(\"test\", test_data, test_target_char, test_target_family)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Ok we now have our dataset! Let's use it!"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Appendix\n",
    "Here is an utility function to draw some samples of the dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training set 60000 items\n",
      "Testing set  10000 items\n"
     ]
    }
   ],
   "source": [
    "# Allow to load packages from parent\n",
    "import sys, os\n",
    "sys.path.insert(1, os.path.realpath(os.path.pardir))\n",
    "import random, math\n",
    "import learn\n",
    "data = learn.load_data()\n",
    "train_data, train_target_char, train_target_family, test_data, test_target_char, test_target_family = data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1008x864 with 20 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "def show_examples(data, target_char, target_family, n_examples=20):\n",
    "    n_rows = math.ceil(n_examples / 5)\n",
    "    plt.figure(figsize=(14, 3*n_rows))\n",
    "    for i in range(n_examples):\n",
    "        ax = plt.subplot(n_rows, 5, i+1)\n",
    "        idx = random.randint(0, len(data))\n",
    "        image = 255 - data[idx]\n",
    "        letter = str([target_char[idx][0]])\n",
    "        family = families[target_family[idx][0]]\n",
    "        ax.set_title(f\"{letter} ({family})\")\n",
    "        plt.axis('off')\n",
    "        ax.imshow(image, cmap='gist_gray')\n",
    "    plt.show()\n",
    "        \n",
    "        \n",
    "show_examples(test_data, test_target_char, test_target_family)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "It is now much harder to distinguish the fonts, isn't it?!\n",
    "\n",
    "Ok let's use this dataset to start our work!"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "### If you like it, star it!\n",
    "\n",
    "The easiest way to show support is just by starring the Repos! This helps raise awareness on this topics and is a precious feedback for the repo maintainers!\n",
    "\n",
    "- [Star the Repo](https://github.com/LaRiffle/collateral-learning)"
   ]
  },
  {
   "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": "3.7.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}